AI research wiki
hardware · 2026-09-30 · Tier 1

General Compute Splits Inference Across Chips, HBM5 Qualifies Early, and GPUs Get Insured

General Compute Splits Inference Across Chips, HBM5 Qualifies Early, and GPUs Get Insured

Source: X home feed, 2026-09-29 (@rohanpaul_ai, @StockSavvyShay on General Compute, on SK hynix HBM5, on Nvidia GPU insurance, on Micron's 14-week quarter); The Information on on-site power financing, Samsung and Helix and Nscale Raw: X feed captures (private) · raw/rss/2026-09-29-the-information-how-investors-are-financing-on-site-power-to-break-ai-b.md

TL;DR

General Compute is building a neocloud for non-Nvidia chips (Cerebras, SambaNova, Etched). Its first move is a large Cerebras purchase funded by $400M of debt. Cerebras systems sit next to Nvidia GPUs and each request is split: GPUs do prefill (reading the prompt, compute-bound), Cerebras does decode (writing the answer, memory-bandwidth-bound), fast because the weights live in on-chip SRAM. This is prefill/decode disaggregation done across vendors, not just across pools of the same GPU.

Same day, from the same feed: SK hynix has validated HBM5 with TSMC CoWoS packaging while HBM4 is only now entering Vera Rubin systems. Nvidia is reportedly working with insurers to protect lenders if a small cloud defaults on GPU-backed loans, which would put GPU residual value into underwriting. The Information described investors financing on-site power (Blackstone-led $5.3B for 49% of five Williams gas projects; Brookfield's Bloom Energy fuel-cell facility up to $25B) and a timing mismatch: power assets last decades, GPUs age in a few years, so investors now want 10- to 15-year commitments.

Each phase goes to the chip that suits it
Prefill is compute-bound. Decode is bandwidth-bound. SRAM chips skip the HBM wall.
flowchart LR
  R["Request<br/><small>prompt arrives</small>"] --> G["Nvidia GPU<br/><small>prefill, compute</small>"]
  G -->|KV handoff| C["Cerebras<br/><small>decode, SRAM weights</small>"]
  C --> O["Tokens out<br/><small>fast decode</small>"]
  D["$400M debt<br/><small>funds the fleet</small>"] -.-> C
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class R input
  class G,C core
  class O exit
  class D err
Blue is the request, purple is the two chip types, green is output, red is the financing cost.

How this relates to prior wiki pages

  • Disaggregation keeps moving down the stack. Disaggregated Quantization (09-24) gave prefill and decode separate weight precisions; Mooncake/DistServe (09-15) separated them into GPU pools. General Compute separates them by silicon vendor. The open question is KV handoff cost between two chip families, which nobody has published.
  • Memory stays the priced resource. Yesterday's SemiAnalysis chart had Rubin Ultra at 8-high HBM4 for more bandwidth per GB. HBM5 qualification this early says suppliers are being designed in two generations ahead.
  • Financing becomes an engineering variable. The compute-economics page (09-28) logged GPU loans pricing at BBB spreads. Insurance on GPU residual value is the next instrument, and the Nscale critique (nearly all of its $103B backlog relies on unbuilt sites) is why lenders want it.
  • Not verified here: the General Compute and Nvidia-insurance claims come from X posts summarizing reports; treat numbers as reported.

Related