Anthropic's $895B compute bill is a bet that AI demand keeps growing, and its IPO filing shows how little room it has

Anthropic has committed to roughly $895 billion in multiyear computing contracts. According to its S-1/A, the amended filing for its stock-market debut reported by Seeking Alpha, that is about seven times its current cash. Some $70 billion of the total falls due within the next twelve months. The filing also shows that the company's first operating profit lasted only one quarter before short-term compute costs pushed it back into losses.

Anthropic is not short of money. CNBC reported last week that it holds $131.7 billion in cash and short-term investments. Among non-banking U.S. companies, only Berkshire Hathaway and Alphabet hold more. Even so, the next year's compute bill alone equals more than half of that reserve. The full commitment is close to seven times larger. The cash doesn't cover the contracts. Future revenue has to.

The list of suppliers is a map of the AI industry. Broadcom leads with $161.2 billion in equipment leases. Google follows with $121.3 billion, plus another $125 billion to $150 billion in planned compute on TPUs, Google's in-house AI chips. The filing also lists up to $120.5 billion with SpaceX's AI unit, $110.0 billion with Amazon, $44.9 billion with Microsoft, up to $44.6 billion with Nscale and $41.3 billion with Lambda. Anthropic has spread its bets across nearly every major supplier. It has also tied its fate to all of them.

The filing contains a second, smaller warning that matters. Anthropic says up to 7% of its revenue came from users who violate its platform rules. It also cautions that an ongoing enforcement crackdown could slow future revenue growth. This is unusually candid. Few companies admit that part of their income depends on customers they intend to remove. It shows that safety enforcement has a direct cost on the income statement. Anthropic is choosing to pay it just as it asks public investors for their trust.

Our reading: this is not the balance sheet of a company in trouble. It is the balance sheet of a company betting that compute will be the scarcest resource of the decade. On our own quality index, Anthropic leads the frontier, with Fable 5 at 90 against GPT-5.6 at 79. That lead depends on training and serving capacity that has to be booked years ahead. The short-term risk is real. One profitable quarter is not a business model, and contracts this large leave little room if demand slows or prices per token fall faster than expected. The concentration risk runs the other way too. Broadcom, Google, Amazon and Microsoft now depend partly on one customer's growth. That links the fortunes of the whole AI supply chain, so a downturn would hit all of them.

The long-term picture is more encouraging than the headline number suggests. Commitments on this scale are how infrastructure gets built. Like the railways and fiber networks before it, much of that capacity will outlast whoever paid for it. It will make AI cheaper and more widely available, which is the abundance we expect. Investors need to look past the $895 billion figure and track three things. Can revenue grow without the rule-breaking 7%? Does the next profitable quarter last? Can Anthropic stay the most capable lab without becoming the most indebted one?