The old venture question – does this company use AI? has stopped being useful. Almost every serious software company now has an AI story. The better question is where the company sits in the new map of control points: does it own the work, the workflow, the data, the governance layer, the distribution channel, the model, or the physical infrastructure underneath it?
That distinction matters because AI is no longer only a product cycle. It is also a capital cycle. The technology can be real and still produce poor venture returns if investors fund the wrong layer, at the wrong price, just as too much capacity arrives and pricing power starts to move elsewhere.
1. From software-as-assistant to software-as-worker
A decade ago, the working assumption in venture capital was simple: software eats the world. Pick a slow, paper-heavy industry, build a tool that helps the humans inside it work faster, sell it as a subscription, repeat. The product assisted. The human still did the job.
That assumption is breaking down in real time. Across 2025 and into 2026, AI captured somewhere between half and three-fifths of global venture capital in the source dataset, and the money was not simply chasing AI features bolted onto existing software. It was rewarding companies that perform work end-to-end, priced against labour or services budgets rather than traditional software seats.
The clearest evidence sits in legal AI, customer service and coding. Harvey [1] raised $200 million in March 2026 at an $11 billion valuation to run AI agents across due diligence, compliance and litigation support inside law firms – work that used to mean associates and billable hours. Cognition [2], maker of the autonomous coding agent Devin, raised $1 billion in May 2026 at a $26 billion valuation, with customers ranging from Mercedes-Benz to Goldman Sachs to NASA. Sierra and Decagon [3] have done the same in customer service, both crossing or approaching unicorn-plus valuations on agents that resolve tickets rather than merely suggesting replies to a human agent.
The durable economics are still unproven. Multi-year retention, pricing power, compute cost and service intensity still have to be tested. But the direction is clear: the best-funded AI companies are increasingly being priced as work-execution platforms, not productivity tools.
2. The capital cycle underneath the AI cycle
The first mistake is to ask whether AI is real or a bubble. It is both. Real technology waves attract speculative capital precisely because the prize is large enough to justify imagination. The internet was real; many internet investments still failed. AI can transform the economy and still disappoint investors who fund the wrong supply curve.
The capital-cycle question is unchanged from every prior boom: what capacity is being built, how fast is it arriving, and who loses pricing power when it does? In AI, capacity no longer means only factories, ships or mines. It means GPUs, data centres, power, frontier-model capability, cost per token, open-source substitutes, funded competitors, enterprise pilots, workflow data, distribution, talent density and governance permissions.
Each layer has its own capital cycle, and those cycles are not moving in sync.
| AI layer | What capital is building | Capital-cycle risk | Return persistence depends on |
|---|---|---|---|
| Frontier models | Model capability, research teams, inference scale and distribution. | Oligopoly formation, very high capital intensity, rapid capability catch-up. | Access to compute, proprietary distribution, enterprise trust and the ability to monetise usage without margin collapse. |
| Compute, chips, data centres and power | Physical capacity for training and inference. | Classic capex cycle: supply arrives late, then too much capacity can compress returns. | Utilisation, long-term contracts, power access, site control and cost position. |
| Vertical workflow AI | Agents embedded inside industry-specific processes such as legal, healthcare, finance or logistics. | Overfunding if companies only automate fragments rather than own the workflow. | Proprietary data, process ownership, switching costs, compliance depth and customer distribution. |
| AI governance and control | Permissioning, audit, identity, policy enforcement and control for autonomous systems. | Category may be crowded before budgets fully mature. | Regulated adoption, machine-readable controls, integration into enterprise identity/security stacks and proof of enforcement. |
| Generic copilots and thin wrappers | Productivity features between someone else's model and someone else's customer. | Fastest route to oversupply and commoditisation. | A move into owned workflow, proprietary data or distribution before the feature is absorbed by incumbents. |
| Fund structures and late-stage capital | Larger follow-on funds, evergreen structures and concentrated ownership of AI winners. | More capital chasing already-consensus assets at elevated prices. | Entry discipline, reserve strategy, patience and the ability to refuse hot rounds without durable control points. |
3. The new investment question: what does the company control?
The most useful AI diligence question is no longer whether a company has adopted AI. It is whether the company has become the place where work actually gets done, where AI activity is governed, where an industry’s workflow is owned end-to-end, where strategic capacity is built, or where capital itself is being structured differently.
Companies that hold one of those positions – a model, workflow, governance layer, distribution channel, data asset, or physical infrastructure such as power and compute – may occupy defensible ground. Companies that sit in the thin layer between someone else’s platform and someone else’s customer do not, no matter how fluent their AI feature appears.
4. Where the money is clustering
Workflow agents
The strongest signal is the move from software that helps a worker to software that performs the work. Legal AI, coding agents and customer service are early proof points because the buyer can compare the AI product against headcount, outsourced services or billable hours rather than against another SaaS seat.
Governance and security
Regulated enterprises cannot allow autonomous systems to act freely without permissioning, audit and enforcement. That requirement is now showing up as an investment signal in its own right: at least seven unrelated venture funds, spanning early-stage and growth investors, have independently backed AI governance and security companies – Noma Security [4], WitnessAI [5], Geordie AI [6] and adjacent names – in the same narrow window, with no sign of coordinating with one another. Independent, uncoordinated conviction from seven-plus investors is a stronger signal than any single fund’s thesis. It suggests the market has concluded that AI governance is becoming its own permanent layer of infrastructure, in the way cybersecurity became permanent infrastructure around the last cloud computing wave – a category worth watching closely for anyone advising regulated-enterprise platforms.
Vertical healthcare and regulated workflows
Healthcare is a clean example of vertical depth beating horizontal breadth. Ambience Healthcare [7] raised $243 million in mid-2025 to own clinical documentation across health systems, and roughly 54% of the $14.2 billion raised by digital health startups in 2025 went to AI-led companies [8] – a sharp pivot away from the old consumer-wellness story toward enterprise and clinical workflow ownership. The same pattern is showing up in legal, fintech onboarding, asset management and logistics: platforms that own a specific regulated industry’s workflow and its underlying data are outcompeting generic productivity tools that any incumbent could absorb as a feature next quarter.
Sovereignty, defence and strategic capacity
Shield AI [9], Helsing [10], Anduril [11] and parallel national efforts around chips and compute point to one broader theme: control over compute, power and strategic capacity is becoming investable in its own right, not merely as infrastructure beneath applications. China is running a parallel, equally aggressive bet on domestic chip and compute sovereignty.
Fund architecture
The investors are changing too. Larger growth funds, permanent-capital ambitions and concentrated late-stage vehicles suggest venture capital is adapting its own structure to an AI cycle that demands larger reserves, longer duration and more concentrated follow-on decisions. Founders Fund [11] closed two growth funds, $4.6 billion and then $6 billion, within about twelve months of each other; Trian Fund Management and General Catalyst’s joint $7.4 billion take-private of Janus Henderson [12] and Sequoia’s $7 billion late-stage fund [13] both point toward permanent-capital, evergreen structures displacing the traditional ten-year venture model at the top of the market.
Step back from the headline narrative and a more disciplined pattern emerges. Funds are barbelling: writing numerous small, exploratory first cheques at one end, and a handful of enormous, concentrated follow-on rounds at the other, while the traditional “normal” growth round – roughly $150–500 million – has noticeably thinned out. A small clique of marquee funds, including Sequoia, Founders Fund, General Catalyst, a16z, Lightspeed, Kleiner Perkins and Coatue, keep showing up together in the same handful of trophy companies, behaving less like venture investors and more like concentrated index holders of three or four AI winners everyone already agrees on. A separate set of funds shows almost no overlap with that group, doing genuinely uncorrelated, original discovery work instead. In coding agents specifically, Founders Fund backed Cognition/Devin while a16z backed the rival Anysphere/Cursor [14] – a clean example of funds hedging across competing bets rather than picking a single winner, a pattern also visible in foundation models.
Encouragingly, the exit window also appears to be reopening – not in one sector, but across healthcare, insurtech, security and AI tooling – an early sign that this build-out may be starting to convert into realised outcomes rather than paper valuations.
Geography
The United States remains the deepest pool of AI venture capital, with AI now representing close to 88–89% of all US venture funding so far in 2026 [17]. Europe has strong company formation in areas such as models and defence sovereignty – Mistral AI [15]’s €1.7 billion raise, Helsing’s sovereignty round, but less capital depth, with AI just over half of European VC, the lowest share of any major region [18]. India is showing platform formation around cloud and AI infrastructure, helped along by roughly $48 billion in planned Amazon [16] investment into India’s AI and cloud infrastructure. China has re-accelerated into space, quantum, fusion and AI, while carrying visible bubble risk [19].
5. Why incumbents face a harder transition than the market assumes
Old SaaS companies are built to sell high-margin subscriptions. They sell access to a tool, often priced per user or per seat. AI agents change the question. If the agent does the work, the buyer may no longer want to pay for ten people to use software. The buyer may want to pay for the job completed.
That can compress margins. Agents may require more compute, more workflow configuration, more customer-specific integration and more service-like support. Incumbents that defend old SaaS gross margins may protect the spreadsheet while losing the workflow to AI-native companies willing to price against outcomes or labour replacement.
The winners will be the companies willing to redesign the product, the pricing unit and the operating model together. The product moves from tool to worker. Pricing moves from seat to usage, task or outcome. The operating model moves from pure software delivery to a more complex mix of compute, service, workflow design and governance. The transition may look worse in the short term but protect control of the customer workflow over the long term.
6. What the evidence does not prove yet
The capital-flow evidence shows where conviction is forming. It does not yet prove which companies will earn durable margins.
Legal AI, customer-support agents, AI governance and vertical healthcare all have strong funding momentum, but the hard questions remain open: retention, pricing power, gross margins after compute and service costs, implementation cycles, regulatory liability, buyer willingness to pay, and whether data or workflow access becomes proprietary enough to defend returns.
Climate has not disappeared as an investment category; it has been reframed into infrastructure, energy and power economics. Europe and India are not necessarily weaker than the public data suggests; they are less completely visible in public capital-flow datasets. China requires an even sharper distinction between strategic state-backed capital formation and private return discipline.
7. Investor checklist for the next eighteen months
- Is the company replacing a labour/services budget, or merely adding another software budget?
- Does the company own a workflow, data asset, permissioning layer, distribution channel, model capability or physical capacity?
- Does each additional dollar of capital deepen the moat, or merely keep the company in the race?
- What new capacity is being built in this layer, and who loses pricing power when that capacity arrives?
- Is the company using OpenAI, Nvidia or a cloud platform as a tool, or is it renting its entire strategic position from them?
- Can the company move pricing toward usage, task or outcome without destroying margins?
- If the round is hot, what is the reason to refuse it?
Conclusion: access is no longer the edge
AI is creating a new capital cycle. Every participant in it is taking the same test under a different name, and most of them have not noticed.
For venture investors, the test is discipline. The opportunity is not getting into the most talked-about AI rounds — that part is easy, and easy things are rarely where the returns live. It is identifying which companies turn capital into a lasting advantage. In a crowded market, access alone is not the edge. The edge is knowing where the advantage survives after more money, more competitors, and more platform changes arrive, because all three are coming whether the thesis accounts for them or not.
For entrepreneurs, the test is durability. Funding should not extend runway or buy attention — those are things money does by default, not things a founder needs a strategy to achieve. It should make the company harder to copy. The best founders will use this moment to build better products, deeper customer relationships, proprietary data, stronger workflows, distribution power, and clear economic value. The rest will use it to extend the runway on a company that was always going to need extending.
For family offices and long-term capital, the test is judgment. AI access is everywhere, which is precisely the problem. Not every opportunity deserves underwriting, and the ones that least deserve it are often the easiest to get into. The question is whether the company owns something that survives hype, competition, and platform dependency — or whether it is simply renting momentum from a cycle that will, in time, move on without it.
The question is the same across all three perspectives: is capital being used to build a stronger business, or to buy another quarter of relevance? Most capital, in most cycles, buys the quarter.
The winners of this AI cycle will not be the companies that raise the most money. They will be the companies that convert money into real advantage — which is a smaller list than the funding announcements would suggest, and always has been.