The most important number in tech this week is not a model benchmark. It is $900.9 billion, and it sits in a regulatory filing.
What the filing shows
As Forbes laid out in detail, Anthropic is preparing to price its IPO on November 18, with shares expected to trade on Nasdaq under the ticker ANT. An amendment to its registration statement discloses $900.9 billion in aggregate purchase commitments and lease agreements, substantially all for compute and infrastructure, running through 2045.
The schedule matters more than the headline. About $70.7 billion is due in the twelve months after September 30, 2026. Cumulative payments reach $451.6 billion by the end of 2030.
The total is not one unconditional obligation. Roughly $186.2 billion is tied to contracts with termination rights or conditions, and another $150 to $175 billion in expected payments sits outside the total because it has no fixed dollar commitment.
The quarter that explains the economics
Anthropic reported about $17 billion in third-quarter revenue and still posted an operating loss of $567 million, after a $974 million operating profit the quarter before. The filing names two causes: about $5 billion of compute bought on short-term commitments, which was significantly more expensive than long-term capacity, and about $2 billion in non-cash charges tied to a donation-matching program.
That first cause is the lesson. Anthropic typically buys compute 12 to 18 months ahead. When demand outruns the plan, the gap gets filled at spot prices, and a profitable quarter turns into a loss.
Every operator recognises this pattern. It is the same as running out of inventory and buying from a competitor's distributor at retail price. The business is healthy, and the margin still disappears.
It also explains the commitment schedule. A company that loses money every time demand surprises it will sign longer and bigger contracts to avoid the next surprise. That is rational for one quarter. Over twenty years, it turns a variable cost into a fixed one, and fixed costs only work when utilisation stays high.
Suppliers, partners and rivals at once
The filing lists minimum commitments of $121.3 billion with Google, $110 billion with Amazon and $47.2 billion with Microsoft, plus large deals with SpaceXAI, Nscale and Lambda. Anthropic runs on four accelerator families, AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs, treated as one pool.
The three hyperscalers supply the compute, distribute Claude through their clouds and compete with their own models. That is an unusual position for any supplier to be in. It means the price of a Claude seat depends partly on negotiations between companies that are also rivals.
I covered the lock-in side of this in Anthropic's filing reads like a lock-in manual. The new amendment adds the cost side. The supplier is locked in too, to a capacity bet that stretches decades beyond the life of any chip.
What this means if you build on Claude
The filing makes AI pricing less mysterious. A frontier lab has to fill enormous committed capacity. When utilisation is high and long-term contracts cover it, prices can stay stable or fall. When demand spikes faster than contracted capacity, someone pays the spot premium.
Historically, the customer pays that premium indirectly. Not through a price increase announcement, but through rate limits, tier changes, priority access fees and quiet routing to cheaper models. I described this pattern in read the prospectus before you renew, and quarterly public reporting will make the pressure more visible.
So treat your AI provider like a critical supplier, not like a software subscription. Know which workflows depend on one model. Know what happens if the rate limit halves or the price per million tokens doubles. Keep one workflow tested on an alternative model, even if you never switch.
Also read the usage reports you already get. Most companies cannot tell you which team spends the most tokens, on which model, for what outcome. When the supplier's costs become public every quarter, the buyer who knows its own consumption is the one who negotiates well.
The bet, stated plainly
Strip away the accounting and the bet is simple. Anthropic is wagering that demand for AI will grow fast enough to use capacity it is committing to now, through 2045. The infrastructure lives for 20 years. The accelerators inside it last a fraction of that.
Markets are nervous about that kind of bet right now. Copper fell below $6 a pound this week for the first time in six months, partly on doubts about data center demand. One week of commodity prices proves nothing, but it shows how quickly sentiment around AI infrastructure can turn.
For a CEO or CMO, the takeaway is not about whether to buy the stock. It is about what your operating plan assumes. If your 2027 plan depends on AI getting cheaper every quarter, you are betting on a supplier's utilisation curve.
That might be a good bet. Just make it deliberately, with a fallback, and with eyes open to the fact that AI is no longer priced like software. It is priced like electricity, and someone is always paying for the power plant.