Every so often the ground under startups moves, and when it does, the rules that governed the last cycle
stop applying. The PC did it. The internet did it. Mobile and the cloud did it. Each time, a platform shift
didn’t just create new companies — it changed what kind of company could exist, how fast it could grow,
what it could charge, and who could build it. We are inside another one of those resets right now, and it is
arguably the largest of them all, because it is not one shift but three arriving at once:
Intelligence became a raw material. For the first time, cognition — reading, writing, reasoning,
deciding, coding, designing — can be summoned on demand at a marginal cost that falls roughly an order of
magnitude every year. Software used to help people do knowledge work. Now software can do the work.
That single change moves the addressable market of technology from the world’s IT budgets (a few
trillion dollars) to the world’s payroll (tens of trillions). The prize is 10× bigger than the one the
cloud generation chased.The physical world got a brain. The same transformer architecture that solved language is now solving
perception and action. Robots that learn instead of being programmed, cars that finally drive themselves,
drones that coordinate, factories that run lights-out — the atoms are catching up to the bits. The bottleneck
is no longer whether a machine can do a physical task; it’s data, reliability, and cost.The old certainties broke. Cheap energy, frictionless globalization, and stable geopolitics — the
quiet assumptions under the last cycle — are gone. Electricity is now the binding constraint on
intelligence. Supply chains are being re-shored and re-armed. Money itself is being re-platformed onto new
rails. Scarcity and sovereignty are back, and scarcity is where fortunes are made.
Stack these three shifts and you get the defining feature of this era: the cost of creating things collapses
while the value of trust, energy, data, and physical execution soars. That asymmetry is the engine under
every one of the 48 topics that follow.
Why “topics” and not “predictions”
Anyone can list technologies. The point of this report is to be useful to a builder — so each topic is
framed not as a prediction but as a buildable opening. For every one of the 48 we answer the same five
questions, because these are the five a founder actually needs:
The shift — what is actually changing, stated plainly.
Why now — the specific unlock in the last ~24 months that makes this the right moment, not five years
ago and not five years from now. Timing is the whole game; a great idea at the wrong time is a graveyard.Why it’s a startup goldmine — the size of the pool, what incumbent or manual process it drains, the
pricing/margin unlock, and — critically — what stays defensible after the obvious version gets copied.Where the opening is — the sharp wedge. Not “build in AI,” but the specific narrow, painful, valuable
beachhead a small team can take and expand from.The signal — real companies, many of them in this very library’s Past Unicorns and Hundred Million Corns folders, that prove the topic is already compounding.
The seven meta-patterns that run through all 48
Read the 48 topics and the same deep structures keep surfacing. If you internalize these seven, you can
evaluate any opportunity — including ones not on this list:
1. Value is priced as outcomes, not access. The defining commercial move of this cycle is charging for
the work delivered — a percentage of the labor replaced, the fraud prevented, the megawatt delivered —
rather than a monthly seat. A SaaS tool that “helps” with a $6,000/month job caps out at $50/seat. An agent
that does that job can charge $2,000/month and still be a bargain. The topics that let you price on outcomes
are the ones that mint the biggest companies.
2. The value migrates down the stack, then back up. Early in every wave the money is in
infrastructure — the models, the rails, the launch vehicles, the picks and shovels. As the wave matures it
moves to applications and the system of record. The winning move is to enter at the layer that is
underbuilt right now and ride it up. Several of these 48 are infrastructure plays that are underbuilt today;
several are application plays where the infrastructure just got cheap enough.
3. The moat is the loop the incumbent can’t cross. In a world where the obvious product is copied in
weeks, defensibility comes from three places: proprietary data generated by your own operations, the
liability and trust the customer will not take on themselves, and becoming the system of record.
Every durable company on this list owns at least one.
4. The best wedge is a boring, expensive, hated job in a big industry. The narrower and more painful the
beachhead, the better. “AI for legal” loses; “an agent that drafts and files the personal-injury demand
letter” wins. The unglamorous, high-friction, regulation-heavy corners are exactly where incumbents are weak
and where a focused team can plant a flag.
5. Regulation and scarcity are tailwinds, not obstacles. Compliance load (DORA, NIS2, the AI Act),
energy scarcity, supply-chain sovereignty, security clearances — these look like walls, but a wall you can
climb becomes a moat once you’re on the other side. Many of the richest topics here are created by
constraints.
6. Verification is the shadow of automation. As AI generates action, code, and content at machine speed,
the scarce resource becomes proof that it was done right. Every wave of automation creates an equal-and-
opposite wave of demand for trust, identity, evals, provenance, and compliance. Whole categories on this list
exist only because the others exist.
7. Distribution is destiny. Cheap creation means the bottleneck is no longer building — it’s getting
found, trusted, and adopted. The topics where a founder can bolt onto an existing distribution loop (a
developer’s editor, a payment flow, a health system’s EHR, a government’s procurement rail) beat the ones
that require building an audience from zero.
How to read this report
The 48 topics are grouped into six clusters:
Cluster 1 — AI Foundations & Agents (topics 1–8): the intelligence layer itself.
Cluster 2 — AI Applications & Knowledge Work (9–16): AI doing the world’s expensive text-and-decision jobs.
Cluster 3 — Physical & Frontier Tech (17–24): robots, mobility, space, and the deep-tech frontier.
Cluster 4 — Energy, Climate & Bio (25–32): the atoms — power, materials, and engineered biology.
Cluster 5 — Money, Fintech & Trust (33–40): programmable money and the verification layer.
Cluster 6 — Platforms, Infrastructure & New Models (41–48): the connective tissue and the new company shapes.
No single founder should chase all six. The clusters are ordered from the softest, fastest-moving,
lowest-capital opportunities (AI software) to the hardest, most capital-intensive, most defensible ones
(energy, bio, deep tech). Where you play should be a function of who you are — your unfair advantages, your
tolerance for capital and time, and the distribution you can credibly reach. The conclusion
draws the through-lines back together and offers a way to choose.
A note on honesty: not all 48 are equally ripe, and several are partly hype. Throughout, the “why now” and
“where the opening is” sections try to separate the durable from the frothy — because the difference between a
trend and a timed trend is the difference between a unicorn and a cautionary tale. Let’s begin.
Cluster 1 — AI Foundations & Agents
The models are commoditizing; the value is migrating to the layer that turns raw intelligence into reliable, autonomous work. This cluster covers the eight foundational shifts — from goal-driven agents and the compute that runs them, down to memory, orchestration, and the sovereignty fight over who owns the weights — that together form the substrate every other startup in this report will stand on.
1. Agentic AI (goal-driven autonomous agents)
The shift: Software stops waiting for prompts and starts pursuing goals. Agentic AI is the move from a chatbot that answers a question to a system you hand an objective — “reconcile last month’s invoices,” “ship this feature,” “book the whole trip” — that then plans, calls tools, observes results, and loops until the goal is met. The unit of interaction is no longer a message; it’s a completed task.
Why now: Three things converged inside 24 months. Models crossed the reliability threshold on multi-step tool use — function calling became structured and dependable rather than a party trick. Context windows blew past the point where a full task history, codebase, or document set fits in working memory. And the emergence of standard tool protocols (MCP and its descendants) meant an agent could reach into arbitrary systems without a bespoke integration for each. Reliability on 10-step chains went from “amusing demo” to “runs unattended overnight.”
Why it’s a startup goldmine: This is the largest TAM in software because it isn’t software’s TAM — it’s labor’s. Agentic systems don’t compete with the $650B SaaS market; they compete with the tens of trillions in global services and payroll. The margin unlock is brutal for incumbents and beautiful for founders: a task that cost a $60/hour analyst now costs cents of inference. Defensibility comes not from the model — everyone rents the same frontier — but from the execution scaffolding: the eval harness that proves the agent is right, the recovery logic when it’s wrong, and the proprietary workflow data that makes your loop converge faster than a generic one.
Where the opening is: Pick a domain where the task is high-frequency, verifiable, and currently done by expensive humans — then own the last mile of reliability, not the model. The wedge is trust: be the agent a CFO or a hospital will actually let run unsupervised, because you’ve solved observability and rollback, not because your demo is flashy.
Signal — startups & scaleups to watch:
OpenAI — Operator & agentic o-series
Anthropic — Claude computer use, Agent SDK
Google DeepMind — Gemini agents, Project Mariner
Microsoft — Copilot agents
Cognition — Devin autonomous engineer
Sierra — enterprise conversational agents
Adept — action models for workflows
Imbue — robust reasoning agents
Manus — general autonomous agent
Genspark — autonomous super-agent
H Company — computer-use web agents
Reflection AI — autonomous coding agents
Factory — autonomous software droids
Lindy — no-code AI assistants
MultiOn — web-action agents
Relevance AI — AI workforce builder
Emergence AI — autonomous web agents
Ema — universal AI employee
Rabbit — consumer action agent
CrewAI — agent-crew framework
xAI — Grok agents
2. Vertical AI agents (labor replacement, priced as work not seats)
The shift: The most important pricing revolution since SaaS: charging for outcomes, not access. A vertical AI agent goes deep into one job function — a legal associate, an SDR, a claims adjuster, a medical coder — and is sold not as a tool that a human uses but as the worker itself, billed per resolved ticket, per booked meeting, per adjudicated claim.
Why now: Horizontal agents proved capability; verticals prove reliability. The unlock was realizing that a narrow domain lets you build the thing generic agents lack — a tight eval loop, domain-specific guardrails, and integrations into the ten systems that job actually touches. Once you can guarantee an SDR agent books meetings at human quality, the customer stops asking “how many seats?” and starts asking “how much per meeting?” That reframing is worth a 10x expansion in contract value.
Why it’s a startup goldmine: Seat-based SaaS caps you at the number of humans in a role. Work-based pricing uncaps you at the total spend on that role — you can capture 30-50% of a fully-loaded salary and still be a bargain. A support SaaS tool charges $100/agent/month; a support agent that resolves tickets can charge $1-2 per resolution and address the entire support payroll. Gross margins run 70-90% because inference is the only real COGS. Defensibility is the workflow moat: the integrations, the domain evals, and the feedback data from millions of resolved cases that a horizontal player will never accumulate in your niche.
Where the opening is: Find a role with (a) a clear success metric, (b) painful labor cost, and (c) a system-of-record you can plug into, then price on the metric from day one. Don’t sell “AI for X”; sell the completed unit of X-work with an SLA.
Signal — startups & scaleups to watch:
Sierra — customer-service agents
Harvey — legal copilot
Decagon — customer-support agents
11x — AI sales reps
Artisan — AI SDR “Ava”
Abridge — clinical documentation
Ambience Healthcare — medical scribe
Nabla — clinical note assistant
Suki — voice clinical assistant
Hippocratic AI — healthcare voice agents
Cresta — contact-center AI
Ada — customer-service automation
Parloa — contact-center agents
Observe.AI — contact-center intelligence
EvenUp — injury-claim drafting
Eve — plaintiff-firm legal AI
Norm Ai — regulatory compliance agents
Rogo — financial analyst AI
Hebbia — finance research agents
Tennr — healthcare referral automation
EliseAI — housing & property agents
DevRev — support & product agents
3. AI inference infrastructure & the compute layer
The shift: The center of gravity in AI economics is moving from training to inference. Training a frontier model is a one-time capital event; running billions of agentic calls a day is a perpetual operating cost — and that recurring spend is where the durable, high-volume business lives. Whoever makes tokens cheaper, faster, and more reliable to serve owns the toll road under the entire agent economy.
Why now: Agents changed the math. A chatbot makes one call per human message; an agentic workflow makes dozens or hundreds of calls per task, many of them reasoning-heavy. Inference demand is exploding super-linearly with adoption, and it’s increasingly latency- and cost-sensitive because agents run in loops. Meanwhile the GPU bottleneck, custom silicon (TPUs, Trainium, Groq/Cerebras-class accelerators), and techniques like speculative decoding, quantization, and KV-cache optimization opened real headroom for specialists to beat the hyperscalers on price-per-token.
Why it’s a startup goldmine: This is a market measured in hundreds of billions of annual compute spend, and it’s structurally recurring — you get paid every time anyone’s agent thinks. The margin unlock is efficiency arbitrage: a 3-5x improvement in tokens-per-dollar through better batching, routing, and hardware utilization flows straight to gross margin or to a price advantage that wins share. Defensibility lives in the systems software — the inference engine, the scheduler, the model-routing layer — and in hardware relationships that let you serve at a cost basis competitors can’t touch.
Where the opening is: Don’t build a foundation model; build the fastest, cheapest, most reliable way to serve everyone else’s. The sharp wedges are inference-optimized serving (own latency-critical workloads), intelligent model routing (send each call to the cheapest model that can handle it), and GPU orchestration for the mid-market that can’t get hyperscaler allocation.
Signal — startups & scaleups to watch:
Together AI — inference & training cloud
Fireworks AI — fast inference platform
Baseten — model serving
Modal — serverless GPU compute
Replicate — run models via API
Fal — fast generative inference
DeepInfra — low-cost model API
Anyscale — Ray-based scaling
Groq — LPU inference chips
Cerebras — wafer-scale AI chips
SambaNova — AI chip & platform
Etched — transformer ASIC
Tenstorrent — RISC-V AI chips
d-Matrix — in-memory inference chips
CoreWeave — GPU cloud
Crusoe — AI-focused cloud
Lambda — GPU cloud
Nebius — AI cloud platform
RunPod — GPU rental cloud
Vast.ai — GPU marketplace
Predibase — fine-tune & serve
4. Small, efficient & on-device models
The shift: Not everything needs a frontier brain. A large share of real-world AI tasks — classification, extraction, routing, structured generation, function-call selection — can be handled by a model 1/100th the size, run locally, for a fraction of the cost and latency. The industry is discovering that the right question isn’t “how big can we go?” but “how small can we get away with?”
Why now: Distillation, quantization, and better training data made 1-8B parameter models shockingly capable — a well-tuned small model now matches a 2023 frontier model on narrow tasks. Simultaneously, on-device silicon (Apple Neural Engine, NPUs in every new laptop and phone) crossed the threshold to run these models locally in real time. The result: a genuine tier of intelligence that is free at the margin, private by default, and works offline.
Why it’s a startup goldmine: The economics invert the whole stack. On-device inference has zero marginal compute cost to the vendor, which enables entirely new price points and business models — including ones the API-metered giants can’t match. The disrupted incumbents are the cloud-inference bills themselves. Defensibility is the hard engineering of squeezing quality into constrained hardware and the vertically-tuned small models that outperform general giants on a specific task while costing nothing to run. Privacy and latency are the durable pull: data that never leaves the device is a feature no cloud model can offer.
Where the opening is: Own a task class where privacy, latency, or offline operation is non-negotiable — healthcare, defense, industrial, personal assistants — and ship a small model that runs on the user’s hardware. Or build the tooling layer: the distillation, fine-tuning, and deployment pipeline that lets everyone else make their own small models.
Signal — startups & scaleups to watch:
Mistral AI — open small models
Meta Llama — open Llama models
Microsoft — Phi small models
Apple — on-device foundation models
Hugging Face — open model hub
Ollama — run local models
LM Studio — desktop local LLMs
Liquid AI — efficient LFM models
Arcee AI — small enterprise models
Nomic AI — open embeddings & models
Nexa AI — on-device model runtime
Cartesia — on-device voice models
Unsloth — fast local fine-tuning
Edge Impulse — edge ML platform
Predibase — small fine-tuned models
Together AI — fine-tuning stack
Fireworks AI — small-model serving
Qualcomm — on-device AI silicon
5. Reasoning models & world models
The shift: From models that pattern-match to models that think — and from models that predict text to models that predict how the world behaves. Reasoning models spend inference-time compute to deliberate, plan, and self-correct before answering, cracking problems that instant next-token prediction never could. World models go further: they learn a predictive simulation of physical or environmental dynamics, letting an agent imagine consequences before acting.
Why now: The reasoning unlock was test-time compute — the discovery that letting a model “think longer” (chain-of-thought at scale, reinforced by verifiable rewards) buys a new scaling axis independent of parameter count. This is why math, code, and multi-step planning capabilities leapt in 2024-2025. World models rode the wave of video and simulation training — models that watch enough of reality start to internalize physics, enabling robotics and autonomous systems to plan in a learned simulator rather than the expensive real world.
Why it’s a startup goldmine: Reasoning is what makes agents trustworthy on hard, high-stakes work — the difference between an agent that drafts a contract and one that can actually reason about liability. That reliability is what unlocks the highest-value verticals (law, finance, engineering, science). World models are the missing piece for the entire robotics and autonomy TAM — a multi-trillion-dollar prize gated on machines that can predict outcomes. Defensibility is in the verifiable-reward training loops, the domain-specific reasoning data, and, for world models, proprietary simulation and sensor data.
Where the opening is: For reasoning, build the verification and reward infrastructure that makes reasoning trainable in a domain — the “gym” for a vertical. For world models, own a physical domain (warehouse robotics, autonomous machines, drug interaction) where a learned simulator collapses the cost of trial and error.
Signal — startups & scaleups to watch:
OpenAI — o-series reasoning
Anthropic — extended thinking
DeepSeek — open reasoning models
Google DeepMind — Gemini thinking, Genie
xAI — Grok reasoning
Mistral AI — Magistral reasoning
Moonshot AI — Kimi reasoning models
Z.ai — GLM reasoning models
Physical Intelligence — robot foundation models
Skild AI — robot foundation model
Wayve — driving world models
Waabi — autonomous-driving simulation
World Labs — large world models
Decart — real-time world models
Runway — general world models
Nvidia — Cosmos world models
Figure — humanoid robot intelligence
1X — humanoid robots
6. Multi-agent systems & orchestration
The shift: One agent is a worker; many coordinated agents are an organization. Complex goals decompose better when specialized agents — a planner, a researcher, a coder, a critic — collaborate, hand off, and check each other’s work, orchestrated by a layer that routes tasks, manages state, and resolves conflicts. The frontier is no longer a smarter single agent but a better-run team of them.
Why now: As soon as agents could reliably complete single tasks, the ceiling became coordination. Practitioners found that a critic agent reviewing a worker’s output, or parallel agents exploring different approaches, beat any monolithic prompt. The tooling to make this practical — durable execution frameworks, agent-to-agent protocols, shared memory and message buses — matured just as the need became acute. Orchestration turned out to be the hard, valuable, and under-built part of the stack.
Why it’s a startup goldmine: Orchestration is the operating system of the agent economy, and OS layers become deeply entrenched, high-switching-cost platforms. The market is every company that wants to run more than one agent — which will be every company. The value unlock is reliability at scale: multi-agent designs with checks and redundancy are how you get from 80% task success (a demo) to 99% (a product). Defensibility is the platform lock-in — once a company builds its agent workflows, evals, and observability on your orchestration layer, ripping it out means rebuilding its digital workforce.
Where the opening is: Build the control plane — the durable orchestration, state management, observability, and inter-agent protocol layer that turns a pile of agents into a governable system. The sharpest wedge is reliability and debuggability: be the layer engineers trust to run agent fleets in production, with tracing, replay, and guardrails they can audit.
Signal — startups & scaleups to watch:
LangChain — LangGraph orchestration
LlamaIndex — data agents framework
CrewAI — multi-agent crews
Microsoft — AutoGen framework
Temporal — durable agent execution
Inngest — durable workflow engine
Prefect — orchestration framework
Orkes — Conductor orchestration
Restack — agent backend framework
E2B — agent code sandboxes
Composio — agent tool integrations
AutoGPT — autonomous agent platform
Sema4.ai — enterprise agent platform
Relevance AI — agent workforce
Griptape — agent framework
Vellum — agent dev platform
Stack AI — no-code agent builder
Lyzr — enterprise agent SDK
Gumloop — agent workflow automation
Dify — LLM app orchestration
Flowise — visual agent builder
n8n — workflow automation
7. AI memory & context engineering
The shift: Intelligence without memory is a brilliant amnesiac. The next unlock isn’t a smarter model but one that remembers — that accumulates knowledge about your company, your preferences, and its own past actions, and surfaces exactly the right context at the right moment. Context engineering — deciding what an agent should know at each step — is emerging as the discipline that most determines whether an agent is useful or hopeless.
Why now: Bigger context windows helped but also exposed the real problem: more context isn’t better context. Stuffing a million tokens degrades reasoning and cost. The field pivoted from “retrieve everything” (naive RAG) to engineered memory — structured long-term stores, hierarchical summarization, learned relevance, and retrieval that understands the task. Agents that run for hours or persist across sessions made memory non-optional: an agent that forgets what it did five steps ago can’t be trusted with anything real.
Why it’s a startup goldmine: Memory is the deepest moat in AI. A model is rentable and swappable; the accumulated context about a user or an enterprise is not — it’s the switching cost that turns a tool into a system of record. Whoever owns the memory layer owns the relationship. The market is every agent and every AI application, all of which need somewhere to remember. The margin story is stickiness: memory compounds, so the product gets better the longer a customer stays, which crushes churn. Defensibility is the proprietary context graph itself — the more it holds, the more painful it is to leave.
Where the opening is: Build the memory layer as infrastructure — a persistent, queryable, permission-aware store that any agent can plug into — or own the context graph inside a specific enterprise so deeply that switching means starting over. The wedge is being the place where an organization’s institutional knowledge accretes.
Signal — startups & scaleups to watch:
Mem0 — agent memory layer
Letta — stateful agents (MemGPT)
Zep — agent memory service
Supermemory — memory API
Cognee — agent memory graphs
Graphlit — context platform
Pinecone — vector database
Chroma — open-source vector DB
Weaviate — vector database
Qdrant — vector search engine
Zilliz — Milvus vector database
LanceDB — embedded vector DB
Turbopuffer — object-storage search
Marqo — vector search engine
Vectara — RAG-as-a-service
Contextual AI — enterprise RAG
LlamaIndex — retrieval framework
Unstructured — data prep for LLMs
Ragie — managed RAG API
Redis — vector & memory store
MongoDB — vector search DB
Neo4j — knowledge graphs
8. Open-weight & sovereign AI models
The shift: Not everyone will — or can — run their intelligence on a handful of American closed APIs. A parallel ecosystem of open-weight models you can download, inspect, fine-tune, and self-host is becoming the foundation for anyone who needs control: nations building sovereign capability, enterprises that can’t send data to a third party, and developers who want to own their stack. Weights are becoming a strategic asset, like a national grid.
Why now: Open-weight models closed the gap. What was a two-year lag behind the frontier compressed to months, with releases from Meta, Mistral, DeepSeek, and Qwen reaching capability tiers that are more than good enough for the vast majority of production workloads. Simultaneously, geopolitics turned AI into infrastructure of national interest — governments now treat dependence on a foreign closed model the way they treat dependence on foreign energy. The combination of “good enough and free to run” plus “strategically necessary to control” created a real market overnight.
Why it’s a startup goldmine: The disrupted incumbent is the closed-API oligopoly and its pricing power. The TAM is every regulated industry (healthcare, defense, finance, government) and every nation that wants its own AI — a category that barely existed two years ago and is now a line item in national budgets. The margin unlock for builders is owning the deployment layer: sovereign clouds, on-prem fine-tuning, compliance and localization services that closed providers structurally cannot offer. Defensibility is trust, data residency, and the integration depth of running inside a customer’s own walls.
Where the opening is: Be the company that operationalizes open weights for those who need sovereignty — the sovereign-AI stack for a nation, the compliant self-hosted deployment for a regulated enterprise, or the fine-tuning and localization layer that turns a generic open model into a domain- or language-specific asset. The wedge is control: sell the customers who legally or geopolitically cannot use a closed API.
Signal — startups & scaleups to watch:
Mistral AI — European open models
Meta Llama — open Llama family
DeepSeek — Chinese open frontier
Qwen — Alibaba open models
Z.ai — Zhipu GLM open models
Moonshot AI — Kimi open models
01.AI — Yi open models
Cohere — enterprise open weights
Nous Research — open fine-tunes
Allen Institute for AI — fully-open OLMo models
EleutherAI — open research models
Prime Intellect — decentralized open training
Reka AI — multimodal models
Poolside — sovereign coding models
Black Forest Labs — open image models (Flux)
Aleph Alpha — German sovereign AI
LightOn — European enterprise models
TII — Falcon open models (UAE)
G42 — Gulf sovereign AI
Sarvam AI — Indian sovereign models
Krutrim — Indian AI stack
AI Singapore — SEA-LION models
Together AI — open-model cloud
Cluster 2 — AI Applications & Knowledge Work
This is where the models cash out. The frontier labs built the engine; the companies below are the ones bolting it to a specific, expensive, hated job and charging for the outcome — and they are the fastest-scaling software businesses in history.
9. AI software engineering / coding agents
The shift: from autocomplete to a colleague that ships. Coding is moving from “the IDE suggests the next line” to “you hand a ticket to an agent and it opens a reviewed pull request.” The engineer becomes an orchestrator and reviewer of machine labour, not the typist of every keystroke.
Why now: code is the ideal agent substrate — it has a compiler, a test suite, and a linter, which means the agent gets a hard, automatic verification signal on every attempt. That closes the loop that every other domain is missing. Add context windows large enough to hold a real repo, cheap enough inference to run hundreds of speculative attempts, and the arrival of asynchronous background agents in 2025, and the reliability crossed the line from toy to load-bearing.
Why it’s a startup goldmine: global developer spend is a trillion-dollar pool of salaries, and this is the one category where AI already visibly moves the needle on output. Cursor (Anysphere) went from zero to a reported $500M+ ARR faster than any SaaS company ever, which tells you the willingness to pay is nearly unbounded. The pricing unlock is that you can charge per seat and per task — and eventually meter the work itself, capturing a slice of a $150k engineer rather than a $20 tool budget; the ceiling isn’t the IT line item any more, it’s the engineering payroll. Defensibility is thin at the model layer, which is exactly why the labs keep leapfrogging each other, but it’s real at the workflow layer: the harness that manages context across a sprawling repo, runs the test-fix loop, integrates with the CI/CD and ticketing stack, and — hardest of all — earns the organizational trust to merge without a human is where the moat lives. Distribution matters too: the tool developers already have open is the one they’ll let write their code.
Where the opening is: own the parts of the SDLC the incumbents ignore — the agent that does migrations, dependency upgrades, flaky-test triage, incident response, security-patching, or brownfield legacy modernization end-to-end. The wedge is the unglamorous maintenance backlog every CTO wants to clear but can’t staff; it is a genuine budget line, it has a crisp definition of done, and the incumbents are too busy chasing greenfield autocomplete to want it. There’s a second wedge underneath: the coding agent aimed at non-engineers — the “prompt-to-app” builders letting a PM or a founder ship a real internal tool — which quietly expands the TAM from developers to everyone who ever filed a ticket with IT.
Signal — startups & scaleups to watch:
Cursor (Anysphere) — runaway AI-native IDE
Cognition (Devin) — autonomous coding agent
Windsurf — agentic coding IDE
GitHub Copilot — incumbent pair programmer
Claude Code — terminal coding agent
OpenAI Codex — lab’s coding agent
Poolside — frontier code models
Magic — long-context code models
Lovable — prompt-to-app builder
Replit — cloud IDE plus agent
Bolt (StackBlitz) — prompt-to-web-app
Vercel v0 — generative UI builder
Sourcegraph (Amp) — code search and agent
Tabnine — enterprise code assistant
Augment Code — enterprise coding agent
Factory — autonomous dev droids
Codegen — software-engineering agents
Qodo — code-integrity and test agents
Continue — open-source IDE assistant
Zencoder — enterprise coding agents
Warp — agentic terminal
Aider — open-source CLI coding
Tessl — spec-driven AI software
10. Generative media (video, voice, music, image, 3D)
The shift: the marginal cost of professional-grade content is collapsing to zero. Voice, image, video, music, and now navigable 3D worlds are becoming things you type into existence rather than shoot, record, or hire a studio to make.
Why now: generation quality crossed the “good enough to ship commercially” line in 2024–2025. Video went from uncanny two-second clips to coherent, direction-following minutes; voice became indistinguishable and real-time; music generation reached radio quality. Diffusion and transformer video models plus cheap GPU inference mean a marketer, game studio, or solo creator can produce in an afternoon what used to take a crew a month.
Why it’s a startup goldmine: the addressable market is the entire creative-services and advertising economy — hundreds of billions in production budgets — plus a long tail of content that was never economical to make at all, from personalized ads to indie-game assets to localized training videos. Margins are software margins once you own the model, and demand is voracious because content is the one input every business needs continuously. The disruption target is stock media, dubbing houses, ad agencies, voiceover talent, and eventually parts of film and games. The catch, and it’s the whole game, is defensibility: raw model quality commoditizes within months, and there is always a free open-weight checkpoint nipping at the leader’s heels. Winners own a proprietary distribution loop, a professional workflow, or a rights/data moat — not just a good model — which is why the smart money in this wave bets on the companies wrapping generation in an editor, a pipeline, or a licensing business rather than the ones racing on raw fidelity alone.
Where the opening is: go vertical and own the workflow. The dubbing-and-localization pipeline for studios, the ad-creative engine wired into ad accounts with closed-loop performance feedback, the game-asset pipeline, the enterprise avatar/training-video platform. Boring, workflow-locked, contract-backed beats another general text-to-video toy. The other durable position is the rights layer — the company that licenses real artists’ voices and likenesses and pays them, which turns the biggest legal liability of the category into a moat and a distribution partnership. Whoever cleanly solves consent and provenance for synthetic media sells a compliance rail to everyone else in the wave.
Signal — startups & scaleups to watch:
ElevenLabs — voice and audio platform
Runway — AI video generation
Pika — AI video generation
Suno — AI music generation
Udio — AI music generation
Synthesia — AI avatar video
HeyGen — AI avatar video
World Labs — 3D world models
Krea — creative generation workflow
Freepik — creator asset platform
Luma AI — video and 3D generation
Midjourney — image generation
Black Forest Labs — Flux image models
Stability AI — open image models
Ideogram — text-in-image generation
Leonardo AI — image and asset generation
Cartesia — real-time voice models
Higgsfield — cinematic AI video
Descript — AI audio/video editing
Captions — AI video creation
Hedra — character video generation
Recraft — design and graphics generation
Kling AI — AI video model
Resemble AI — voice cloning
Photoroom — AI image editing
11. AI in clinical care (ambient scribing, care ops)
The shift: the computer stops being the thing between the doctor and the patient. Ambient AI listens to the visit and writes the note, the order, the referral, and the billing code — dissolving the documentation burden that drives half of physician burnout.
Why now: speech recognition finally handles cross-talk, accents, and medical jargon in noisy rooms, and LLMs turn a messy transcript into a structured, billable, defensible clinical note. The ROI is immediate and measurable — clinician time saved, more patients seen, cleaner coding — so hospital systems that never move fast are signing enterprise deals in months, not years. This is the AI use case healthcare actually trusts because a human still signs off.
Why it’s a startup goldmine: US healthcare is a $4.5T system whose administrative overhead is a national scandal; documentation and revenue-cycle work alone is a massive line item. Ambient scribing is the beachhead, but the real prize is expanding into care operations — coding, prior authorization, referral management, quality reporting. Pricing is per-clinician-per-month with land-and-expand into the whole revenue cycle. Defensibility comes from EHR integration depth (Epic/Cerner), health-system trust, clinical accuracy data, and the compliance moat that keeps generic chatbots out.
Where the opening is: move past the note into the money. The agent that closes the loop on coding and prior authorization — the workflows that literally determine whether a hospital gets paid — is stickier and higher-value than transcription, which is racing toward commodity as the labs offer scribing for free. There is a parallel opening on the patient side: the ambient system for nursing, home health, and behavioral care, where documentation burden is just as brutal but the incumbents have barely arrived. Land on the ambient mic, expand across the entire revenue cycle, and you become the operating layer of the clinic rather than a feature.
Signal — startups & scaleups to watch:
Abridge — enterprise ambient scribing
Ambience Healthcare — scribing plus coding
Suki — AI clinical assistant
Nuance (Microsoft DAX) — incumbent ambient scribe
OpenEvidence — clinical decision support
Commure — health-system operating layer
Nabla — ambient clinical notes
DeepScribe — ambient medical scribe
Corti — clinical conversation AI
Heidi Health — ambient scribe
Freed — AI scribe for clinicians
Augmedix — ambient documentation
Tortus — clinical AI assistant
Hippocratic AI — patient-facing care agents
Navina — AI clinical copilot
Notable — healthcare workflow automation
Innovaccer — health data and AI
Cohere Health — prior-authorization AI
CodaMetrix — autonomous medical coding
Rad AI — radiology reporting AI
Aidoc — radiology imaging AI
Viz.ai — care coordination AI
Ellipsis Health — vocal-biomarker behavioral AI
12. AI drug discovery & techbio
The shift: biology becomes a design problem. Instead of screening millions of compounds by brute force, models propose molecules, proteins, and antibodies that are likely to work — turning the wet lab from a search engine into a fabrication line for pre-designed hypotheses.
Why now: AlphaFold was biology’s GPT moment, and the generation of protein-design and molecular models that followed can now generate binders and structures that hold up experimentally. Simultaneously, lab automation and cheap sequencing/synthesis close the design–build–test–learn loop fast enough that the AI actually learns from its own experiments. The bottleneck shifts from “can we imagine this molecule” to “how fast can we test and feed the data back.”
Why it’s a startup goldmine: pharma R&D is a $250B+ annual spend defined by a brutal statistic — most drug candidates fail, and each failure costs years and hundreds of millions. Even a modest improvement in hit rate or timeline is worth a fortune, which is why the platform deals with big pharma are enormous. The margin unlock is owning IP on the molecules themselves, not just selling software. Defensibility is the AI + wet-lab flywheel: proprietary experimental data generated in-house that no competitor and no foundation model can replicate.
Where the opening is: don’t try to be a full-stack pharma company on seed money. The sharp wedge is a specialized model-plus-data platform for a hard modality — antibody design, protein engineering, delivery, or a specific disease area — sold to pharma as a discovery engine while keeping equity or royalties in the assets it generates. The lean-founder version is even narrower: sell the design tool to the hundreds of biotechs and academic labs that can’t build their own model, generating proprietary usage data from every customer’s experiments. That data flywheel, not the model architecture, is what compounds into a moat while the underlying models commoditize.
Signal — startups & scaleups to watch:
Xaira Therapeutics — model-native drug discovery
Isomorphic Labs — DeepMind drug-discovery spinout
EvolutionaryScale — protein language models
Chai Discovery — molecular structure models
Cradle — protein-engineering platform
Generate Biomedicines — generative protein therapeutics
Recursion — AI drug-discovery platform
Insilico Medicine — generative drug discovery
Genesis Therapeutics — molecular AI models
Iambic Therapeutics — AI drug design
Absci — generative antibody design
Profluent — protein-design AI
Latent Labs — generative protein design
Basecamp Research — biodiversity data plus AI
Enveda — AI natural-product drugs
Bioptimus — biology foundation model
Nabla Bio — AI antibody design
Charm Therapeutics — 3D deep-learning drugs
Atomwise — structure-based AI screening
Deep Genomics — RNA-therapeutics AI
Relay Therapeutics — protein-motion drug design
13. AI for science / research automation
The shift: the scientific method gets an autonomous operator. AI moves from analyzing data to running the loop — generating hypotheses, designing experiments, reading the entire literature, and in robotic labs, executing and iterating without a human at each step.
Why now: models can read and synthesize far more literature than any human researcher, reason across disciplines, and now plan multi-step experimental campaigns. Cloud labs and lab robotics make physical execution programmable. Reasoning models that can check their own work make the difference between a plausible-sounding hallucination and a real, reproducible result — the gating requirement for anything scientific.
Why it’s a startup goldmine: R&D is a multi-trillion-dollar global activity — pharma, materials, chemicals, energy, academia — and it is stunningly inefficient, bottlenecked on scarce PhD attention. Automating even the literature-review, hypothesis-generation, and experiment-design layers compresses discovery timelines that translate directly into value. This category also sells to a rare buyer: national labs, universities, and corporate R&D with real budgets and a mandate to move faster. Defensibility is proprietary experimental data and integration into the physical lab.
Where the opening is: start with the layer that has a verification signal and a clear buyer — automated literature synthesis and the “AI research assistant” for a specific field (materials, chemistry, bio), then earn the right to design and run experiments in a cloud lab. The wedge is the grunt work that burns most of a scientist’s week: reading, summarizing, coding analysis, and searching for what’s already been tried. Own that, accumulate the proprietary result data no one else has, and you graduate from tool to autonomous discovery platform — the highest-value position of all.
Signal — startups & scaleups to watch:
FutureHouse — autonomous biology research
Lila Sciences — scientific superintelligence lab
Periodic Labs — AI-run physical experiments
Emerald Cloud Lab — programmable wet lab
Orbital Materials — AI materials discovery
Elicit — AI research assistant
Consensus — AI literature search
Undermind — AI literature co-researcher
SciSpace — research reading and writing AI
Scite — citation-analysis AI
ResearchRabbit — literature discovery
Causaly — biomedical research AI
Iris.ai — scientific research assistant
Scholarcy — paper summarization
Julius — AI data analysis
Semantic Scholar (Ai2) — scholarly search engine
14. Legal & professional-services AI
The shift: the billable hour meets its solvent. Legal, tax, audit, and consulting are made of expensive, structured, text-and-judgment work — exactly what agents do — and the industry is moving from “AI helps a lawyer research faster” to “an agent drafts, reviews, and diligences the document.”
Why now: models became good enough at long-document reasoning, citation, and structured drafting to be trusted on real matters — with a professional in the loop for liability. The economics are irresistible: professional services runs on leveraging junior labor at enormous markups, and AI attacks that leverage model directly. Elite firms, historically the slowest adopters, are now buying because their clients demand it and competitors are.
Why it’s a startup goldmine: legal services alone is a ~$1T global market; add tax, audit, and consulting and you’re looking at the richest per-hour knowledge work on earth. The pricing unlock is you can charge as a fraction of the labor replaced — a fortune relative to a per-seat SaaS license — because the alternative is a $1,000/hour associate. Defensibility is proprietary workflow depth, becoming the system of record for a matter type, the trust and security posture that elite firms require, and the liability structure clients won’t take on themselves.
Where the opening is: own a specific high-value document workflow end-to-end — M&A diligence, contract lifecycle, patent prosecution, tax provisioning, compliance filings. Or skip the incumbent firms entirely and go direct to the in-house corporate legal and finance teams who’d rather buy the agent than the outside counsel — a route that also sidesteps the billable-hour incentive problem, since a law firm is structurally conflicted about a tool that shrinks its own hours. The most disruptive version of all is the agent sold to the underserved: the litigation platform for plaintiffs’ firms, the tax agent for small businesses, the compliance agent for companies that never could afford a Big Four engagement.
Signal — startups & scaleups to watch:
Harvey — legal AI for elite firms
Legora — transactional legal AI
Robin AI — contract review AI
Hebbia — finance and pro-services document AI
EvenUp — personal-injury demand automation
Eudia — in-house legal AI
Crosby — AI-powered law firm
Spellbook — contract drafting AI
Luminance — contract intelligence
Ironclad — contract lifecycle plus AI
Leya — legal workflow AI
Supio — litigation and personal-injury AI
Darrow — legal intelligence for plaintiffs
LegalOn — contract review AI
Clio — legal practice management plus AI
Norm Ai — regulatory compliance agents
Rogo — finance research AI
Basis — accounting AI agents
Definely — legal drafting tools
Genie AI — legal document automation
Filevine — legal case management
Wordsmith — in-house legal AI
CoCounsel (Casetext) — legal research AI
15. Customer support & sales AI (CX/GTM agents)
The shift: the front office gets automated, not assisted. Customer support and sales development are moving from “AI suggests a reply to the human agent” to “the agent resolves the ticket or books the meeting itself” — end-to-end, measured on outcomes.
Why now: support is a near-perfect agent domain — high volume, repetitive, with a clear success signal (was the issue resolved?) and existing knowledge bases to ground on. Models crossed the reliability bar for multi-turn resolution with tool use in 2024–2025, and the willingness of enterprises to route real customers to an AI followed once resolution rates and guardrails proved out. Sales followed the same curve.
Why it’s a startup goldmine: support and contact-center is a multi-hundred-billion-dollar labor market — much of it outsourced to BPOs precisely because it’s costly and painful — and it’s the clearest place in all of software to price AI as labor replaced. Sierra’s outcome-based pricing — charging per successful resolution rather than per seat — is the template for the whole cycle: your revenue scales with the payroll you remove, not with the license count, which means a single enterprise logo can be worth what a hundred SaaS seats used to be. The disruption target is the BPOs, the call centers, and the per-seat helpdesk software that only assists the human. Defensibility is integration depth into the systems of record (CRM, ticketing, billing, order management), the resolution-quality data flywheel that makes your agent better than a competitor’s on the same model, and — decisively — owning the outcome and the liability that comes with letting an AI speak for the brand.
Where the opening is: go vertical or go outcome-priced. The support agent purpose-built for a hard domain (healthcare, fintech, telco) with the compliance and integrations that generic bots lack — or the GTM agent that runs the full outbound-to-booked-meeting motion, not just drafts emails. The voice channel is a specific, wide-open front: the AI phone agent that actually holds a multi-turn call, handles interruptions, and takes real actions is only now becoming reliable, and it attacks the single largest pool of front-office labor — the call center — head-on. Whoever owns the outcome, the integration, and the resolution-quality data becomes impossible to rip out.
Signal — startups & scaleups to watch:
Sierra — outcome-priced CX agent
Decagon — enterprise support automation
Intercom (Fin) — incumbent support agent
Clay — GTM data and enrichment
11x — AI sales reps
Cresta — contact-center AI
Parloa — voice contact-center AI
Ada — customer-service automation
Forethought — support AI
Lorikeet — complex support agent
Crescendo — CX outcome service
PolyAI — voice customer assistants
Cognigy — enterprise conversational AI
Artisan — AI sales BDR
Unify — GTM automation
Qualified — pipeline and AI SDR
Bland AI — AI phone calls
Retell AI — voice-agent platform
Vapi — voice AI developer platform
Observe.AI — contact-center intelligence
Maven AGI — enterprise support agent
Ema — universal AI employee
Replicant — voice contact-center automation
16. AI-native consumer apps & companions
The shift: software you talk to instead of operate. A new generation of consumer products is being built with the model as the core interaction — companions, tutors, therapists, coaches, assistants — creating relationships and habits, not just features. This is the first genuinely new consumer paradigm since the mobile app.
Why now: models became conversational, memory-capable, multimodal, and cheap enough to run in a consumer free tier. Voice made them ambient. And crucially, a generation now defaults to talking to AI for companionship, learning, and emotional support — behaviors that didn’t exist three years ago and are now measured in hours per day. ChatGPT itself became the fastest consumer app in history, proving the appetite.
Why it’s a startup goldmine: consumer is winner-take-most with venture-scale upside — a single breakout app can reach hundreds of millions of users and command real subscription revenue at software margins. The categories in play (companionship, education, wellness, personal productivity) are each enormous, and AI makes previously impossible experiences — an infinitely patient tutor, an always-available companion — suddenly real. The disruption target is everything from Duolingo to dating to therapy. Defensibility is the hardest question: distribution, brand, personalization data, memory, and network effects, because the model alone is rentable by anyone.
Where the opening is: own a specific relationship and the data that deepens it over time. Companionship, an AI tutor that actually moves measurable outcomes, a wellness/therapy companion, or a personalized life assistant — where accumulated memory and habit make switching painful. The wedge is emotional stickiness, not feature count: a product the user talks to daily, that remembers, and that gets more valuable the longer you use it. The other opening is native multimodality — the always-on voice or wearable companion that lives in your ear or on your body, a form factor the incumbent chat apps aren’t built for and that a focused startup can define.
Signal — startups & scaleups to watch:
Character.AI — companionship at scale
Replika — original AI companion
ChatGPT (OpenAI) — default consumer assistant
Perplexity — AI answer engine
Grok (xAI) — consumer AI assistant
Google Gemini — consumer AI assistant
Meta AI — assistant and companion
Pi (Inflection) — personal AI companion
Speak — AI language tutor
Praktika — AI avatar language tutor
Duolingo — AI language learning
Khan Academy (Khanmigo) — AI tutor
Synthesis — AI tutor for kids
Slingshot AI (Ash) — AI therapy companion
Ada Health — AI symptom and health
Wysa — mental-health chatbot
Headspace — AI-assisted wellness
Friend — wearable AI companion
Tolan (Portola) — voice AI companion
Nomi — AI companion
Kindroid — AI companion
Chai — AI chat companions
Cluster 3 — Physical & Frontier Tech
The last decade of venture capital hid from atoms. This decade runs straight at them: the same AI that conquered language is now conquering motion, the West is re-industrializing under geopolitical duress, and compute itself is being rebuilt out of new physics. These are harder, more capital-intensive, and slower than SaaS — which is exactly why the winners will be near-impossible to copy.
17. Humanoid robots & embodied AI
The shift: Robotics is moving from programmed to learned — a general-purpose body that can be taught tasks the way you’d teach a person, rather than hand-coded for one.
For fifty years a robot could only do what an engineer explicitly scripted inside a fixed cage. The vision-language-action (VLA) model breaks that ceiling: a single neural net that maps camera pixels and a spoken instruction directly to motor commands, and — crucially — generalizes to objects and tasks it never saw in training. The body becomes a peripheral; the value migrates to the policy running on it.
Why now: Three curves crossed at once. The transformer that solved language turned out to solve action too, given enough demonstration data. Teleoperation, simulation, and internet-scale human video finally supply that data at the volume VLAs need. And China’s supply chain collapsed the cost of actuators, harmonic drives, and motors by an order of magnitude, so a capable humanoid platform now costs tens of thousands, not millions. On the demand side, structural labor shortages in logistics, elder care, and manufacturing have made “a body that shows up” worth real money.
Why it’s a startup goldmine: The addressable market is not the robotics budget — it’s the global wage bill for physical labor, tens of trillions of dollars. Pricing flips from capex to robot-as-a-service: charge a monthly rate below the loaded cost of the human shift it replaces, and the ROI math sells itself. Defensibility lives in the data flywheel — every deployed robot streams manipulation data back into the foundation policy, so the leader’s robots get smarter faster than a follower can catch up. Fleet learning, not the sheet metal, is the moat.
Where the opening is: Be honest — building the humanoid itself is brutally capital-hungry and slow, a game for the well-funded. The sharper wedges: the robot foundation model / data layer (be the “OpenAI for actions” that others license), teleoperation-to-autonomy pipelines that convert cheap human remote-operation into training data, simulation and eval infrastructure, and hands/tactile-sensing components. Or go narrow: one dexterous task in one industry where the ROI is undeniable.
Signal — startups & scaleups to watch:
Figure — humanoids for logistics
1X — home-oriented humanoids
Physical Intelligence — robot foundation models
Skild AI — cross-embodiment robot brain
Tesla — Optimus vertically integrated humanoid
Apptronik — Apollo warehouse humanoid
Agility Robotics — Digit bipedal robot
Sanctuary AI — dexterous humanoid hands
Boston Dynamics — Atlas humanoid
Unitree — low-cost humanoids, quadrupeds
Covariant — foundation-model manipulation
Field AI — foundation models for robots
Neura Robotics — cognitive service robots
Robust AI — collaborative warehouse robots
Mentee Robotics — AI-native humanoid
Wayve — embodied driving intelligence
18. Industrial, warehouse & logistics automation
The shift: The boring, high-ROI cousin of humanoids — purpose-built machines that pick, sort, move, and inspect inside the four walls of a warehouse or factory, reaching production scale today while humanoids are still demos.
This is where embodied AI actually ships revenue. A picking arm that clears bins, an autonomous forklift, a mobile robot that ferries totes — narrow, unglamorous, and deployed by the thousand. The AI unlock (grasping novel SKUs, adapting to messy real environments) is the same VLA advance driving humanoids, but the form factor is constrained enough to be reliable now.
Why now: E-commerce permanently raised throughput expectations while warehouse labor got scarcer, more expensive, and harder to retain — turnover in fulfillment centers routinely tops 100% a year. Perception models finally handle the “any object, any orientation” problem that defeated a decade of rules-based vision systems. And post-COVID, operators treat automation as resilience, not just cost-cutting.
Why it’s a startup goldmine: Warehouse automation is a multi-hundred-billion-dollar market with a labor problem that only worsens. The pricing unlock is again RaaS — Amazon-scale operators will pay per-pick or per-hour rather than buy capex, converting robotics into a recurring, high-margin software-like revenue stream. Defensibility comes from integration depth: once your robots and warehouse-execution software are wired into a customer’s operations and SLAs, ripping them out is a nightmare. Switching costs are the moat.
Where the opening is: Own a specific station in the flow — depalletizing, induction, each-picking, truck unloading — with a system that beats human unit economics on day one. The best wedge is often the software orchestration layer that makes heterogeneous robot fleets work together, since most warehouses will run mixed hardware from many vendors. Capital intensity is real but far lower than humanoids: you’re selling a bounded, provable ROI.
Signal — startups & scaleups to watch:
Symbotic — warehouse system automation
Dexterity — robotic palletizing, picking
Ambi Robotics — AI parcel sortation
Locus Robotics — warehouse mobile robots
Exotec — Skypod goods-to-person
Geek+ — autonomous fulfillment robots
GreyOrange — fulfillment robotics, software
Berkshire Grey — pick-and-pack robotics
RightHand Robotics — piece-picking arms
Nimble Robotics — autonomous fulfillment
Pickle Robot — autonomous truck unloading
Mujin — intelligent robot controllers
Plus One Robotics — parcel induction robots
Vecna Robotics — autonomous forklifts, pallet jacks
inVia Robotics — goods-to-person automation
Third Wave Automation — autonomous forklifts
Dexory — warehouse inventory robots
Verity — autonomous inventory drones
Agility Robotics — Digit material handling
19. Autonomous vehicles & mobility
The shift: After a decade of overpromising, autonomy quietly started working — robotaxis now run driverless, paid, at scale in real cities, and the question flipped from “can it work” to “how fast does it scale.”
Waymo crossed the threshold from science project to service: millions of paid, fully driverless rides, expanding city by city. That changes the entire narrative. Autonomy is no longer a bet on a future — it’s an operating business whose only remaining variables are cost curve, geographic expansion, and regulatory pace.
Why now: Foundation models and vastly better perception cracked the long tail of edge cases that stranded the field for years. Sensor costs (especially lidar) fell dramatically. And an end-to-end, learning-based stack — pixels to controls — is proving more robust than the brittle modular pipelines of the 2010s, echoing the same VLA thesis driving robotics writ large.
Why it’s a startup goldmine: The prize is trillions in global mobility and freight spend. But the robotaxi race itself is a capital bloodbath won by giants (Waymo, and Tesla’s camera-only bet). The startup money is in the arms dealers and the adjacent lanes: the simulation and validation tooling every AV program must buy, autonomous trucking on the simpler highway domain, and the sensor/compute components. Margins in tooling are pure software; margins in the fleet are a decade of capex away.
Where the opening is: Sell picks and shovels to the AV industry — Applied Intuition’s playbook of simulation, validation, and developer tooling that every OEM and autonomy team needs regardless of who wins. Or attack a constrained domain (highway trucking, yard trucks, mining, ports) where the driving problem is bounded and the labor shortage is acute. Avoid trying to out-Waymo Waymo in open urban robotaxi — that ship has sailed for the underfunded.
Signal — startups & scaleups to watch:
Waymo — proven driverless robotaxi
Applied Intuition — AV simulation, tooling
Aurora — autonomous trucking
Kodiak — driverless trucks
Waabi — generative AI trucking
Nuro — licensing its autonomy driver
Zoox — purpose-built robotaxi
Gatik — middle-mile autonomous delivery
Pony.ai — robotaxi and trucking
WeRide — robotaxi, robobus
Wayve — end-to-end driving AI
Mobileye — ADAS and autonomy
May Mobility — autonomous shuttles
Oxa — autonomy software platform
Einride — autonomous electric freight
Plus — autonomous trucking software
Torc Robotics — self-driving trucks
Helm.ai — AV foundation software
Foretellix — AV verification, validation
20. Drones & counter-drone systems
The shift: Ukraine turned the cheap FPV drone into the defining weapon of modern war and, simultaneously, exposed the West’s near-total lack of affordable defenses against them — creating twin gold rushes in autonomous drones and counter-drone systems.
The battlefield of the 2020s is a drone battlefield: swarms of sub-$1,000 quadcopters destroying multimillion-dollar armor, ISR drones providing persistent overwatch, and loitering munitions replacing artillery. Every military on earth is now scrambling to both field these systems at scale and defeat the enemy’s. It’s the fastest-moving procurement shift in a generation.
Why now: Ukraine is a live, three-year proof of concept that rewrote doctrine. Chinese commercial drone dominance (DJI) became a national-security liability the West must domestically replace. Cheap autonomy — onboard AI for navigation, targeting, and jamming resistance — turned drones from remote-controlled toys into autonomous agents. And defense budgets are re-opening to fast, cheap, attritable hardware after decades of exquisite, unaffordable platforms.
Why it’s a startup goldmine: Defense procurement is a trillion-dollar arena finally willing to buy from startups, thanks to Anduril proving the model. Attritable drones invert the classic defense cost curve — you want the unit to be cheap and expendable, which favors software-defined, mass-manufacturable designs over gold-plated primes. Counter-drone (detection, tracking, kinetic and electronic kill) is an almost greenfield market with desperate, immediate demand. Defensibility is autonomy software plus defense-grade manufacturing plus the security clearances and program relationships that take years to build.
Where the opening is: Two clean wedges — autonomous drone platforms built for attritable mass (not exquisite one-offs), and the counter-UAS layer where Western militaries are most exposed. Software-defined autonomy that works under jamming and GPS-denial is the durable edge. Capital intensity is high and sales cycles are long, but non-dilutive government funding and committed offtake soften the curve.
Signal — startups & scaleups to watch:
Anduril — Lattice autonomy, drones
Shield AI — Hivemind GPS-denied autonomy
Saronic — autonomous surface vessels
Skydio — autonomous reconnaissance drones
Helsing — AI defense, strike drones
Quantum Systems — reconnaissance UAS
AeroVironment — Switchblade loitering munitions
Mach Industries — attritable munitions, drones
Auterion — open drone operating system
Zipline — autonomous delivery drones
Percepto — autonomous industrial drones
Wingtra — mapping and survey drones
Dedrone — counter-drone detection
DroneShield — counter-UAS defeat systems
Fortem Technologies — DroneHunter counter-UAS
Epirus — directed-energy counter-drone
21. The space economy (launch, satellites, defense space)
The shift: Cheap, reusable launch collapsed the cost of reaching orbit by roughly 20x, turning space from a government-only frontier into a commercial platform layer — and a contested military domain.
When it costs a fraction of what it used to per kilogram to orbit, everything downstream becomes economically viable that wasn’t before: massive satellite constellations, in-space manufacturing, Earth observation at continuous cadence, and orbital defense. SpaceX didn’t just win launch — it created the enabling infrastructure for an entire industry, the way AWS enabled a generation of software companies.
Why now: Reusability is proven and being pushed toward full reuse, driving cost toward a further order-of-magnitude drop. Starlink demonstrated that a space-based service can be a real, cash-generating consumer and defense business, not a science mission. And great-power competition has reopened defense space budgets — resilient constellations, space domain awareness, and missile tracking are urgent national priorities.
Why it’s a startup goldmine: The space economy is on a trajectory to $1T+ by the 2030s, and the application layer is wide open now that launch is commoditized. Earth observation with AI analytics, satellite communications, in-space logistics, and defense payloads all have real customers and, increasingly, real revenue. The margin unlock is that launch access is now a purchasable commodity — you no longer need to be a nation-state. Defensibility comes from constellation network effects, exclusive data, and defense contracts.
Where the opening is: Don’t build a rocket unless you’re SpaceX-scale funded. Build on top of cheap launch: Earth-observation data and analytics, satellite servicing and logistics, defense-space payloads and space domain awareness, and ground-segment/software infrastructure. The picks-and-shovels layer — components, propulsion, software — is capital-lighter than launch and still riding the same wave.
Signal — startups & scaleups to watch:
SpaceX — Starship, Starlink platform
Rocket Lab — launch and space systems
Firefly Aerospace — launch and lunar landers
Relativity Space — 3D-printed rockets
Stoke Space — fully reusable rocket
K2 Space — large low-cost satellites
Impulse Space — orbital transfer logistics
Vast — commercial space stations
Varda Space — in-space manufacturing
Astranis — small GEO comms satellites
Apex Space — productized satellite buses
Muon Space — Earth-observation constellations
Planet Labs — daily Earth imaging
Albedo — ultra-high-resolution VLEO imagery
ICEYE — SAR Earth observation
Capella Space — SAR satellite imagery
Umbra — high-resolution SAR
True Anomaly — space-superiority spacecraft
Turion Space — space domain awareness
Astroscale — debris removal, servicing
Axiom Space — commercial space station
Sierra Space — Dream Chaser, stations
22. Advanced manufacturing & Western reshoring
The shift: Geopolitics ended the era of frictionless offshoring — the West is now re-industrializing under duress, and software-defined, AI-native factories are the only way to make domestic manufacturing economically competitive.
Decades of moving production to China created a strategic dependency the West is now urgently unwinding: semiconductors, batteries, pharmaceuticals, defense components, critical minerals. But reshoring to high-wage countries only pencils out if the factory is radically more automated and software-driven than its offshore predecessor. The reshoring wave and the automation wave are the same wave.
Why now: COVID and the Taiwan risk exposed supply-chain fragility as an existential vulnerability. Industrial policy — the CHIPS Act, the IRA, European sovereignty pushes — is pouring hundreds of billions into domestic capacity. Simultaneously, AI-driven robotics, generative design, and modern manufacturing software finally make high-mix, low-volume domestic production viable. Labor scarcity forces the automation that makes reshoring affordable.
Why it’s a startup goldmine: Manufacturing is a multi-trillion-dollar sector running on decades-old software and manual processes — a target-rich environment for AI-native disruption. The disruption is replacing brittle legacy MES/ERP, manual quality inspection, and human-limited process control with software and robotics. The margin unlock: software attach to physical production, and the ability to win contracts that must be domestic for security reasons. Defensibility is deep process expertise, regulatory qualification, and integration into physical supply chains that are painful to switch.
Where the opening is: AI-native factory software (MES, quality, scheduling), machine-vision inspection that replaces manual QA, generative and simulation-driven design tools, and vertically integrated “new-prime” manufacturers that own production of a critical component (chips, batteries, munitions, drones). Capital intensity varies wildly — software-thin at one end, fab-heavy at the other; pick your altitude deliberately.
Signal — startups & scaleups to watch:
Hadrian — automated aerospace parts factories
Machina Labs — robotic sheet-metal forming
Anduril — Arsenal defense manufacturing
Instrumental — AI manufacturing inspection
Bright Machines — software-defined assembly
Path Robotics — autonomous robotic welding
Standard Bots — affordable robotic arms
Divergent — digital metal manufacturing
Gecko Robotics — industrial inspection robots
Tulip — frontline operations apps
Chef Robotics — food assembly robotics
Nominal — hardware test, telemetry
Ursa Major — rocket propulsion manufacturing
MP Materials — rare earths and magnets
Niron Magnetics — rare-earth-free magnets
Redwood Materials — battery materials recycling
Sila — silicon battery anodes
Form Energy — iron-air grid batteries
23. Quantum computing
The shift: Quantum computing crossed from physics experiment toward engineering roadmap — error correction is finally working, and the first logical, fault-tolerant qubits are appearing, putting commercial utility on a credible timeline.
For years quantum was a perpetual “ten years away.” That’s changing: demonstrations of quantum error correction that improves as you scale, credible fault-tolerance roadmaps from multiple hardware approaches, and real capital committing to specific architectures. The machine that breaks encryption and simulates molecules is no longer purely hypothetical — it’s an engineering program with milestones.
Why now: Error correction is the whole game, and 2024–2025 produced the first convincing evidence that adding physical qubits actually reduces logical error rates — the threshold theorem working in practice. Photonic, neutral-atom, and superconducting approaches each have a path to scale. And the security implications (”harvest now, decrypt later”) are forcing governments and enterprises to fund the field seriously regardless of exact timing.
Why it’s a startup goldmine: The prize is enormous and winner-take-few: molecular simulation for drugs and materials, optimization, and cryptography represent markets worth hundreds of billions. Whoever reaches useful fault-tolerance first owns a genuine platform monopoly. The margin structure is extreme — a working fault-tolerant machine is close to irreplaceable IP. Defensibility is the hardest and deepest in all of tech: physics, fabrication, and error-correction expertise that takes a decade and a fortune to assemble.
Where the opening is: Be honest — the core hardware race is a capital-and-physics marathon for the deep-pocketed (PsiQuantum’s photonic bet, QuEra’s neutral atoms, IBM/Google internally). The startup-friendly wedges: the enabling layer (control systems, cryogenics, error-correction software, quantum-classical middleware), post-quantum cryptography for the security transition happening now, and quantum-algorithm/application software that will be valuable the moment hardware matures. Sell shovels into the race.
Signal — startups & scaleups to watch:
PsiQuantum — photonic fault-tolerance at scale
QuEra — neutral-atom logical qubits
IonQ — trapped-ion quantum computers
Quantinuum — trapped-ion quantum systems
Rigetti — superconducting quantum
Atom Computing — neutral-atom qubits
Pasqal — neutral-atom quantum
Infleqtion — cold-atom quantum tech
Alice & Bob — cat-qubit error correction
Oxford Ionics — trapped ions on chips
Diraq — silicon spin qubits
Quantum Brilliance — diamond quantum
Nord Quantique — error-corrected qubits
SEEQC — digital quantum control
Quantum Machines — quantum control systems
Q-CTRL — quantum control software
Classiq — quantum algorithm software
Multiverse Computing — quantum-inspired software
Xanadu — photonic quantum computing
SandboxAQ — post-quantum crypto, sensing
PQShield — post-quantum cryptography
QuSecure — post-quantum security
24. Photonics & new compute substrates
The shift: The AI compute crunch — power, heat, and bandwidth walls that silicon can no longer scale past — is forcing a rethink of the substrate itself, and light-based (photonic) computing and interconnect is the leading candidate to break through.
AI’s insatiable compute demand is colliding with physics: data centers are constrained by power delivery, cooling, and the bandwidth of moving data between chips. Electrons over copper are hitting hard limits. Photonics — computing and, especially, moving data with light — offers dramatically lower energy per bit and higher bandwidth, exactly where the bottleneck now is. The substrate, not just the model, becomes a frontier.
Why now: The AI buildout made data-center power and interconnect the limiting factor, with training clusters constrained by how fast chips can talk to each other. Silicon photonics manufacturing matured enough to integrate optics with electronics at scale. And energy cost became a first-order economic driver of AI — anything that cuts joules per token or per training step is worth billions. The pain is acute and the buyers are the deepest-pocketed companies on earth.
Why it’s a startup goldmine: The AI-infrastructure market is measured in the hundreds of billions and every hyperscaler is desperate for more efficient compute and interconnect. Optical interconnect and co-packaged optics disrupt the electrical networking that dominates data centers today; optical compute could eventually disrupt the GPU itself for specific workloads. The margin unlock is enormous — even incremental efficiency at data-center scale translates to gigawatts and billions saved. Defensibility is hard photonics IP, foundry relationships, and integration know-how.
Where the opening is: Optical interconnect and co-packaged optics are the near-term, revenue-now wedge — the bottleneck is bandwidth between chips, and that’s where photonics wins first and most cleanly. Longer-dated but larger: analog optical compute for matrix multiplication (the core of AI). Also adjacent substrates — analog in-memory compute and novel materials. Capital intensity is high and foundry-dependent, but hyperscaler demand and strategic investment de-risk the path.
Signal — startups & scaleups to watch:
Lightmatter — photonic interconnect, compute
Ayar Labs — optical I/O chiplets
Celestial AI — photonic memory fabric
Lightelligence — photonic computing
POET Technologies — optical interposer platform
d-Matrix — digital in-memory compute
EnCharge AI — analog in-memory AI
Mythic — analog compute chips
Groq — deterministic inference LPUs
Cerebras — wafer-scale AI compute
SambaNova — reconfigurable dataflow chips
Tenstorrent — RISC-V AI compute
Etched — transformer-specialized ASIC
Axelera AI — in-memory edge AI
Extropic — thermodynamic computing
BrainChip — neuromorphic Akida processor
Innatera — neuromorphic sensing chips
Cluster 4 — Energy, Climate & Bio
The two hardest problems on Earth — where the electrons come from and how the body decays — are becoming venture problems, because AI turned power into the binding constraint on the economy and because biology finally became programmable. This cluster is where atoms, not just bits, get repriced, and where the winners look less like the cleantech flameouts of 2010 and more like software companies wrapped around a physical asset. The unifying lesson across all eight topics: the market is no longer the bottleneck — the demand is contracted, subsidized, or desperate — so the game is execution, defensibility, and finding the capital-light wedge into a capital-heavy domain.
25. Nuclear (SMRs) & fusion
The shift: After thirty years of stagnation, firm zero-carbon baseload is being re-engineered as a manufactured product — small modular reactors built on an assembly line, and fusion moving from a physics experiment to an engineering roadmap with commercial off-take.
The old nuclear model was a one-off, decade-long, over-budget megaproject. The new model is factory-built reactors of 50–300 MW that ship as modules, plus fusion machines whose economics finally close on high-temperature superconducting magnets. Nuclear stopped being a climate story and became an electricity-supply story.
Why now: AI data centers created a step-change in electricity demand that no other source can serve cleanly, firmly, and 24/7. Hyperscalers signed the first-ever corporate nuclear off-take deals in 2024–2025 — Microsoft reviving Three Mile Island, Amazon and Google contracting SMR fleets before a single unit is built. Regulatory tailwinds (ADVANCE Act, DOE loan guarantees), a decade of high-temperature-superconductor magnet progress, and desperate utilities have collapsed the willingness-to-pay barrier. For the first time in a generation, the customer is pulling the technology forward rather than the technology begging for a market — buyers now sign contracts for power that does not yet exist.
Why it’s a startup goldmine: The global power market is measured in the tens of trillions; firm clean baseload is the scarcest slice of it and commands premium pricing precisely because intermittent renewables can’t provide it. A working SMR or fusion plant disrupts the entire merchant-power and PPA stack, and the margin unlock is the shift from bespoke construction to learning-curve manufacturing — each unit cheaper than the last. Defensibility is brutal and real: nuclear licenses, fuel supply chains, magnet IP, and multi-decade off-take contracts are the deepest moats in energy.
Where the opening is: Most founders shouldn’t build a reactor. The wedge is the picks and shovels — advanced fuel (HALEU) supply, digital-twin and simulation software that compresses licensing timelines, reactor-grade component manufacturing, and the software layer that matches new firm generation to data-center load. Whoever de-risks NRC licensing with software owns a chokepoint.
The deeper point: the last time capital tried nuclear, it drowned in construction risk and public fear. This time the demand is contracted, the units are factory-repeatable, and the buyers are the most creditworthy companies on Earth. That inversion of who bears the risk is what makes it investable.
Signal — startups & scaleups to watch:
Commonwealth Fusion Systems — HTS-magnet tokamak (SPARC)
Helion Energy — direct-electricity fusion
TAE Technologies — beam-driven fusion
Zap Energy — sheared-flow Z-pinch fusion
Type One Energy — stellarator fusion
Tokamak Energy — spherical-tokamak fusion
Proxima Fusion — stellarator fusion, Germany
Marvel Fusion — laser inertial fusion
Xcimer Energy — laser inertial fusion
General Fusion — magnetized target fusion
First Light Fusion — projectile inertial fusion
Kairos Power — molten-salt SMR
X-energy — pebble-bed HALEU SMR
Oklo — fast microreactors
TerraPower — Natrium sodium reactor
NuScale Power — light-water SMR
Terrestrial Energy — molten-salt SMR
Last Energy — factory-built micro-SMR
Radiant — portable microreactor
Aalo Atomics — modular microreactor
Valar Atomics — gas microreactor
newcleo — lead-cooled fast reactor
26. Grid-scale & long-duration energy storage
The shift: Storage is graduating from four-hour lithium batteries that shave daily peaks to multi-day systems that let renewables act like baseload — turning intermittent wind and solar into firm, dispatchable power.
Lithium-ion solved the daily arbitrage problem. It cannot economically solve the multi-day and seasonal problem — the windless, cloudy week. Long-duration energy storage (LDES) using iron-air, flow chemistries, thermal, and gravity systems targets 10–100+ hour discharge at a fraction of lithium’s cost per kilowatt-hour.
Why now: Grids are hitting the point where adding more solar and wind without storage is worthless — curtailment and negative pricing are now routine in California and Texas. AI-driven load growth means utilities need firm capacity fast, and new transmission takes a decade, so storage becomes the fastest deployable firm resource. IRA manufacturing credits and capacity-market reforms finally pay for the duration that grids actually need.
Why it’s a startup goldmine: The storage TAM runs to trillions as the grid electrifies transport and heat. LDES disrupts the peaker-plant and gas-turbine business that has owned grid balancing for a century. The pricing unlock is capacity payments and multi-hour arbitrage that four-hour batteries physically cannot capture. Defensibility comes from novel electrochemistry and — critically — using cheap, earth-abundant inputs (iron, salt, air) that sidestep the lithium and critical-mineral supply crunch entirely.
Where the opening is: Beyond the cell chemistry itself, the software wedge is enormous: bidding-and-dispatch optimization that turns a battery into a trading desk, virtual power plants aggregating distributed storage, and the AI that forecasts prices and degradation. A founder can build a capital-light software layer on top of everyone else’s steel.
The tell of the winner is not who has the cheapest cell in a lab but who can finance, site, and cycle it profitably across thousands of grid events. The company that pairs a novel cheap chemistry with a merchant-trading software stack captures both the asset margin and the market spread.
Signal — startups & scaleups to watch:
Form Energy — 100-hour iron-air storage
ESS Inc. — iron flow batteries
Eos Energy — zinc-based long-duration storage
Antora Energy — thermal battery for industry
Rondo Energy — heat-brick thermal battery
Fluence — grid storage software + integration
Tesla — Megapack grid storage
Energy Vault — gravity and hybrid storage
Ambri — liquid-metal battery
Highview Power — liquid-air storage
Quidnet Energy — geomechanical pumped storage
Invinity Energy Systems — vanadium flow batteries
EnerVenue — nickel-hydrogen batteries
Natron Energy — sodium-ion batteries
Peak Energy — sodium-ion grid storage
Alsym Energy — non-flammable sodium-ion cells
Nostromo Energy — ice-based thermal storage
27. Data-center power & behind-the-meter energy
The shift: The hyperscaler build-out has broken the grid’s ability to connect new load, so compute is going behind the meter — data centers now come with their own dedicated generation, and power availability, not chips, has become the binding constraint on AI.
Interconnection queues stretch five to seven years in the key markets. A gigawatt data center can’t wait. So the industry is bringing power on-site — gas turbines, fuel cells, on-site solar-plus-storage, and eventually SMRs — and co-locating compute next to stranded generation. The unit of competition in AI is quietly becoming megawatts delivered per quarter.
Why now: Data-center electricity demand is projected to double or more by 2030, and the grid simply cannot connect it on the required timeline. This is the single sharpest “why now” in energy: an unlimited-budget buyer (hyperscalers, neoclouds) with an acute, right-now scarcity and no patience for the utility interconnection process.
Why it’s a startup goldmine: Every gigawatt of AI compute needs roughly a gigawatt of firm power, and the buyers have effectively infinite balance sheets and zero price sensitivity relative to the value of the compute. This disrupts the regulated utility monopoly on new connections. The margin unlock is selling speed — behind-the-meter power that comes online in 18 months instead of 84. Defensibility is site control, grid-interconnection rights, and long-term generation contracts locked up before competitors move.
Where the opening is: The consumer-and-SMB analog is the sharpest wedge — a “Base Power” model that installs distributed generation and storage and sells power-as-a-service, or software that finds and permits stranded-power sites, manages behind-the-meter microgrids, and arbitrages on-site generation against the grid. The founder wedge is turning power procurement into a SaaS-and-services product for anyone who can’t wait in the queue.
This is the purest arbitrage in the whole cluster: the value of a megawatt to an AI company is set by the compute it unlocks, not by the wholesale price of electricity, so whoever delivers firm power fastest captures a spread that will not compress for years.
Signal — startups & scaleups to watch:
Base Power — home battery power-as-a-service
Crusoe — energy-first AI data centers
Bloom Energy — on-site fuel cells
Fervo Energy — enhanced geothermal power
Sage Geosystems — geothermal power and storage
Quaise Energy — deep geothermal drilling
Exowatt — solar-thermal for data centers
Mainspring Energy — linear-generator power
Enchanted Rock — on-site backup microgrids
VoltaGrid — mobile natural-gas generation
Emerald AI — grid-flexible AI compute
Oklo — microreactors for data centers
28. Carbon removal & industrial decarbonization
The shift: Carbon is becoming a managed commodity — captured, removed, and re-priced — while the hardest-to-abate industries (steel, cement, chemicals) get rebuilt around clean processes, turning emissions from an externality into a line item with a market.
There are two moves here. Carbon removal (direct air capture, enhanced weathering, biochar, mineralization) creates durable, verifiable tonnes buyers will pay a premium for. Industrial decarbonization replaces the coal and gas at the heart of heavy manufacturing — green hydrogen for steel, novel cement chemistries, electrified process heat.
Why now: A durable-removal demand market appeared almost overnight — Frontier’s ~$1B advance-purchase commitment, Microsoft’s multi-megatonne offtakes, and the maturing 45Q tax credit created guaranteed buyers before the supply existed. Corporate net-zero commitments plus tightening carbon border adjustments (CBAM) turned decarbonization from voluntary PR into a regulatory and trade requirement.
Why it’s a startup goldmine: The addressable market is every tonne of hard-to-abate emissions on Earth — gigatonnes at premium prices as CBAM and net-zero mandates bite. It disrupts the incumbent carbon-intensive commodity producers who cannot decouple from their emissions. The pricing unlock is that early durable tonnes sell for hundreds of dollars, funding the cost-down curve. Defensibility is process IP, cheap energy contracts, and verified, high-integrity carbon accounting.
Where the opening is: Hardware DAC is capital-hungry; the software-and-services wedges are where lean startups win — MRV (measurement, reporting, verification) that makes a tonne trustworthy and bankable, carbon-removal marketplaces, and the accounting-and-procurement layer corporates need to buy credibly. Whoever becomes the trust layer for a tonne of carbon owns the market.
The uncomfortable truth is that most removal today is expensive and unverifiable, which is exactly why the trust-and-measurement layer is where durable value accrues — the market cannot scale until a tonne is as bankable as a barrel of oil, and someone has to build that standard.
Signal — startups & scaleups to watch:
Climeworks — direct air capture
Heirloom Carbon — limestone-based DAC
CarbonCapture Inc — modular DAC
Charm Industrial — bio-oil sequestration
Lithos Carbon — enhanced rock weathering
Mati Carbon — enhanced weathering on farms
Terradot — enhanced rock weathering
Vaulted Deep — biomass slurry injection
Graphyte — biomass carbon casting
Living Carbon — engineered carbon-capturing trees
44.01 — mineralizing CO2 in rock
Captura — ocean carbon capture
Ebb Carbon — ocean alkalinity removal
Equatic — seawater carbon removal
Isometric — carbon-removal registry and MRV
Sylvera — carbon-credit ratings
Frontier — advance-market demand aggregation
Sublime Systems — electrochemical clean cement
Brimstone — carbon-negative cement
Fortera — CO2-mineralized cement
CarbonCure — CO2 injection into concrete
Boston Metal — green-steel electrolysis
Electra — low-temperature clean iron
Stegra — green-hydrogen steel
Twelve — CO2-to-fuels and chemicals
LanzaTech — carbon-recycling fermentation
Svante — carbon-capture filters
29. Critical minerals & battery materials
The shift: The electrification and AI build-out runs on lithium, copper, nickel, and rare earths — and the West’s near-total dependence on Chinese processing has turned materials from a mining afterthought into a matter of national security and a venture-scale opportunity.
Every battery, motor, transformer, and data center is a claim on refined minerals. Demand is exploding while supply is concentrated in geopolitically fraught hands — China controls the majority of processing for most critical minerals. The response is a full-stack rebuild: new extraction, domestic refining, and recycling to close the loop.
Why now: Export controls became a live weapon in 2023–2025 as China restricted gallium, germanium, graphite, and rare-earth processing. Simultaneously, EV and grid-storage demand curves went vertical and Western governments opened enormous subsidy and offtake programs (IRA, Defense Production Act, EU Critical Raw Materials Act). The buyer of last resort is now the government itself.
Why it’s a startup goldmine: The materials market is a multi-hundred-billion-dollar flow growing with every electrified thing, and secure Western supply commands a strategic premium. It disrupts the incumbent Chinese processing monopoly and legacy mining majors. The margin unlock is novel extraction (direct lithium extraction, e-waste and battery recycling) that produces at lower cost and footprint than conventional mining. Defensibility is resource control, proprietary separation chemistry, and government-backed offtake.
Where the opening is: The software wedges are underrated — AI-driven mineral exploration that finds deposits from geological data, battery-recycling logistics and pre-processing, and supply-chain-traceability software that proves provenance for compliance. The hard-tech wedge is direct lithium extraction and rare-earth separation that doesn’t need China.
The strategic frame matters: this is one of the few venture categories where the state is an active co-investor and guaranteed customer, which de-risks the capital intensity that killed earlier materials startups. A founder who aligns a technical edge with that policy tailwind is building with a wind at their back.
Signal — startups & scaleups to watch:
KoBold Metals — AI mineral exploration
Redwood Materials — battery recycling
Li-Cycle — lithium-ion recycling
Ascend Elements — cathode from recycled batteries
Lilac Solutions — direct lithium extraction
EnergyX — direct lithium extraction
Standard Lithium — brine lithium extraction
Ioneer — lithium-boron mine
Mangrove Lithium — modular lithium refining
MP Materials — rare-earth mine-to-magnet
Niron Magnetics — rare-earth-free magnets
Noveon Magnetics — recycled permanent magnets
Cyclic Materials — rare-earth recycling
Phoenix Tailings — rare earths from mine waste
Magrathea Metals — clean magnesium from brine
Nth Cycle — critical-metal refining
6K Energy — plasma battery materials
Group14 Technologies — silicon battery anodes
Sila Nanotechnologies — silicon anode materials
Novonix — synthetic graphite anodes
30. Longevity & healthspan
The shift: Aging is being reframed from an inevitability into a treatable process — a set of biological mechanisms (cellular senescence, epigenetic drift, mitochondrial decline) that can be measured, slowed, and potentially reversed, turning healthspan into a consumer and clinical market.
The old model treated diseases one at a time as they appeared. The longevity thesis targets the upstream driver — aging itself — on the logic that slowing biological aging prevents the whole downstream cascade of cancer, heart disease, and dementia at once. It spans serious science (partial reprogramming, senolytics) and a booming consumer wellness layer.
Why now: Aging biomarkers matured — epigenetic clocks now measure biological age, giving a quantifiable target and endpoint. Ozempic and the GLP-1 revolution proved that a metabolic drug can produce broad, systemic health benefits and that consumers will pay cash for healthspan. AI-driven biology and cheap sequencing made the underlying research tractable, and an aging-and-affluent demographic is desperate to spend.
Why it’s a startup goldmine: The longevity-and-wellness market is already tens of billions and structurally uncapped — everyone ages, and the willingness-to-pay for more healthy years is nearly infinite among the affluent. It disrupts reactive sick-care with proactive healthspan. The pricing unlock is cash-pay consumer medicine that bypasses insurance entirely, plus eventual pharma-scale therapeutics. Defensibility is clinical data, proprietary biomarkers, and brand trust.
Where the opening is: The consumer-and-data wedge is sharpest — full-body diagnostic and biomarker platforms (the “annual physical, reinvented”), longevity clinics with recurring memberships, and the software that turns continuous biological data into personalized protocols. On the therapeutics side, partial cellular reprogramming is the moonshot. The wedge is owning the measurement layer of aging.
The category has a credibility problem — it sits between rigorous science and snake oil — which is precisely the opening: the company that brings clinical-grade rigor and beautiful consumer experience to healthspan will define the market the way Peloton or Oura defined theirs, but with a far larger and more durable willingness to pay.
Signal — startups & scaleups to watch:
Neko Health — full-body preventive scanning
Function Health — cash-pay biomarker testing
Fountain Life — preventive diagnostics clinics
Human Longevity — genomic health screening
Superpower — consumer healthspan platform
Levels — continuous metabolic monitoring
Tally Health — epigenetic-age testing
Elysium Health — longevity supplements
Altos Labs — cellular reprogramming
Retro Biosciences — cellular reprogramming
NewLimit — epigenetic reprogramming
Turn Biotechnologies — mRNA reprogramming
Shift Bioscience — reprogramming discovery
Life Biosciences — partial reprogramming therapies
BioAge Labs — aging-biology drugs
Unity Biotechnology — senolytic therapies
Cambrian Bio — aging drug pipeline
Gordian Biotechnology — in-vivo aging screens
Rejuvenate Bio — gene-therapy aging
Loyal — canine longevity drugs
Gero — AI aging biology
Insilico Medicine — AI drug discovery for aging
Juvenescence — longevity drug developer
Calico Labs — aging research (Alphabet)
31. Programmable biology / synthetic biology
The shift: Biology is becoming an engineering discipline — cells programmed like computers to manufacture chemicals, materials, foods, and drugs, with AI now designing the genetic code and proteins directly rather than discovering them by trial and error.
The first synbio wave over-promised and stumbled on the gap between designing a cell and manufacturing at scale. The new wave is different because AI closed the design loop: models now generate novel proteins, enzymes, and genetic circuits computationally, and lab automation tests thousands of designs per week. The bottleneck moved from “what do we build” to “how fast can we iterate.”
Why now: AI protein and genome models (the AlphaFold lineage, generative protein design) turned biology into a design problem software can attack. DNA synthesis and sequencing costs kept collapsing, foundation models for biology arrived, and automated “self-driving labs” made the design-build-test-learn loop fast and cheap. Biology finally has its own scaling laws — more data and compute reliably yield better molecules — which is the signal that a field is about to compound the way software did.
Why it’s a startup goldmine: The addressable market is essentially all of chemistry, materials, and manufacturing — a multi-trillion-dollar substitution of engineered biology for petrochemical and extractive processes, plus every drug. It disrupts industrial chemistry and traditional drug discovery. The margin unlock is designing a molecule in silico and brewing it in a vat instead of a refinery. Defensibility is proprietary design models, strain libraries, and manufacturing know-how.
Where the opening is: The AI-drug-design layer is the hottest wedge — companies applying generative models to design therapeutics faster and cheaper than the pharma pipeline. Adjacent wedges: bio-manufacturing infrastructure (the “AWS for biology”), enzyme and materials design, and the tooling-and-data layer feeding the models. Software founders can enter biology through the design and data stack without owning a fermenter.
The lesson of the first synbio cycle is that design is necessary but not sufficient — scale-up is where value is won or lost. The winners of this cycle will be the ones who treat manufacturing as a first-class engineering problem, not an afterthought, and who own the proprietary data that makes their design models compound over time.
Signal — startups & scaleups to watch:
Xaira Therapeutics — AI-native drug design
Isomorphic Labs — AI drug design
Generate Biomedicines — generative protein design
Chai Discovery — AI antibody design
Profluent — AI protein design
Cradle — AI protein engineering
EvolutionaryScale — protein foundation models
Basecamp Research — biodiversity protein data
Dyno Therapeutics — AI-designed gene-therapy capsids
Absci — generative antibody design
Recursion — AI drug discovery via phenomics
Arzeda — computational enzyme design
Ginkgo Bioworks — cell-programming foundry
Solugen — enzymatic green chemicals
LanzaTech — gas-fermentation manufacturing
Twist Bioscience — synthetic DNA
Synthego — CRISPR genome engineering
bit.bio — programmed human cells
Pivot Bio — engineered nitrogen microbes
Culture Biosciences — cloud bioreactors
32. Precision & preventive medicine
The shift: Medicine is moving from population-average, reactive treatment to individualized, predictive care — diagnosing and intervening before disease manifests, using genomics, multi-omics, continuous monitoring, and AI to tailor treatment to the single patient.
Standard medicine treats the average patient with the average drug after they’re already sick. Precision medicine uses your genome, your biomarkers, and your continuous physiological data to predict your risk and pick your therapy — and preventive medicine pushes the whole intervention upstream, catching cancer and cardiovascular disease years earlier when they’re cheap and curable.
Why now: Whole-genome sequencing fell below the cost of a routine lab test, multi-cancer early-detection blood tests (liquid biopsy) reached clinical validation, and wearables turned every patient into a continuous data stream. AI can finally integrate genomics, imaging, and longitudinal data into a real risk model. The economic logic is overwhelming: preventing disease is vastly cheaper than treating it.
Why it’s a startup goldmine: Healthcare is a $4-trillion-plus market in the US alone, and the value of shifting spend from late-stage treatment to early detection is enormous. It disrupts the reactive fee-for-service model. The pricing unlock spans cash-pay screening, value-based contracts that share the savings from prevention, and diagnostics with software margins. Defensibility is clinical validation, proprietary datasets, and payer and provider integration.
Where the opening is: The wedge is the screening and risk-stratification layer — multi-cancer early detection, AI diagnostics that read scans and pathology better than humans, and preventive platforms that combine genomics with continuous monitoring into an actionable risk score. The software founder enters through the data-integration and decision-support layer that sits on top of the diagnostics.
The reimbursement question is the whole game: prevention saves money in aggregate but the savings and the costs land on different balance sheets, so the winners will be those who crack either cash-pay demand or value-based contracts that let a payer share in the downstream savings. Solve the business model and the clinical value is already proven.
Signal — startups & scaleups to watch:
GRAIL — multi-cancer early detection
Tempus — AI precision oncology
Guardant Health — liquid-biopsy cancer tests
Freenome — blood-based cancer screening
Exact Sciences — Cologuard cancer screening
Delfi Diagnostics — blood-based lung screening
Harbinger Health — early cancer detection
Natera — molecular residual-disease testing
Foundation Medicine — tumor genomic profiling
Neko Health — full-body preventive scans
Function Health — cash-pay biomarker panels
Prenuvo — whole-body MRI screening
Ezra — AI-guided MRI screening
Nucleus Genomics — consumer whole-genome testing
Color Health — population genomics and screening
Cleerly — AI coronary plaque analysis
HeartFlow — AI coronary artery analysis
Viz.ai — AI acute-care imaging
Aidoc — AI radiology triage
PathAI — AI pathology diagnostics
Paige — AI cancer pathology
Owkin — AI biomarker discovery
Karius — genomic infection diagnostics
Q Bio — whole-body digital-twin scans
Cluster 5 — Money, Fintech & Trust
Money is the oldest software category and the one most violently re-plumbed by AI and crypto at once. Two forces collide here: value that finally moves like data (stablecoins, embedded rails, programmable money), and a trust stack that has to be rebuilt from scratch now that anyone — human or agent — can forge, transact, and impersonate at machine speed. The eight topics below are the arbitrage between those two forces. The through-line: the winners will not be the ones with the flashiest consumer app but the ones who own a rail, a system of record, or a standard — the boring, load-bearing infrastructure that every other player has to route through and can never cheaply leave.
33. Stablecoins & programmable money
The shift: Dollars stopped being a banking-hours, batch-settled, correspondent-bank artifact and became an internet-native object that moves 24/7, settles in seconds, and carries code. Stablecoins turned the dollar into an API call — and in 2025 that call became legal.
Why now: The GENIUS Act gave the US its first federal stablecoin framework, converting a regulatory grey zone into a licensed asset class with reserve, audit, and redemption rules. Regulatory clarity is the single most powerful unlock in fintech, because it moves a product from “compliance can’t sign off” to “compliance requires a plan” overnight. The moment banks, card networks, and Fortune 500 treasuries got legal cover, adoption stopped being a crypto story and became a payments story — Visa, Mastercard, PayPal, and Stripe all shipped stablecoin settlement, and Walmart-scale merchants started evaluating stablecoin checkout to escape interchange. Stablecoin transfer volume now rivals or exceeds the card networks on a settled-value basis, and the marginal cost of moving a dollar cross-border collapsed toward zero.
Why it’s a startup goldmine: Cross-border B2B payments alone is a multi-trillion-dollar flow paying 3–7% in fees and days of float to correspondent banks. A stablecoin rail collapses that to sub-1% and sub-minute — a classic 10x cost-and-speed unlock where the disruptor keeps a fat margin and still undercuts incumbents 5x. The float alone — the money that today sits idle in the correspondent-banking system for days — is a prize worth billions once it’s freed and settled instantly. Defensibility comes from the on/off-ramp licences (genuinely hard, slow, and jurisdiction-by-jurisdiction to acquire), the treasury and compliance tooling wrapped around the token, and being the settlement layer other fintechs build on so that switching means re-plumbing their money movement. Programmability — escrow, streaming payroll, conditional release, agent-initiated micropayments — is a whole new product surface that legacy rails structurally cannot offer, and it turns a payment from a one-time transfer into a piece of software you can attach logic to.
Where the opening is: Not issuing another stablecoin (Circle and Tether won that) but building the orchestration layer — the Stripe for stablecoins that lets any company accept, convert, and settle in stablecoins without touching a blockchain, plus the treasury/FX/compliance middleware that makes CFOs comfortable. The agentic wedge is the sharpest of all: stablecoins are the only rails an autonomous agent can actually hold and spend. When millions of agents start transacting — buying data, compute, and services from each other in sub-cent increments — they will not open bank accounts or hold Visa cards. They will hold programmable dollars. The company that becomes the wallet-and-settlement layer for agent-to-agent commerce is underwriting a machine economy that doesn’t exist yet but is arriving fast.
Signal — startups & scaleups to watch:
Circle — USDC issuer, now public
Tether — largest stablecoin issuer
Stripe — payments giant, stablecoin push
Bridge — stablecoin orchestration (Stripe-owned)
Privy — embedded wallet infrastructure
BVNK — stablecoin settlement rails
Sphere — global stablecoin payments API
Mesh — crypto payments network
Paxos — regulated stablecoin infrastructure
Zero Hash — embedded crypto/stablecoin rails
Fireblocks — digital-asset custody and infra
Ripple — RLUSD, cross-border settlement
Coinbase — exchange, USDC co-issuer
Ondo Finance — tokenized treasuries
Agora — white-label stablecoins (AUSD)
Catena Labs — AI-agent-native financial institution
Conduit — stablecoin cross-border payments
Brale — stablecoin issuance platform
Rain — stablecoin-powered card issuing
Skyfire — agent identity and payments
Payman — agent-controlled payments
Crossmint — agent wallets and payments
34. AI-native financial services (AI CFO, AI underwriting)
The shift: Finance functions built as workflows for humans to operate are being rebuilt as agents that operate the function. The AI CFO doesn’t dashboard your numbers — it closes the books, forecasts, flags anomalies, and drafts the board deck. AI underwriting doesn’t score an application — it reads the full evidentiary record and issues the decision.
Why now: Financial work is the ideal agent substrate — structured, text-and-number heavy, auditable, and expensive per hour. Models crossed the reliability bar for multi-step reconciliation and document reasoning in 2024–2025, and the data (ledgers, bank feeds, invoices) is already digital and API-accessible via Plaid-style connectivity. Underwriting in particular is a decision with a clean feedback loop, which is exactly where AI compounds.
Why it’s a startup goldmine: This is a payroll-sized TAM, not a software-sized one. A finance team is a $500k–$5M/year cost centre; an AI CFO priced at a fraction of that with agent-level throughput is an obvious buy, and the buyer measures value in heads not saved but redeployed. Underwriting is even sharper — faster, cheaper, and more accurate decisions directly expand a lender’s approvable population and cut loss rates, so the vendor can price on outcomes (basis points of loans underwritten) rather than seats. That is the durable moat: an underwriting agent that sees more loans gets better at pricing risk, which wins more lenders, which feeds it more loans — a data flywheel that compounds and cannot be bought off the shelf. Defensibility beyond the flywheel: owning the system of record, the audit trail, and the regulatory sign-off that a bank’s risk committee will actually accept.
Where the opening is: Vertical, liability-bearing agents that own one financial job end-to-end — an AI controller for SMBs, an AI underwriter for a specific asset class (SMB loans, insurance, trade credit), an AI FP&A analyst — priced as labour and standing behind their output. The winning wedge is the unglamorous, high-frequency, rules-heavy job that a mid-market company currently staffs with two or three analysts: monthly close, invoice reconciliation, expense audit, credit decisioning. Start where the ground truth is machine-checkable and the mistakes are recoverable, earn the right to the judgment calls, and expand outward from the ledger you now own.
Signal — startups & scaleups to watch:
Ramp — spend into autonomous finance ops
Brex — corporate cards and finance ops
Basis — AI accounting agents
Puzzle — AI-native accounting
Zest AI — AI credit underwriting
Upstart — AI lending marketplace
Concourse — AI agents for finance teams
Rillet — AI-native ERP
Casca — AI loan origination
Digits — AI accounting automation
Pilot — bookkeeping and finance ops
Truewind — AI bookkeeping and close
Numeric — AI-assisted accounting close
Taktile — automated credit decisioning
Ocrolus — document automation for lending
Vic.ai — AI accounts-payable automation
FloQast — accounting close automation
Trullion — AI accounting and audit
35. Embedded finance & B2B payments
The shift: Financial products stopped being things you go to a bank for and became features embedded inside the software you already use to run your business. The vertical SaaS you use for your dental practice or trucking fleet now issues the cards, extends the credit, and moves the money — and captures the economics.
Why now: BaaS infrastructure matured past its 2023 compliance reckoning; the survivors (properly bank-partnered, KYC-native) made it genuinely turnkey to embed accounts, cards, and lending without a startup having to become a chartered bank. Simultaneously, B2B payments — still shockingly stuck on paper checks and ACH in the US, where roughly a third of business payments still move by check — became the last giant analog flow ripe for software, and AI finally made invoice-to-pay automation actually work end-to-end rather than dumping edge cases back on a human. The two trends reinforce: once a vertical platform moves the money, automating the accounting on top of it becomes trivial.
Why it’s a startup goldmine: Embedded finance can 2–5x the revenue-per-customer of a vertical SaaS company, turning a $200/month tool into a $1,000/month payments-and-lending relationship — the highest-leverage business-model unlock in software. B2B payments is a multi-trillion-dollar flow with take-rate economics attached. Defensibility comes from owning the workflow the money flows through (the system of record) plus the compliance and risk infrastructure that’s genuinely hard to build.
Where the opening is: Two wedges. First, be the embedded-finance infrastructure for a specific vertical where generic BaaS is too generic (construction, healthcare, logistics have unique money-movement needs). Second, attack B2B accounts-payable/receivable with AI agents that read invoices, reconcile, and initiate payment — the “AP is now autonomous” pitch.
The deeper play is to combine both: a vertical software company that lands with a workflow, then embeds the money movement, then automates the AP/AR with agents — each layer raising switching costs and net revenue retention.
Signal — startups & scaleups to watch:
Stripe — payments platform layer
Adyen — global payments platform
Ramp — corporate cards plus AP
Mercury — startup banking stack
Parafin — embedded lending infrastructure
Pipe — embedded capital for platforms
Toast — restaurant embedded finance
Squire — barbershop embedded payments
Unit — banking-as-a-service
Marqeta — card issuing platform
Modern Treasury — payment operations
Melio — SMB B2B payments
BILL — AP/AR automation
Tipalti — payables automation
Highnote — embedded card issuing
Lithic — card issuing API
Increase — banking API
Column — bank infrastructure
Treasury Prime — banking-as-a-service
Finix — payment processing infrastructure
Slope — B2B payments and BNPL
Settle — AP plus working capital
Rainforest — embedded payments for software
36. Emerging-market fintech & financial inclusion
The shift: The next billion financial customers are being onboarded mobile-first, cash-out, and leapfrogging the entire branch-banking era — the way they leapfrogged landlines for mobile. The bank branch is being skipped, not digitized.
Why now: Smartphone penetration and cheap data crossed the threshold across Latin America, Southeast Asia, Africa, and India simultaneously. Public digital-payment rails — India’s UPI, Brazil’s Pix, and their imitators now spreading across Africa and Asia — created instant, free, interoperable money movement that private incumbents in the West still lack, giving builders a public rail to innovate on top of the way US fintechs once built on ACH but faster and cheaper. Stablecoins add a dollar-access layer where local currencies are unstable, letting a Nigerian or Argentine hold and transact in dollars from a phone. And AI slashes the cost of serving low-ARPU customers profitably — support, underwriting, and fraud that were uneconomic to staff for a $3/month customer become viable when an agent handles them, which is what finally makes the bottom of the pyramid a real business rather than a development project.
Why it’s a startup goldmine: Billions of underbanked people and tens of millions of underserved SMBs represent a greenfield of the size the West saw a century ago — but compressed into a decade. Because there’s no legacy to rip out, unit economics can be structurally better than incumbents from day one. Nubank proved a fintech can reach 100M+ customers and serious profitability in this terrain. Defensibility: distribution, local trust and compliance, and the low-cost-to-serve model AI enables.
Where the opening is: SME banking and credit in specific markets (the underbanked-business gap is even wider than the consumer one), remittances and dollar-access rebuilt on stablecoins, and cross-border commerce infrastructure for the region-to-region trade the West ignores.
The pattern to copy from Nubank: pick a market where incumbents are lazy, expensive, and hated; enter with one wedge product (a fee-free card, a merchant account); use radically lower cost-to-serve to grow virally; then cross-sell the full financial stack once you own the relationship and the data.
Signal — startups & scaleups to watch:
Nubank — LatAm neobank archetype
Mercado Pago — LatAm payments giant
Flutterwave — African payments infrastructure
Moniepoint — Nigerian business banking
Wave — West Africa mobile money
Chipper Cash — pan-African payments
OPay — Nigerian fintech super-app
PalmPay — Nigerian payments app
Kuda — Nigerian neobank
Ualá — Argentine neobank
Clip — Mexican payments
Konfío — Mexican SME lending
dLocal — emerging-market payments
Bitso — LatAm crypto exchange
Yellow Card — African stablecoin rails
Tala — emerging-market credit
Belvo — LatAm open finance
Pomelo — LatAm card infrastructure
Djamo — Francophone Africa neobank
LemFi — immigrant remittances
TymeBank — South African digital bank
37. Cybersecurity for the AI era
The shift: The attack surface and the attacker both changed. AI lets adversaries generate exploits, phishing, and polymorphic malware at machine scale — and every AI application a company deploys is itself a new, poorly-understood attack surface. Security has to defend at machine speed against machine-speed attacks.
Why now: Two curves crossed. Offense got cheap — an attacker with an LLM runs personalized, adaptive campaigns that used to require a team, and the volume of machine-generated phishing and exploit code jumped an order of magnitude. And defense got a new frontier — prompt injection, model exfiltration, poisoned training data, and agent hijacking are live threats with no mature tooling and no established playbook. Boards that treated AI security as theoretical in 2024 saw real incidents in 2025, and the asymmetry is brutal: defenders still operate at human speed against attacks that now scale like software, which is precisely why defense itself has to become autonomous.
Why it’s a startup goldmine: Security spend is non-discretionary and grows with every new surface, and AI just minted two of them (AI-powered attacks, and securing AI itself). Wiz’s $32B exit to Google shows the ceiling. The pricing unlock is that AI-native defense can be sold as an autonomous SOC analyst — outcome-priced labour replacing a $150k security hire — rather than another dashboard. Defensibility: proprietary threat data, the model tuned on it, and deep integration into the customer’s stack.
Where the opening is: The autonomous SOC (agents that triage, investigate, and remediate alerts end-to-end), and the entirely new category of AI application security — scanning for prompt injection, testing agents adversarially, and runtime-monitoring what deployed models actually do.
The structural tailwind: as companies deploy their own agents, every one of those agents is a new privileged insider that can be socially engineered, prompt-injected, or turned into an exfiltration vector — so the security budget grows in lockstep with the agent rollout it is meant to protect.
Signal — startups & scaleups to watch:
Wiz — cloud security (Google exit)
CrowdStrike — endpoint plus AI SOC
Palo Alto Networks — platform security
SentinelOne — autonomous endpoint security
Abnormal Security — AI email security
Snyk — developer and AI code security
Protect AI — AI/ML security (Palo Alto acq.)
Lakera — LLM prompt-injection defense
HiddenLayer — ML model security
Cyera — AI-era data security
Torq — hyperautomation SOC
Dropzone AI — autonomous SOC analyst
Prophet Security — AI SOC agent
Simbian — AI SOC agents
Prompt Security — GenAI runtime security
Pillar Security — AI application security
Zenity — agent security and governance
Noma Security — AI and agent security
Aim Security — GenAI security (Cato)
Straiker — agentic AI security
Mindgard — AI red-teaming
CalypsoAI — AI model security
Harmonic Security — GenAI data protection
Chainguard — secure software supply chain
38. Identity & access control for AI agents
The shift: The entire identity stack was built for two actors — humans and static service accounts. Now there’s a third: the autonomous agent that acts on a human’s behalf, spins up sub-agents, and needs credentials, permissions, and an audit trail. None of the existing plumbing knows what an agent is.
Why now: As agents move from demos to production and start touching money, data, and systems, the unsolved question becomes acute: how does an agent prove who it is, what it’s allowed to do, and on whose authority? You can’t hand an agent a human’s password and full access — that’s an unbounded liability the first time it’s phished or goes off-script. Enterprises piloting agents in 2025 hit this wall immediately, and their security teams blocked production rollouts until the delegation problem was solved. The agent explosion of 2025–2026 makes agent IAM the single most load-bearing piece of infrastructure for the whole agentic economy — it is the gate every other agentic product has to pass through, and right now that gate barely exists.
Why it’s a startup goldmine: This is the Okta-and-Auth0 opportunity replayed for a new, faster-growing class of identity — and the agent population will dwarf the human one. Identity is the stickiest, highest-margin layer in software (it sits under everything and is agony to rip out). Whoever becomes the standard for agent authentication, scoped delegation, and per-action authorization owns a toll booth on every agent transaction. Defensibility is the network and standard effect: once agents and services speak your protocol, switching is systemic.
Where the opening is: The “Okta for agents” — issuing agent identities, scoped and time-boxed delegation tokens (”this agent may spend up to $500 on flights for the next hour”), per-action authorization, and the immutable audit log of what every agent did on whose behalf.
This is the rare infrastructure category where being early to the standard matters more than being best at any feature — the money and the moat accrue to whoever the ecosystem adopts as the default way agents prove themselves.
Signal — startups & scaleups to watch:
Okta — identity leader, agent push
Auth0 — developer identity (Okta)
Stytch — auth and agent identity
WorkOS — enterprise auth and AuthKit
Descope — auth plus agent identity
Astrix Security — non-human identity
Oasis Security — non-human identity management
Aembit — workload identity and access
Token Security — machine and agent identity
Clutch Security — non-human identity
Britive — cloud privileged access
Teleport — infrastructure access
SPIRL — workload identity
Corsha — machine identity authentication
Natoma — governed agent access (Snowflake)
ConductorOne — identity governance
Veza — access and authorization graph
Arcade — agent auth and tool-calling
Scalekit — agent auth stack
Cerbos — authorization service
Oso — authorization as a service
Permit.io — fine-grained authorization
SGNL — privileged access and policy
39. Content provenance & deepfake defense
The shift: We crossed the line where you can no longer trust that a video, voice, document, or face is real. When generation is free and perfect, the scarce, valuable thing becomes proof of authenticity — provenance flips from a nice-to-have to the foundation of digital trust.
Why now: Generative models made convincing deepfakes a consumer commodity in 2024–2025, and the attacks got expensive fast — voice-cloned CEO fraud that moved millions, fake-video KYC bypass, synthetic identities opening accounts at scale. A single real-time video deepfake defeating a bank’s onboarding is no longer a research demo; it is a line item in fraud losses. Financial institutions and governments now face real losses and regulatory pressure, and the elections and information-warfare dimension adds a second, sovereign buyer with a bottomless budget. The response is a two-sided market forming: cryptographic provenance for authentic content (C2PA), and detection/verification for everything else.
Why it’s a startup goldmine: Every institution that authenticates people or content — banks, insurers, governments, media, marketplaces — now needs deepfake defense, and it’s non-discretionary the moment they take a loss. The KYC/identity-verification market was already large; deepfakes just made the old document-and-selfie approach obsolete and forced a full re-buy. Defensibility: proprietary detection data (an arms race that rewards scale and feedback loops) and becoming the embedded trust layer inside onboarding flows.
Where the opening is: Liveness and deepfake-resistant identity verification for financial onboarding (the highest-value, highest-pain wedge), and provenance infrastructure — the layer that signs, tracks, and verifies whether media is authentic or AI-generated across the content supply chain.
The recurring-revenue quality is unusually good: because detection is a live arms race against ever-better generators, this is not a one-time integration but a permanent subscription to staying ahead — the customer can never stop paying without going blind.
Signal — startups & scaleups to watch:
Reality Defender — deepfake detection
Sensity — deepfake detection
GetReal Security — deepfake defense
Truepic — cryptographic content authenticity
Persona — identity verification
Socure — identity verification and fraud
Hive — AI content detection/moderation
Sumsub — KYC plus deepfake defense
Onfido — identity verification (Entrust)
iProov — biometric liveness
Incode — identity verification
Veriff — identity verification
Jumio — identity verification
Pindrop — voice deepfake detection
Nametag — deepfake-resistant identity
Loti — likeness protection
Vermillio — AI content provenance and licensing
Digimarc — digital watermarking
Resemble AI — voice and deepfake detection
Attestiv — media authenticity
40. Compliance-as-code / RegTech
The shift: Compliance is moving from a binder of policies interpreted by humans after the fact to executable rules enforced by code in real time — and increasingly to AI agents that read regulation, map it to controls, and produce the evidence automatically. Compliance becomes continuous and machine-checked, not annual and manual.
Why now: Two pressures compound. Regulatory complexity is exploding — financial rules, data-privacy regimes, and now a wave of AI regulation (the EU AI Act, sectoral AI rules) each demand documented, auditable controls, and the volume of rule-making has outrun any human compliance team’s ability to track it. And AI finally made it tractable to parse dense regulation, map it to a company’s actual systems, and generate the audit artifacts on demand. The cost of manual compliance became untenable exactly as the tools to automate it arrived — the classic condition for a category to flip from services to software.
Why it’s a startup goldmine: Compliance is a massive, non-discretionary, growing cost centre — enterprises spend enormous sums on GRC, audits, and compliance headcount. Automating it is a labour-replacement pricing story (an AI compliance analyst versus a team) with the stickiness of being wired into audit and regulatory workflows. Vanta and Drata proved the SOC-2-automation motion; the frontier is broader and deeper — every regulated industry and the entirely new surface of AI governance (proving your models are compliant). Defensibility: the regulatory content library, the integrations that gather evidence, and being the system auditors trust.
Where the opening is: AI-governance-as-code (the compliance layer for companies deploying AI — model inventories, risk assessments, audit evidence for the AI Act and its successors) and vertical RegTech agents that own compliance end-to-end for a specific regulated industry.
The meta-point ties the whole cluster together: every trend in this section — stablecoins, AI finance, agent identity, deepfake defense — generates new regulation, and compliance-as-code is the layer that turns each new rule into enforceable code. It is the tax collector on all the other waves.
Signal — startups & scaleups to watch:
Vanta — security compliance automation
Drata — continuous compliance automation
Secureframe — compliance automation
Thoropass — audit and compliance
Anecdotes — enterprise GRC automation
Sedric — financial marketing compliance
Greenlite — AI AML/compliance agents
Norm Ai — regulatory AI agents
Credo AI — AI governance
Holistic AI — AI governance
Delve — AI compliance automation
Sardine — fraud and AML compliance
Alloy — identity risk and compliance
ComplyAdvantage — AML screening
Hummingbird — financial-crime compliance
Unit21 — fraud and AML monitoring
Hawk — AI AML and fraud
Lucinity — AML/financial crime
Feedzai — financial-crime prevention
Trustible — AI governance
Enzai — AI governance and compliance
Fairly AI — AI governance
Cluster 6 — Platforms, Infrastructure & New Models
If the vertical agents are the applications of the AI cycle, this cluster is the ground they stand on: the data plumbing, the developer substrate, the sovereign rails, the marketplaces, and the new economic surfaces where value is captured. These are the picks-and-shovels and the platform shifts — where the money migrates first when a wave crests, and where it settles again as the wave matures.
41. Data infrastructure for AI (vector DBs, pipelines, context)
The shift: The bottleneck in AI stopped being the model and became the context you feed it — and that has spawned a whole new data stack.
A foundation model is a stateless genius with amnesia. Everything useful it does in production depends on what you retrieve, embed, chunk, rank, and inject into its context window at inference time. The database of the AI era is not a place you store data — it’s a place you serve meaning from. Retrieval, memory, and context engineering have quietly become the hardest problems in applied AI.
Why now: RAG went from a research trick to the default enterprise pattern in 2024–2025; every serious AI app now needs vector search, hybrid ranking, and a memory layer. Embedding costs collapsed, context windows grew, and MCP standardised how agents pull context from tools — creating a durable new category of infrastructure spend between the data warehouse and the model.
Why it’s a startup goldmine: This is a data-infrastructure market measured in tens of billions, and it re-runs the Snowflake/Databricks playbook one layer up. It disrupts the assumption that your existing warehouse is enough — it isn’t, because analytical stores were never built for low-latency semantic retrieval at agent speed. The margin unlock is classic infra: usage-based pricing on a system that every AI app is structurally forced to depend on. Defensibility comes from being the system of context — once your embeddings, indexes, and memory graph live in one place, migration is brutal.
Where the opening is: Not another bare vector index — those are commoditising into Postgres extensions. The wedge is the context/memory layer for agents: durable, queryable, permission-aware memory that spans sessions and tools, plus the pipeline that keeps it fresh. Own retrieval quality and governance, not raw similarity search.
The tell of the winners: they sit on the critical path of every inference request, so their revenue grows with usage, not with seat count, and they become impossible to rip out without re-embedding a company’s entire knowledge base.
Signal — startups & scaleups to watch:
Databricks — lakehouse + AI platform
Snowflake — cloud data warehouse
Pinecone — managed vector database
Weaviate — open-source vector DB
Qdrant — vector search engine
Chroma — open embedding database
LanceDB — multimodal vector DB
Turbopuffer — serverless vector search
Zilliz — Milvus vector cloud
Supabase — Postgres + pgvector
Neon — serverless Postgres
MotherDuck — DuckDB analytics cloud
LlamaIndex — RAG data framework
Unstructured — LLM data preprocessing
Vectara — RAG-as-a-service
Voyage AI — embedding + rerank models
Cohere — enterprise embeddings
Mem0 — memory layer for agents
Zep — agent memory store
Letta — stateful agent memory
Fivetran — managed data pipelines
dbt Labs — data transformation
Confluent — real-time data streaming
Redis — in-memory vector search
42. Developer tools & platform engineering in the AI era
The shift: Software development is being re-tooled from the editor down — and the platform that orchestrates armies of coding agents is a bigger prize than the editor itself.
When a single developer commands multiple AI agents writing, testing, and shipping code in parallel, the constraint moves from typing to coordination, review, and trust. The dev-tools stack is being rebuilt around agents as first-class contributors: the IDE, the CI pipeline, the review loop, the observability layer, and the internal developer platform all have to assume non-human authors.
Why now: Coding is the single most proven agentic use case — models are best at it, feedback is instant and verifiable, and Cursor/Anysphere reached escape velocity on it. As agent-written code volume explodes, the human bottleneck shifts to review, verification, and orchestration, opening entirely new tool categories that didn’t need to exist 18 months ago.
Why it’s a startup goldmine: Developer tools sell to the highest-density-of-value population on earth, with land-and-expand economics and eye-watering retention. The disruption is that the legacy DevOps/CI/observability stack was designed for human-paced commits; agent-paced software breaks it. Pricing moves from seats to compute + outcomes (per agent-run, per merged PR). The moat is workflow lock-in plus a proprietary signal loop — every agent run teaches your platform what “good” looks like.
Where the opening is: The orchestration and verification layer above the coding agent — managing fleets of agents, gating their output, tracking provenance, and giving platform teams a control plane. Also: agent-native CI, testing, and observability built to assume machine authorship rather than retrofitting it.
The European note: platform engineering is a lean-technical-founder’s game — small teams have always punched above their weight in dev tools, because the buyer is the builder and word-of-mouth is the go-to-market.
Signal — startups & scaleups to watch:
Cursor — AI code editor (Anysphere)
GitHub — Copilot pair programmer
Vercel — AI-app deploy platform
Replit — full-stack coding sandbox
Warp — agentic terminal
Windsurf — AI coding IDE
Sourcegraph — code search + agents
Cognition — Devin coding agent
Poolside — code foundation models
Magic — code-gen models
Augment Code — AI coding assistant
Tabnine — private code completion
Zed — collaborative code editor
Railway — app deployment platform
Render — cloud hosting
Fly.io — edge app hosting
Netlify — web deploy platform
Gitpod — cloud dev environments
Coder — self-hosted dev environments
Temporal — durable workflow orchestration
Pulumi — infrastructure as code
Sentry — error monitoring
Grafana Labs — observability stack
CodeRabbit — AI code review
Graphite — code review platform
All Hands AI — open coding agents
43. Sovereign & defense-grade agentic infrastructure (Europe angle)
The shift: Nations and regulated institutions will not run their most sensitive workloads on someone else’s cloud and someone else’s models — creating demand for sovereign, auditable, defense-grade agentic stacks.
The agentic era forces a hard question: who controls the model, the data, and the audit trail when an autonomous system acts on behalf of a government, a bank, or a hospital? For Europe especially — squeezed between US hyperscalers and Chinese platforms, and armed with a regulatory instinct — sovereignty is not paranoia, it’s policy. The result is a market for infrastructure that is self-hostable, jurisdiction-bound, and provably compliant.
Why now: Geopolitics rewired capital toward sovereignty; the EU AI Act made compliance a shipping requirement; and defense budgets are rising for the first time in a generation. Open-weight models good enough to self-host arrived, making a genuine sovereign stack technically feasible rather than aspirational.
Why it’s a startup goldmine: The buyers are governments, defense ministries, and regulated enterprises — few in number, deep in pocket, desperate for alternatives to American clouds. It disrupts the assumption that serious AI means calling a US API. Margins are strong because sovereignty commands a premium, and defensibility is the strongest kind: certifications, clearances, a working product, and switching costs measured in political capital. This is the “trust rail” for autonomous software in the places that can least afford to outsource it.
Where the opening is: The European Anduril/Palantir of agents — a sovereign agentic control plane that runs open-weight models on jurisdiction-bound infrastructure, with end-to-end audit, identity for agents, and compliance-as-code baked in. Start where the pain is sharpest: defense, critical infrastructure, and public administration.
This is arguably the single sharpest wedge for a technical European founder in the whole report: a genuine home-field advantage, a policy tailwind, and incumbents (the US clouds) who structurally cannot follow you into true sovereignty.
Signal — startups & scaleups to watch:
Anduril — defense autonomy systems
Palantir — data + ops platform
Mistral AI — European open models
Helsing — European defense AI
Aleph Alpha — sovereign enterprise LLMs
Comand AI — battlefield command software
Quantum Systems — reconnaissance drones
Tekever — surveillance drone systems
Nebius — sovereign AI cloud
Scaleway — European cloud
OVHcloud — European cloud
Silo AI — European AI lab
Kyutai — open research lab
Edgeless Systems — confidential computing
SiPearl — sovereign European CPUs
Northern Data — AI compute infrastructure
Shield AI — autonomous defense
Vannevar Labs — defense intelligence
Saronic — autonomous naval vessels
Applied Intuition — defense simulation
Destinus — hypersonic defense aircraft
DeepL — European AI translation
44. Agent marketplaces & the agent economy
The shift: As agents become the unit of work, an economy forms around them — discovery, hiring, payment, and coordination of agents that transact with each other and with humans.
Once an agent can be handed a goal and paid for a result, it becomes a market participant. You need a place to find the right agent for a job, a way to pay it (or its owner) per outcome, a reputation system so you know which to trust, and protocols so agents can delegate to and transact with one another. This is the marketplace and settlement layer of the coming agent economy.
Why now: MCP and agent-to-agent protocols standardised how agents expose and consume capabilities; stablecoins gave machines a native way to pay per transaction 24/7; and the sheer proliferation of narrow agents created a discovery problem worth solving. The pieces for machine-to-machine commerce are, for the first time, all on the table.
Why it’s a startup goldmine: Marketplaces are the most defensible business model ever invented — liquidity begets liquidity, and the winner takes the network effect. The TAM is the transaction volume of automated work itself. It disrupts labor marketplaces and app stores alike, and the pricing unlock is a take-rate on outcome-based transactions rather than a subscription. The moat is the network: buyers, sellers, reputation, and settlement, compounding.
Where the opening is: The payment-and-trust rail for agent-to-agent commerce — identity, reputation, escrow, and settlement so agents can safely hire and pay other agents. Owning the trust and money layer beats owning any single directory of agents.
The caution: most “agent marketplace” pitches are directories with no liquidity and no lock-in. The durable version owns settlement and trust, because that is the part that compounds and the part nobody wants to rebuild.
Signal — startups & scaleups to watch:
OpenAI — agent platform + protocols
Anthropic — MCP + agents
Sierra — outcome-priced agents
Decagon — AI support agents
Skyfire — agent payment rails
Payman — agent-controlled payments
Catena Labs — AI-native bank for agents
Bridge — stablecoin infrastructure
Stripe — agent commerce payments
Coinbase — x402 agent payments
Crossmint — agent wallets
Nevermined — agent payment infra
Virtuals Protocol — agent tokenization
Fetch.ai — agent economy network
Olas — decentralized agents
Lindy — no-code AI agents
Relevance AI — AI agent workforce
LangChain — agent framework (LangGraph)
CrewAI — multi-agent orchestration
Composio — tools for agents
Toolhouse — agent tool infrastructure
Zapier — agent automation
Salesforce — Agentforce platform
Agent.ai — agent directory
45. Vertical SaaS re-platformed by AI
The shift: The incumbent vertical-software winners built systems of record; the next generation embeds agents that do the work the record was merely tracking.
ServiceTitan digitised the trades; Toast digitised restaurants; Procore digitised construction. Each won by becoming the system of record for an industry. But a record is passive — it stores what humans did. The re-platforming is when the software itself schedules the job, dispatches the tech, writes the estimate, chases the invoice, and books the revenue. AI turns vertical SaaS from a filing cabinet into a workforce.
Why now: AI collapsed the cost of building deep, industry-specific software and of automating the messy back-office work that was previously too bespoke to touch. Incumbents have the data and distribution but are architecturally slow to embed agents; new entrants can go AI-native from line one and reach 10× the value with a fraction of the headcount.
Why it’s a startup goldmine: Vertical SaaS already proved it mints multi-billion-dollar outcomes one industry at a time. The re-platforming expands the TAM from IT budget to payroll, because you’re now selling the work, not the tool. It disrupts every incumbent whose product stops at record-keeping, and the pricing unlock is outcome- or labor-based rather than per-seat. The moat is the classic trifecta — workflow depth, proprietary industry data, and system-of-record ownership — now compounded by an operations loop only you can see.
Where the opening is: Pick a boring, expensive, hated back-office job in a big vertical and own it end-to-end as an agent, then expand into the system of record from there. Or arm the incumbents’ unserved long tail — the trades and industries too small for the last cycle’s winners to bother with.
The strategic read: in a maturing wave, value migrates back up to applications and the system of record — and this is exactly that migration, playing out one industry at a time, with the added twist that the software now performs the labor rather than merely recording it.
Signal — startups & scaleups to watch:
ServiceTitan — trades operating system
Toast — restaurant platform
Procore — construction management
Harvey — legal AI
Abridge — clinical documentation
Ambience Healthcare — AI medical scribe
OpenEvidence — medical decision AI
EvenUp — injury-law AI
Legora — legal AI workspace
Clio — legal practice management
Overjet — dental AI
Rilla — trades sales coaching
Sixfold — insurance underwriting AI
Corti — healthcare voice AI
Rogo — financial-analyst AI
Hebbia — knowledge-work AI
Truewind — AI accounting
Fieldguide — audit + advisory AI
Salient — loan servicing AI
Vooma — freight operations AI
HappyRobot — logistics voice AI
Tennr — healthcare referral automation
Commure — healthcare operations AI
Cedar — patient billing platform
46. The creator economy & new monetization models
The shift: Generative AI collapses content-production cost to near zero and hands every creator a studio — reshaping who creates, how much, and how they get paid.
When one person can generate voice, video, music, and interactive media at the quality that once needed a team, the constraint moves from production to distribution, taste, and monetization. The creator economy stops being about editing tools and becomes about the new economic surfaces: AI personas that scale a creator infinitely, licensing of likeness and voice, and outcome-based monetization that doesn’t route through ad networks.
Why now: Generative media crossed the “good enough to ship commercially” line in 2024–2025; AI voice, video, and avatars became indistinguishable enough to monetize; and platforms are opening native rails for AI-generated content. Simultaneously, creators are hunting for income beyond volatile ad revenue.
Why it’s a startup goldmine: The creator economy is a multi-hundred-billion-dollar market with a structural monetization deficit — enormous attention, thin and fragile income. AI disrupts the studio-and-agency cost structure and the ad-only revenue model at once. The pricing unlock is new surfaces: licensing, subscriptions, tips, and AI-persona interactions that creators own directly. Defensibility comes from a proprietary distribution loop or a rights/data moat — owning the audience relationship and the likeness rights, not just a good generator.
Where the opening is: The monetization-and-rights layer for AI-native creators — tools that let a creator license, protect, and scale their voice/likeness/persona and get paid per interaction, not per impression. Own the creator’s economic relationship with their audience.
The trap to avoid: the generative models themselves commoditise fast, so a company built purely on “better output” gets outrun. The winners wrap the generation in a distribution loop, a rights framework, or an audience relationship the model alone can never replicate.
Signal — startups & scaleups to watch:
ElevenLabs — AI voice generation
Suno — AI music creation
Udio — AI music generation
HeyGen — AI video avatars
Synthesia — AI avatar video
Runway — AI video generation
Pika — AI video creation
Luma AI — video + 3D generation
Captions — AI video editing
Descript — audio/video editing
Higgsfield — creator video infra
Midjourney — image generation
Ideogram — image generation
Krea — creative AI studio
Delphi — creator digital clones
Character.AI — AI personas
Fanvue — creator + AI-persona monetization
Patreon — creator subscriptions
Substack — creator publishing
Whop — digital product marketplace
Passes — creator monetization
Vermillio — likeness licensing + protection
Loti AI — likeness protection
47. Spatial computing / AR & the robotics-AI convergence
The shift: The same spatial-perception AI that lets a headset understand a room lets a robot navigate a warehouse — spatial computing and embodied AI are converging on one stack.
For decades, AR was a display problem and robotics was a control problem, solved separately. The vision-language-action wave unifies them: the hard part in both is a machine that perceives 3D space, understands it semantically, and acts in it. A world model that reconstructs and reasons about physical environments is the shared substrate for the next headset and the next warehouse robot alike.
Why now: VLA models brought generalizable manipulation and navigation; 3D world-model generation matured; China’s supply chain crushed sensor and actuator cost; and labor shortages in logistics and manufacturing created desperate pull. The perception stack that both AR and robotics need finally works well enough to build on.
Why it’s a startup goldmine: This spans the AR/spatial-computing market and the far larger robotics-and-automation market — both measured in the hundreds of billions. It disrupts screen-bound computing and cage-bound automation together. The margin trap is hardware’s capital intensity, which is precisely why the unicorns cluster in the shared “brain and data” layers — spatial perception, world models, sim-to-real, and fleet learning — rather than the metal. Defensibility is the proprietary spatial data flywheel.
Where the opening is: Sell the spatial-intelligence layer — perception, world models, and simulation — as infrastructure to both the AR and robotics ecosystems, rather than betting the company on a single device. Or take a narrow, high-ROI embodiment (warehouse, inspection, agriculture) that reaches payback before the general humanoid does.
The convergence is the whole thesis: a founder who builds spatial intelligence once can sell it into two of the largest hardware waves of the decade at the same time, hedging device-level bets while owning the layer both are forced to depend on.
Signal — startups & scaleups to watch:
Apple — Vision Pro spatial computing
Meta — AR/VR headsets
World Labs — large world models
Physical Intelligence — robot foundation models
Skild AI — general robot brain
Field AI — embodiment-agnostic robot AI
Applied Intuition — simulation substrate
Figure — humanoid robots
1X — humanoid robots
Agility Robotics — Digit humanoid
Apptronik — humanoid robots
Nvidia — robotics + Omniverse
Covariant — robot manipulation AI
Dexterity — warehouse robots
Collaborative Robotics — mobile manipulation robots
Standard Bots — affordable robot arms
Waabi — physical-AI driving
Wayve — embodied driving AI
Matterport — 3D spatial capture
Polycam — 3D scanning
Varjo — high-fidelity VR/AR
Ultraleap — hand tracking
Snap — AR Spectacles
Gecko Robotics — inspection robots
48. Neurotech & brain-computer interfaces
The shift: Brain-computer interfaces are crossing from science project to medical product — and, on a longer arc, toward a new human-machine bandwidth layer.
Reading and writing neural signals with enough fidelity to restore movement, speech, and sensation has moved from lab demos to implanted humans doing real tasks. The near-term market is unambiguously medical — paralysis, ALS, blindness, neurological disease — where the value is life-changing and the willingness to pay is total. The longer arc, non-medical augmentation, is where the imagination (and the eventual scale) lives.
Why now: Electrode density and biocompatibility jumped; AI decoders — the same sequence models powering the rest of this report — dramatically improved signal-to-intent translation; surgical robotics de-risked implantation; and the first regulated human trials produced credible results. The decoding bottleneck fell to AI, not to hardware alone.
Why it’s a startup goldmine: The addressable medical population runs to tens of millions with severe unmet need, and neurotech more broadly is a deep-tech frontier where success is category-defining. It disrupts assistive technology and, eventually, the input layer of computing itself. Margins in medical devices are high once approved, and defensibility is as strong as it gets: patents, regulatory moats, surgical partnerships, and proprietary neural datasets that compound with every implant.
Where the opening is: The near-term wedge is medical restoration with a clear regulatory path — restoring speech or movement for a defined patient population — plus the decoding and software layer that turns noisy neural signal into reliable intent (a picks-and-shovels play usable across hardware platforms).
The honest caveat: this is the longest-horizon bet in the cluster — regulatory and surgical timelines are measured in years, not quarters. But it is also the one with the highest ceiling, and the decoding-software layer lets a non-surgical team participate in the upside without carrying the full weight of the hardware and the trials.
Signal — startups & scaleups to watch:
Neuralink — high-bandwidth brain implant
Synchron — endovascular BCI
Precision Neuroscience — cortical-surface BCI
Paradromics — high-data-rate implant
Blackrock Neurotech — implantable BCI
Science Corporation — visual + neural prosthesis
Onward Medical — spinal-cord stimulation
INBRAIN Neuroelectronics — graphene BCI
Neurosoft Bioelectronics — ultra-soft brain electrodes
Forest Neurotech — ultrasound BCI
Iota Biosciences — neural-dust implants
Cognixion — non-invasive AR-BCI
Kernel — non-invasive neuroimaging
Emotiv — consumer EEG
Neurable — everyday EEG
OpenBCI — open BCI hardware
Conclusion — from 48 topics to one decision
Forty-eight topics is a map, not a route. If you try to hold all of them in your head at once they blur into
noise — “everything is the future.” So let’s compress. Underneath the 48 there are really three super-currents,
and almost every topic is a tributary of one of them:
Cognition becomes labor (most of Clusters 1, 2, 6). AI stops assisting and starts doing, and the
addressable market moves from software budgets to payroll. This is the largest, fastest, lowest-capital
current — and therefore the most crowded. Winning here is not about having AI; it’s about owning a workflow,
a system of record, and a proprietary data loop that a wrapper can’t copy.Atoms get intelligence and energy gets scarce (Clusters 3 and 4). Robots, mobility, space, manufacturing,
and the entire energy stack that must power the AI build-out. Slower, more capital-hungry, far more
defensible. The AI demand for electricity has turned sleepy energy into the hottest hard-tech arena in forty
years, and the physical-AI stack is where the most durable moats of the next decade will be dug.Trust becomes the scarce resource (Cluster 5 and the verification threads throughout). Every unit of
automation manufactures an equal unit of demand for identity, security, provenance, compliance, and
programmable money. This is the connective tissue that every other current has to pay for — which makes it
quietly one of the best places to build.
A way to choose
The mistake founders make with a list like this is to pick the hottest topic. The right move is to pick the
topic where your unfair advantage × the market’s readiness is highest. Three filters:
Readiness (why-now strength). Some of these are ripe today — agentic vertical software, AI in
clinical and legal workflows, stablecoin infrastructure, AI security and agent identity, data-center power.
Some are early and will reward patience — humanoid robots, fusion, quantum, neurotech, BCIs. Match your
capital and time horizon to the topic’s clock. A great company in a not-yet-ready market dies of exposure;
a mediocre one in a ripe market can still win.Defensibility (does the moat compound?). Rank your candidates by how quickly the obvious version gets
commoditized. Pure model-wrapper apps decay fastest. Companies that own a system of record, a regulatory
license, a security clearance, proprietary operational data, or the customer’s liability decay slowest. If
your only edge is “we used a good model,” you do not have a company — you have a feature.Distribution (can you get found and trusted?). Cheap creation means adoption is the bottleneck. Favor
topics where you can bolt onto an existing loop — a developer’s editor, a payment flow, an EHR, a
government procurement rail, a security stack — over ones that require conjuring an audience from zero.
Run those three filters and the 48 collapse into a personal shortlist of three or four. That shortlist is
your actual opportunity space.
The through-line no one should miss
If there is a single sentence to carry out of this report, it is this: the winners of this cycle will not be
the companies with the best AI — they will be the companies that own the loop the incumbent cannot cross.
The model is a commodity you rent. The moat is everything around it: the proprietary data your own operations
generate, the trust and liability your customer refuses to hold, the workflow you become the system of record
for, the regulatory or physical barrier you cross that keeps the next entrant out. AI lowers the cost of
building the product to near zero — which is exactly why the product is no longer where the value is. The
value is in the loop.
For a technical founder with a governance, security, or public-sector edge — the profile this library was
built to serve — the sharpest expression of that principle is the convergence of four topics on this list:
agentic vertical software (2), AI security and agent identity (37–38), compliance-as-code (40), and
sovereign/defense-grade agentic infrastructure (43). They are not four ideas. They are one company’s
expanding surface: the trusted, auditable, sovereign platform on which regulated institutions actually run
their agents — entered through a single painful compliance or government workflow, and expanding into the
identity-and-trust layer beneath every agent they will ever deploy. That is the thesis the whole Forecast
folder keeps arriving at, and these 48 topics are the terrain it sits on.
The last word
Platform shifts don’t reward the people who see them coming — everyone sees them coming. They reward the
people who time the wedge: who pick the narrow, painful, valuable beachhead at the exact moment the
technology crosses the line from “impressive demo” to “cheaper and better than the status quo,” and who build
the loop that compounds while everyone else is still admiring the model. The 48 topics in this report are the
places where that line is being crossed right now. The map is drawn. The rest is execution.
— End of report. See the macro-waves essay for the deeper structural forces, and the
20 buildable ideas for concrete, foundable theses derived from them.
Great analysis of the startup trends shaping the next generation of businesses. I especially liked the focus on AI agents and distributed teams. As startups work across different U.S. time zones, accurate scheduling becomes important too. A PST to EST converter can help teams coordinate meetings and calls more easily: https://psttoest.com/