Training AI models and serving inference on them carries a very large upfront cost.

True if industry and financial evidence as of 2026-06-04 (company filings, credible reporting, analyst estimates) shows that training frontier AI models and serving inference on them requires very large capital or compute spending, incurred ahead of revenue. Mostly true if training costs are clearly very large and incurred up front but inference costs are mostly ongoing rather than up front. Mixed if only one of the two is shown to carry a large upfront cost. Mostly false if the costs are large but not substantially incurred in advance. False if the evidence shows these costs are small.

Set before the research began.

1. E1 Anthropic's S-1 makes clear it's on the hook for billions in compute. Its customers can walk at any time | Fortune Fortune · 26 October 2026

   “It warned, however, that its infrastructure build-out will increasingly require significant upfront payments before computing capacity comes online, threatening to eat into cash reserves even as financial performance improves.”

2. E2 Anthropic's S-1 makes clear it's on the hook for billions in compute. Its customers can walk at any time | Fortune Fortune · 26 October 2026

   “The company also spent a whopping $9.7 billion on computing and infrastructure costs for research and development of future models.”

3. E3 Anthropic's S-1 makes clear it's on the hook for billions in compute. Its customers can walk at any time | Fortune Fortune · 26 October 2026

   “The company has committed at least $475 billion in future cloud compute purchases with its major partners, including at least $389.9 billion pledged to Amazon, Google parent Alphabet, SpaceXAI, and Microsoft alone, alongside deals with Nscale and Lambda.”

4. E4 Anthropic's IPO filing shows most of its compute goes to R&D, not serving customers Tom's Hardware · 29 October 2026

   “Anthropic appears to do so easily: it spent $2.4 billion serving customers in the first half of the year, versus $16.3 billion in revenue.”

E1 and E2 show large compute spending on developing future models and significant upfront infrastructure payments. E1 does not split those payments between training and inference, and E4 compares serving costs with revenue without saying when those costs are incurred. So the evidence does not show whether inference costs are upfront or ongoing, and the criteria need that to tell true, mostly true and mixed apart.