VC funding in AI: infra, vertical, physical — or nothing

VC funding is open in AI Infra, Vertical AI, and Physical AI — and effectively closed everywhere else. That's not a cycle. It's a structural reset. AI captured roughly 80% of global venture dollars in Q1 2026, yet the vast majority of AI startups are raising nothing. The capital is concentrating into three buckets, and if your company doesn't fit one of them, you're not in a slow market — you're in a dead one.

The Three Buckets VCs are Still Writing Checks for in 2026

The data is unambiguous. General Intuition raised a $320 million Series A — a round size that would have been extraordinary for a frontier lab, let alone a first institutional round. RunPod closed $100 million in a Series A for GPU cloud rental. Assort Health raised $120 million in a Series C for healthcare-specific AI workflows. Taktile pulled $110 million in a Series C for fintech decisioning software.

None of these are horizontal foundation-model plays. Every single one is a vertical or infrastructure bet with demonstrated customer traction. The pattern is deliberate: investors are rewarding defensible niches and real revenue with growth-equity-sized checks. The three surviving categories are:

1. AI Infrastructure — GPU clouds, agent orchestration layers, inference optimization, the picks-and-shovels layer that every AI application depends on.
2. Vertical AI — AI built specifically for healthcare, legal, fintech, life sciences — domains where the data moat, compliance complexity, and switching costs create genuine defensibility.
3. Physical AI — Robotics, autonomous systems, and the hardware-software integration required to make AI operate reliably in the real world.

Everything else — horizontal SaaS wrappers, general-purpose chatbots, broad productivity tools without a defensible vertical — is competing for a shrinking pool of capital that increasingly isn't there.

Why "everything Else" Isn't Just Underfunded — It's Structurally Frozen

Nikesh Arora called it directly: a "Darwinian moment for AI providers." That framing is precise. This isn't a correction where patient founders wait 18 months for sentiment to shift. The structural logic is working against horizontal AI permanently.

Here's why. The four largest hyperscalers committed a combined $650 billion in AI infrastructure spend by end of 2026. When the infrastructure layer is that well-capitalized, the commodity risk for anything built on top of it is existential. A horizontal AI product that could be replicated by an OpenAI feature drop or a Claude API update has no durable moat. VCs know this. They've stopped pretending otherwise.

As Grant C. Simon, SVP & GM of the Venture Capital Group at Comerica, put it: "AI infrastructure is the new cloud war. Today's battleground is system control, who can shape, monitor, and enforce the rules on how agents operate at scale — which requires new infra tools for the next decade."

The implication for founders is brutal: if your AI product doesn't own a specific workflow in a specific vertical, or doesn't sit in the infrastructure layer that other AI products depend on, you're building on land that the hyperscalers are about to pave over.

Vertical AI is No Longer a Small-check Category

The conventional wisdom 18 months ago was that vertical AI was the "safe" but smaller bet — you'd raise a $5–10M seed, build for a niche, and grow slowly. That's dead. General Intuition's $320M Series A is 16–32x the typical 2026 Series A range of $10–20M. Assort Health's $120M Series C is infrastructure-round sizing applied to a healthcare workflow company.

What changed: investors realized that vertical AI companies with genuine traction have better defensibility than frontier labs. A healthcare AI company embedded in hospital billing workflows has switching costs that GPT-5 can't eliminate overnight. A fintech decisioning tool that's been trained on a specific lender's historical data isn't replaceable by a general model.

The catch — and this is what separates the funded from the frozen — is that "vertical AI" now means demonstrated retention and revenue, not a pitch deck with a vertical TAM slide. As the Value Add Pulse analysis noted, investor expectations for growth rate and defensibility have risen to match the new check sizes. You don't get $120M for a vertical thesis. You get it for a vertical business.

This is also where the OpenAI vs. Anthropic product philosophy split becomes relevant for founders: the race to ship vertical-specific capabilities is accelerating, and the window to establish a data moat in your niche is narrowing fast.

Physical AI is the Hardest and Best-funded Bet Right Now

VC Eclipse launched a $1.3 billion fund in April 2026 specifically to back and build physical AI startups. Wing Venture Capital published a thesis in June 2026 titled "Software Ate the World. Now Hardware Is Eating Software." The capital is moving.

Physical AI is expensive precisely because the integration challenges are non-trivial. Consider what autonomous vehicle development actually requires at the component level: Tesla's engineering teams transformed the windshield into an active optical shield with wire-grid-free transparent-film heating — a materials science problem, not a software problem — to meet the sensory requirements of the AI systems operating behind it. That's the nature of physical AI: every real-world constraint requires a novel engineering solution, and each one requires capital.

Nikesh Arora's prediction is direct: expect LLMs to chase more vertical profit pools in legal and life science, and expect physical AI companies to follow. The convergence of robotics, autonomous systems, and AI-driven manufacturing is where the next decade of infrastructure spend goes. For founders, the question is whether you have the hardware-software integration capability and the capital runway to compete — because the check sizes required to play here are not seed-stage numbers.

The Monetization Problem is Why Horizontal AI is Dying

Nikesh Arora's sharpest observation isn't about funding — it's about the underlying cause: "It's not a demand problem — it's a monetization problem." Demand for AI capabilities is real and growing. The problem is that most AI companies can't convert that demand into durable revenue.

Horizontal AI products face a specific version of this: they're often the first tool a user tries, but not the last tool a user keeps. Churn is high because the switching cost is zero. The product that replaces you is one API call away. This is why VCs have stopped funding the category — not because the products don't work, but because the revenue doesn't compound.

Vertical AI solves this structurally. When your AI is embedded in a legal firm's document review workflow, trained on their historical case data, and integrated with their billing system, the switching cost is measured in months of re-training and workflow disruption. That's a moat. Horizontal AI has no equivalent.

As Hila Rom, Venture Capital Investor and Serial Venture Builder, observed: "As AI agents shift from assisting workers to replacing them, the per-seat pricing model is dying. The future belongs to those who can monetize actual outcomes." Outcome-based pricing is structurally easier to defend in verticals where the outcome is measurable — a claim processed, a contract reviewed, a loan underwritten — than in horizontal productivity tools where "outcome" is diffuse.

Hyperscalers and Public Markets are Replacing Private VC at the Frontier

For founders watching the frontier lab funding rounds and wondering when that capital trickles down: it won't. Nikesh Arora was explicit that new frontier labs will need to go public, because there isn't enough private market capital available to support their capex needs. The actions of hyperscalers in terms of their capital commitments will be the key variable as 2026 proceeds.

This matters for mid-stage founders because it clarifies the competitive landscape. Frontier labs are not going to be funded by the same pool of VC capital that funds Series A and B rounds. They're going to be funded by Microsoft, Google, Amazon, and eventually public markets. That means the VC capital that exists for private companies is concentrated in the vertical and infrastructure layers — not at the frontier, and not in horizontal applications.

The practical read: if your AI company's competitive moat depends on having access to better foundation models than your competitors, you don't have a moat. The foundation model layer is becoming a commodity funded by hyperscaler capex. Your defensibility has to come from data, workflow integration, or domain expertise — not model quality.

Open-source is a Real Counter-strategy, Not a Consolation Prize

There's a counter-narrative gaining traction that founders outside the three funded buckets should take seriously. Nikesh Arora's own projection: pure models will continue to see arbitrage with open source. That's not a rounding error — that's a structural shift in how AI gets deployed.

Synthetic Sciences (YC W26) made the case directly: "Scientific AI should be open. One company shouldn't own the tools the rest of us discover with, or decide who gets to." The argument isn't just ideological. For founders in research-adjacent verticals — life sciences, drug discovery, materials science — open-source models eliminate vendor lock-in and allow fine-tuning on proprietary datasets without exposing that data to a third-party API.

The strategic implication: if you're building in a domain where data is your moat, open-source infrastructure plus proprietary training data is a more defensible position than proprietary model plus generic data. This is a viable path for founders who can't raise at the valuations the three funded buckets command — but it requires a clear answer to the monetization question that open-source alone doesn't solve.

What the Per-seat Pricing Collapse Means for Your Revenue Model

The per-seat SaaS model is structurally broken for AI products where agents are replacing workers rather than assisting them. If your product automates a task that used to require five people, per-seat pricing means your revenue goes down as your product gets better. That's not a business model — it's a self-destruct mechanism.

The funded vertical AI companies have already solved this. Assort Health prices on healthcare workflow outcomes. Taktile prices on loan decisions processed. RunPod prices on compute consumed. None of them are charging per seat.

For founders still on per-seat pricing: this is the first thing to fix before your next fundraise. Investors in 2026 are pattern-matching on outcome-based or consumption-based models. A per-seat AI product signals that the founder hasn't thought through the agent-replacement dynamic — and that's a fast no.

If you're unsure where your funnel is leaking revenue before you even get to a pricing conversation, doableclaw.com runs the same diagnosis a growth consultant would — RCA tree, ICP gaps, funnel leaks — in 2 minutes, without a ₹50K retainer.

How to Diagnose Whether Your AI Startup Fits the Surviving Categories

Three questions that map directly to what VCs are funding:

1. Do you own a workflow, or do you sit on top of one? If your product is a layer on top of an existing workflow that could be replaced by a ChatGPT plugin, you don't own the workflow. Funded vertical AI companies own the workflow — they're the system of record, not the assistant.

2. Is your defensibility in the data or in the model? Model quality is a commodity. Data is not. If your defensibility comes from proprietary training data, domain-specific fine-tuning, or compliance-driven data moats (HIPAA, SEBI, RBI), you're in the fundable category. If it comes from "we use GPT-4 better than competitors," you're not.

3. Does your revenue compound as the product gets better? Vertical AI and infrastructure products have compounding revenue dynamics — more usage generates more data, which improves the model, which drives more usage. Horizontal AI products often have the opposite: commoditization drives prices down as the underlying models improve. If your revenue model doesn't compound, fix that before you pitch.

Grant C. Simon's observation is worth anchoring here: "The most interesting companies are building economy agents that live inside payments, memory, AI sandboxes, insurance, telecommunications, and legal workflows." The word "inside" is doing the work. Not adjacent to — inside.

This diagnostic connects directly to how AI agents are reshaping the infrastructure layer — the companies that survive will be the ones embedded in the workflow, not orbiting it.

The One Move Founders Outside These Three Buckets Should Make Now

If your AI startup doesn't fit AI Infra, Vertical AI, or Physical AI, you have one real option: pick a vertical and go deep before your runway runs out. Not "we serve SMBs" — that's horizontal with a size filter. A real vertical means a specific industry, a specific workflow, a specific data moat, and a pricing model tied to a measurable outcome.

The window is narrowing. Vertical AI Series A rounds are now sized like growth equity — $100M+ — which means the bar for what counts as "demonstrated traction" is rising fast. The founders who move now, while the category is still being defined in legal, life science, fintech, and industrial automation, will set the defensibility benchmarks that make the next wave of entrants uncompetitive.

As Enrique Peláez put it: "AI architecture isn't a stack diagram. It's a governance spine." The founders who understand that — who build AI into the governance and compliance layer of a specific industry rather than bolting it on top — are the ones raising $120M Series C rounds in 2026.