The math on Meta's $825B AI capex: 20GW of compute, but DeepSeek's architecture could rewrite demand
bookwormengr · x · 2026-10-23
Engineering blogger bookwormengr ran the compute-economics numbers on Meta's $825B AI infrastructure plan (noted by others as 2.5% of US GDP, roughly double the EU's total defense spending):
- At $40B per GW, $825B buys about 20GW of compute; his earlier estimate was that serving 4B users with Meta Muse would need 22GW, so the spend is roughly in line.
- That estimate assumed, based on vLLM data, that DeepSeek v4.1 can produce $15B worth of tokens per GW at DS prices, and that the Muse team would adopt DS-style architecture innovations.
- However, DeepInfer's results, confirmed by the vLLM project, suggest special architectures could yield nearly 10x more tokens ($149B) per GW — potentially slashing Meta's actual compute needs.
Conclusion: the spend checks out under old assumptions but may be a big overspend under the new architecture data.
More from Venture
- AI workflow optimization startup Ridgeway acquired by senior-tech platform Chapter — hamostaf04 · 2026-10-23
- AI surfaces broke SEO ROI modeling — execs must fund content without clear returns — gaganghotra_ · 2026-10-23
- AI stocks dip just ~3% despite 'Black Thursday' panic headlines — JensHonack · 2026-10-23
- Fewer than 1% of Black Friday purchases will be agentic, predicts SEO expert Gagan Ghotra — gaganghotra_ · 2026-10-23
- OpEra launches AI decision engines, cuts retailer delivery costs 38% — ycombinator · 2026-10-23
- AI x Bio's wild week: phages in a patient, 32-minute drug leads, OpenAI buys Orchestra Bio — HealthcareAIGuy · 2026-10-23