Before NVIDIA GTC 2026: what is already claimed, and what buyers should check

NVIDIA's GTC conference opens in San Jose next week, and the announcements will arrive wrapped in superlatives. Before the keynote, it is worth writing down what the company has already said on the record, so that anything new can be judged against it.

The business backdrop

On February 25 NVIDIA reported fourth-quarter fiscal 2026 revenue of $68.1 billion for the quarter ended January 25, of which Data Center was $62.3 billion; full-year revenue was $215.9 billion ^[1]. It guided first-quarter fiscal 2027 revenue to $78.0 billion, plus or minus 2% ^[1].

The framing from the CEO in that release is explicit: "Computing demand is growing exponentially — the agentic AI inflection point has arrived," and "Enterprise adoption of agents is skyrocketing" ^[1]. Those are statements of position, not measurements, but they signal what GTC will be selling: agents as the reason to buy the next hardware generation.

What Vera Rubin already claims

At CES in January NVIDIA announced the Rubin platform as six co-designed chips: the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch ^[2]. The flagship Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs ^[2].

The headline claims, all NVIDIA's own and measured against its Blackwell generation, are up to 10x lower cost per token for inference on large mixture-of-experts models and one-quarter the GPUs for training them ^[2]. NVIDIA says Rubin is in full production, with partner products arriving in the second half of 2026 and early cloud instances planned from AWS, Google Cloud, Microsoft and Oracle, plus CoreWeave, Lambda, Nebius and Nscale ^[2].

The software side: models and agents

NVIDIA is not only selling silicon. On March 11 it released Nemotron 3 Super, an open-weight hybrid mixture-of-experts model of about 120 billion parameters with roughly 12 billion active at inference and a 1-million-token context window ^[3]. The company claims up to 5x higher throughput than the previous Nemotron Super and says it is available through build.nvidia.com, Hugging Face, OpenRouter, Google Cloud Vertex AI and Oracle Cloud, with AWS Bedrock and Azure listed as coming ^[3].

More speculative is the agent platform. On March 10 a report citing Wired said NVIDIA is preparing NemoClaw, an open-source platform for enterprises to deploy agents that perform tasks for employees, with built-in security and privacy tools ^[4]. The report said NVIDIA had held talks with Salesforce, Cisco, Google, Adobe and CrowdStrike, none of which had confirmed any collaboration, and that the platform could be used whether or not a company's products run on NVIDIA hardware ^[4]. The same report noted that security concerns have limited enterprise adoption of OpenClaw, an open-source agent tool that runs locally on personal computers ^[4]. Treat all of this as unconfirmed until NVIDIA says it.

Who the growth depends on

The $78.0 billion guidance ^[1] depends heavily on the cloud providers and system builders NVIDIA names as early Vera Rubin deployers ^[2]. If those buyers are the growth, their pricing and availability decisions are what ultimately reach other businesses, which is why cloud announcements around GTC matter more to most readers than rack specifications.

How to read the announcements next week

1. Separate "announced" from "orderable." A rack that ships to hyperscalers in late 2026 is not a planning input for a mid-sized business this year.

2. Ask who the 10x is measured against. Cost-per-token claims depend on the model, the batch size and the baseline; look for named workloads and independent measurements.

3. Watch the power and cooling requirements. Rack-scale systems change facility assumptions, and those costs rarely appear in a per-token headline.

4. Treat agent platforms as alpha until proven. A reported platform with unconfirmed partners is a signal of direction, not a product.

For most organizations the hardware news will reach them through cloud pricing and instance availability rather than purchase orders. The practical response is to avoid long commitments on today's per-token rates that assume nothing changes, and to revisit model and hosting choices once independent benchmarks of the new systems exist.