Anthropic's IPO Filing Shows the AI Money Problem. Could It Cause a Financial Crisis in 2027?

Short answer: Anthropic's IPO filing shows an AI lab whose revenue is growing faster than almost any company in history, and whose computing bills are growing faster still. Across the industry, the biggest tech companies are now spending close to everything they earn on AI, and borrowing for the rest. That makes a sharp market correction a real risk for 2027. A 2008-style banking crisis is less likely, but there are specific weak points worth watching.

The filing matters because it's the first time we can see a frontier AI lab's books in full. Below: what it shows, how the wider AI build-out is being paid for, what central banks are warning about, and the signals that would tell you which way 2027 is going. This is an explainer, not investment advice.

What Anthropic's filing shows

Anthropic filed to list on Nasdaq in late October and amended the filing on 2 November. The figures below come from MLQ's coverage of the amended filing; we couldn't download the filing itself from the SEC's site, so treat them as reported rather than checked line by line.

Read those numbers two ways. The bullish reading: revenue is up more than twelvefold in a year, the business was roughly break-even at the operating level, and the huge net loss is mostly an accounting effect of how its earlier funding was structured. The worrying reading: computing costs rose from 48% to 58% of revenue in a single quarter, partly because short-term capacity cost far more than planned, and the company has signed up for close to $900bn of future spending. Revenue has to keep compounding for years to cover that.

The amended filing also added an unusual line. As reported by DataStudios, Anthropic says its expectations about future models indicate it "may imminently meet" its own safety policy's threshold for dramatically accelerated, AI-driven progress, while stressing that no current model has been designated as exceeding it. Investors will read that as a growth story. It's also a reminder of why these companies feel they can't stop spending: whoever falls behind on computing power risks falling behind for good.

The bigger bill: Big Tech's spending

Anthropic's costs are other companies' revenue. Its compute is bought from Amazon, Google and Microsoft, and those companies are building data centres at a pace with no modern precedent. From their October earnings (Amazon, Meta, CNBC):

Two things stand out. Cloud revenue really is growing fast, so this isn't spending into a void. But three of the richest companies in the world had negative free cash flow last quarter. According to Quartz, UBS estimates 2026 capex will consume about 100% of these companies' operating cash flow, against a ten-year average of about 40%. The rest has to come from somewhere.

Where the money comes from

• Bonds. Morgan Stanley expects around $570bn of AI-related debt issuance in 2026, per Quartz. Amazon and Meta have each sold tens of billions in bonds this year.
• Leases and commitments. Nikkei Asia estimates the off-balance-sheet AI obligations of Alphabet, Microsoft, Amazon, Meta and Oracle at about $3.2tn, up about 90% in six months. These are mostly leases and purchase commitments disclosed in footnotes, not secret loans, but they're obligations all the same.
• Off-balance-sheet vehicles. Meta's Hyperion data centre is financed through a joint venture with Blue Owl that carries $27.3bn of debt off Meta's balance sheet, backed by a Meta guarantee on the value of the site (Bisnow).
• Circular deals. AMD gave OpenAI warrants for up to about 10% of AMD in exchange for a huge chip order (Implicator), and Nvidia has put more than $40bn into AI companies this year, many of them its own customers (The Next Web). When suppliers fund their buyers, demand can look stronger than it is.
• Stretched balance sheets. Oracle, with roughly $125bn of borrowings, announced a $20bn raise in stock and bonds on 2 November, days after promising to cap new financing (BNN Bloomberg).

And OpenAI, the biggest buyer of all, is reported to be heading for about $50bn of annualised revenue against around $1.4tn of compute commitments, with the FT reporting projected cash burn of about $278bn through 2030 (via The Star).

What central banks are saying

The Bank of England's Financial Policy Committee said in September that the risk of these vulnerabilities crystallising has risen, and that a sharp repricing of AI companies could spill into funding markets. It cited a JPMorgan estimate of about $4.1tn of debt-financed AI investment over 2026 to 2030. The IMF's October Global Financial Stability Report listed an AI correction among its main risks; as reported by ResultSense, it estimated a moderate correction could knock about 0.4 percentage points off global growth.

So, a financial crisis in 2027?

It helps to separate a market crash from a financial crisis. In 2000, the dot-com bubble burst and wiped out trillions in share value, but it was mostly shareholders' money, and the recession that followed was mild. In 2008, the losses sat inside banks that had borrowed heavily, so they froze lending to everyone.

The AI build-out has features of both:

• Like 2000: Most of the spending comes from enormously profitable companies, and from shareholders in Nvidia, Microsoft and the rest. If AI stocks fall, a lot of that is paper wealth disappearing, painful for pensions and savers but not a banking collapse.
• Like 2008, in places: The fastest-growing part is debt: bonds, data-centre leases, private-credit loans and off-balance-sheet vehicles. If AI revenue disappoints, the pressure lands on the most leveraged borrowers and their lenders, and private credit is less transparent than bank lending.

Our reading, and it is only that: on today's evidence, a sharp correction in AI stocks within the next year or two is quite plausible; a systemic financial crisis would need the losses to land on leveraged lenders faster than the giants' cash flows can absorb them. That's possible, not the base case. The companies to watch are the ones without a profitable core business to fall back on.

Five signals to watch in 2027

1. Compute as a share of revenue. If Anthropic's compute bill keeps rising faster than revenue, the economics get worse with scale, not better.
2. Big Tech free cash flow. Whether Amazon, Alphabet and Meta return to positive free cash flow, or keep borrowing to build.
3. Credit spreads. Rising yields on Oracle's bonds or AI-related private credit would be the earliest sign that lenders are getting nervous.
4. Capex guidance cuts. The first big company to cut its AI spending plan would tell you demand isn't keeping up.
5. Anthropic's IPO reception. A strong debut for ANT would keep the money flowing. A weak one would make every other AI fundraise harder.

What it means if you just use AI

If the money gets tighter, expect AI companies to push harder on revenue: higher prices, tighter free tiers, more ads and more use of your data to improve their models. Tools that run on your own device are less exposed to that, because nobody has to pay a data-centre bill each time you use them. That's one reason we build our apps that way, but it applies to any on-device tool, ours or not.

Frequently asked questions

How much revenue does Anthropic make?

According to coverage of its amended IPO filing, Anthropic made about $16.96bn in revenue in the three months to 30 September 2026, up from $11.56bn in the previous quarter and $1.38bn a year earlier.

Is Anthropic profitable?

Not on paper. It reported a net loss of about $23.3bn for the quarter, but most of its losses this year are non-cash changes in the value of convertible notes and similar instruments. Its operating result was close to break-even: a $567m operating loss in the third quarter after a $974m operating profit in the second.

What are Anthropic's compute commitments?

About $895bn of purchase commitments and leases, with roughly $447.5bn due before the end of 2030, according to coverage of the amended filing.

Is there an AI bubble?

Nobody knows until afterwards. The IMF and the Bank of England have both warned that AI valuations and AI-related borrowing are a financial-stability risk. Revenue at the big cloud businesses is also growing fast, which is the strongest argument that this isn't pure speculation.

Could AI spending cause a financial crisis in 2027?

A sharp correction is a real possibility. A 2008-style banking crisis is less likely on the evidence so far, because most of the spending is funded by very profitable companies and by shareholders rather than by heavily indebted households and banks. The weak points to watch are the companies and lenders that have borrowed most against future AI demand.

Is this investment advice?

No. This is an explainer, not financial advice. We're a small software studio, not financial advisers.

More from Offgrid Studio

Why LLM Prices Fell 80%: The Cache Is the Bill
Why Cloud AI Bills Are So Hard to Budget
AI Agents Broke Out of Their Tests: What We Know