NVIDIA GTC 2026: everything Jensen Huang will unveil on March 16
NVIDIA GTC 2026 runs March 16-19 in San Jose with the Rubin platform, Jetson T4000, physical AI breakthroughs, and robotics partnerships with Boston Dynamics and more.
Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchAI, startups & growth insights. No spam.
TL;DR: NVIDIA GTC 2026 runs March 16-19 in San Jose with Jensen Huang unveiling the Rubin platform and six new chips for AI infrastructure. The Jetson T4000 edge module launches at $1,999, targeting physical AI deployments. Partners including Boston Dynamics, Caterpillar, and LG are already building on the new stack, making this the most consequential GTC since the original Blackwell reveal.
30,000+ attendees. 1,000+ sessions. Six new chips. One keynote that will define AI infrastructure for the next two years. Jensen Huang takes the stage at SAP Center in San Jose on March 16 at 11 AM PT — and the industry has not seen a lineup this dense since the original Blackwell reveal. Here is everything you need to know before the curtain goes up.
GTC started as a GPU technology conference for developers. It has become something closer to an AI industry summit where product roadmaps, research milestones, and capital allocation decisions get made in real time. The 2025 edition drew roughly 20,000 attendees. This year the number crosses 30,000, with 1,000+ sessions, 150+ research presentations, and 240+ Inception startups on the floor.
The shift is structural. NVIDIA is no longer primarily selling discrete GPUs into consumer markets. It is selling what Huang calls "AI factories" — end-to-end systems that combine chips, networking, storage, software, and models into infrastructure that produces tokens the way a power plant produces electricity. GTC 2026 is where that factory metaphor gets its fullest articulation yet.
Three themes dominate every session preview:
Each theme maps to a product announcement. The architecture of this keynote is unusually legible from the outside.
NVIDIA announced the Vera Rubin platform at CES in January 2026 as the direct successor to Blackwell. It ships in the second half of 2026. GTC is where Huang will give the platform its full public showcase — with production partners, deployment commitments, and likely a pricing framework.
The Rubin platform is not a single chip. It is a six-chip system engineered to operate as a unified AI supercomputer:
| Chip | Role |
|---|---|
| NVIDIA Vera CPU | 88 custom Olympus cores, full Arm compatibility, optimized for AI factory control plane |
| NVIDIA Rubin GPU | Two reticle-sized dies, 288 GB HBM4 memory per unit |
| NVIDIA NVLink 6 Switch | High-bandwidth chip-to-chip and GPU-to-GPU interconnect |
| NVIDIA ConnectX-9 SuperNIC | Network interface for AI factory scale-out |
| NVIDIA BlueField-4 DPU | Data processing and security offload |
| NVIDIA Spectrum-6 Ethernet Switch | Fabric for rack-scale AI systems |
The flagship configuration is the Vera Rubin NVL72: 72 Rubin GPUs combined with 36 Vera CPUs, NVLink 6 switching, ConnectX-9 SuperNICs, and BlueField-4 DPUs in a single rack-scale system. The NVL144 CPX variant — optimized for the prefill phase of LLM inference — delivers 8 exaflops of AI performance per rack alongside 100 TB of fast memory and 1.7 PB/s of bandwidth.
Manufacturing is on TSMC's 3 nm process. Memory is HBM4. Production is already underway — Huang confirmed this at CES. Microsoft has committed to deploying NVL72 systems in its next-generation Fairwater AI superfactories, scaling to hundreds of thousands of Vera Rubin Superchips.
The raw headline figures from NVIDIA's own comparisons:
| Metric | Blackwell | Rubin | Rubin Ultra (2027) |
|---|---|---|---|
| FP4 performance (per GPU) | 20 PFLOPS | 50 PFLOPS | 100 PFLOPS |
| Inference token cost reduction | baseline | 10x lower | TBD |
| MoE training GPU count reduction | baseline | 4x fewer | TBD |
| Memory per GPU | HBM3e | 288 GB HBM4 | HBM4e (expected) |
50 petaflops of FP4 performance per GPU is the number that gets cited most often. It represents a 2.5x increase over Blackwell on a per-unit basis. When you scale that across the NVL72's 72 GPUs, the system-level numbers become difficult to contextualize against anything that came before.
The 10x inference token cost reduction is the number that matters most to enterprise buyers. Cost-per-token is the unit economics of AI deployment. A 10x drop does not just make existing use cases cheaper — it makes economically unviable use cases viable. That is a market expansion argument, not just a performance argument.
Rubin Ultra arrives in 2027, doubling Rubin's per-GPU performance to 100 PFLOPS. Beyond that, NVIDIA has placed Feynman on the roadmap — though no specifications are public yet.
While Rubin anchors the data center stack, the Jetson T4000 anchors the edge. Announced at CES in January, the module brings Blackwell architecture to autonomous machines and general robotics at $1,999 at 1,000-unit volume.
Key specifications:
The positioning is deliberate. NVIDIA is not trying to win the robotics market with the most powerful chip. It is trying to win it with the most deployable chip. The combination of Blackwell-class inference performance at 70W in a module priced under $2,000 at volume removes the primary cost barrier for commercial robotics integrators.
At GTC, expect to see the Jetson T4000 demoed inside production-ready robot hardware from NVIDIA's partner ecosystem — likely including systems from Boston Dynamics, NEURA Robotics, and Franka Robotics. The JetPack 7.1 SDK and Isaac Lab-Arena evaluation framework will get extended coverage in developer sessions.
Hardware without models is infrastructure without purpose. NVIDIA's answer to the model layer for physical AI is a two-part stack: Cosmos for world simulation and reasoning, and GR00T for robot-specific learning and control.
Cosmos is NVIDIA's open world foundation model platform. The flagship release is Cosmos Reason 2, described as a leaderboard-topping reasoning VLM that enables robots and AI agents to see, understand, and interact with higher accuracy in the physical world. Cosmos models are trained on NVIDIA's multimodal dataset, which includes:
The key use case is sim-to-real transfer: training robots in photorealistic simulation, validating behaviors there, then deploying to physical hardware with high confidence. This shortens the physical testing cycle dramatically and reduces the cost of gathering edge-case training data.
GR00T is the humanoid robot foundation model. At GTC 2025, NVIDIA introduced the original Isaac GR00T model. GTC 2026 will showcase the expanded GR00T ecosystem, including GR00T N1 — a model designed to enable generalizable manipulation and locomotion across robot morphologies. The OSMO edge-to-cloud compute framework, which simplifies robot training workflows from Isaac Lab to physical deployment, will also feature prominently.
For developers, the open model strategy is significant. NVIDIA is releasing Cosmos and GR00T weights publicly, mirroring the approach that made Llama a de facto standard in language AI. If GR00T achieves similar adoption among robotics developers, NVIDIA's data flywheel compounds faster than any competitor could match through proprietary development alone.
Physical AI systems need an intelligence layer that can reason, plan, and use tools — not just perceive and respond. That is where Nemotron fits.
The Nemotron 3 family, unveiled at CES, is purpose-built for agentic AI: autonomous systems capable of multi-step reasoning, tool use, and complex decision-making across extended workflows. Unlike previous Nemotron releases that focused on chat and summarization, Nemotron 3 is architected for agents — systems that persist state, call external APIs, and coordinate across sub-agents to complete tasks.
NVIDIA's open model strategy extends to Nemotron. The company has contributed training frameworks and one of the largest collections of open multimodal datasets available. The breadth of the dataset matters: agentic systems that can reason over language, code, scientific structures, and sensor data simultaneously are categorically more capable than systems trained on any single modality.
At GTC, Huang will likely frame Nemotron as the reasoning engine that sits above the physical AI stack — the model that coordinates Cosmos perception, GR00T motor control, and Alpamayo vehicle reasoning into integrated autonomous systems. Watch for enterprise announcements around Nemotron licensing and deployment on NVIDIA DGX Cloud.
Alpamayo is NVIDIA's open model family for autonomous vehicle development, released at CES 2026 alongside Cosmos and Nemotron. It represents NVIDIA's most direct play for the autonomous driving software stack — a market where it has long provided compute but has not owned the intelligence layer.
The centerpiece is Alpamayo 1: the first open, large-scale reasoning Vision-Language-Action (VLA) model for autonomous vehicles. It enables vehicles to understand their surroundings, reason about what to do next, and explain their actions in natural language — the latter being increasingly important for regulatory approval processes.
Mercedes-Benz is the first automaker to deploy Alpamayo in production, with the new Mercedes-Benz CLA featuring AI-defined driving built on NVIDIA's Drive platform, scheduled to reach the United States market in 2026.
The supporting infrastructure is also significant:
At GTC, expect additional automaker partnerships to be announced alongside deeper technical disclosure on Alpamayo's architecture and benchmark performance. The autonomous vehicle session track is one of the most densely scheduled at GTC 2026.
NVIDIA's physical AI strategy is not a solo effort. At CES, six hardware partners debuted new robots and autonomous machines built on NVIDIA technologies:
| Partner | What They Showed |
|---|---|
| Boston Dynamics | Next-generation robots using Cosmos and GR00T; expanded AI collaboration with Google DeepMind on Atlas platform |
| NEURA Robotics | New humanoid robot built on NVIDIA Jetson and Isaac stack |
| Franka Robotics | Manipulation arm with GR00T-based learning for industrial assembly |
| Caterpillar | Autonomous heavy equipment with Cosmos-based terrain reasoning |
| LG Electronics | Home service robot platform on NVIDIA AI |
| Humanoid | Unnamed general-purpose humanoid prototype |
At GTC, these partners will have expanded presence. Boston Dynamics' collaboration is particularly notable — the company's Atlas humanoid is the most capable general-purpose robot in commercial production, and integrating it with NVIDIA's Cosmos simulation and GR00T learning stack creates a uniquely well-resourced development environment. The partnership with Google DeepMind adds another dimension: DeepMind's robotics research (Gemini Robotics, RT-2) combined with NVIDIA's infrastructure is a combination that no other partnership in the space currently matches.
Boston Dynamics expanded its NVIDIA collaboration specifically to accelerate AI capabilities in Atlas, covering perception, locomotion, and manipulation. At GTC, expect live demos and likely a technical session from Boston Dynamics engineers on the integration architecture.
Caterpillar's presence in this list is underappreciated. Industrial autonomy in mining, construction, and agriculture represents a market that dwarfs consumer robotics by capital deployed, and Caterpillar's machines are already in hundreds of thousands of worksites globally. Cosmos-based terrain reasoning applied to a fleet of autonomous excavators is not a research demo — it is a near-term revenue event.
Huang's framing device for GTC 2026 is the "AI factory" — a facility purpose-built to produce AI tokens the way a semiconductor fab produces chips. The five-layer stack he will describe maps neatly to NVIDIA's product portfolio:
| Layer | What it means | NVIDIA product |
|---|---|---|
| Energy | Power infrastructure for AI compute density | Partner ecosystem (not a direct NVIDIA product) |
| Chips | The compute substrate | Vera Rubin, Jetson T4000, Rubin CPX |
| Infrastructure | Networking, storage, and system design | NVLink 6, ConnectX-9, BlueField-4, Spectrum-6 |
| Models | Foundation models for reasoning and action | Cosmos, GR00T, Nemotron, Alpamayo |
| Applications | Domain-specific deployments | Partner ecosystem, NVIDIA Inference Microservices |
The five-layer framing is not just marketing. It is a land-grab across the entire AI value chain. Every layer where NVIDIA owns the standard is a layer where switching costs accumulate. The chips are already near-inescapable for serious AI workloads. The infrastructure layer is cementing with NVLink becoming a preferred interconnect for training clusters. The model layer is newer territory, and the open model strategy is NVIDIA's primary weapon there.
DRAM pricing is an indirect signal worth watching. A Digitimes analysis published March 3 noted that DRAM prices are forecast to surge up to 70% in Q2 2026 — driven in part by HBM4 supply constraints directly related to Rubin production ramp. Memory supply is tight because demand for Rubin-class systems is pulling forward procurement cycles.
In a pre-GTC comment reported by VideoCardz, Jensen Huang said he would unveil "several new chips the world has never seen before." This language is deliberately provocative. The Rubin platform is already known. The Feynman architecture is acknowledged but unspecified. What sits between them — or alongside them — has not leaked in a meaningful way.
Candidate categories for the surprise:
A dedicated inference chip. Rubin CPX is optimized for the prefill phase of inference. An equivalent product for the decode phase — which has a very different compute and memory profile — would complete the inference-specific stack. Pure inference workloads, which now dominate production AI traffic, have different requirements than training. A chip purpose-built for decode would directly compete with custom silicon efforts at Google (TPU), Amazon (Trainium/Inferentia), and Microsoft (Maia).
A confidential compute chip. Regulated industries — healthcare, finance, government — require AI compute that can operate on encrypted data. A purpose-built chip for confidential AI inference would unlock procurement at organizations that currently cannot deploy NVIDIA hardware due to data handling requirements.
A sovereign AI chip. Multiple governments have announced AI sovereignty initiatives in the past 12 months. A chip variant with supply chain characteristics acceptable to non-US governments (separate from the existing export-controlled product line) would open a large addressable market currently blocked by regulatory constraints.
None of these are confirmed. But the "surprise the world" framing suggests something beyond an incremental spec bump on an existing product line.
GTC 2026 includes a formal investor Q&A on March 17 at 9 AM PT — the morning after the keynote. Three questions are likely to dominate:
1. Rubin production trajectory. NVL72 is in production. The question is volume ramp timing and customer allocation. The same HBM4 supply constraints that are driving DRAM price increases create a ceiling on how fast Rubin can scale. Analysts will want specifics on H2 2026 shipment cadence.
2. China revenue exposure. Export restrictions have cut NVIDIA off from a significant portion of its historical data center revenue in China. GTC announcements around sovereign AI or any regulatory developments will be read directly against this exposure.
3. Software revenue conversion. NVIDIA Inference Microservices (NIM), the DGX Cloud platform, and the open model ecosystem are all vehicles for recurring software revenue that does not carry chip margins but does carry subscription economics. At what scale does software become material to NVIDIA's financial model? GTC is typically where the clearest signals on this emerge.
When and where is NVIDIA GTC 2026? March 16-19, 2026, at the San Jose Convention Center and SAP Center in San Jose, California. Jensen Huang's keynote is March 16 at 11 AM PT at SAP Center. The keynote will be livestreamed at nvidia.com — no registration required.
What is the Vera Rubin platform and when does it ship? Vera Rubin is NVIDIA's next-generation AI compute platform, succeeding Blackwell. It comprises six chips — the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch. The flagship NVL72 rack-scale system is in production and will be available from partners in the second half of 2026.
What is physical AI and why is NVIDIA focused on it? Physical AI refers to AI systems that perceive, reason about, and act in the physical world — robots, autonomous vehicles, and industrial machines. Unlike software-only AI, physical AI requires tight integration between simulation, sensor processing, and actuation. NVIDIA's Cosmos (world simulation), GR00T (robot learning), and Jetson (edge compute) platforms form its physical AI stack.
What is the Jetson T4000 and who is it for? The Jetson T4000 is NVIDIA's latest edge AI module, bringing Blackwell architecture to robots and autonomous machines at $1,999 per unit at 1,000-unit volume. It delivers 1,200 FP4 TFLOPS within a 70W power envelope — a 4x performance improvement over the previous generation. It is designed for commercial robotics integrators who need high inference performance at robot-compatible power budgets.
What robotics partners will be at GTC 2026? NVIDIA has confirmed partnerships with Boston Dynamics, NEURA Robotics, Franka Robotics, Caterpillar, and LG Electronics, among others. Boston Dynamics' expanded collaboration — which also involves Google DeepMind — is expected to receive significant stage time during the keynote.
What is Alpamayo and how does it relate to autonomous vehicles? Alpamayo is NVIDIA's open model family for autonomous vehicle development. Alpamayo 1 is a Vision-Language-Action model that enables vehicles to reason about their environment and explain their decisions in natural language. Mercedes-Benz is the first production automaker to deploy it, in the new Mercedes-Benz CLA launching in the United States in 2026.
How does Rubin compare to Blackwell in performance? Rubin delivers 50 PFLOPS of FP4 performance per GPU versus Blackwell's 20 PFLOPS — a 2.5x per-unit increase. At the system level, Rubin NVL72 reduces inference token costs by 10x and cuts the GPU count required to train Mixture-of-Experts models by 4x compared to equivalent Blackwell configurations.
Can I watch the GTC 2026 keynote online for free? Yes. The keynote will be livestreamed at nvidia.com on March 16 at 11 AM PT. No registration is required. It will also be available on demand after the event.
German robotics startup Neura Robotics closed approximately €1 billion in funding from Tether Holdings, valuing the company at €4 billion as it prepares to fill nearly €1 billion in existing orders for cognitive humanoid machines.
NVIDIA launches the Rubin computing architecture at MWC Barcelona with six new chips spanning agentic AI, physical AI, autonomous vehicles, robotics, and biomedical.
Alphabet's robotics software company Intrinsic joins Google. What this means for AI-powered robots and Google's automation strategy.