Jensen Huang walked onto the stage at SAP Center in San Jose on Monday with the energy of someone who knows exactly what he’s holding. Three hours later, he had dropped more announcements than most companies make in a year. Welcome to NVIDIA GTC 2026: chips, robots, AI agents and a revenue forecast with twelve zeroes.
Here’s what you actually need to know about the March 16 keynote.
A Trillion-Dollar Forecast. Yes, Really.
In NVIDIA’s keynote recap, Huang says he now sees at least $1 trillion in revenue from Blackwell and Rubin over 2025–2027. That is a forecast spanning those years, not a statement that NVIDIA has already collected a trillion dollars or disclosed a matching pile of firm purchase orders.
The actual revenue numbers are enormous enough without giving them a promotion. NVIDIA’s February 25 financial results reported $68.1 billion for the latest quarter, up 73% from a year earlier. The $78 billion figure is its outlook for the following quarter, with a 2% margin either way. Results in one column, expectations in another. The spreadsheet appreciates the distinction.
Vera Rubin and the Machine That Drinks Power Wisely
Named after astronomer Vera Rubin, NVIDIA’s next platform combines processors, networking and storage rather than simply swapping one graphics card for another. The company says its seven new chips are in full production, while partner products are expected in the second half of 2026. Chips entering production and customers receiving complete systems are different milestones.
In the March 16 platform announcement, NVIDIA claims up to ten times the inference throughput per watt for the NVL72 rack versus Blackwell, at one-tenth the cost per token. That is a claim about serving AI workloads, not a promise that every computing task becomes ten times faster or that a data center’s electricity bill automatically shrinks by 90%.
Fitting the whole platform together still sounds like a job requiring the precision of a Swiss watchmaker and the ambition of a Saturn V engineer.
Groq Technology Joins the Lineup
The December deal was not an acquisition of the entire Groq company. Groq’s December 24 announcement describes a non-exclusive technology licence and the move of founder Jonathan Ross, president Sunny Madra and other staff to NVIDIA. Groq said it would remain independent and keep GroqCloud running.
At GTC, NVIDIA introduced Groq 3 LPX racks to complement Vera Rubin for low-latency inference. Each LPX rack contains 256 language processing units. The company advertises up to 35 times higher inference throughput per megawatt for the combined approach, aimed at very large models and long contexts. It is a specific platform claim, not a 35-times upgrade button for every Rubin workload.
The attraction is fairly straightforward: an AI service needs both the capacity to handle lots of work and the responsiveness to avoid leaving everyone staring at a blinking cursor. The announced systems target the second half of this year.
Kyber: There Is Always Another Rack
Huang also looked beyond the immediate launch. NVIDIA’s keynote recap connects the future Feynman generation with a new Rosa CPU and Kyber rack architecture. The roadmap is about increasingly integrated systems, including how processors communicate, rather than just a succession of faster individual chips.
Roadmaps explain the direction of travel; they are not a warehouse inventory. For anyone still digesting Blackwell, being shown the generations beyond Rubin is rather like ordering lunch and receiving the restaurant’s five-year expansion plan.
Disney Brought a Robot. Of Course Disney Brought a Robot.
The snowman on stage was Olaf, the robotic character developed by Walt Disney Imagineering Research & Development. Disney’s March 16 interview with the team describes the work behind his upcoming Disneyland Paris appearance, including learning to balance on a boat.
The interesting part is movement: Disney uses simulation and reinforcement learning to train a very unusual body with snowball feet. That is impressive without treating the demonstration as proof of an unrestricted conversational character that can improvise with any guest. A snowman earning his sea legs is already quite a pitch.
NemoClaw: An Agent Stack With Boundaries
NVIDIA also announced NemoClaw for OpenClaw. It installs Nemotron models and the OpenShell runtime, adding controls around how agents execute tasks and access data or networks.
The announcement describes local models, a privacy router for cloud models and support for dedicated platforms; it is not simply an agent that only runs on NVIDIA hardware. Nor does a one-command installation mean every agent is ready to be handed the company accounts. Setup convenience and sensible permissions remain separate jobs.
If you have been following AI-assisted coding, this is part of the same broader shift: software that takes several steps on your behalf, with the surrounding tools becoming increasingly important.
Autonomous Cars: Plans With Places and Dates
The DRIVE Hyperion announcement names BYD, Geely, Isuzu and Nissan among companies developing vehicles ready for Level 4 systems. NVIDIA and Uber also plan a rollout across 28 cities on four continents by 2028, beginning in Los Angeles and the San Francisco Bay Area in the first half of 2027.
Those are development and deployment plans, not 28 fleets already waiting outside. Level 4 automation handles the driving within its defined operating conditions without relying on a human driver to intervene. It does not mean the car can drive anywhere, in any conditions. The geographical small print matters rather more when it is carrying passengers.
Data Centers in Space (Yes, Really)
This was more than an offhand suggestion about orbital servers. NVIDIA announced the Space-1 Vera Rubin Module alongside computing products for space applications. It is a hardware direction, not evidence that the difficult business of powering and cooling orbital data centers has already been solved.
After GPUs, robots and self-driving cars, apparently the next frontier is quite literally above the next frontier.
The Big Picture
GTC 2026 showed the breadth of NVIDIA’s ambitions: processors, complete computing systems, software, vehicles and robotics. The trillion-dollar forecast is the headline magnet, but the connecting idea is to provide more of the infrastructure around AI, wherever it runs.
Whether all that investment delivers the returns its backers expect remains an open question. Monday supplied plenty of announcements and projections; the deployments and results will have to follow. Either way, Jensen Huang’s leather jacket isn’t going anywhere.
