Three Nvidia CPUs, One Licensor — The ARM v9.2 Toll Threading Vera, Losamis, and Rosa

Three CPUs, One Kernel

In the past 72 hours Nvidia disclosed three CPU projects: Vera (AI factory), Losamis (next-gen server), and Rosa (v9.2 with expanded L2). Different names, one microarchitecture — codenamed 'Rigel' — and one ISA: ARM v9.2.

The market read the reveal as Nvidia entering the CPU market in earnest. AMD's immediate repositioning — 'our AI strategy is infrastructure harmony, not CPU competition' — hints at the size of that shift. HSBC continues to see server-CPU upside for Intel, but that thesis pre-dates Nvidia arriving with its own CPU stack. The story's real royalty layer is quieter.

Perplexity's 1.5x — not a benchmark, a validation

While Nvidia's CPUs sat as specs on paper, one real workload landed: Perplexity ran agentic coding on Vera and reported a 1.5x speedup. Two things matter about that number. First, Nvidia's architectural bet — that single-threaded performance is the axis for agentic AI, workloads that orchestrate sequential tool calls — got its first commercial-customer validation. Second, Perplexity announcing 'adoption' means pipeline, not demo.

CoreWeave layered on: Vera Rubin deployed with training and inference on one stack. Demand-side (Perplexity) and infrastructure-side (CoreWeave) validation arrived within three days of the CPU reveal.

The $7.8M rack — memory is the biggest line

A Vera Rubin rack costs $7.8M, and the cost decomposition is the interesting part: memory (HBM-heavy) is the single largest line item. In the same week Meta announced CXL adoption for memory expansion and rebuilt its storage stack to cut GPU idle time by 97%. The rack BOM and the datacenter architecture are being redesigned around the same bottleneck.

Overlay today's DDR5 16Gb spot at $47.067. As memory spot recovers, memory's share of the rack BOM rises; Micron is committing $9.3B to Hiroshima to expand 1γ DRAM and HBM output; SK Hynix filed for a $28B US IPO. Three capital events, one direction: memory capacity is the ratelimiter and the market is funding to unblock it.

Yet the market read this oddly. Samsung's Q2 profits surged 1,800% YoY, topping Nvidia's quarterly record, but the stock fell and the disappointment cascaded — Intel -10%, Applied Materials -10%, AMD -8%. Consensus had already priced a shock, so a mere new record was not enough. The surprise threshold has moved.