San Jose, 17–21 March. Jensen Huang delivered his 2025 GTC keynote in a packed SAP Center with the composure of a man who has already sold you next year’s roadmap. He spent two hours doing something that in any other industry would be considered impolite: he told the room what NVIDIA is shipping in 2025, what it is shipping in 2026, and what it is shipping in 2027, and then he told them to plan their datacentres accordingly.
The subtext, for anyone who has been watching the hyperscaler capex numbers, was not subtle. NVIDIA is moving to an annual product rhythm. Blackwell Ultra ships in the second half of this year. The Vera Rubin platform lands in the second half of 2026. Rubin Ultra follows in 2027. The compute buyers in the room — Microsoft, Google, Meta, AWS, Oracle, CoreWeave, sovereign AI programmes — now have three years of purchasing calendars pinned to Huang’s slides. That is not an accident; that is the strategy.
If you wanted the single-line takeaway from GTC 2025, this is it: the AI build-out is not slowing down, NVIDIA is not slowing down, and Huang has just made it structurally very difficult for anyone else to catch up.
Blackwell Ultra is the workhorse for the reasoning era
Blackwell — the current generation that spent most of 2024 shipping to hyperscalers — is now in what NVIDIA calls “full production.” Huang cited the familiar 40x-Hopper performance number, which is the framing for training and increasingly for large-model inference. The more consequential news was Blackwell Ultra, the platform that arrives in the second half of 2025 and is explicitly positioned for reasoning-era AI workloads: agentic systems, chain-of-thought inference, test-time compute, and the physical AI applications that were the closing act of the keynote.
The reasoning-era framing matters more than it sounds. Through 2024 the story was training scale. The story for 2025 is that inference — specifically the very heavy inference of reasoning models running long deliberation chains — is now the workload that eats compute. Blackwell Ultra is NVIDIA’s answer to the fact that customers who thought they were finished shopping after their 2024 training buildouts now have to shop again for inference capacity. It is a very expensive way to be right, but Huang is unmistakably right.
Vera Rubin: NVIDIA books the 2026 datacentre in March 2025
The bigger reveal, for anyone thinking about NVIDIA’s competitive moat rather than its next earnings print, was the Vera Rubin architecture: a new GPU family paired with a new NVIDIA-designed Vera CPU (the Grace successor), landing in the second half of 2026. Huang quoted roughly three times the performance of Blackwell for the Rubin GPUs. Rubin Ultra, the mid-cycle upgrade, is on the calendar for 2027.
Announcing 2026 silicon in March 2025 is a supply-chain move as much as a marketing move. It gives NVIDIA’s largest customers permission to place forward orders now for what would ordinarily be an eighteen-month planning cycle, and it gives NVIDIA visibility into utilisation and pricing for the next two generations. It also, more quietly, tells AMD, Broadcom, Cerebras and the custom-silicon programmes at Google, AWS and Meta that they are competing against a moving target with an unusually well-lit roadmap.
The tell is that when Huang said the roadmap will now cadence annually — new GPU family, new CPU family, new networking every year — the room didn’t gasp. The room nodded. Everybody had spent the past twelve months trying to buy a datacentre and understood exactly what he meant.
Robotics and physical AI came out of the corner
The section of the keynote that will read best a year from now was the physical-AI push. NVIDIA formally opened the Isaac GR00T humanoid robot foundation-model programme, launched Cosmos — a family of world foundation models for physical AI — and demoed live robotics work with a small Disney-imagineered droid that walked out on stage next to Huang. It was a set piece, but the underlying stack is not: GR00T + Cosmos + Isaac Sim + Omniverse forms an end-to-end pipeline for training embodied AI in simulation and deploying it into hardware, which is the honest version of the “humanoid robotics” story that Chinese national programmes and a handful of Silicon Valley startups have been chasing.
Two things matter here. First, NVIDIA is not building robots — it is doing to robotics what it did to AI training, which is selling the platform layer while every hardware maker builds on top. Boston Dynamics, Agility, Figure, 1X, Apptronik and the Chinese cohort will all use pieces of this stack whether they announce it or not. Second, this positions robotics as an AI workload — training, sim-to-real, edge inference — that fits neatly into the Blackwell Ultra / Rubin planning story. It is another demand vector for the same silicon.
Agentic AI in the enterprise: platform, not product
Huang leaned hard into the agentic AI theme, which by GTC 2025 has moved past the definitional argument (are agents just chained prompts?) and into deployment. NVIDIA’s angle is characteristically infrastructural: NIM microservices, the NeMo tooling, and reference architectures for enterprise agentic workflows that plug into ServiceNow, SAP, Snowflake, Palantir and the usual enterprise data platforms. The vendor keynotes and partner sessions were full of variations on the same theme: enterprise agentic AI is being productised, and NVIDIA wants to be the reference platform underneath all of it.
For enterprise CIOs actually listening — and there were a lot of them in San Jose this year, more than in previous GTCs — the practical story is that agentic AI has a shape now. It is a planner model, tool-use interfaces, structured memory, guardrails, and observability. The vendors that can articulate all five parts are getting bought. The vendors that still lead with “our model is bigger” are not.
Sovereign AI: a product line, not a slogan
The sovereign AI conversation, which had been a talking point through 2024, has now become a defined go-to-market segment. NVIDIA references sovereign deployments across France, Japan, Saudi Arabia, India, UAE and several others, most anchored on national compute programmes and national telco or systems-integrator partners. The reference architecture — Blackwell-class silicon, on-prem inference, localised model families, national data residency — has enough repeat customers that NVIDIA now sells to it as a category with dedicated go-to-market motion.
For the European and Middle Eastern buyers in the room, this was probably the most consequential undercurrent of the show. Sovereign AI is no longer a policy debate; it is a purchase order.
The compute-supply war is the actual story
The unsaid subject of GTC 2025 was the physical reality of building this. Power is the constraint. Grid access is the constraint. Cooling water is the constraint. The site-selection conversation for the next generation of AI datacentres is now indistinguishable from a utility negotiation. Huang gestured at it without dwelling: the Rubin generation is designed with denser rack architectures, liquid cooling assumed, and interconnect fabrics that expect the whole datacentre to behave like one large machine.
The hyperscalers and the sovereigns are both playing this game. Microsoft is buying nuclear. Google is buying nuclear. Amazon is buying nuclear. Meta is signing decade-long PPAs. The uncomfortable truth Huang did not need to say is that if you cannot get the megawatts, you cannot buy the GPUs. The compute buyers who prospered in 2025 were the ones with the utility teams.
What GTC 2025 actually moved
- Blackwell Ultra makes inference — specifically reasoning-era inference — the workload that funds the next capex cycle.
- Vera Rubin on the 2026 calendar changes the buyer’s planning horizon. Hyperscalers have to place multi-generation orders now.
- Annual cadence is the strategy. Every AI vendor upstream and downstream now has to plan around a yearly NVIDIA cycle.
- Physical AI has a platform. Isaac GR00T + Cosmos + Omniverse is the reference stack for humanoid and industrial robotics.
- Sovereign AI is a product line. Not a slide.
Open questions
Do power, grid and cooling scale fast enough to actually field the Rubin generation at the volumes NVIDIA is planning? Does the custom-silicon cohort (Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia, plus Cerebras and the Chinese programmes) close enough of the gap on inference economics to force a price response before Rubin ships? And can NVIDIA sustain an annual cadence without a manufacturing hiccup — TSMC advanced packaging, HBM memory supply, cooling supply — that resets everyone’s calendar?
Bottom line: GTC 2025 was Jensen Huang locking in three years of AI infrastructure demand and telling the room to plan around it. Blackwell Ultra is the workhorse for the reasoning era. Vera Rubin is the platform buyers already have to underwrite. Physical AI became a product line rather than a demo. And sovereign AI became a procurement track. Everybody else in the AI supply chain — competitors, customers, regulators, utilities — is now planning around a NVIDIA calendar they did not get to write.
