AI Bubble Warnings Flash Red: Which Investments Will Survive
The wrong question in the AI bubble investment conversation is whether valuations are too high. Plenty of people are asking that. The right question is whether the unit economics of generative AI can support the capital structure being built on top of them, and there is already an answer sitting in OpenAI’s own financials.
OpenAI generated $13.1 billion in revenue in 2025. It spent approximately $22 billion to do it. That is $1.69 out the door for every dollar that came in. According to internal documents reviewed by The Information, the company is projecting a $14 billion operating loss for 2026 and does not expect to break even until 2029.
The AI bubble debate assumes the technology will eventually become a real business. OpenAI’s numbers describe a business that is burning toward that assumption faster than it is earning its way there.
The AI bubble’s wrapper problem
The market structure that emerged from the 2021–2023 AI funding boom split into two layers. A handful of foundation model labs raised billions. Several thousand application-layer companies built products on top of their APIs, adding workflow integrations, industry-specific prompts, and user interfaces, then charged enterprises a margin above API cost. The AI bubble conversation usually focuses on the first layer. The second is where the structural problem lives.
That model has a structural problem. Inference cost per million tokens dropped by 80% from 2023 to 2025, which is good for users and fatal for startups whose only moat was the margin between API costs and customer prices. OpenAI’s own product cadence made it worse. GPT Store, Operator, Tasks, Canvas, and Search each absorbed functionality that had funded startups built around. IdeaProof’s analysis, drawing on CB Insights data, documented at least 200 funded GPT wrapper companies cannibalised by these feature drops in 2024 alone.
According to Digital Silk’s startup failure analysis, the 2022 cohort of AI startups burned through $100 million in three years, double the cash-burn rate of earlier tech generations. Most of those runways expire in late 2025 or early 2026. Some have already failed.
Where the VC money actually went
Crunchbase data show that 60% of global and 70% of US venture capital in 2025 went to $100 million-plus rounds, a concentration level not seen before.
Of the approximately $430 billion deployed in 2025, $202.3 billion targeted AI, with $90 billion concentrated in just six companies, each raising more than $5 billion. As seed deals declined, megarounds dominated. The distribution of venture capital funding during the dot-com era spanned hundreds of categories, including food tech, health tech, robotics, and fintech. Now it funnels toward a small number of foundation model labs and late-stage AI infrastructure plays.
That concentration is itself a signal worth reading. When capital stops spreading across a sector and starts piling into the same five or six companies, it usually means investors have concluded that most of the sector will not produce returns. The money going to OpenAI, Anthropic, xAI, Mistral, and Cohere is a bet that the foundation model layer captures most of the value.
That is the AI bubble thesis in its clearest form: not that all AI is overvalued, but that most of the money below the foundation model layer is.
The Anthropic contrast
The profitability argument is not uniformly bad across all AI companies. TechCrunch reported that Anthropic projects gross margins reaching 77% by 2028, up from approximately 50% in 2025.
The path runs through inference cost reduction and a shift toward more efficient deployment architectures. As of May 2026, TechCrunch reported Anthropic told investors it expects to deliver its first operating profit in Q2 2026, ahead of earlier projections. If that trajectory holds, Anthropic has a real business case.
The Information noted that Anthropic’s inference costs on Google and Amazon cloud infrastructure came in 23% above internal estimates. That gap matters. Projected and actual margins have been diverging across the sector, and the gap between what OpenAI forecast and what it spent suggests that AI companies’ internal numbers warrant scrutiny before they become investment theses.
The deal AI companies struck with users was always a quiet one: subsidise access, flood the market, figure out profit later. That arrangement now has a deadline. The IPO queues of OpenAI, Anthropic, and xAI, in sequence, are forcing a reckoning. Public markets do not absorb “we’ll figure it out” as a line item.
What survives and what doesn’t
Not all AI investments carry the same risk. The table below maps five categories by where AI bubble pressure is most concentrated, based on the capital flow and unit economics data available as of mid-2026.
| Category | Bubble Risk | Why |
|---|---|---|
| Foundation model labs (OpenAI, Anthropic, xAI) | Moderate: dependent on capital markets staying open | Burning billions annually, the path to profit requires margin assumptions that have not held |
| AI infrastructure (Nvidia, cloud providers) | Low: actual revenue from actual customers | Selling picks and shovels regardless of which AI company wins |
| Vertical AI with proprietary data | Low to moderate | Moat from data nobody else has, not competing directly with foundation models |
| Horizontal AI wrappers (no data moat) | High | Commoditised by foundation model providers’ own features; no defensible margin |
| Enterprise AI with demonstrated ROI | Low to moderate | Real contracts, real renewals; AI bubble risk contained to valuation multiples |
The table describes current risk, not certainty. Foundation model labs could still turn profitable if their margin projections hold. The pattern that most resembles the dot-com collapse is the widespread funding of companies with no path to revenue, competing in markets that the infrastructure layer will eventually dominate.
That applies most directly to the horizontal wrapper category. For enterprise IT leaders evaluating vendor risk within that wrapper layer, the open-source AI model alternative changes the calculus: a startup building on open weights avoids the platform-cannibalisation risk that closed-API wrappers face.
Distilled
The AI investment story is increasingly splitting into winners and losers. Capital continues to concentrate around foundation model providers and the infrastructure companies supporting them, while many application-layer startups face shrinking margins and growing competition from the platforms they depend on.
At the same time, profitability remains an open question for several of the sector’s biggest names. Revenue growth is strong, but long-term success will depend on whether AI companies can translate scale into sustainable margins rather than continued dependence on investor funding.
The AI bubble question is usually framed as, “Are valuations justified by fundamentals?” The more useful framing is, “Which fundamentals?” OpenAI’s revenue is real and growing. Its unit economics describe a business that loses more money the more it grows. Whether those unit economics ever flip is the variable the entire valuation structure rests on.
