Jensen Confirms the Inference Inflection

NVIDIA invested $2 billion in Marvell today, built around NVLink Fusion — a rack-scale platform enabling semi-custom AI inference infrastructure. Marvell supplies custom XPUs and scale-up networking; NVIDIA contributes Vera CPU, ConnectX, BlueField, NVLink interconnect, Spectrum-X switches.

Jensen's framing is precise: 'The inference inflection has arrived. Token generation demand is surging, and the world is racing to build AI factories.' Not training. Inference. Token generation. That's agents consuming compute in production.

This confirms the structural thesis behind the Closelook Lab tools we launched today:

1. The Agentic Winners 25 tracks the demand side — the 25 software companies that agents cannot operate without. If inference demand is surging, every AW25 sector benefits: Toll Booth (NET, DDOG, CRWD) collects rent per API call. Platform (NOW, CRM, SAP) becomes the operating layer. Commerce + Orchestration (PLTR, SHOP) coordinates the workflows.

2. The Money Temperature dashboard reads 38 — cool regime. SPY and QQQ both at 36. But GLD at 62 and UUP at 58. Gold hot, dollar strong, equities cold. The market hasn't priced the inference inflection yet. When it does, the rotation signal will show in AW25 sector temperatures before it shows in headlines.

3. All 6 cointegration pairs are BREAKING — cascade status CRITICAL. The market is repricing structural relationships. SPY/GLD (p=0.99), TLT/SPY, VEU/SPY — the assumptions behind 60/40, equity-bond correlation, US vs. non-US equilibrium are all dissolved in the current 60-day window.

The Marvell deal adds a hardware lock-in dimension. Custom XPUs that only work within the NVIDIA NVLink ecosystem create switching costs at the silicon level. This is the NVIDIA Revenue Multiplier thesis extending from training into inference.

MRVL +11% on the news. The stock is a Rubin Build-Out 100 constituent (Layer 3 — Networking). Silicon photonics collaboration addresses the interconnect bottleneck that Rubin identifies as the next constraint after packaging and memory.

What we're watching: does today's AW25 basket temperature rise from 38 on the next update? If inference inflection is real and being priced, S3 Toll Booth and S7 Commerce should heat up first.