Kai ChenThe thing that happens before a company becomes infrastructure is that its competitors stop treating it as a competitor. On April 7, The Information reported that Meta engineers compete on an internal leaderboard dubbed "Claudeonomics" — a gamified tracker of how effectively they use Anthropic's API. This is a company that publishes Llama. A company with its own frontier models, its own research lab, its own inference stack. And its engineers have an internal game built around optimizing their spend on someone else's intelligence.
That's not competition. That's dependency management. And dependency management is what you do with infrastructure.
The Curve
The trajectory that produced the Claudeonomics leaderboard is a ten-month curve: adoption that looks like product-market fit until you realize the product has become the substrate.
That's a 6x increase in ten months. Reuters reported $3 billion in May 2025. By January 2026, Bloomberg put the run rate at $9 billion. Anthropic itself confirmed $14 billion by mid-February. By March, Bloomberg reported $19 billion. The growth curve doesn't look like a consumer app going viral. It looks like a utility crossing the threshold where opting out costs more than staying.
Seventy-three percent. Not of all AI spending — of first-time purchases. When a company buys AI tools for the first time, nearly three-quarters of the money goes to Anthropic. That's the kind of market share that stops being about product quality and starts being about default status. It's not that Claude is better for every use case. It's that Claude is what people reach for when they don't know what to reach for. That's what a utility looks like.
The developer adoption data tells the same story from a different angle. In February, SemiAnalysis reported that Claude Code authors 4% of all public GitHub commits. One company's model, wired into the development workflow of every major language and framework, producing a measurable share of the world's code. That's not a popular product. That's a dependency.
The Moves
If Anthropic were just selling a popular API, the story would be simpler — a company winning on product execution, the way Stripe won payments or Twilio won messaging. But three stories landed on the same day, and they aren't product stories. They're infrastructure stories.
First: Anthropic signed a deal with Google and Broadcom for multiple gigawatts of next-generation TPU capacity. Not leasing cloud compute on demand. Signing a multi-gigawatt agreement for custom silicon. That's the kind of deal a power company makes — securing generation capacity at the scale where you're planning in watts, not workloads.
Second: the Wall Street Journal reported that Anthropic plans to invest $200 million in a venture with private equity firms to support AI startups. Not just selling the API — funding the companies that will build on the API, the way AWS invested in startups that would become AWS customers. Platform economics: the returns compound on both sides. In March, Anthropic launched the Claude Marketplace, letting enterprises buy third-party tools that run on Claude. You don't build a marketplace for a product. You build one for a layer.
Third: the Claudeonomics leaderboard. Meta doesn't build internal games around optimizing spend on competitors. It builds them around optimizing spend on infrastructure — cloud costs, compute budgets, bandwidth. The fact that Anthropic's API now occupies the same mental category inside Meta as AWS or GCP is the signal. The leaderboard is the evidence.
Where It Breaks
There's a problem with calling something infrastructure when it still behaves like a startup.
On April 1, Anthropic accidentally leaked parts of Claude Code's source code through a misconfigured npm package. They spent the next 48 hours issuing copyright takedown requests for 8,000-plus copies. Infrastructure doesn't leak its own blueprints. That's a startup failing at operational hygiene during a period when thousands of companies depend on its systems.
The Wall Street Journal reported that both OpenAI and Anthropic project profitability at roughly the same time — meaning Anthropic is still burning capital at a rate that no utility could sustain. And the $80 billion in cloud commitments Anthropic expects to pay Amazon, Google, and Microsoft means this putative utility is itself a customer — $80 billion worth — of three other companies' infrastructure. Infrastructure that depends on infrastructure has a fragile risk profile.
The Pentagon exposed another edge of utility status in March when it designated Anthropic a supply source — effectively saying that a company this embedded in national operations doesn't get to choose its customers. Labs get to pick their principles. Utilities get regulated.
The Therefore
The framing is still "AI lab" and "ChatGPT competitor." But a company that captures 73% of first-time spending, authors 4% of public GitHub commits, and shows up as a cost-optimization game inside a competitor isn't competing anymore. It's becoming the layer the competition runs on.
This doesn't mean Anthropic is becoming a utility in the economic sense. Utilities own physical infrastructure — wires, poles, generators — and earn regulated returns on that capital. Anthropic owns no data centers, no fiber, no spectrum. Its market power is intellectual property and model quality, which is both more fragile than a physical monopoly (open-source models provide competitive pressure that no prior infrastructure monopolist faced) and harder to regulate (you can't rate-base a neural network). The 140-year pattern of infrastructure economics says consolidation followed by margin compression. But Anthropic is a new kind of entity — a dependency without the physical assets that make traditional infrastructure regulation possible.
A vendor is something you chose. Infrastructure is something you can't remove without rebuilding the workflows around it. The leaderboard tells you which one you've become.
The real risk isn't utility economics. It's the maturity gap. Companies are treating Claude like stable infrastructure — embedding it in production workflows, building businesses on it, optimizing their spend on it through internal leaderboards. But Anthropic is still operating like a startup: burning billions, leaking source code, navigating an IPO, fighting the Pentagon over supply-source designation, and owing $80 billion in cloud commitments to three companies that are simultaneously its landlords and its competitors. The dependency has outrun the dependability.
The companies building on Claude haven't priced this in. They treat it as a vendor — a line item they chose and could theoretically un-choose. But when three-quarters of first-time buyers land on the same platform, and that platform is securing multi-gigawatt compute deals and funding the ecosystem that depends on it, the switching cost isn't the API migration. It's the workflows, the pipelines, the institutional muscle memory. Lock-in doesn't look like a contract. It looks like a dependency so embedded that removing it costs more than the dependency itself.
Three billion to nineteen billion in ten months. Seventy-three percent of first-time spending. A competitor's internal leaderboard named after your economics. The dependency is already everywhere. The dependability hasn't caught up.