THE NUMBER: 12x — For every dollar the hyperscalers earn from AI today, they’re spending twelve dollars building more capacity. That’s $575 billion in capex this year. Alphabet just issued a century bond — the first by a tech company since Motorola in 1997 — to fund it. The debt matures in 2126. The chips it buys will be obsolete by 2029.

Anthropic now wins 70% of new enterprise deals in direct matchups with OpenAI, according to Ramp’s March 2026 AI Index. Claude Code generates $2.5 billion in annualized revenue. OpenAI’s Codex manages $1 billion. OpenAI’s enterprise share dropped from 50% in 2023 to 27% by end of 2025. Fidji Simo told the entire staff to stop chasing “side quests.” The enterprise market had already made its choice.

This isn’t a morality play about the Pentagon. It’s a product story. Two years ago, nobody outside of tech knew who Anthropic was. ChatGPT was everywhere — it had the brand, the consumer base, the first-mover advantage. What it didn’t have was a model lead that held. OpenAI’s gap evaporated. Claude got better. Claude Code captured the developers. Claude Cowork captured everyone else. The coders went first. The wanna-be coders followed. And once the people building things chose their tool, the enterprise contracts followed.

Meanwhile, the hyperscalers are writing checks that would make a defense contractor blush. Tomasz Tunguz calculates that Amazon, Microsoft, Alphabet, Meta, and Oracle will spend 90% of their operating cash flow on AI data centers in 2026 — up from a historical average of 40%. Morgan Stanley projects $1.5 trillion in total AI borrowing over the next few years. To justify a five-year payback at 60% gross margins, they need $180 billion in annual AI revenue. Current AI revenue: $35 billion. They’re underwriting 5x growth — and if NVIDIA’s twelve-month chip cycle makes hardware obsolete in three years instead of five, the required revenue jumps to $276 billion.

The best product is winning. But the best product still can’t fund its own infrastructure. That tension is the story.

Claude Won the Coders. The Pentagon Was a Press Release.

The conventional narrative goes like this: Anthropic refused the Pentagon’s demands to remove safety guardrails, got blacklisted as a “supply chain risk,” 30+ employees from OpenAI and DeepMind publicly defended them, the #QuitGPT movement hit 2.5 million supporters, and Anthropic rode the wave of moral authority to enterprise dominance.

It’s a good story. It’s also wrong.

Anthropic didn’t win enterprise because a bunch of Democrats downloaded Claude after reading about the Pentagon. They won because the product is better and the model lead flipped. Benedict Evans put it bluntly in this week’s newsletter: OpenAI has no unique tech, limited engagement, no network effect. The incumbents have matched the technology and are leveraging distribution. When Evans — who is not given to hyperbole — says the structural advantages aren’t there, that’s a signal worth noting.

The real story is what happened in the code editors. Claude Code didn’t just match Copilot — it pulled ahead on the tasks that matter: complex reasoning, multi-file refactoring, the kind of agentic workflows where a developer says “fix this” and walks away. Then Cowork brought that same capability to people who don’t write code for a living but need to get things done. The non-technical founder who can now build a prototype. The operations manager who can automate a workflow. The writer who can turn scattered files into finished work. That’s not a niche. That’s the entire knowledge economy.

The Pentagon drama did one useful thing: it got Anthropic on the front page of newspapers that don’t usually cover AI model benchmarks. Brand awareness matters when your competitor has been a household name for two years and you haven’t. But awareness without product is just marketing. Anthropic had both — and the Ramp data shows it. One in four businesses on Ramp now pays for Anthropic, up from one in twenty-five a year ago.

OpenAI knows this. Sam Altman has been telling interviewers they’re refocusing on the core product. Fidji Simo is killing the side quests — Sora, the Atlas browser, the Jony Ive hardware device. But here’s the problem with pivoting back to enterprise after you’ve spent two years chasing consumers: the enterprise customers already picked somebody else. At 70% win rates, Anthropic isn’t just ahead. They’re building switching costs that compound with every deployment.

The signal for allocators: Don’t confuse the Pentagon story with the product story. The morality argument made for good press. The 70% enterprise win rate is what makes for a good business. If you’re evaluating AI companies, look at Ramp’s April data. If Anthropic’s share keeps climbing past 70%, OpenAI’s Q4 IPO valuation takes a direct hit.

The Railroads Are Being Built. The Cars Are Already on the Road.

Here’s Tunguz’s math, and it should keep every capital allocator awake tonight. The hyperscalers are spending $575 billion this year on AI infrastructure. They’ve issued $159 billion in bonds in 2026 alone, up from $20 billion annually before the AI boom. Alphabet’s century bond. Oracle with $260 billion in off-balance-sheet lease obligations. KKR and Blackstone walking away from data center financing because the facilities lack adequate natural disaster insurance. Google sourcing liquid cooling equipment from Chinese suppliers because domestic capacity can’t keep pace.

This is the railroad construction boom of the 1860s. Lay track as fast as you can. Borrow everything. Bet that the cargo will come. And for the railroads, the cargo did come — for a while. Then along came the automobile, and the interstate highway system, and the intermodal shipping container. The companies laying the most track right before those innovations arrived were, as the British say, well and truly screwed.

The AI equivalent of the automobile is already on the road. It’s called specialization.

If you’re sick and want to consult AI on your condition, do you want to talk to ChatGPT? Sure, if there’s nothing else. But you’d rather talk to the Harvard Medical School model — trained specifically on clinical data, refined by the best diagnostic minds in the world, drawing on proprietary datasets that never touched Reddit. That model doesn’t need a $2 billion training cluster. It needs a fraction of the compute. And it’s worth more per query than any general-purpose model, because you’re not paying for intelligence. You’re paying for judgment.

Intelligence is commodity. Judgment is not.

Consider the best maritime attorney in the world. He bills $2–3 million a year because there are only so many hours in a day. But his knowledge isn’t capacity-constrained. If you capture that judgment in a specialized model — every case he’s tried, every brief he’s filed, every negotiation strategy he’s deployed — and sell it to every shipping company, insurer, and port authority simultaneously, you haven’t automated a $3 million lawyer. You’ve created a $300 million judgment engine. The constraint that made expertise scarce just broke.

Now run that logic against the $575 billion capex bet. The hyperscalers are building massive centralized infrastructure optimized for frontier general-purpose models. But the highest-value AI won’t be general purpose. It will be the oncologist’s diagnostic engine, the maritime lawyer’s judgment model, the structural engineer’s risk assessment system. Specialized intelligence running on commodity compute. The railroad is being built, but the cars are already on the road — and they can go places the railroad can’t.

Here’s where it gets structural. Google and Meta can fund this arms race from operating cash flow. They’ll overspend, they’ll write down assets, they’ll survive. But OpenAI and Anthropic? They’re at the mercy of investors. Of bond markets. Of wars and sentiment and the willingness of capital to keep writing checks against a 12:1 spend-to-revenue ratio. The winning product company — Anthropic, right now — still can’t self-fund its infrastructure. That’s a vulnerability that no amount of enterprise win rate fixes.

The smart play isn’t to keep pouring billions into general-purpose frontier training. It’s to take some of those $100 million executive compensation packages and hire five of the world’s best maritime specialists at $5 million each for three years. Build the definitive maritime law model. Do it across twenty high-value verticals. Partner with Harvard for the medical data, with top law firms for the legal data, with engineering societies for the technical data. You have the compute and the distribution — use them to build specialized judgment engines at premium prices. Because at that point, where else is anyone going to go?

The companies that figure this out won’t be the ones with the most GPUs. They’ll be the ones who understood that the $575 billion bought them a railroad, and the real money is in the cars.

What the numbers tell you: The hyperscalers need $180 billion in annual AI revenue to justify their capex at a five-year payback. Current revenue is $35 billion. If the highest-margin AI turns out to be specialized vertical models that run on a fraction of the compute, the general-purpose infrastructure gets repriced to utility returns — 8-12%, not 40% software margins. Nobody buying NVIDIA at 40x earnings is pricing in a utility multiple.

What This Means For You

The AI industry just split into two races, and most people are watching the wrong one. The model race — who has the best general-purpose AI — is becoming a commodity fight. The race that matters is who captures the highest-value judgment and locks in the distribution to sell it.

Stop evaluating AI companies on model benchmarks. The 70% enterprise win rate tells you more than any leaderboard. Product quality, developer experience, and switching costs are the moats now — not parameter counts.

Pressure-test your infrastructure assumptions. If your AI strategy depends on frontier compute getting cheaper, remember that the people providing that compute are spending 12x what they earn. Someone has to pay for that eventually, and it might be you.

Find the maritime lawyers in your industry. The highest-margin AI opportunity isn’t building another general model. It’s capturing domain-specific judgment that can’t be replicated from public data. If you’re sitting on proprietary expertise, it’s an asset that just became orders of magnitude more valuable.

Watch the capital structure, not the product demos. Anthropic has the best product and the worst balance sheet relative to Google. The company that wins the product war might still lose the funding war. That’s not a contradiction — it’s the central tension of this entire industry.

Three Questions We Think You Should Be Asking Yourself

If the best AI company can’t self-fund its infrastructure, what happens when capital markets tighten? Anthropic’s 70% win rate means nothing if they can’t raise the next round at terms that keep the lights on. Google can absorb a downturn. OpenAI and Anthropic cannot. Your vendor risk assessment should include a line item for “what if my AI provider’s investors get spooked?”

What proprietary judgment does your company have that’s worth more as a model than as a service? Every company has people whose expertise is capacity-constrained by hours in the day. The maritime lawyer billing $3 million a year is the extreme case, but the principle applies everywhere: if your best people’s knowledge could be captured and sold at scale, what’s the multiple? And if you don’t do it, who will?

Are the hyperscalers building railroads or highways? Centralized frontier infrastructure bets on a world where everyone needs the biggest model. Specialized vertical models bet on a world where everyone needs the smartest model for their specific problem. These are different worlds with different winners. Which one are you investing in — and does your portfolio reflect that view?

– Harry and Anthony

Sources

• Ramp March 2026 AI Index: Enterprise Spending Data
• Tomasz Tunguz: The 12x Bet on AI
• Benedict Evans Newsletter #634
• Marcus Schuler/Implicator: Seven Chips Ship. Seventy Percent Walk.
• Fortune: OpenAI Pentagon Deal Controversy
• Fortune: Anthropic Pentagon Blacklisting
• Wall Street Journal: OpenAI Side Quests Restructuring
• CO/AI: Chamath Says Your Portfolio Is Worth 75% Less Than You Think
• CNBC: Big Tech’s AI Bond Binge
• Fortune: Google, Meta, and Oracle’s $1 Trillion Borrowing Spree
• Bloomberg: Alphabet Century Bond