At CES in January, Jensen Huang announced Vera Rubin: six co-designed chips that ship as one rack-scale system. Custom Arm CPU (88 Olympus cores), Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-X. NVIDIA claims 4x fewer GPUs to train equivalent models and 10x lower inference token costs versus Blackwell. Full production shipping H2 2026.

The structural shift matters more than the specs. NVIDIA stopped selling components and started selling complete AI infrastructure. Every major cloud provider has committed capacity. Every major AI lab is on the partner list. The Vera CPU for agentic workloads means NVIDIA now competes with Intel and AMD on the CPU side too. This is a systems company that happens to make great GPUs.

The open question in January was whether NVIDIA could make Vera Rubin chips fast enough to back the platform shift. HBM4 spec revisions in Q3 2025 pushed mass production timelines, and Blackwell demand kept pulling resources forward. Two months later, GTC answered.

First customer samples shipped in late February. Colette Kress confirmed it on the earnings call: "We shipped our first Vera Rubin samples to customers earlier this week." Production shipments remain on track for the second half of the year. Quanta's executive VP expects the first customer racks by August.

The demand signals are hard to overstate. Nebius Group and Meta signed a $27 billion infrastructure deal on March 16, with $12 billion in dedicated Vera Rubin capacity starting early 2027. Micron hit high-volume production of HBM4: 36GB stacks delivering 2.8 TB/s bandwidth, a 2.3x improvement over HBM3e. The memory bottleneck that could have stalled Rubin production is clearing.

NVIDIA expanded the platform since CES. Seven chips across five rack configurations now, not six chips and one rack. The "AI factory" framing dominated GTC: data centers as token production facilities, not general-purpose compute. Jensen Huang described a split - 25% of data center compute on Groq LPUs for low-latency decode, 75% on Vera Rubin GPUs for prefill and training.

The Vera CPU is now positioned as a standalone product. CoreWeave expects to be among the first cloud providers to deploy it in production in H2 2026. A standalone NVIDIA CPU for agentic AI workloads was theoretical in January. It has a deployment timeline now.

The thesis holds. NVIDIA is not a GPU company anymore. And the supply chain is keeping up.