AI research wiki
hardware · 2026-11-04 · Tier 1

SpaceX as the biggest AI cloud, and the compute-supply squeeze around it

SpaceX as the biggest AI cloud, and the compute-supply squeeze around it

Sources: SemiAnalysis: "Elon Is Doing It Again" · The Information: memory smuggling · The Decoder: Huawei · The Information: Nvidia and China cloud · SambaNova SN50 + Supermicro · raw: raw/rss/2026-11-03-semianalysis-elon-is-doing-it-again-spacex-will-be-the-world-s-bigge.md and others

TL;DR. SemiAnalysis argues that SpaceX (which now runs the xAI Colossus sites) is on track to be the largest AI cloud in the world. SpaceX's 2027 target went from 10GW to about 13GW. SemiAnalysis already sees 3.6GW under construction and has identified more than 9GW of candidate capacity. That is roughly all of Google's datacenter capacity built from 1999 to 2026, in about 1.5 years. The edge is build speed. Retrofits went from 122 days (150MW, Colossus 1) to 91 days (210MW, Colossus 2). A new greenfield site ("Gun Range", Mississippi) is targeting about 4 months. Starbase rocket campuses are being converted into datacenters. The TeraFab chip-fab site likely includes a datacenter for third-party renting. Power is decided on the fly: more than 8GW of behind-the-meter power is tracked, with a 5.9GWh battery (BESS) build at Colossus 2. Labour runs at about 3k workers per GW, and Tesla likely makes modular MEP (mechanical, electrical and plumbing) gear.

The supply side around it (same day)

  • HBM leaks to China through a packaging loophole. US rules rate a package by its processor, not its memory. So HBM ships legally when it is mounted on plates with sub-threshold server CPUs, then gets separated on site. Huawei now tells customers to source HBM for its AI chips themselves, which in practice pushes them to this grey market.
  • Nvidia tightens the rental route. Nvidia reportedly told cloud customers to cut exposure to Chinese companies. One report puts the rule at 51% or more of capacity sold to non-Chinese buyers. That closes the overseas GPU-rental route Chinese labs used around export controls.
  • Disaggregated inference ships on new silicon. Supermicro will build racks that pair NVIDIA GPUs (prefill: "reading the prompt") with SambaNova SN50 RDUs (decode: "writing the answer"), starting with 10,000 chips, on standard power and cooling. SambaNova also closed the first $1B of a financing round at an $11B valuation.
  • SK hynix buys into KV-cache offload. SK hynix took a 9% stake in Penguin Solutions by buying SK Telecom's $377M preferred position. Penguin works on CXL and SSD-based KV cache.

Relation to prior wiki

  • Extends compute economics: the 11-03 memory-squeeze page (SK hynix 80% operating margin, GPU lead times of 45 weeks) explained why capacity is scarce. SpaceX's answer is vertical integration of power, construction and eventually fabs. The 10-27 HBM4 note (8-high stacks for Rubin Ultra because 12-high yields are poor) is the component ceiling this build-out will hit.
  • Confirms the behind-the-meter thesis SemiAnalysis first made for Colossus 2 in 2025 (truck-mounted turbines that bypass the grid).
  • Disaggregation is now a product SKU. Prefill/decode split has been a serving-paper idea in the wiki since mid-2026 (see memory hierarchy). SN50 + GPU racks make it a hardware purchase decision.