The AI Crash Has Already Started. Almost Nobody Noticed

$581 billion invested. $20 billion lost by the market leader. 120,000 jobs cut. The data tells a more complicated story than either the bulls or the bears want you to believe.

I’ve been sitting with a set of numbers for the past two weeks that I genuinely cannot reconcile with each other, and the more I look at them the more I think that’s not a problem with my analysis — that’s the actual story.

Let me show you what I mean by holding two true things simultaneously.

Global AI investment hit $581.7 billion in 2025, up 130% in a single year, more than double the prior record, according to Stanford’s 2026 AI Index. Anthropic closed a $65 billion funding round in May 2026, making it the most valuable private company on earth at $965 billion. AI startup funding in Q1 2026 hit $242 billion, roughly 80% of all global venture capital that quarter — meaning four in five venture dollars went into AI in a single quarter.

Those are the bull numbers. Now the bear numbers.

OpenAI — the company that started this cycle — lost $20.92 billion from operations in 2025 on $13.07 billion in revenue. Their own internal projections show $74 billion in operating losses in 2028 alone before reaching profitability by 2030. ChatGPT’s web traffic share fell from 86.7% in January 2025 to 64.5% twelve months later. The company missed its own internal targets for weekly active users and monthly revenue.

Both sets of numbers are from the same year. Both are verified by primary sources. Both are simultaneously true.

And that’s exactly why the “AI crash” framing — which is everywhere right now — is the wrong lens. What’s happening isn’t a crash. It’s a sorting. And the difference matters enormously for how you should be thinking about any of this.

After completing this, I recommend reading this…

[embed]AI Is Quietly Replacing Junior Analysts. The Data Is No Longer Debatable. 120,000 tech jobs cut. Entry-level employment down 20%. $581 billion invested. The data on what’s actually happening to…medium.com

I. The numbers that look like a bubble — and the ones that don’t

Let me be precise about what the investment data actually shows, because most of the coverage collapses important distinctions.

Global corporate AI investment reached $581.7 billion in 2025, up about 130% from $253 billion in 2024, surpassing the prior record of $360 billion set in 2021. Of that, private investment was $344.7 billion, with generative AI capturing $170.9 billion — nearly half of all private AI funding.

Those numbers look like a bubble from the outside. But two things make them structurally different from classic bubble dynamics.

First, concentration. Five companies raised 20% of all AI VC funding in 2025. OpenAI and Anthropic alone absorbed roughly 14% of every venture dollar invested worldwide. This isn’t broad-based speculation — it’s extremely concentrated capital flowing to a small number of bets on foundational infrastructure. The dot-com bubble had thousands of companies receiving speculative capital. The current AI cycle has a handful of foundation model labs receiving most of the serious money, with everything else fighting over scraps.

Second, the consumer value is real. The value US consumers get from generative AI hit roughly $172 billion a year by early 2026, up from $112 billion twelve months earlier, with the median value per user tripling. This isn’t like the dot-com era, where traffic numbers were used as a proxy for imagined future value. People are genuinely extracting value from AI tools right now, today, at scale.

But here’s the uncomfortable counterweight, and it’s the number the bulls keep dancing around: AI company revenue is climbing at historically fast rates, but compute costs are climbing right alongside it. Google alone reported more than $150 billion in capital expenditure in 2025. Revenue scale is real. The meter running on that revenue is also real. And the meter does not stop when you add users.

This is the structural problem at the centre of the “crash” narrative. Not that AI isn’t valuable. It is. But enormous value is being created, and most of it is landing with consumers rather than with the companies spending the most to create it.

II. OpenAI: the most important loss statement in tech history

The OpenAI financials, verified by the Financial Times in June 2026, tell the story of the entire AI economy in one company.

OpenAI lost $38.53 billion attributable to the company in 2025, on $13.07 billion in revenue against $34 billion in total costs and expenses, with a $20.92 billion operating loss. Revenue grew 250% year-over-year. Operating losses grew 138%. The faster OpenAI grows, the more money it loses.

This is not the standard startup narrative of burning cash to acquire users who will eventually pay. The cost structure is the problem. Internal projections show $14 billion in losses for 2026 alone, with cumulative losses of $115 billion through 2029 before reaching profitability sometime in the 2030s.

For comparison: the Manhattan Project cost roughly $30 billion in today’s dollars. OpenAI expects to lose nearly four times that before it breaks even.

The circular financing problem runs deeper than the headline losses. Nvidia has committed up to $100 billion to OpenAI, money that, as OpenAI’s own CFO acknowledged, “will go back to Nvidia” in GPU purchases. Nvidia is a prominent investor in CoreWeave, which supplies cloud capacity to OpenAI and has spent billions buying Nvidia chips. Every dollar in the system flows through Nvidia. Nvidia is, in a meaningful sense, the only entity in the AI economy that is structurally guaranteed to profit regardless of which foundation model wins.

OpenAI’s financial tension centres on its infrastructure commitments — approximately $600 billion committed to building data centres over coming years. Key deals include a $300 billion cloud deal with Oracle over five years and an $11.9 billion contract with CoreWeave. The company is building the infrastructure for a future it’s betting will arrive before it runs out of capital.

What makes OpenAI’s situation interesting rather than simply catastrophic is the revenue trajectory. Revenue surpassed $20 billion by year-end 2025, a milestone that took Google seven years and Facebook six years to reach. The question is not whether OpenAI has a real product. It clearly does. The question is whether the economics of that product can ever justify the infrastructure required to run it.

III. The layer that’s actually crashing: AI wrappers

Here’s where the “crash” framing has real merit, and where most of the carnage is actually occurring.

Between 2022 and 2024, thousands of startups raised money on a simple thesis: take an LLM API, build a clean interface on top of it, charge subscription fees. AI copywriters. AI customer support tools. AI meeting summarisers. AI legal research assistants. AI email drafters. Hundreds of companies selling access to the same underlying model wrapped in a slightly nicer UI.

The most vulnerable segment of the market consists of “AI wrappers” — startups that provide a functional layer or user interface on top of third-party LLM APIs. In the 18-month window following early 2026, approximately 80% of these firms are expected to disappear as their lack of defensible moats is exposed.

The mechanism of death is straightforward. OpenAI adds a feature natively to ChatGPT. The startup that charged $60/month for that feature in a wrapper loses its value proposition overnight. Capital in the VC landscape has matured — investors are no longer impressed by technical complexity alone. They are seeking startups with unique data assets, proprietary training methodologies, and exclusive access to specific market segments. The era of “growth at all costs” has definitively ended.

This is the legitimate crash. Not a crash of AI itself. A crash of undifferentiated AI product businesses that were always going to be commoditised when the underlying models improved and the platform providers noticed the opportunity.

The dot-com parallel is instructive here. After 2001, most “.com” companies died. But Amazon didn’t. Google was just getting started. The crash eliminated the businesses with no moat while accelerating the ones that had built something real. The AI wrapper crash is following the same pattern — brutal for undifferentiated products, irrelevant for companies with genuine distribution, data advantages, or structural moats.

IV. The US-China gap that has effectively closed

The US still invests 23 times more than China in private AI capital — $285.9 billion versus $12.4 billion in 2025. But the investment gap has not produced the capability gap people expected.

The US-China model performance gap has effectively closed. The top US model now leads by just 2.7% as of March 2026, down from a 17.5 to 31.6 point gap in May 2023. Chinese open-weight models are now viable enterprise alternatives. DeepSeek-R1, released in February 2025, briefly matched the top US model and sparked a market sell-off in NVIDIA shares that wiped hundreds of billions in market cap in a single day — because it demonstrated that competitive frontier models could be trained at a fraction of the assumed cost.

The geopolitical implications are significant and the article you’re reading doesn’t have space to do them full justice. But the analytical implication for anyone building AI-dependent businesses is this: the assumption that US models maintain a durable performance lead is no longer supportable by the data. Competition has arrived. It arrived faster than most expected.

V. What’s not crashing: the infrastructure layer

If wrappers are the crash and foundation models are the uncertain bet, the infrastructure layer is the clearest winner in the current cycle — and it’s worth understanding why.

NVIDIA’s position is unlike any company in the dot-com era. During the internet bubble, Cisco sold the routers the internet ran on. When the bubble popped, companies stopped buying routers. NVIDIA’s situation is different: even if every application-layer AI company fails, someone will still need to train the next frontier model. And training frontier models requires NVIDIA GPUs, because the software ecosystem has been built around CUDA for twenty years and switching costs are prohibitive.

Global corporate AI investment hit $581.7 billion in 2025, more than doubling the prior year. The United States accounted for $285.9 billion. On SWE-bench Verified, which measures autonomous software engineering, model performance rose from 60% to near 100% of the human baseline in a single year. Each of those capability improvements requires another round of training runs. Each training run requires GPU capacity. NVIDIA benefits regardless of which model wins.

The cloud hyperscalers — Microsoft, Google, Amazon — are in a similar structural position. They supply the compute that runs AI. Major cloud providers have accelerated capital expenditures, with Google reporting more than $150 billion in annual capex in 2025. That’s not money at risk from an AI application crash. That’s money building the pipes through which all AI value flows.

VI. The real question — who captures the value?

Here’s the framework I keep coming back to when I look at all of this data together.

The consumer surplus from AI is real and growing. Estimated US consumer surplus reached $172 billion annually by early 2026, up from $112 billion a year earlier, with the median value per user tripling over the same period. Most of these tools remain free or close to it.

That last sentence is the whole story. The people generating the most value from AI in 2026 are users paying nothing for it. The companies spending the most to create that value are losing tens of billions of dollars annually creating it. The only entities reliably capturing value are the chip makers and the cloud providers whose infrastructure enables everything else.

This isn’t a crash. It’s a redistribution — from investors and company employees to users and hardware suppliers. The question for everyone else in the ecosystem is whether that changes before the capital runs out.

The $400 billion structural gap between AI infrastructure spending ($527 billion consensus estimate for 2026) and enterprise revenue realised from those investments (approximately $100 billion) is described as the single biggest risk to the AI investment thesis.

Four hundred billion dollars. That’s the gap between what’s being spent and what’s being earned. That’s not a gap you close quickly. And unlike the overbuilt fiber optic networks of the dot-com era, the current GPU infrastructure is heavily utilised — which means the spending isn’t obviously irrational. It’s just not yet profitable.

VII. Three scenarios for what happens next

The data supports three plausible trajectories, and I don’t think any analyst who looks honestly at the numbers can tell you with confidence which one plays out.

Scenario 1: The productivity payoff arrives

The current capital deployment is a rational bet on productivity gains that haven’t shown up yet in aggregate economic statistics. A National Bureau of Economic Research study from February 2026 found that despite 90% of firms reporting no current impact of AI on workplace productivity, executives still project a 1.4% increase in the near future. If the productivity gains arrive — if AI agents genuinely begin compressing knowledge work at scale — the revenue that currently looks inadequate to justify the spending will grow rapidly enough to close the gap. This is the bull case, and the historical parallel is electrification: years of capital spending before productivity shows up in the data.

Scenario 2: Commoditisation destroys margins everywhere

The US-China capability convergence continues. Model performance becomes essentially equivalent across providers. Inference prices collapse further (they’ve already fallen 99.5% since GPT-4). The only way to compete is on price, and nobody can compete on price while spending $150 billion per year on compute. The wrapper companies die. The foundation model labs consolidate or fail. Only the infrastructure layer survives profitably. This is the bear case, and it’s not implausible given the current trajectory.

Scenario 3: The agents unlock the revenue

AI agent deployment was in the single digits across nearly all business functions as of March 2026. Every major AI company is betting that agents — AI systems that can take autonomous action across tools and workflows — are where the value capture happens. Not in answering questions but in doing work. Cowork usage data shows 33.4% of sessions are business process and operations tasks — compiling reports, reconciling spreadsheets, building trackers. Only 8.7% is software development. If AI agents can genuinely handle the “work around the work” at scale, the revenue opportunity is significantly larger than the current subscription model implies. This is the middle case, and it’s why companies like Anthropic are building Cowork alongside their foundation models.

The close: what the crash conversation gets wrong

The binary framing — crash or no crash — is the wrong question.

What’s happening is a differentiation of an asset class that was briefly treated as monolithic. “AI” is not one thing. It never was. It is NVIDIA’s chips and OpenAI’s losses and Anthropic’s enterprise contracts and 80,000 wrapper startups dying and 170 million new jobs projected by the WEF and 120,000 tech layoffs in 2026 alone.

Some of those things are crashing. Some are compounding. The question isn’t whether to believe the bulls or the bears. It’s which part of the system you’re trying to understand.

The crash that is happening is real, selective, and probably necessary. Undifferentiated products with no moat are being commoditised by the platform providers who created them. That’s not a failure of AI — it’s AI working exactly as economic theory would predict.

The crash that isn’t happening is the one most people imagine when they use the word. The infrastructure investment is accelerating, not slowing. The capability improvements are continuing. The consumer adoption is real. The value being created is measurable and growing.

What’s uncertain — genuinely uncertain, in a way that the current valuations don’t fully price — is whether the companies spending the most to create that value will ever capture enough of it to justify the spending.

That’s not a crash question. That’s a business model question.

And it’s the most important open question in technology right now.

Primary Sources

• Stanford HAI AI Index 2026 — hai.stanford.edu/ai-index/2026-ai-index-report
• OpenAI audited financials 2025, verified by Financial Times, June 16 2026
• Layoffs.fyi tracker (ongoing), cross-referenced TechCrunch running list
• Crunchbase AI funding data, 2025 annual and Q1 2026

Research Reports

• National Bureau of Economic Research, February 2026 — productivity paradox study
• Cornell University — AI adoption hiring reduction study
• Goldman Sachs — entry-level employment decline analysis

Company Reports

• Google FY2025 annual capex disclosure
• Microsoft AI revenue reports 2025
• Similarweb Global AI Tracker 2025–2026

News Sources

• Financial Times, June 16 2026 — OpenAI financial verification
• Wall Street Journal — OpenAI revenue miss, user targets, $600B commitment
• TechCrunch — major tech layoffs naming AI, July 2026

I write about data, AI, and careers in plain language. Follow @analystuttam for more.