Amazon put $13 billion into Anthropic and locked in $110 billion back

I read an IPO filing on Sunday night. My wife was asleep. The dog was asleep. I sat on the couch with a cup of coffee I should not have been drinking at 11 p.m.

I do this too often. I know that.

But this one was different. Anthropic, the company behind the Claude AI models, filed its S-1 with the SEC on October 26. It wants to list on Nasdaq under the ticker ANT. Goldman Sachs and Morgan Stanley are leading the deal. Reports say the company could be valued at $1.5 trillion to $2 trillion.

The revenue growth is staggering. Anthropic brought in $16.3 billion in the first six months of 2026. A year earlier, that number was $1.16 billion. In three years the company went from $41 million in annual revenue to a $67.2 billion annual run rate.

That is the headline. Every outlet ran it Monday morning.

I kept reading.

On pages 171 and 172 of the prospectus, Anthropic lists its compute contracts. It has committed roughly $637 billion in spending to seven partners over the next decade. At least $114.5 billion goes to Google Cloud. At least $110 billion goes to Amazon Web Services. Another $44.9 billion goes to Microsoft.

Now look at the other side. Amazon has invested $13 billion in Anthropic and committed up to $20 billion more. Google has committed up to $40 billion. Microsoft and Nvidia together committed up to another $15 billion.

So the companies investing money into Anthropic are the same companies Anthropic is committed to pay that money back to… and then some.

Amazon puts in $13 billion. It gets $110 billion in committed cloud revenue back. Google puts in $40 billion. It gets $114.5 billion back. Reuters reported that 47% of Anthropic's 2025 sales flowed through Amazon and Google as distribution channels. The S-1 says 20% of first-half 2026 revenue came from just two customers.

In other words, the investors are also the suppliers, the distribution partners, and the biggest customers. The money goes in a circle.

This is not unique to Anthropic. Nvidia invested roughly $30 billion in OpenAI. OpenAI uses the money to buy Nvidia chips. Microsoft invested over $13 billion in OpenAI. OpenAI committed to $250 billion of Microsoft Azure cloud services. Amazon invested $50 billion in OpenAI in 2026 and got a commercial agreement for OpenAI to draw on Amazon computing capacity.

The research firm IDC described it as a pattern where the same capital flows simultaneously as vendor payment and equity stake. The BIS—the central banks' central bank—named it as one of four pressure points facing the global economy in its June 2026 annual report. It compared the setup to prior technology booms that ended in economy-wide recessions.

I have no idea when or if this breaks. The companies writing the checks—Microsoft, Alphabet, Amazon, Meta—have over $1 trillion in combined annual revenue. They can afford it. For now.

But the structure itself is worth watching. So that is what we are doing this week.

When large tech companies invest in AI startups and those startups commit to buy cloud services back, the arrangement creates spending obligations that stretch years into the future. In August 2026, the Wall Street Journal analyzed the footnotes of securities filings from nine major tech companies. It found roughly $3 trillion in off-balance-sheet commitments tied to AI infrastructure.

That is about five times the roughly $600 billion in capital expenditures those same firms reported.

The $3 trillion includes $1.2 trillion in leases and uncommenced leases, plus additional purchase and capacity commitments. A separate Nikkei analysis in July 2026 had already estimated $1.65 trillion in off-balance-sheet AI debt across five major tech companies.

These are not hypothetical. Many are non-cancelable take-or-pay contracts. If revenue growth slows at the AI labs, the commitments still come due. And the companies on the other side of those contracts—the cloud providers—are also the equity investors.

The BIS compared the AI investment boom to prior technology cycles that ended in economy-wide recessions.

So far, the revenue growth has kept pace. Anthropic went from $41 million in 2023 revenue to a $67.2 billion run rate. But pace is not permanence.

And here is where the strain shows up in credit markets. On July 9, S&P Global Ratings cut Oracle's credit rating from BBB to BBB−. One more notch and it becomes junk.

The reason was AI spending. Oracle reported fiscal 2026 capital expenditures of $55.7 billion—82.6% of its total revenue. Its free cash flow came in at negative $23.7 billion. Its long-term debt rose from $85.3 billion to $122.3 billion in a single fiscal year. S&P forecasts an even wider cash flow deficit of nearly $42 billion for fiscal 2027.

Oracle's five-year credit default swap reached 322 basis points, up from 144 basis points at the start of 2026. Its stock sits 63% below its high.

Oracle is not alone. Nvidia's five-year CDS reached 101 basis points—the highest since the contract started trading in late 2025. Meta's hit 108 basis points. These are not distressed levels. But the direction is clear. The credit market is pricing in more risk on the companies building the AI circle.

The bond market tends to notice problems before the stock market does. It is worth paying attention.

Meanwhile, the companies on the other side of those circular deals are feeling it too. In the second quarter of 2026, Alphabet reported capital expenditures of $44.9 billion—double the year-earlier quarter. That pushed its free cash flow to negative $5.86 billion.

It was the first time Alphabet had ever posted negative free cash flow as a public company. The company has been public since 2004. Through recessions, antitrust fights, and a global pandemic, it always generated cash. Until now.

Management guided 2026 full-year capex to $195 billion to $205 billion. Alphabet's long-term debt grew from $46.5 billion to $98.2 billion. It suspended its share buyback program.

Alphabet is one of the most profitable companies ever built. It has the revenue to support this level of spending. But the fact that even Alphabet—with over $350 billion in annual revenue—is now cash-flow negative on AI investment tells you something about the scale of commitments running through this system.

The question is simple. What happens to the smaller players when the biggest company in the circle is already burning cash?

Three more things worth keeping track of this week.

Office CMBS delinquencies keep climbing. The office commercial mortgage-backed securities delinquency rate reached 13.2% in August 2026, according to CRED iQ. That is the highest reading since at least 2019. The special servicing rate hit 15.7%, also a record. About 71% of the distressed office loan balance is tied to failed or looming refinancing rather than missed monthly payments. This is a slow grind, not a sudden break. The maturities keep coming.

European natural gas is back. The Dutch TTF benchmark traded near €96 per megawatt hour this week, up about 234% since January. Storage sits at roughly 75% of capacity, well below the 90% seasonal target, according to GIE AGSI+ data. The last time TTF was near these levels was December 2022, when Europe was still scrambling after cutting off Russian pipeline gas. Winter is eight weeks away.

US pending home sales fell to a record low. The National Association of Realtors' index dropped 7.3% in September to 66, the lowest reading in the survey's history. The 30-year fixed mortgage rate averaged 7.32% as of October 22, per Freddie Mac. At that cost of borrowing, the math does not work for most buyers. We will see.

You could look it up if you wanted. Lucent Technologies. It traded under the ticker LU.

Most people have forgotten it.

In 1999, Lucent Technologies was one of the ten most valuable companies in America. It had spun out of AT&T with the prestige of Bell Labs behind it. Its stock ran to about $84. Its market capitalization reached roughly $258 billion. If you owned a mutual fund in the late 1990s, you owned Lucent.

Lucent sold networking equipment to telecom companies that were building the internet. Business was extraordinary. Revenue hit $38 billion in fiscal 1999.

But there was a problem with how it was making those sales. Many of Lucent's customers were young telecom startups that did not have the cash to buy the equipment they needed. So Lucent lent them the money. It extended credit and loan guarantees to its own customers—up to $8.1 billion in total commitments, with about $2.1 billion outstanding at the peak.

The supplier was financing its own customers so they could afford to buy its products. The revenue looked great on paper. The cash was going in a circle.

When the dotcom bubble broke, those customers went under. WinStar Communications filed for bankruptcy. So did others. Lucent took $2.2 billion in bad-debt provisions in 2001 and another $1.3 billion in 2002. It wrote off $700 million from a single customer bankruptcy.

The SEC later found that Lucent had improperly recognized $1.148 billion in revenue and roughly $470 million in pre-tax income for fiscal 2000. The revenue growth that made the stock a Wall Street favorite was artificial. Lucent pulled revenue forward, booked it before the cash arrived, and tied some of it to customers who could never pay.

The stock fell from $84 to under $1. The company merged with Alcatel in 2006 at a fraction of its peak value. More than $250 billion in shareholder wealth was gone.

I am not saying today's AI companies are Lucent. The hyperscalers have real revenue, real profits, and real balance sheets. The comparison is not one-to-one.

But the structure rhymes. A supplier invests in its customers. The customers commit to buying the supplier's products. Revenue looks strong. Growth looks inevitable. The commitments pile up off-balance-sheet. And nobody asks what happens if the customers cannot keep up.

Lucent taught us that circular revenue is not the same as organic demand. We will see if we have to learn that lesson again.