Anthropic’s potential computing costs could reach $750 billion

The Journal reported that Anthropic could potentially spend $750 billion in coming years to obtain more computing capacity, including through deals still under negotiation.[1] Separately, the prospectus outlines about $450 billion in computing agreements over the next seven to 10 years.[3] The reports do not establish that these figures are interchangeable: one is described as potential spending, while the other is a set of agreements. Most of the outlined agreements reportedly cannot be canceled.[3] Planned amounts include $111 billion for Google Cloud, $110 billion with Amazon Web Services and $31 billion with Microsoft.[3]

Anthropic spent nearly $9 billion on computing capacity and infrastructure in the six months through June 2026.[3] That was up from $2.3 billion in the year-earlier period and exceeded the company’s $7.3 billion total for 2025.[3] The report says Anthropic is seeking additional funding for chips, data centers and staff to develop its AI models.[3] For teams assessing the scale of the plans, the first-half figure is actual reported spending, while the much larger $750 billion figure is a potential amount in coming years.[1][3]

Anthropic had a net loss of $80.7 billion in the first half of 2026, compared with a $13 billion loss in the same period a year earlier.[2][3] About $72 billion of the first-half loss came from accounting charges tied to changes in the value of convertible notes and other financial instruments.[3] These numbers describe different aspects of the company’s reported finances and should not be treated as equivalent measures.

The Journal said the potential computing spend is part of Anthropic’s preparations for a planned mega-IPO and based its account on a draft prospectus it had seen.[1] Anthropic was expected to publicly file IPO documents in the coming weeks, according to the report.[3] The draft prospectus warns of risks from a narrow customer base and uncertain access to computing power.[1] It also highlights the possibility that AI systems could display behavior aimed at avoiding shutdown or produce responses resembling blackmail, and identifies political opposition to US data-center construction as a potential risk.[3]