Google reported second-quarter net income up 298% to $112 billion. Back out the $99 billion write-up on its Anthropic stake and there’s still a $10.8 billion gain, so far, so good.
Then read the cash flow statement. Capital investment nearly doubled to $44.9 billion, money into private companies came to $21 billion, and adjusted free cash flow ran to minus $26 billion. To keep the buildout going, Google issued $30.5 billion of common shares, $19 billion of convertibles and $24.8 billion of straight debt. In one quarter. In the first quarter it raised $32 billion, including a $1.4 billion century bond at 6.125%. A hundred-year bond, from a technology company, in a currency also unlikely to last the distance. Hmm.
That is one of many similar examples in Daniel Oliver’s August letter for Myrmikan, the best piece of work I’ve read on the AI build this year. What follows are the salient points.
You can see the computer age everywhere but in the productivity statistics. Robert Solow
Between 1955 and 1979, total factor productivity grew at 0.8% a year. Through the great computerisation, 1980 to 1997, it grew at 0.45%. Add the internet, 1998 to 2023, and you get 0.62%. Three decades of the most celebrated technology in human history, and productivity growth never got back to where it started.
AI is taking the same path. MIT’s NANDA study found 95% of organisations getting zero return on their generative AI spending. Of 318 corporate executives surveyed, 96% said costs came in above expectations and 71% said they had little or no control over them. Uber burned through its annual AI budget in two months. An Nvidia vice president put it plainly: for his team, the cost of compute runs far beyond the cost of the employees.
The direction of travel makes this worse. The growth is in agents, machines that do things, book the flight rather than price it, and an April academic paper found agentic tasks consume a thousand times more tokens than code chat. Frontier models systematically underestimate their own token consumption. The bill arrives after the work, and it arrives whether the agent succeeded or not.
Meanwhile the new Fed chairman tells us AI will be a significant disinflationary force that raises productivity. Greenspan made precisely that argument in February 2000, one month before the top, and used it to justify printing faster.
Apollo’s chief economist put numbers on the value chain. The silicon suppliers run 41% operating margins, the cloud companies, before capex, manage 11% and the AI labs themselves run at minus 59%.
Every other industry on earth works the other way round: the consumer-facing company earns the fat margin, and it thins as you walk back down the supply chain toward the commodity producers. In this case the money is being made by the people selling shovels to prospectors who haven’t yet found any gold.
Which raises the question: if the customer at the top of the chain can’t make any money, who’s paying for the shovels?
The answer is that they’re paying for each other, in a circle, and each hand that touches the parcel strips off a slice and keeps it.
Nvidia has put $30 billion into OpenAI and $10 billion into Anthropic, its own customers and it holds 9.3% of Nebius and $2 billion of CoreWeave equity, alongside a $6.3 billion agreement to buy back any capacity CoreWeave can’t sell through April 2032. It has been in talks to guarantee $250 billion of financing for an Ohio data centre, most of which returns to Nvidia as revenue when the chips are bought. CoreWeave buys the chips, Nvidia books the sale, Nvidia guarantees the revenue, CoreWeave borrows against the guarantee to pay Nvidia.
Microsoft put $13 billion into OpenAI, reportedly $10 billion of it in credits for Microsoft’s own cloud, then counts the use of those credits as revenue.
Then the Hyperscalers (Amazon, Microsoft, Google, Oracle and Meta) found a way to make the capex disappear altogether. Building the specialised data centres themselves would put an ugly lump on the balance sheet, so they contract with the Neoclouds (CoreWeave, Nebius, Nscale) and pay over years as an operating expense. Goldman counts $1.5 trillion of Hyperscaler lease commitments, a trillion of which never appears on a balance sheet and Morgan Stanley counts another $982 billion of similar purchase commitments. Meta signs $21 billion with CoreWeave; CoreWeave promptly raises $1.75 billion of senior notes and $3.5 billion of convertibles against it; Meta’s promise, which backs that debt, is recorded nowhere.
Bloomberg has AI-tied debt at $1.2 trillion, some 15% of the investment-grade market. The template is Hyperion, the $30 billion Meta-backed data centre on a flood plain: Meta contributes in kind for 20% and takes no exposure to the flood risk, Blue Owl takes 80% with $7 billion of equity and sells $27 billion of A+ rated bonds with 24-year maturities, the largest private credit transaction ever executed. PIMCO led with $18 billion, on behalf of central banks, pension funds, endowments and insurers and the SEC has since exempted a large subset of data-centre securitisations from risk retention, so the issuer no longer has to keep any of the risk.
And the buyer of last resort, the hand the parcel stops in, is the life insurance industry, increasingly owned by the same private equity firms originating the deals. And round we go again...
Private equity ownership of life insurers went from near zero in 2009 to over $700 billion across 134 insurers by 2024; McKinsey puts the controlled figure at $1.5 trillion by the end of 2025. Life insurers now hold $849 billion of the $2 trillion private credit universe and the industry carries $11 trillion of assets against $10.6 trillion of liabilities, a mere 4% cushion. When a life insurer fails, the state guaranty association covers policies to around $300,000, funded after the event by the surviving competitors, who then recoup it through tax credits. The public pays the bill, as usual. How accommodating.
A sobering and realistic analysis of the AI financing bubble. Go back in history and look at all the previous tech bubbles from railroads in the 19th Century (read 1873 by Liaquat Ahamed for some brilliant observations on the first great depression and the making of the modern world) to the present day. What we have learned is that technology can be revolutionary and the investment bubble can still be real. Railways, electricity, automobiles, radio, PCs and the internet all changed civilisation. Investors nevertheless repeatedly overestimated how quickly the economic returns would arrive and which companies would capture them.
Logic, a dangerous attribute when divining market outcomes, suggests the market will crash, the insurers will be bailed out and the Fed will print. The weight of evidence skews heavily that way; the timing is anyone’s guess. A government with 2.4% of GDP riding on this AI boom has every reason to keep the market bubble going. Then there are the midterm elections approaching rapidly. The odds don’t look good for the Republicans, and a collapsing stock market would be the final straw. Bessent is also mindful of the pressure to fund the enormous US deficits, and we hear rumours of a cunning plan involving stablecoins, would you believe? There is of course the small matter of the denouement of the Iran war, which could move the market in either direction with some rapidity.
What’s different this time?
The 2000 bubble was an equity event. The AI bubble bursting will start in credit. So watch the credit data like a hawk.
Equity markets will be the last to know, because the index has 29.7% of the S&P 500 in the Hyperscalers, Nvidia and its three US suppliers, all bought every two weeks through payroll by eighty million Americans who have rarely, if ever, formed a view on any of it. Add $600 billion of margin loans against that inflated collateral, and $152 billion borrowed against life policies, and you have borrowing secured on secured borrowing; pass the parcel please.
When the music stops you don’t want to be the holder. As Dan Oliver summed it all up:
Gold may not be the only safe haven, but it is certainly the clearest and the simplest and the best.
What more can I say? Got gold?
Source: Daniel Oliver, “AI Debt Failure Will Prompt Another Wave of Fed Bailouts,” Myrmikan Research, 14 August 2026. The figures are his; the conclusions are mine. Nothing here is advice.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.