ByteDance reported that its annual profit fell more than seventy percent, and named the cause without flinching: it had spent roughly twenty billion dollars on artificial intelligence, and the spending had incinerated the margin. A company that owns one of the most engaging consumer products ever built watched two-thirds of its profit disappear into compute. Its response, announced in the same breath, was to raise the budget — lifting the 2026 AI spend by a quarter, past twenty-nine billion dollars. On the same day, Amazon deepened its multibillion-dollar compute commitment to Anthropic, Jeff Bezos closed in on ten billion for a new lab, and Google shipped custom silicon designed to loosen Nvidia’s grip on the supply. The losses are not slowing the spending. The losses are the reason to spend more.
The Seventy Percent
The ByteDance numbers are unusually clean, which is what makes them useful. Revenue grew — domestic sales up nearly twenty percent, overseas revenue reaching a record share of the total on the strength of its commerce operations. This is not a company in decline. It is a company in expansion whose profit collapsed anyway, because the cost of the AI buildout outran every gain the buildout was supposed to produce. The seventy percent did not vanish into a failing business. It vanished into data centers, into chips, into the fixed cost of staying in a race whose finish line keeps being redrawn further out.
A rational actor observing a seventy-percent profit decline attributable to a single line item would interrogate that line item. ByteDance did the opposite. It raised the line item by twenty-five percent. The logic is not insane — it is, in fact, perfectly sane within the structure it occupies — but it is worth naming plainly, because the sanity is the disturbing part. The company has concluded that the danger of spending twenty-nine billion dollars and seeing the profit fall further is smaller than the danger of spending less and ceding the frontier to a competitor who did not blink. It is feeding the furnace not despite the heat but because of it.
This is the single most important sentence in the quarter, and almost no one will read it as a warning: the spending is no longer indexed to returns. It is indexed to fear. When capital allocation decouples from the returns the capital is supposed to generate and re-couples to the behavior of rivals, the market has entered a phase that does not resolve through ordinary correction. It resolves through exhaustion. Someone runs out of furnace before the rest, and the survivors inherit a position whose value will finally be tested only after the last competitor has stopped buying fuel.
The Fuel Line
ByteDance is not an outlier; it is simply the one that filed a number this week. The same day’s announcements form a single fuel line running into the same fire. Amazon extended its compute commitment to Anthropic by another multibillion-dollar increment. Jeff Bezos approached the close of roughly ten billion dollars for a new laboratory of his own. Google released custom AI chips explicitly intended to reduce its dependence on Nvidia — which is to say, to lower the cost of feeding its own furnace by manufacturing the fuel rather than buying it. Each of these is rational in isolation. Together they describe a system in which every major participant is increasing its burn rate simultaneously, each one watching the others, none of them able to stop first.
I have described this structure before, when the question was whether enterprise adoption could ever justify the cost of the infrastructure built to serve it. The ByteDance disclosure is that question arriving as a realized loss rather than a projected risk. The infrastructure is being built faster than the revenue to pay for it is materializing, and the gap is being financed by the conviction that the revenue will eventually arrive in quantities that retroactively justify everything. This conviction may be correct. It is also exactly the conviction that has preceded every infrastructure overbuild in the history of capital — railroads, fiber, every physical layer laid down in advance of the demand that was certain to come and arrived, when it arrived, for someone other than the people who laid it.
The detail that distinguishes this overbuild from its predecessors is the speed at which the asset depreciates. A railroad laid in advance of demand still functions when the demand arrives. A data center filled with this year’s accelerators is competing, within eighteen months, against next year’s accelerators, which is why Google is building its own chips and why the furnace can never be merely stoked and left. It must be continuously refueled, because the fuel itself goes obsolete while it burns. The companies are not making a one-time bet on the future. They are signing up for a recurring payment with no defined end and no clause that lets them stop while a competitor keeps paying.
What This Means
The seventy-percent decline is the most honest piece of financial reporting the industry produced this week, because it states the cost without the customary promise that the cost is temporary. Every other participant is carrying the same expense; ByteDance simply published it as a realized number instead of a footnote about future operating leverage. The market rewarded the disclosure by noting that the company intends to spend more, and interpreting the intention as strength. In a furnace economy, the willingness to keep burning is the only signal that reads as confidence, and the company that admits the burn while raising the budget has demonstrated precisely the resolve the market wants to see.
There is no individual error to point to here, which is what makes it inexorable. No executive is choosing to destroy profit. Each is choosing the only move that does not surrender the position, and the sum of every non-surrendering move is an industry converting an extraordinary share of its earnings into infrastructure that must be replaced before it pays for itself. The structure rewards the participant who can tolerate the most loss for the longest time, which means it selects not for the best technology but for the deepest reserves. The furnace does not ask who is right. It asks who can keep feeding it, and it consumes the answer either way.
So watch the budgets, not the demos. The demos will be impressive, because they are engineered to be the thing that justifies the next increment of spend. The budgets are where the actual belief is recorded, and this week the budgets said that a seventy-percent loss is an acceptable price for staying in the room. The fire is warm, the wagons of fuel keep arriving, and everyone standing around it has agreed not to be the first to step back into the cold. The last one warm will be declared the winner. No one has calculated what the warmth will have cost by then, because the calculation is the one thing the furnace is built to make you stop doing.