Anthropic's amended IPO filing paints an uncomfortable picture of the company's operating economics. According to the document, compute (processing power) costs reached 58% of revenue last quarter, up from 48% the quarter before. Over the same period, revenue rose from $11.6 billion to $17 billion. In other words, while AI revenue is growing fast, the cost of producing it is growing even faster.
What happened?
A chart in Anthropic's filing shows the revenue curve and the compute-cost curve rising together. Normally, at a software company, you would expect infrastructure costs to shrink as a share of revenue as revenue grows — because fixed investments are spread across a wider revenue base. Here the opposite happened: the share climbed from 48% to 58%.
Why does this matter so much?
The classic appeal of the software business model lies in economies of scale. You write the code once and deliver it to millions of users at near-zero marginal cost; as you grow, margins widen. AI breaks this logic. Every new query, every new user consumes real GPU time, electricity and cloud capacity. The marginal cost doesn't fall to zero; on the contrary, it can rise as the model grows larger and usage intensifies.
This weakens the narrative AI companies have repeated for years: "We're losing money for now, but once we reach scale we'll be profitable." Anthropic's numbers say the opposite, at least for this quarter. Scale grew, but margins didn't improve — they got worse. The problem isn't just the pace of growth; it's that growth itself is expensive.
Behind this lies a question of position in the value chain. Anthropic and model providers like it depend on Nvidia's chips on one side and on the capacity of cloud providers such as Amazon, Google and Microsoft on the other. A significant portion of the value added stays with these suppliers. The model company is starting to look like a relatively low-margin intermediate layer, squeezed between expensive hardware and the end user. It is no coincidence that Nvidia posts record profits while model companies report losses.
Who does this affect, and how?
The picture carries different meanings for different players:
| Party | Impact |
|---|---|
| Model companies | Pressure to raise prices or cap usage; a constant need for capital |
| Chip/cloud suppliers | Continue to hold on to the bulk of the margin |
| Enterprise customers | Potential price increases; risk of dependence on a single provider |
| Investors | A fresh test of the "profitability through scale" thesis |
On the user side, this pressure has two classic release valves: raising prices or curbing usage on free and cheap tiers. Both paths help explain the usage limits and tiered-pricing moves seen across the industry lately.
Parallels with the dot-com boom
This dynamic isn't new. In the late 1990s, the enormous capital that poured into internet infrastructure often enriched not those who built fiber and data centers, but those who sold the equipment. Today, Nvidia plays the role of "selling the shovels." The difference: back then, once infrastructure was laid it could be used for many years; GPUs, by contrast, both age quickly and keep incurring electricity and cooling costs as long as they run. In other words, the cost isn't one-off — it's a permanent, recurring burden.
The entry of players like Broadcom is a sign that model companies are seeking custom chips to reduce their dependence on Nvidia. This could ease the cost side over the medium term; but in the short term, the transition itself also requires serious investment.
What's next?
There are a few things to watch. First, the direction of that 58% share in the coming quarters: will it stabilize, or keep climbing? Second, whether custom chips and more efficient model architectures actually bring costs down. Third, pricing: how companies will adjust their enterprise API prices or consumer subscription fees.
Ultimately, these numbers aren't a picture of disaster; they could just as easily be an expected cost curve for a company in a high-growth phase. But the assumption that "we'll automatically become profitable as we grow" now has to wait for evidence. AI's business model looks far more like a capital-intensive industry than like software; and any valuation that ignores this distinction carries the risk of being wrong.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.