The Most Expensive Bet in Corporate History — And the Spreadsheet Nobody Wanted to Open

ChatGPT reached 100 million users faster than any product in history. Anthropic is generating $30 billion in annualized revenue. These are real numbers.

What's cracking isn't the technology. It's the financial architecture built around it — the pricing models, the capital structures, the ROI assumptions that enterprise finance teams wrote into 2026 budgets without fully understanding what they were buying.

And the evidence is now sitting in enough spreadsheets, earnings calls, and analyst reports that it is no longer possible to call it contrarian noise.

Here's what the numbers actually say.

Here is the number that should be pinned on every CFO's monitor in 2026.

AI-related services delivered roughly $25 billion in revenue in 2025. Against more than $250 billion in infrastructure spending that year. Ten cents of revenue returned per dollar of capital deployed. Goldman Sachs ran the numbers and concluded that for hyperscalers to maintain returns on this investment, they would need to generate over $1 trillion in annual AI profit. The consensus estimate for 2026 AI profit is $450 billion. Less than half.

The Big Five — Amazon, Microsoft, Google, Meta, Oracle — are projected to spend over $600 billion on infrastructure in 2026 alone, a 36% increase over last year. Roughly $450 billion of that is directly tied to AI. To fund the gap between capex and free cash flow, hyperscalers raised $108 billion in bonds during 2025 alone. Amazon, Microsoft, and Google — companies that have historically funded operations from internal cash — now have aggregate capital expenditure that exceeds their projected cash flows. For the first time in their histories.

Larry Page, Google's co-founder, was quoted saying he is "willing to go bankrupt rather than lose this race." That sentence, from one of the most financially disciplined technology executives in history, tells you everything about where this market is psychologically right now.

$600 billion spent on AI infrastructure in 2026. $25 billion in AI revenue returned in 2025. That is ten cents on every dollar deployed. Goldman Sachs says you need $1 trillion in profit to justify the bet. The consensus says $450 billion. Do the math.

The invoices are no longer hypothetical. They've arrived. And some of the companies receiving them have started talking.

In December 2025, Microsoft opened access to Claude Code — Anthropic's command-line AI coding agent — to thousands of its engineers, product managers, and designers. The rollout was framed internally as a learning exercise. By spring 2026, Claude Code had become the preferred tool inside Experiences and Devices: the division that builds Windows, Microsoft 365, Outlook, Teams, and Surface.

In May 2026, emails went out. Engineers were told to stop using Claude Code and migrate to GitHub Copilot CLI by June 30 — the last day of Microsoft's fiscal year. The official explanation was toolchain unification. The real explanation fits on a single row of a budget spreadsheet.

Wait. Microsoft Is the Investor Who Just Got Billed by the Competition.

Here's the detail that almost every headline about this story has quietly skipped over. And it's the most important one.

Microsoft has invested $13 billion into OpenAI. That is not a rounding error. That is a strategic, multi-year commitment to the idea that OpenAI's models — not Anthropic's, not Google's, not anyone else's — are the foundation of Microsoft's AI future. GitHub Copilot, the tool Microsoft now wants all its engineers to use, runs on OpenAI's models. The entire Microsoft AI product stack is built around that $13 billion bet.

And yet — in December 2025 — Microsoft gave its own engineers Claude Code. Anthropic's product. The direct rival to the company they funded. Not by accident. Not by a rogue team going off-script. By choice.

What happened next is the part nobody's saying out loud.

The engineers chose Claude Code. Over Copilot. Over the product built on their own company's $13 billion investment. By spring 2026, Claude Code was the preferred tool inside the division building Windows and Microsoft 365.

So Microsoft found itself in a position so structurally absurd it almost reads like satire: paying monthly bills to Anthropic — the competitor to their $13 billion investment — for a product that was winning against the product their $13 billion was supposed to have already made dominant. The bills grew large enough that they cancelled it. They then sent their own engineers back to Copilot. The tool nobody had chosen when given a free choice.

There is a charitable read and an uncharitable read here. The charitable one: this was a genuine experiment, and Microsoft learned what the competitive landscape looks like from the inside before making a rational financial decision to consolidate on a platform they can shape. The uncharitable one: their investment's product couldn't win a free market inside their own offices, and the only fix was to end the free market.

Both are probably partially true. What is not in dispute: when the company with the most negotiating leverage of any enterprise on earth walks away from a vendor because the bills got too large — that is not a vendor quality problem. That is a pricing model problem that nobody has solved.

Microsoft invested $13 billion in OpenAI. Then paid Anthropic — OpenAI's biggest rival — to run inside their own offices. Their engineers chose the competitor. The bills got too high. They banned it. Then sent engineers back to the tool nobody picked when given a free choice.

How Microsoft ended up paying its own competition — the full chain:

"Claude Code was an important part of that learning. At the same time, Copilot CLI has given us something especially important: a product we can help shape directly with GitHub."

— Microsoft internal communication, May 2026

Read between the lines: we learned what the best available tool looks like. Now please use ours.

The Numbers Microsoft Won't Say Out Loud — But Uber Will

Microsoft hasn't published what it spent on Claude Code. But Uber has — and the disclosure is the most instructive corporate admission of 2026.

Uber rolled out Claude Code to its roughly 5,000-engineer organization in December 2025. Adoption climbed from 32% of engineers in February to 84% by March. By April, Uber's CTO Praveen Neppalli Naga was telling The Information something that stopped boardrooms across the enterprise world: the company had burned through its entire 2026 AI coding budget in four months.

(Edit: I originally wrote that Uber's total 2026 AI budget was "$3 billion" in my notes. The correct figure for Uber's full-year 2025 R&D spend was $3.4 billion — and the point isn't the total R&D number. The point is that whatever specific allocation was made for AI coding tools was completely exhausted before May. I've corrected this.)

Individual engineers were spending between $500 and $2,000 per month on tokens. Naga himself spent $1,200 in a single two-hour demo session. Not a production workload. A personal demo. Around 70% of code committed at Uber by spring originated from AI tools, and approximately 11% of live backend updates were being shipped by agents with no human in the loop at all.

Uber made things worse by ranking engineers on internal leaderboards based on Claude Code usage. That turned a productivity tool into a token consumption competition with no budget ceiling. The teams driving adoption were not the same teams watching the spend.

Reminds me of when AWS launched EC2 in 2006 and every enterprise IT director I knew called it reckless — "you can't run serious workloads on rented servers with no cost ceiling." They were right about the cost ceiling part. They just underestimated how fast adoption would outpace the governance frameworks. Companies spent the next decade building FinOps practices to manage the AWS bill. We are watching the exact same movie right now, with tokens instead of compute instances. And most enterprises are still in the 2007 chapter — where the adoption is real but the governance doesn't exist yet.

What Gartner, McKinsey, Forrester, MIT, and BCG Are All Saying — At the Same Time

When five of the most cited research firms in enterprise technology all publish data pointing the same direction within a six-month window, that is not a bear case. That is a consensus.

Gartner placed generative AI in the Trough of Disillusionment on its 2026 Hype Cycle. Their January press release put a specific number on the damage: 25% of planned 2026 AI budget will slip into 2027 as proofs of concept fail procurement review. AI, Gartner said, "will most often be sold to enterprises by their incumbent software provider rather than bought as part of a new moonshot project." ROI predictability has to improve before any real scaling happens.

Forrester's 2026 predictions matched independently: enterprises will defer a quarter of AI spend to 2027 as financial rigor kills proofs of concept before they reach production.

McKinsey's November 2025 State of AI survey: over 80% of organizations report no meaningful EBIT impact from AI adoption. BCG's September 2025 research: 60% of companies generate no material value despite continued investment. MIT's Nanda group: 95% of organizations investing in generative AI are getting zero measurable return.

A separate Gartner read in April 2026: only 28% of AI infrastructure projects fully deliver against their stated business case.

Not 50. Not 60. 28.

And the most damning Gartner data point of all, from their May 2026 survey of 350 global executives: 80% of organizations using autonomous AI report workforce reductions — but those reductions do not translate into ROI. Companies are cutting people to prove AI is working. And then discovering it isn't working any better than before they cut anyone.

"Many CEOs turn to layoffs to demonstrate quick AI returns; however, this disposition is misplaced. Workforce reductions may create budget room, but they do not create return."

— Helen Poitevin, Distinguished VP Analyst, Gartner — May 2026

The Quiet Part — That Nobody's Saying In the Earnings Call

The AI bubble is not going to pop the way the dot-com bubble popped. There will not be a single day, a single company, a single front page that marks it. What is happening is more like a slow puncture — a series of quiet emails, missed budget lines, cancelled licences, deferred projects, and fiscal-year-end decisions that nobody announces with a press release.

Microsoft's email to its Windows and Surface engineers. Uber's CTO conversation in April. GitHub's pause on new sign-ups in November. Anthropic's new terms of service clause. Gartner's trough of disillusionment call. Forrester's 25% deferral forecast. MIT's 95% zero-ROI finding.

These are not isolated incidents. They are data points on the same curve. And together they describe a market that is repricing itself — not from technology failure, but from assumption failure.

The assumptions that are failing are specific: that token-based pricing would scale like SaaS. That enterprise engineers would naturally self-govern their consumption. That productivity gains would show up in the same P&L quarter as the AI investment. That the infrastructure build-out would generate returns faster than the debt that funded it.

None of those assumptions were crazy. They just weren't tested at this scale, at this speed, against finance teams that were built for a different kind of software cost.

The technology will survive this repricing. The companies with real revenue, real governance, and a CFO who can explain the AI line item to a board — those companies will come out of this stronger. What will not survive is the assumption that the rules of enterprise finance somehow don't apply to AI.

They do. They always did. We just needed a few very large invoices to prove it.

The AI bubble isn't bursting. It's bleeding — slowly, quietly, in spreadsheets that CFOs open on Friday afternoons and don't share with their teams on Monday morning.

Worldwide AI Spending Will Total $2.5 Trillion in 2026 — Official Press Release

Uber Burns Its 2026 AI Budget In Four Months On Claude Code — Forbes, May 2026

Forrester 2026 Predictions: Enterprises Will Defer 25% of Planned AI Spend — BusinessWire

Why AI Companies May Invest More Than $500 Billion in 2026 — Goldman Sachs

Microsoft banned Claude Code despite investing $13B in OpenAI. Uber burned its AI budget in 4 months. 95% of enterprises report zero ROI. The AI bubble isn't bursting — the money around it is cracking.


The AI Bubble Isn't Bursting. It's Bleeding — And Nobody's Talking About the Bill.

Want More AI Insights?

Related Articles