San Jose, 16–19 March. Jensen Huang walked out to a sold-out arena, 450 sponsors, 2,000 speakers, and a room full of hyperscaler CFOs already trying to figure out where they were going to find the megawatts to run what he was about to sell them. GTC 2026 was Huang’s most disciplined performance to date — less new-product theatre than in Blackwell year, more accounting.
The two-hour keynote broke into four beats. Vera Rubin is in production and shipping to customers, and it is designed for the workloads Huang has been telling everybody the AI industry is moving toward — reasoning, agentic AI, and inference at scale. NVIDIA has bought the Groq team and asset base for roughly $20 billion, and the first Groq-under-NVIDIA silicon — the Groq 3 Language Processing Unit — was on stage. NVIDIA is opening a full agentic AI framework it is comparing, more than a little pointedly, to Linux. And Huang told the room to plan for roughly one trillion dollars of purchase orders across Blackwell and Vera Rubin through 2027.
That last number was the one that mattered. Everything else on the slides was in service of it.
Vera Rubin ships, and inference gets its purpose-built silicon
The Vera Rubin platform — new Rubin-family GPUs paired with the NVIDIA-designed Vera CPU that succeeds Grace — is in production and moving to first customers. Huang framed the platform explicitly around agentic AI and reasoning-era inference, which is a materially different sales pitch than the training-first framing that anchored the Hopper and Blackwell cycles. The customer type NVIDIA is now designing for is not just the frontier lab training the next foundation model. It is the enterprise, the sovereign cloud, and the hyperscaler running an increasingly heavy inference book of business.
The rack-scale design assumes liquid cooling, dense interconnect, and — reading between Huang’s slides — power envelopes that put site selection into utility-negotiation territory. The Vera Rubin platform is not a chip you buy; it is a datacentre wing you build. That is the honest version of the sales conversation happening in 2026.
The Groq acquisition: NVIDIA buys the inference layer
The single biggest surprise of the show — although the roughly $20 billion asset purchase had been telegraphed in December — was the debut of the Groq 3 LPU, NVIDIA’s first Groq-branded silicon since the acquisition. The Groq 3 LPX rack holds 256 LPUs and is designed to sit alongside the Vera Rubin rack-scale system as a specialised inference tier for very-low-latency workloads. Chatbot inference, agentic tool-use loops, and voice-native applications are the target profile.
Strategically, the move is transparent. NVIDIA had two problems the Groq acquisition addresses. First, its GPU-based inference economics, while world-class for training-adjacent workloads, are increasingly being challenged at the low-latency end by purpose-built inference silicon from Groq, Cerebras, Etched, and a handful of hyperscaler internal projects. Second, the reasoning-era shift toward heavy test-time compute changes the inference cost curve in ways NVIDIA’s roadmap had not fully answered. Buying Groq gives NVIDIA a category-leading LPU story, a specialised rack alongside the general-purpose Vera Rubin one, and a message to CFOs that they do not need to build a heterogeneous silicon stack from the market: they can buy it all from one vendor.
For customers who had been carving out an inference budget line for non-NVIDIA silicon, the acquisition is uncomfortable. NVIDIA now sells you the training tier, the general-purpose inference tier, and the specialised LPU inference tier from one channel with one integration story. That is not a market NVIDIA is trying to keep open.
NemoClaw and OpenClaw: NVIDIA’s Linux move on agentic AI
The other headline product story was the release of OpenClaw and NemoClaw, an open-source agentic AI framework Huang compared, without much modesty, to Linux. NemoClaw sits above the existing NeMo tooling and provides the reference architecture for enterprise agentic AI: planner-plus-tool-use models, structured memory, guardrails, evaluation harnesses, and the identity and audit primitives that enterprise IT security teams have been demanding for eighteen months. OpenClaw is the open-source substrate; NemoClaw is the enterprise-secure reference design NVIDIA will ship to regulated buyers.
The comparison to Linux is doing more work than the words themselves suggest. NVIDIA is making an explicit bet that agentic AI will consolidate around one dominant open-source substrate — the way container orchestration consolidated around Kubernetes, and the modern application stack consolidated around Linux itself. It is trying to become the vendor of the reference implementation, on the theory that the reference implementation gets bought.
Whether OpenClaw actually wins that consolidation battle, over the medium term, is one of the two or three most consequential product questions in enterprise AI. The rivals are formidable — LangChain, Anthropic’s agent frameworks, the OpenAI Agents SDK, and various hyperscaler-native offerings from AWS, Google Cloud and Azure. But NVIDIA has the distribution the framework vendors do not, and it is prepared to spend to close the ecosystem gap.
A trillion dollars of purchase orders through 2027
The number Huang wanted the market to remember was $1 trillion. That is his stated expectation for combined Blackwell and Vera Rubin purchase orders through 2027. Whether the number is analytically precise is beside the point. It is a target for investor communication, an anchor for hyperscaler capex planning, and, uncomfortably for competitors, an argument that the AI capex cycle has not peaked and will not peak on this generation of silicon.
The hyperscaler customers in the room did not push back publicly, which is its own tell. Microsoft, Google, Meta, AWS, Oracle, CoreWeave, and the sovereign programmes that increasingly sit alongside them have their own capex numbers to defend to their own boards; the Huang trillion-dollar frame gives them cover.
The automotive lock-up: Uber, BYD, Geely, Nissan, Hyundai
The most quietly consequential part of the keynote was the automotive section. Huang announced that Uber will launch an autonomous fleet powered by NVIDIA’s Drive AV software across 28 cities on four continents by 2028. Alongside that, Nissan, BYD, Geely, Isuzu, and Hyundai — a Japanese, two Chinese, a Japanese-commercial-vehicle, and a Korean OEM — were named as Drive Hyperion-programme partners for Level 4 autonomous vehicles.
NVIDIA has spent most of a decade turning its Drive stack into the reference platform for automotive autonomy. In 2026 the lock-up looks meaningfully more complete than it did in 2024. The three volume markets outside the US — China, Japan, Korea — are now visibly aligning behind Drive Hyperion for L4. The market that is not aligning is the two US OEMs with the strongest internal autonomy stacks (Tesla and, to a lesser degree, GM), but the international volume outweighs the domestic exceptions. If autonomy actually clears the last-mile productisation problem this cycle — and there are more credible reasons than at any prior GTC to believe that it will — NVIDIA is positioned to sit under a lot of it.
Power is the real constraint
The subject Huang gestured toward without dwelling on was the physical reality of running the Vera Rubin generation at the volumes the trillion-dollar order book implies. Grid access, cooling, and interconnect fabric are the binding constraints in most markets. Microsoft, Google, Meta, and Amazon have all publicly written cheques for nuclear and long-duration renewables in 2026 not because they want to be in the utility business but because they cannot buy GPUs they cannot power.
The Vera Rubin generation, and the Groq 3 rack that sits alongside it, are designed with those constraints in mind — denser packaging, liquid cooling assumed, thermal-envelope negotiations built into the site plan. But the underlying industrial problem is not going away. Utilities in the US, Europe, and East Asia are quietly the most important suppliers to the AI industry in 2026.
What GTC 2026 actually moved
- Vera Rubin is in production. The reasoning-era inference workload has purpose-built silicon.
- Groq is inside the fence. NVIDIA now sells the specialised LPU tier alongside the general-purpose GPU tier.
- OpenClaw is NVIDIA’s bid to own the agentic framework layer. The Linux comparison is intentional.
- $1 trillion of purchase orders through 2027. The AI capex cycle has not peaked.
- Uber, BYD, Geely, Nissan, Hyundai are on the Drive stack. International automotive autonomy is consolidating around NVIDIA.
Open questions
Do the utility-side commitments in the US and Europe scale fast enough to actually field the trillion-dollar Vera Rubin order book by 2027? Does OpenClaw win the agentic-framework consolidation battle against LangChain, the hyperscalers, and the frontier labs, or does the market fragment along vendor-specific stacks? And does the Groq absorption produce the low-latency inference cost curve NVIDIA needs to hold its share of the reasoning-era workload, or does the specialised-silicon cohort continue to peel off inference share?
Bottom line: GTC 2026 was Jensen Huang locking in the next two years of AI infrastructure demand and telling the room to plan around it. Vera Rubin is real. Groq is now part of the offer. Agentic AI has a reference implementation NVIDIA wants to standardise. The automotive story is quietly closing. And the AI capex cycle, on the trillion-dollar frame Huang set out, has room to run. The follow-through is industrial — power, cooling, interconnect, utilities — and it is now, more visibly than at any prior GTC, everybody’s problem, not just NVIDIA’s.
