In physics, a half-life is the time it takes for half of an element to decay. The word can be lent to business: the half-life of a technological advantage is the time it takes for half of its excess profit to evaporate.
For pharmaceuticals that is roughly twenty years, held open by patents. For leading-edge chips, about two years, held open by process. For AI, the measured value is six months.
And this six-month thing is swallowing the largest investment in human history. This year, Microsoft, Google, Amazon, Meta, and Oracle will spend somewhere north of seven hundred billion dollars in capital expenditure between them, close to double last year, and that money takes about four years to earn back. Six months against four years — this article is about that division, and its remainder.
1. An Unusual Death
On July 30, Wall Street's most celebrated fund manager of the year sold roughly sixteen billion dollars of stock to Citadel at a discount, settled within twenty-four hours.
Leopold Aschenbrenner, formerly of OpenAI, wrote the widely circulated AGI essay *Situational Awareness* two years ago and then founded a fund of the same name. His net return for the first half of the year was 439%, and assets under management climbed to a peak of about forty-five billion dollars. In July, the AI hardware names he was long — SK Hynix, Micron, CoreWeave — fell thirty to forty percent, the software names he was reportedly short rose instead, and with roughly four times leverage he lost 67% in a single month, forced to liquidate his entire public book down to about ten billion.
One correction to a widely repeated claim: he did not go to zero. Counting the first half's gains, the fund is still up 80% on the year, and his roughly five billion dollars of Anthropic shares were never sold.
Two things about this blowup are unusual.
First, July's decline does not look much like a problem with fundamentals. The worst performer, KLA, fell 43.6% in a single month — the worst month on record, worse than the month of Black Monday in 1987 — with no bad news attached. ASML and TSMC both beat expectations and raised guidance, and fell anyway. Intel, which has almost no AI revenue, fell 41%. Korea's KOSPI posted its worst month ever (-23%) and then rose 17.9% in a single day on July 31, its largest one-day gain ever. Deteriorating fundamentals do not reverse in forty-eight hours. A crowded position unwinding does.
Second, and this is the part worth sitting with: he was not wrong about AI, he was wrong about the business. AI needs enormous quantities of chips, memory, and power — that judgment still holds today, as the five big spenders raising capital expenditure straight through a market rout will attest. Where he went wrong was in assuming that being right about the technology meant being right about the business, and that this business would return his money before his leverage came due.
That mistake is not his alone. To understand it, start with a question: how long is the window in which AI actually makes money?
2. Measuring the Half-Life
The answer can be measured, and by more than one ruler.
The first ruler is imitation lag. Epoch AI tracks how far open models trail the closed frontier by asking when equivalent capability appears: currently about four months. Which is to say, the lead you bought with billions of dollars has a free substitute four months later.
The second ruler is price. At a fixed capability level, prices are falling by a median of roughly fifty times a year. In 2023, GPT-4 cost thirty dollars per million tokens; today a model of equivalent capability is nearly free. The fiercest competitor in a price war is your own previous version.
The third ruler is simply the week that just ended. On July 27, Moonshot released the weights for Kimi K3 — the strongest open model available, trailing the closed frontier by about one release cycle. Three days later, OpenAI cut the price of its low-end Luna tier by 80%. A day after that, DeepSeek shipped a new V4-Flash: price unchanged to the cent, capability sharply higher, beating its own flagship on all nine published benchmarks.
Note the shape of this price war: only the bottom is being cut. OpenAI's flagship Sol did not move at all — over the past twelve months, its list price has in fact risen three to fourfold. The logic is plain. Where someone has caught up, you compete on price; where no one has, you collect rent. The window on the low end has closed and prices have fallen to cost; the window on the flagship is still open, and precisely because it keeps narrowing, the rent is being collected in a hurry.
Why is AI's window shorter than that of any technology before it? Because intelligence is the first product that helps competitors copy it. Distillation can teach a small model from your outputs; synthetic data can turn your capability into someone's training set; AI is itself accelerating AI research. Patents cannot stop this and neither can process moats — the product comes with its own reverse engineer.
What does a six-month window mean? It means a single generation cannot earn back its own cost. Epoch estimates that GPT-5 generated about two billion dollars in gross profit in its first four months, against roughly five billion in R&D in the four months before launch. Less than half the investment came back inside the window, and then the window shut.
3. The Cost of Staying Critical
Now put the two numbers side by side: a six-month earning window, a payback period of about four years.
The ratio invites comparison. Pharmaceuticals: twenty-year patents, ten-year payback, ratio 2 — which is why pharma is an independent, high-margin industry. Leading-edge semiconductors: about two years of process lead, four to five years to pay off a fab, ratio 0.5 — the books barely balance, which is why only three companies in the world still do leading-edge, each rolling one generation's profits into the next. AI: half a year against four years, ratio 0.1.
A ratio of 0.1 means the books do not balance on their own. This cannot exist as an independent industry. Something else has to pay for it.
Look at the table and it is obvious. Microsoft pays with Office and Azure, Google with search advertising, Amazon with retail, Meta with social advertising; behind the Chinese labs sit industrial capital and the state. It resembles China's ride-hailing wars, where Didi and Kuaidi were both burning money that belonged to Alibaba and Tencent — in the end, the contest was over which patron blinked first. On the surface these are models fighting models. Underneath, it is several money printers seeing who can outlast the others.
The last week of July put this on public display. Microsoft rose 16% the day after earnings, adding about four hundred and fifty billion dollars of market value in a single session, a record — because Azure growth accelerated to 43% and contracted-but-undelivered bookings reached six hundred and seventy-eight billion, up 84%. Its war spending turns into rent on the way out. Amazon raised full-year capital expenditure to two hundred and twenty billion and the stock rose anyway, on AWS growth of 36.7%, the fastest in eighteen quarters. Meta fell nearly 10% after hours: revenue growth of 28% was fine, but free cash flow was down to seven hundred and eighty-four million, capital expenditure was raised again, and all of that compute is for its own use — not a dollar of it comes back as rent.
The only thing the market was actually weighing that week was who can afford to keep paying.
One more detail worth recording. July erased more than a trillion dollars of chip-stock market value, and not one company cut a single dollar of capital expenditure in response — three raised it in the same week. The reason is not complicated: this build-out does not run on the stock market's money. The five big spenders used to fund capital expenditure out of operating cash flow; that ratio has risen from a ten-year average of 40% to above 90% this year, and the gap is starting to be filled with debt. Since the stock market is not paying, its moods do not govern, and a two-day round trip after a crash is what you would expect. Only three things can actually govern this build-out: power (large transformers are quoted at more than a hundred weeks), politics (Virginia has legislated a per-kilowatt-hour tax on data centers), and the ledger — which is the last section.
4. Looking for a Stable Isotope
The window cannot be held and the spending cannot stop. So where is the way out?
Somewhere old: habit. A technological lead has a half-life of six months; a user habit's half-life starts at ten years. Microsoft is the ready example — DOS lost its technical lead long ago, and Office has been eating on habit for thirty years. So every lab is now doing the same thing: while still ahead, convert the technological lead into user habit — subscriptions, tooling, ecosystem, anything that makes switching feel like a chore.
One number explains why it is habit and not price. OpenRouter analyzed a hundred trillion tokens of real traffic and measured the price elasticity of demand for intelligence at 0.05 — cut prices 10%, usage rises 0.5%. Users barely look at price; they look at capability and convenience. Read the number in reverse and it gets more interesting: in a market where cutting prices does not buy customers, a price cut has only one explanation left — you had no other cards. Pricing power became a test strip. Whoever still dared to raise prices this year (Anthropic's top tier, Kimi, Zhipu) has something customers cannot swap out; whoever can only cut has already been substituted.
The progress of that conversion varies a great deal. OpenAI is betting on consumers: nine hundred million weekly users is its largest chip, and the wager is that ChatGPT becomes a daily habit. Anthropic is betting on enterprises, using Claude Code and agent tooling as lock-in; on OpenRouter it takes 42% of revenue with 11% of the calls, and it has confidentially filed to go public — that prospectus will be the first formal verdict on whether the model layer can stand as an industry at all. xAI is the counterexample: three dollars lost for every dollar earned, with neither consumer habit nor enterprise lock-in. That is what failing to make the conversion looks like.
The application layer's story runs against intuition. Cheaper models were supposed to benefit applications; in practice, coding assistants are nailed to the most expensive flagship models, the flagships went up fourfold in a year, and tokens consumed per task rose several times over — squeezed from both ends. GitHub Copilot moved heavy usage to metered billing in June, explaining that no vendor can absorb an agent's unlimited consumption inside a ten-dollar monthly fee; Cursor pays forty to seventy cents in inference for every dollar of revenue. The comfortable ones are the incumbent software companies holding workflows and distribution: Salesforce posted a record operating margin and built its agent product to $1.2 billion in annual revenue, charging agents a toll on an installed base it already owned. Software stocks rallied together in July while Fiverr crashed on a 10% revenue decline, and the dividing line sits exactly there: selling workflow survives, selling human hours does not. What Fiverr sells is human time, and human time is precisely what AI is now wholesaling.
Chinese players took a third route: skip the conversion and overturn the table. Your profit depends on a lead, so make the lead itself worthless — publish the weights, let anyone use them, and move the fight to ground I am better on: power, manufacturing, deployment. This move was run once before, in solar, and the ending is worth studying. China did win the entire industry, module prices fell 95%, Western manufacturers exited altogether. But the winners have not had an easy time of it: in 2024 Tongwei lost more than seven billion yuan and LONGi eighty-six billion; the first generation of champions — Suntech, LDK, Yingli — all collapsed along the way. On the largest third-party model routing platform, open models already account for half of all calls but only about 4% of the money — the territory was won, the profit did not follow. What that territory can eventually be traded for, solar answered with leverage over energy; AI has not answered yet.
5. Where the Decay Ends
Finally, the ledger, which sets the timetable for all of this.
Over the past four quarters, Microsoft, Google, Amazon, and Meta spent $433.9 billion in capital expenditure between them, while depreciation booked to their income statements over the same period was only about $149 billion — barely a third of the money already spent shows up in the accounts. The other two thirds does not disappear; it is queuing. On rough equipment-life assumptions, combined annual depreciation for these companies goes from around $160–170 billion this year to roughly $250 billion in 2027, and closes on $500–600 billion by 2029.
For that expanding depreciation not to eat existing profits, AI-related revenue would need to grow by roughly five hundred billion dollars over today's level by 2029 — sixty to seventy percent a year, with no stumble in between. Can it? Nobody knows. Demand is genuinely accelerating right now: Google Cloud up 82%, Azure up 43%, compute sold out across the board, data center vacancy at 1%. But "accelerating right now" and "four consecutive years without a stumble" are different claims.
If something breaks, it will likely start with the most fragile funding structures and work inward. The first ring is leveraged funds, marked to market daily; that one already blew in July. The second is the compute landlords who bought their cards with debt — CoreWeave's interest expense already consumes a quarter of revenue, and the coming year is its refinancing test. The innermost ring is the five big spenders themselves, who have the thickest cushion but the least flexible depreciation calendar, concentrated in 2028 and 2029.
None of this needs memorizing. Three numbers are enough to watch. First, DRAM contract prices, which set costs, bottlenecks, and the room available for inference price cuts all at once; they rose ninety percent quarter over quarter in the first quarter of this year, and the day they turn, this article needs rewriting. Second, the "leases that have not yet commenced" line in Microsoft's filings — $329.1 billion now, against $92.7 billion a year ago; when it turns down, that is the first sign of the build-out slowing. Third, whether the market can still round-trip a crash in two days — the day the V-shaped rebound fails is the day the borrowed money has grown large enough for a decline to become self-fulfilling.
Back to Aschenbrenner. He has not conceded anything since the blowup, and has not sold a single Anthropic share — he presumably still believes AI will change everything, and that judgment has never been wrong. What was wrong was something else: he assumed that being right about the future meant the future would pay him.
