Neocloud: The Noun, the Workload, and the Balance Sheet

A single-topic deep dive into "neocloud" — the GPU-as-a-Service business. The word fuses three things worth separating: a genuine engineering cost edge, a depreciating asset financed with debt, and a marketing label. We argue the durable move is to ignore the noun, date the workload's half-life, watch the balance sheet, and watch for the next silicon swing — today GPU, tomorrow TPU and inference ASICs.

In this episode

  • The noun vs. the workload. Allbirds becoming "NewBird AI" — a shoe company that six-x'd its stock on a GPU pivot funded by a $50M facility that had delivered just $3.25M — as the comic tell that the label has decoupled from the business.
  • Why the cost edge is real. Bare-metal servers, stripped bills of materials, and 3.2 Tbps InfiniBand fabric explain the 2x–7x price gap over hyperscalers — structural, not magic. Even Microsoft rents from neoclouds.
  • ...and why it compresses. Post-depreciation bare-metal margins of ~14–16%, advertised-not-transacted pricing, the 66% utilization breakeven, and the migration of durable value toward reliability, orchestration, and software.
  • The financing layer. CoreWeave's ~$30B debt and Microsoft concentration, $20B+ in GPU-collateralized lending, the Nvidia–CoreWeave–Nebius circular-financing critique, collapsing H100 resale values, and the Michael Burry depreciation fight — steelmanned on both sides.
  • The power floor. Interconnection queues as the binding constraint (2.3 TW backlog; Texas at 1.8% operational), why a PPA isn't physical delivery, and the Bitcoin-miner pivot (IREN–Microsoft, CoreWeave buying Core Scientific).
  • The silicon swing. Inference overtaking training, TPUs/Trainium/Maia, Broadcom's design concentration, Anthropic's three-platform reality behind the "de-Nvidification" headline, and neocloud as one rung on cloud's 15-year re-bundling ladder.
  • The architect's durable questions. A buyer's checklist for evaluating any "new cloud," including financial due diligence and asking when firm interconnection capacity physically exists.

Sources & References

Primary / originating sources (operator-provided — ground zero)

Research & critique (financing, economics, depreciation)

Industry & implementation (pricing, market structure, deals)

Power & grid (the binding constraint)

Geopolitics & sovereign demand

Note on figures: several pricing and balance-sheet numbers cited above originate in secondary analysis and vendor blogs; advertised GPU rental rates are offered, not transacted, prices; and the "$200B" Anthropic–Google aggregate and the IREN/Core Scientific deal terms trace to secondary reporting pending primary confirmation. We've flagged these in the episode where they carry weight.

Have questions about this episode? Reach out.