5 min read · 23 sources

The scarcest compute of 2026 turns out to be the boring kind - and whoever already owns it wins without buying a single new GPU.

The scarce resource today is a CPU

The headline-grabbing items - an orbital data center test satellite, a $500 ChatGPT tier - matter less than a quieter one: CPU shortages across all three hyperscalers, with spot pricing gone and reservations requiring months of lead time, driven by AI workloads like reinforcement learning. This is a second-order story worth sitting with. Everyone is watching GPU supply; the GPU is now so dominant in the box that the CPU around it has become the binding constraint. That inverts the usual hierarchy: hyperscalers with large owned fleets and strong scheduling software can route around scarcity, while customers on smaller providers simply wait. It also quietly strengthens the case for optimisation tooling that squeezes more from what you already have.

Google shipped a bundle of exactly those tools: GKE Pod snapshots cut AI inference cold starts by up to 89% (70B models loading in 37 seconds), and a multi-cluster GKE Inference Gateway benchmark across 17,000 nodes in three regions achieved near-linear throughput with under 1% routing overhead. Both are utilisation plays - when capacity is the constraint, the vendor who can pack the fleet wins share without building anything. The orbital TPU test satellite, carrying four Planet Labs-built TPUs launching October 1, is a decade-out experiment rather than a 2026 earnings item, but it signals Google treats compute scarcity as a permanent condition worth escaping the planet for.

Oracle's backlog meets physics

Oracle sent a force majeure notice on its 2.4-gigawatt New Mexico Jupiter project, delayed to 2028, powered by Bloom Energy fuel cells, with $18 billion in stressed debt - and the stock fell 3%. The bull case on Oracle has always been “enormous contracted backlog will convert to revenue.” A force majeure notice on a flagship project is the first crack in the conversion assumption, and it lands just as CPU scarcity makes everyone else’s timelines worse too. This is the kind of event that deserves a re-rating of backlog quality, not just backlog quantity.

The fat middle, confirmed

Anthropic and OpenAI cut prices within 90 minutes of each other, open models now carry a majority of token volume at an 86% discount to closed models, and frontier models fell from 53% to 45% of corporate token consumption. The demand curve for AI is a fat middle, not a pyramid. Two implications: first, pricing power at the frontier is eroding faster than anyone’s model shows; second, the beneficiaries are the tooling and infrastructure layers who monetise volume regardless of which model runs it. Datadog fine-tuned Qwen3.5-9B on GLM-5.3 investigation traces to hit 87% of the bigger model’s recall at $0.003 per investigation versus $0.06 via API - a 20x cost cut running on one A100. That is the fat-middle thesis executed inside a monitoring vendor, and it is a small demonstration of why observability pricing cannot stay where it is. NVIDIA’s NVFP4 quantisation work (1.30x vLLM throughput on Qwen3.6-35B-A3B, 5.9x decode throughput on Nemotron 3 Ultra) is the same squeeze from the hardware side.

The rest of the day

AWS had a busy but mixed day: CloudWatch Omni for AI observability, and - more substantively - first-hyperscaler approval for NATO Restricted classified information across 15 member-nation regions, a genuine moat in sovereign cloud that Microsoft and Google must now match. Cloudflare’s “Agent Development Lifecycle” and Ando’s $20m raise for human-and-agent messaging are the same bet from different altitudes; Databricks buying Row Zero addresses the spreadmart problem that has resisted three generations of BI vendors. The SemiAnalysis ClusterMAX 3.0 ratings - Nebius to Platinum alongside CoreWeave, Azure down to Silver, Crusoe to Bronze - will matter more for neocloud financing terms than for hyperscaler economics.

Friday note; most of this is positioning, not news. The CPU shortage and the Oracle notice are the two items we would actually revisit next week.

Scoreboard

The book

Trades decided 2026-09-25

Filled at the next session’s open, not today’s close.