Last October I wrote about Sam Altman repeating "bubble" three times in 15 seconds, knowing exactly what the headlines would say. Bezos offered the correction: industrial bubble, not financial bubble. Industrial bubbles leave infrastructure. Society benefits. The companies that built the infrastructure go bankrupt.

I concluded that both were playing narrative control.

Seven months later, I think the entire debate was wrong.

$1.69. That's how many dollars OpenAI spends for every dollar it earns.

The company generated $13.1 billion in revenue in 2025 and spent approximately $22 billion to do it. The projected loss for 2026 is $14 billion. Cumulative losses through 2029: $44 billion. Current valuation: $852 billion.

Anthropic hit $30 billion in annualized revenue by April 2026 — up from $9 billion four months earlier. The fastest revenue growth of any company in recorded history. Valuation: $380 billion. Reportedly pursuing a round above $900 billion. Gross margins improved from negative 94% in 2024 to 40% in 2025 — better than OpenAI, but still $5.2 billion in EBITDA losses.

Behind both companies stands the infrastructure layer funding the entire cycle. Google plans $175 to $185 billion in AI capex for 2026 — double what it spent in 2025 — while simultaneously investing $40 billion in Anthropic. Meta: $115-135 billion. Microsoft: $190 billion. Combined hyperscaler AI infrastructure spend this year alone exceeds $600 billion. That's more than the military budget of every country on Earth except the United States.

Read those numbers again.

Revenue is exploding. Losses are exploding faster. Valuations are exploding fastest. And yet — nobody is panicking. Investors are lining up. Capital keeps flowing in.

If OpenAI raises prices, Anthropic takes the market.

If Anthropic raises prices, DeepSeek and Llama take the market.

If both raise prices, open-source models — now matching frontier performance on most benchmarks at a fraction of the cost — absorb the demand.

If state-subsidized Chinese alternatives keep publishing weights under MIT licenses, the floor stays at zero regardless.

Every player's dominant strategy is to keep subsidizing. Defecting — raising prices, cutting investment — means losing share to whoever doesn't. This is a textbook prisoner's dilemma. The Nash equilibrium is permanent margin compression.

And unlike the classic version, there's no way out. In a two-player game, repeated interaction can produce cooperation. Here you have OpenAI, Anthropic, Google, Meta, DeepSeek, Mistral, Qwen, plus state actors, plus open source. The MMLU benchmark gap between open-source and proprietary models narrowed from 17.5 to 0.3 percentage points in a single year.

Too many players. Too many incentives. No coordination mechanism.

The equilibrium holds. The subsidy continues. And it's not accidental. It's the only rational move available.

Now look at where the money is actually coming from.

Anthropic's Claude Code: $2.5 billion in annualized revenue by February 2026. From zero. In nine months. More than half of all enterprise spending on Anthropic products. OpenAI's coding tools, Cursor, GitHub Copilot — same trajectory. Coding assistants are the killer app of the current AI era.

They are also the product that destroys its own customer base.

Think about what a coding assistant does. It makes every developer more productive — at building SaaS platforms, internal tools, custom applications, everything that involves writing code. The agency building client projects needs fewer developers. The in-house team needs a smaller headcount. The freelancer is competing against a tool that costs $20 a month.

The Stanford AI Index 2026 already measured the result: employment for software developers aged 22-25 dropped nearly 20% since 2024. Older developers' headcount grew 6-12% in the same period. Entry-level tech job postings fell 67%.

The product works. It just works against itself.

Right now this paradox is invisible because coding assistants are still propagating. Adoption is expanding faster than the compression it produces. But propagation has a ceiling. Compression does not.

And if the trajectory holds — if AI moves from assisting developers to replacing the development task itself — the coding assistant loses its customer entirely. The developer was always an intermediary in a chain being compressed from both ends.

Which raises the question: if coding assistants cannibalize the developer market, and the developer market is what makes coding assistants valuable — who exactly is the long-term customer?

Follow the capital through every layer:

Every layer subsidizes the destruction of the next.

And the startups built on top of frontier models? They're burning their own investors' money to pay API prices that are already subsidized by the lab's investors. One pool of capital, draining from both ends. Nobody in the chain is self-sustaining.

On May 4, 2026, OpenAI and Anthropic launched private-equity-backed deployment companies. On the same day.

OpenAI's DeployCo: $4 billion, $10 billion valuation, 19 investors — TPG, Bain Capital, Brookfield. And McKinsey. And Bain & Company. And Capgemini.

Anthropic's venture: $1.5 billion. Blackstone, Goldman Sachs, Hellman & Friedman.

Read the investor lists carefully.

And the AI labs — whose model layer is commoditizing beneath them — are pivoting into consulting and deployment services as the next margin opportunity.

There's a problem with that plan.

McKinsey cut roughly 10% of its non-client-facing workforce over the past 18 months. Headcount dropped from 45,000 to 40,000. The firm deployed thousands of internal AI agents to automate tasks once handled by junior consultants. Bain, Deloitte, KPMG, Accenture, PwC, EY — all restructuring. All citing the same reason.

Consulting is the next market the same technology commoditizes. The surviving firms lean on scale and incomplete AI adoption. Not on immunity.

The labs are not escaping the chain by moving into services. They are entering another layer of it.

So if the model layer commoditizes and the service layer follows — what's left?

Position.

Eight hundred million weekly users depending on your infrastructure. That's not a business model in the traditional sense. It's a political asset. Regulatory influence. Defense contracts. National security relevance. Standard-setting power. The ability to sit in every room where AI policy is written.

Amazon ran AWS at thin margins for a decade to become infrastructure nobody could leave. Google owned search at 90%+. Meta owned the social graph. Those were winner-take-all markets. AI is not.

Six frontier families. Dozens of open-weight alternatives. State-backed efforts from China, the EU, and others. The competitive structure guarantees permanent fragmentation. No single player will achieve AWS-level dominance.

But being among the four or five that matter — in the US, in Europe, globally — still means something. Not $852 billion per company. Not monopoly returns. Not the next Google. But a seat at the table where the rules get written.

Political leverage with commodity economics attached.

Here is what I think is actually happening. It's not a bubble — bubbles pop. It's not a correction — corrections restore equilibrium. It's investment-driven commoditization: a game-theoretic structure where rational actors, funded by unprecedented capital, permanently compress every margin they touch.

No collapse is coming. No winner-take-all is coming either.

What's coming is a permanent repricing of intelligence toward utility economics. The AI labs know this. They can't stop it. They can only position themselves to be among the survivors — and surviving at commodity margins with massive scale is a real outcome. Just not the outcome the current valuations are pricing in.

This won't play out like AWS, Google, or Meta. The timing is different. The competitive structure is different. The margins will be lower — if they exist at all. But owning a major frontier model still means something. The same way owning critical infrastructure has always meant something. Not because of the returns. Because of the leverage.

I wrote in The AI Agent Paradox that AI does not generate expertise. It amplifies existing expertise. Pieter Levels built a $3 million solo portfolio because he had ten years of audience before the tools arrived. Peter Steinberger got acquired by OpenAI because he had thirteen years of building software companies, not because of a Friday night hack.

That observation was about individuals. This article is about the entire industry structure. The conclusion is the same, viewed from the other end.

The game-theoretic trap compressing every margin in AI — models, SaaS, development, consulting — is the best thing that has ever happened to anyone with the competence to use these tools. The subsidy means the cost of intelligence is approaching zero. The competition guarantees it stays there. Six frontier models, dozens of open-weight alternatives, state-backed efforts across continents — all fighting to give you cheaper, better tools.

European SMEs are structurally closer to the AI-native micro-team model than to the enterprise adoption model. You don't have the data silos. You don't have thirty years of legacy systems. You don't have change management for hundreds of people. You're small enough to redesign how you work around these tools. The enterprises with 40% failure rates on agentic AI projects are playing a different game than you are.

But the tools alone change nothing — not without domain expertise and the willingness to change how you operate. The question is no longer whether to adopt. The economics made that decision for you. The question is whether you're willing to change the way you work in response.

OpenAI financials and valuation Investing.com: OpenAI's Real IPO Risk Is Financial Transparency | Fortune: OpenAI Plans Stunning Annual Losses Through 2028 | Yahoo Finance: OpenAI Forecast — $14 Billion Loss by 2026 | TechCrunch: OpenAI $100B Deal at $850B+ Valuation | Sacra: OpenAI Revenue & Funding

Anthropic financials and Claude Code VentureBeat: Anthropic Hits $30B Revenue Run Rate | Sacra: Anthropic Revenue & Funding | MindStudio: Claude Code $2.5B Annualized Revenue | European Business Magazine: Anthropic Margins and Losses | SaaStr: Anthropic Passed OpenAI in Revenue

Hyperscaler AI infrastructure spending CNBC: Alphabet Resets the Bar for AI Spending | Fortune: Microsoft, Meta, Google AI Capex | DCD: Google Estimates 2026 Capex Up to $185B

DeployCo and AI consulting ventures Axios: OpenAI DeployCo Launch | Axios: Private Equity and the AI Giants | AI Business: OpenAI Consulting Company | Nerd Level Tech: OpenAI and Anthropic vs. Consulting

Developer employment and SaaS repricing Stanford HAI: 2026 AI Index — 12 Takeaways | SaaStr: The SaaS Rout of 2026

Open-source competition MindStudio: DeepSeek V4 Closes the Gap | Swfte AI: Open Source Models Rival Proprietary Giants | AI Productivity: Open Source Models 2026 Compared

Consulting industry restructuring Fast Company: McKinsey Layoffs as Warning Signal | Entrepreneur: McKinsey Plans 10% Cuts | Metaintro: McKinsey Layoffs 2026

Fabio Lauria

CEO & Founder, ELECTE

Every week, we explore AI without the hype—using data, analysis, and an independent perspective.