Anthropic could spend $750 billion on compute: the IPO reveals the real cost of leading in AI

By Bloomberg · October 20, 2026.

Anthropic could spend $750 billion on additional computing power over the coming years, according to the Wall Street Journal, as reported by Bloomberg. The company is preparing for an initial public offering (IPO).

The same outlet reports that Anthropic posted a net loss of $80.7 billion in the first half of 2026. According to the WSJ, the figure comes from a draft of the IPO prospectus. Neither the planned spending nor the loss has been confirmed by the company in the material we have, so they should be treated as leaked, unaudited figures.

Bloomberg's text is a newsletter summary and does not detail the timing of that spending, the suppliers or how the loss breaks down. We don't know, for example, how much is model training costs, how much is running services for customers, or whether there are non-cash accounting items. Without that breakdown, any detailed conclusion would be speculation.

Our take: what matters is not the headline figures, but what they reveal about the business model. Leading the AI frontier today requires committing capital on a scale befitting national infrastructure: energy, data centers and chips. A loss of that magnitude in six months indicates that revenue growth, however strong, still does not cover the cost of compute.

Broadly, this fits a pattern in the industry we have already pointed out: the bottleneck is shifting from training to the continuous running of agents, and with it to available energy and per-task efficiency. A compute commitment of this size is, in practice, a bet that demand for agents will grow enough to absorb it.

There is also a methodological caveat. Revenue at the various labs is not always accounted for using the same criteria, so it is not directly comparable. An IPO prospectus should be the first chance to see more consistent figures. Until then, comparing losses or valuations across labs without that adjustment is misleading.

Winners and losers: compute and energy providers win, receiving highly visible long-term demand. The risk falls on investors who buy into the IPO, exposed to spending running ahead of revenue, and on smaller labs, which cannot finance a race like this.

In the short term, this means financial strain and an even greater concentration of power in a few companies able to pay for the frontier. In the long term, the same investment is what, if it works, makes intelligence cheaper and brings closer advances in medicine and abundance. The question the final prospectus will answer is not whether a lot is being spent, but whether each dollar of compute turns into recurring revenue and useful capability.