# Sixteen Years of Compute Forecasts Held Up. The Money Kept Breaking: Reading The Infinity Machine

> Will AI capital spending overshoot? After reading Sebastian Mallaby's new book on DeepMind, I swapped the question. Across sixteen years of records, the compute curve tracked the forecast and every crisis was about money. So the question is not total spending, it is whose money can sit unpaid for ten years. Educational reading notes and reflections, not investment advice.

Published: 2026-09-06
Locale: en
Tags: book notes, DeepMind, AI capex, Google, supply chain, education

TL;DR: DeepMind's 2010 business plan drew a compute curve for 2025. Sixteen years on, it landed close. The money in the same plan nearly broke at every stage: first round, third round, the sale to Google, a ten-year fifteen-billion-dollar pledge that evaporated. Compute was never the bottleneck. Who can pay was. Today that reads as: check whose pocket your AI supply-chain stock's orders come from, earned cash or borrowed.


![Fauvist oil painting cover: a colossal machine fills the upper canvas, and below it a small figure turns both pockets inside out; flat slabs of vermilion and cobalt collide](/covers/infinity-machine-who-pays-for-compute.png)

> *All the world hustles for profit; all the world jostles for gain.*<br/>
> —— Sima Qian, *Records of the Grand Historian*, "Biographies of Merchants" (Western Han); translation mine

## The question I brought to the book

The question friends have asked me most this year: will AI capital spending overshoot? A handful of companies are pouring hundreds of billions of dollars a year into data centres. One day it is all built, nobody needs it, and then what?

This weekend I finished Sebastian Mallaby's new book on Demis Hassabis and DeepMind, *The Infinity Machine* (2026, [books.com.tw](https://www.books.com.tw/products/F01b490580)). I did not come out with an answer. I came out with a different question.

A word on the book first. Mallaby is a former Financial Times contributing editor, twice a Pulitzer finalist, now a senior fellow at the Council on Foreign Relations, and the author of *More Money Than God* on hedge funds and *The Power Law* on venture capital. Penguin Press published this one on 31 March 2026, four hundred-odd pages covering DeepMind's sixteen years from its founding in 2010 to late 2025: how it raised money, how Google bought it, how OpenAI overtook it and how it clawed back. Mallaby had more than thirty hours with Hassabis and interviewed over a hundred people around him, including the co-founder he fell out with. A Traditional Chinese edition came out from CommonWealth Publishing on 27 April 2026.

## The compute line held for sixteen years

The page that stopped me is in chapter five. In 2010, months after DeepMind was founded, Hassabis wrote a thirty-page business plan. One chart projected that by 2025 a supercomputer would reach ten to the twentieth operations per second, a hundred thousand times what existed then. The same chart put human-level general intelligence around 2030.

Mallaby, writing in 2026, says the compute line came out "almost exactly in line" and the intelligence line looks "just slightly conservative."

A thirty-four-year-old with no product, in 2010, drew a compute curve that came true fifteen years later. I copied that page out before reading on. Every few months someone announces that compute is peaking or scaling is hitting a wall. The book's record: in sixteen years, the wall was never compute.

## Every crisis was money

The money line in the same plan was nothing like as smooth.

First round: Founders Fund put in 2.3 million dollars for a little under half the company. Peter Thiel's verdict was "A-plus on the science, F on the business model." The third round targeted 65 million and raised about 25. The sale to Google in 2014 was 650 million dollars, and the first year of payroll afterwards was 260 million, six times the company's entire spending over the previous three years.

The sale did not end it. In early 2018 Google drafted a term sheet promising a spun-out DeepMind fifteen billion dollars of research funding over ten years. Then Apple poached one senior executive, seats got reshuffled, and the whole scheme vanished along with the fifteen billion. In 2025 Google paid 2.7 billion dollars to bring Noam Shazeer back, for one man and his team.

Sixteen years, five money crossroads. Not one of them was about computers being too slow. Every one was about who would pay, and for how long.

![Two stacked timelines: the upper compute line rises smoothly from 2010 to 2026 with forecast and actual almost overlapping, while the lower money line over the same period is broken by five gaps, each marked with a funding round or acquisition.](/figures/compute-line-versus-money-line-en.svg)

Google's then chief financial officer, Patrick Pichette, has a line in the book that I read twice. Public companies look ninety days ahead, he says, but some infrastructure takes ten years, and Google has the balance sheet for that. Then: "You can bury anything in that sort of balance sheet. You can bury Wisconsin, and nobody would know about it."

That is the most useful sentence in the book for me. It flips the overcapacity question. Whether the compute gets built depends on whether there is a balance sheet big enough to bury the money in, for ten years.

## Where this belief comes from

"Money goes out first, technology grows later" is not new. A century ago Schumpeter put banks at the centre of innovation: somebody has to lend to things that do not earn yet. He was right at the time, and it explains how railways and electricity got built.

The difference today is who is lending. In the railway age it was banks and bond markets, and when the money stopped, a whole cohort failed together. The AI age splits in two. One group, Google and Meta, funds this out of operating cash flow and buries it in its own balance sheet. The other, OpenAI and its kind, raises round after round. Chapter twenty describes OpenAI in 2025 rolling out "a new, desperate fundraising gimmick every few weeks," becoming exhibit A for anyone predicting an AI valuation crash. Mallaby does not call it a bubble. He calls it an exhibit.

The same chapter has one more line: American AI builders have promised trillions of dollars of capital spending without a clear account of where the electricity will come from.

## Running the numbers on my own data

I went back to our own Alphabet financial summary from late June (data pulled 27 June 2026, trailing four quarters). Operating cash flow 174.4 billion dollars. Free cash flow 27.9 billion. The 146 billion in between is the money Pichette was talking about. Eighty-four percent of operating cash flow went into building.

![A single horizontal bar represents operating cash flow, with a long left segment showing the money spent building data centers and only a short right segment left as free cash flow.](/figures/alphabet-cashflow-buried-en.svg)

Two things in that number. First, Google can afford it without asking anyone. Second, eighty-four percent cannot last many years unless cloud and advertising revenue grow with it. So for anyone watching Google, the thing to track is not how much it builds, it is whether operating cash flow keeps pace with capital spending. In the book, Google falls behind three times and catches up three times on money and its own chips, and that cash flow is what pays for it.

Now Taiwan. The island's AI supply chain sits downstream of all this spending. Two companies can win the same size of order, but if one customer is Google and the other lives on fundraising, order visibility differs by an order of magnitude. The first order hangs on operating cash flow. The second hangs on the next funding round. Funding news moves faster than earnings, and it shows up more often.

![Two panels side by side: on the left, an order and operating cash flow form a closed loop back to itself; on the right, an order leads to a next funding round and then a dashed line opens outward, ending in a question mark.](/figures/order-follow-the-money-en.svg)

## Two orders of bubble

Yesterday I listened to Yu Ting-hao's quarterly session for subscribers. His view: an AI bubble bursts under one condition only, the data centres get finished and the compute turns out to be surplus. That is one ordering. Compute overshoots first, money breaks after.

This book offers another. In DeepMind's history compute was never in surplus. The money broke several times first, and each time somebody caught it. Whether somebody catches it next time, nobody knows.

Both orderings can happen. The difference is whether you can see it coming. Compute surplus only shows once everything is built. Money breaking is visible: fundraising headlines, capital spending as a share of operating cash flow, whether the power contracts got signed. That second set of indicators updates every quarter.

![Two stacked timelines: the upper one has a single observation point only at its far end, while the lower one is dotted with several observation points along the way.](/figures/two-orders-of-the-bubble-en.svg)

I no longer ask "will it overshoot." I ask: for the companies I hold, whose money are the customers spending, earned or borrowed? And if borrowed, borrowed until when?

## One thing to take with you

Whether AI capital spending overshoots, you cannot see. Whose money breaks first, you can.

Something I have tried: the next time an AI name hits limit-up, close the chart and open its latest annual report to find the largest customer. Copy that customer's last four quarters of operating cash flow and capital spending onto a piece of paper, work out the ratio, and write today's date on it. Three months later, take the same sheet and run it again, and see which way the ratio moved.