NVIDIA GTC 2026 San Jose Global AI Industry Mapping

GTC's organizing logic differs from a neutral trade show: it is a single vendor's roadmap launch, and the entire ecosystem is arranged by its interface position relative to NVIDIA's technology stack. Jensen Huang summarized this structure in his keynote as "AI's five-layer cake." This mapping subdivides it further into ten layers, to locate companies layer by layer.

Field One · Influence Tier: T0 category-definer · T1 first tier · T2 effective participant. The judgment is based on the company's global standing within its specific sub-track.

Field Two · Company Stage: Mature publicly listed or profitable at scale · Growth shipping at scale, Series B or later, or on an IPO track · Early Angel–Series A · Absent a global leader in its track that does not appear in the official record — flagged with ➕ and added in place to keep the track's competitive landscape complete.

• The company sample is drawn from NVIDIA's official channels' individually verifiable public record — keynote and pre-show speakers and their affiliations, partners and customer cases named in official press/blog posts, the official descriptions of the eight Innovation Pavilions, and officially published product-adoption and procurement lists.
• This is a vendor-run conference, and the official narrative is fundamentally marketing material. What "partner" actually means varies enormously — from a multi-billion-dollar procurement commitment to a single technical-adoption statement, and official language does not distinguish between the two. Wherever possible, this mapping restores the concrete action (how much was deployed, what was procured, which layer was adopted) rather than repeating the official "joined hands in partnership"-style phrasing.
• "Absent — added for completeness" means only that a company does not appear in the official record above, not that it sent no one to the event. Note in particular that GTC's absences are strongly structural — direct competitors and upstream capacity holders are almost never on the list — which is a fundamentally different absence logic from a neutral trade show (see Section 12).
• Figures such as attendance, sponsor counts and order-visibility are NVIDIA's or media's own reporting, not independently audited. Financial figures are drawn from public reporting and have not been verified by this mapping.
• Company tiering and stage assessments are this mapping's own analytical judgment and do not constitute investment advice.

The conference ran March 16–19, 2026, with the main venue at the San Jose Convention Center and the keynote held at the 17,000-seat SAP Center. NVIDIA organized the show floor into eight Innovation Pavilions, and this breakdown is itself an industry-chain map — it reflects the directions in which NVIDIA believes its own technology stack currently has an established ecosystem:

One asymmetry in the pavilion breakdown is worth noting: five of the eight pavilions are industry verticals (automotive, healthcare, finance, telecom, quantum), and only one is infrastructure (DSX). That doesn't mean infrastructure isn't important — quite the opposite, infrastructure is the absolute main thread of the keynote. It reflects something else: the infrastructure layer's ecosystem is already highly concentrated among a few dozen large OEMs, cloud vendors and storage vendors, and doesn't need a pavilion to organize it — while the industry verticals are still at a stage that needs education and aggregation.

Before reading this mapping, it helps to grasp the keynote's six main threads first — the position of every company in every layer that follows is, in essence, a response to one of these six threads.

① A step-change reassessment of compute demand. Jensen Huang said compute demand has grown roughly a million-fold over the past several years, and raised the order/revenue visibility for the Blackwell and Vera Rubin product generations across 2025–2027 from a previous \$500B figure to roughly \$1 trillion. Per public reporting, NVIDIA's fiscal 2026 revenue (through January 2026) was approximately \$215.9B, up roughly 65% year over year. This number is the explanatory variable behind every ecosystem participant's behavior at this conference — everyone is making capacity, product and financing decisions against this curve.

② The Vera Rubin platform, and a preview of Feynman. Vera Rubin is a complete platform spanning 7 chips, 5 rack-scale system types and 1 supercomputer, including the new Vera CPU and the BlueField-4 STX storage architecture, explicitly designed for inference and agentic workloads. The next-generation architecture, Feynman, has already been previewed: a new CPU named Rosa (after Rosalind Franklin), paired with the next-generation LPU "LP40," BlueField-5 and CX10, scaling up via Kyber copper cabling and co-packaged optics and scaling out via Spectrum-class optics. The cadence of product naming is itself a commitment-management tool: publishing the roadmap two generations out effectively asks the ecosystem to lock in investment early.

③ The absorption of Groq. NVIDIA acquired Groq's technology through a roughly \$20B asset purchase in December 2025 — its largest deal ever — and launched the Groq 3 LPU at this edition, expected to ship in the third quarter, alongside a rack system built specifically to house the accelerator. Groq founder Jonathan Ross and other core team members joined NVIDIA. This is the single most consequential industry event of this edition: inference-specific architecture was once the most-favored "non-GPU path," and NVIDIA has folded the leading architectural challenger into its own product line with one transaction.

④ OpenClaw and NemoClaw — the fight over agentic computing's operating system. Jensen Huang spent considerable time on developer Peter Steinberger's open-source project OpenClaw, calling it "the most popular open-source project in human history" (it drew more than 100,000 GitHub stars and over 2 million visitors in its first week), and declared that "every company in the world today must have an OpenClaw strategy." NVIDIA immediately followed with the NemoClaw open-source stack and the OpenShell runtime, providing policy enforcement, network guardrails and privacy routing, positioned as "a policy engine for every SaaS company in the world." Elevating a community open-source project into an industry standard, then supplying its enterprise-grade governance layer, is a classic platform play.

⑤ Six open model families and the Nemotron Alliance. Nemotron (language and reasoning), Cosmos (world and vision), Isaac GR00T (general-purpose robotics), Alpamayo (autonomous driving), BioNeMo (biology and chemistry), Earth-2 (weather and climate) — and the formation of the Nemotron Alliance, joining with AI labs worldwide to advance open frontier models. NVIDIA's purpose in building open models is not to compete with frontier labs, but to give every layer of the ecosystem a directly fine-tunable starting point — anchoring development activity to CUDA.

⑥ Standardizing physical AI and the AI factory. On one side, the Vera Rubin DSX AI-factory reference design and the Omniverse DSX digital-twin blueprint — turning "building an AI factory" into a replicable standard part; on the other, pushing AI into the physical world: the Robotaxi platform added automaker partners including BYD, Hyundai, Nissan and Geely and partnered with Uber, while the robotics side partnered with ABB, Universal Robots and KUKA. The keynote closed with Disney's Frozen character Olaf walking live on stage, with NVIDIA stressing that every demo was real-time simulation rather than pre-rendered.

What's distinctive about this layer: there is only one lead actor, and everyone else is supporting cast. GTC's L0 layer has no neutral display of architectural competition — direct competitors don't attend, so this square shows NVIDIA's own product line, plus the semiconductor suppliers that don't directly conflict with it.

The creation of the DSX pavilion is this edition's clearest structural change. Once per-rack power density rises to the point of needing dedicated power generation, distribution and liquid cooling, building a data center stops being an IT procurement problem and becomes an energy and engineering problem. Most companies at this layer come from heavy-industry and energy backgrounds, with almost no overlap with traditional AI-conference exhibitors.

This layer is the densest and most complete in the GTC roster — nearly every global server OEM and Taiwanese ODM is here. The reason is direct: their product-definition authority is largely set by NVIDIA's reference designs, so the keynote's cadence is their product cadence.

NVIDIA disclosed this edition: NVIDIA Cloud Partners (NCP) have cumulatively deployed more than 1 million GPUs and roughly 1.7 gigawatts of AI capacity, double the prior GTC's 400,000 GPUs / 550 megawatts. The structure of this layer is clear as a result: the three hyperscalers are the dominant force, followed by a group of regional and sovereignty-oriented NCPs.

This layer got far more space this edition than in the past, because the launch of the BlueField-4 STX storage architecture and the AI Data Platform reference design elevated storage from a supporting component to a first-class citizen of the reference architecture. A second thread is the broad adoption of the cuDF and cuVS acceleration libraries by mainstream data engines.

Jensen Huang hosted a 90-minute, two-round back-to-back open-model roundtable this edition, bringing together 11 ecosystem leaders. His framing: "proprietary versus open source isn't the real question — it's proprietary and open source." Nearly every company at this layer entered the official record through this roundtable and the pre-show — they are independent forces on the model side, not NVIDIA's suppliers.

This edition's most drastically changed layer. NVIDIA explicitly positioned "agentic computing" as the next operating-system-level battleground, and built an enterprise-grade governance layer around the community open-source project OpenClaw. Companies at this layer fall into two groups: tool vendors that give agents capability, and enterprise-software vendors being reshaped by agents.

Jensen Huang said he "can't think of a robotics company that isn't working with NVIDIA." This layer's actual structure: NVIDIA supplies simulation (Isaac Sim / Isaac Lab / Omniverse / Newton), foundation models (GR00T) and edge compute (Jetson / IGX Thor); robotics companies supply the embodiment and the setting.

The pre-show's host lineup is itself a capital map — hosted by three investors rather than NVIDIA executives, an arrangement worth noting: it hands the conference's opening voice to the buy side.

The table below aggregates every Absent entry in this mapping. GTC's absence logic is strongly structural, and fundamentally different from a neutral trade show — reading it correctly requires separating four distinct causes:

• ① Direct competitors — AMD, Broadcom, Cerebras, Tesla, Waymo, Physical Intelligence. They compete with NVIDIA for the same budget at the same layer, and could never appear at each other's launch events. This means GTC cannot be used to assess the competitive architecture landscape.
• ② Upstream capacity holders — TSMC, SK hynix, Samsung. They set the physical ceiling on the entire roadmap, yet never appear at their customers' launch events.
• ③ Customers' in-house silicon — Google TPU, AWS Trainium, Microsoft Maia. All three hyperscalers are present in force and invested heavily, yet their own accelerators are entirely absent from GTC's own narrative. This is this year's most subtle structural silence.
• ④ Geopolitics and export controls — Huawei Ascend, Cambricon, Moore Threads, MetaX, Unitree and other Chinese compute and robotics players. Worth noting, though: Chinese open-weight models (Qwen, DeepSeek, Kimi) do appear on the officially supported model list — the models made it into the ecosystem, the companies did not.

① The main competitive battleground is shifting from chips to the agent runtime and governance layer. NemoClaw and OpenShell's positioning — "a policy engine for every SaaS company in the world" — signals that NVIDIA is trying to define the default compliance architecture for enterprises deploying autonomous agents. If a standard forms at this layer, its lock-in effect will be stronger than CUDA's: CUDA binds developer habits, while a policy engine binds a company's audit and compliance systems — an order of magnitude higher switching cost. This thread is also the most uncertain — it puts NVIDIA in direct conflict of interest with the three hyperscalers' own agent platforms, and those three are precisely NVIDIA's largest customers.

② The absorption of Groq has rewritten the competitive map for inference chips. A roughly \$20B asset acquisition bought inference-specific architecture outright, and within three months NVIDIA launched the Groq 3 LPU and a matching rack system, while writing LP40 into the next-generation Feynman roadmap. Inference-specific architecture was once the furthest along commercially among "non-GPU paths" — now it's an NVIDIA product line. Two trackable inferences follow: the remaining non-GPU architecture companies (Cerebras, SambaNova, Tenstorrent, Etched) become strategically scarcer; and cloud vendors' incentive to build in-house ASICs strengthens in parallel — the market has one fewer independent, purchasable alternative.

③ The AI factory's bottleneck has already shifted to power and construction timelines. The creation of the DSX pavilion, Mitsubishi Heavy Industries showcasing on-site power generation and distribution, and Caterpillar and EPRI both appearing on power topics — these signals all point to the same thing: the constraint has moved from "can you buy the chips" to "can you get power on time." Data from the NCP side corroborates the scale: more than 1 million GPUs cumulatively deployed, roughly 1.7 gigawatts of capacity, doubling in a year. The trackable physical quantities are per-rack power density, the time from breaking ground to power-on, and the choice of cooling scheme — these three numbers are set by heavy-industry companies, yet they directly determine whether every compute plan upstream can be fulfilled.

① This is a vendor-run conference, and the official narrative is marketing material. The word "partner" covers wildly different actual commitments — from AWS's procurement commitment to deploy more than a million GPUs, to a company merely stating a technical adoption like "trained robots using Isaac Sim" — and official language does not distinguish between the two. The first question to ask when reading any partnership should be: is money actually moving in this relationship, and in which direction?

② The \$1 trillion figure needs to be unpacked. This number is framed as order/revenue visibility for the Blackwell and Vera Rubin product generations across 2025–2027, double the prior \$500B figure. But "visibility" is neither an in-hand order nor confirmed revenue — how much of it comes from a handful of hyperscalers versus emerging neoclouds, what those customers' financing structures look like, and whether any circular-transaction component exists, are all outside the disclosure. This is the single number in today's AI infrastructure narrative most in need of independent verification.

③ Physical AI's narrative intensity far outpaces its commercial maturity. The official framing at the humanoid-robotics layer runs mostly to "trained using Isaac Sim" and "demonstrated dexterous manipulation," while the automaker side shows verifiable platform adoption and integration into Uber's network. There is an order-of-magnitude gap between "adopted a given simulation tool" and "how many units shipped, in what setting, with what repeat-purchase rate", yet the official narrative uses the same verbs for both — readers have to separate them themselves.

• Five analog and embedded-semiconductor giants collectively connected into Isaac Sim (ADI, Infineon, NXP, ST, TI) — physical AI's deployment in the field depends on actuator and sensing chips, not on GPUs. This group has higher certainty than humanoid platforms, and is almost unaffected by which platform maker wins or loses.
• AI storage has become a first-class citizen of the reference architecture — the launch of BlueField-4 STX and the AI Data Platform pushed VAST Data, WEKA, DDN, Hammerspace and Cloudian from supporting components into a bargaining position. As cluster scale grows, data-pipeline throughput is the real bottleneck.
• The most solid evidence on the enterprise side all points to cost, not capability — Snap's 76% cut in data-processing cost, Nestlé's 83% cost reduction in its proof of concept, Heidi Health's 64% cut in operating expense. In an environment of tightening enterprise AI budgets, cost-side evidence drives real procurement more than capability-side demos do — and it's also the more reliable indicator of which vendors will get renewed.

• Whether OpenShell is adopted by cloud vendors as the default agent-governance layer, or whether the three hyperscalers each launch competing policy engines instead — this will determine who owns the standard at the L6 layer;
• Whether the Feynman architecture's timeline (Rosa CPU + LP40) holds, and the actual shipment volume and customer mix for the Groq 3 LPU — this is the first window for verifying whether the "absorb Groq" deal is paying off;
• Whether NCP deployment volume keeps doubling (currently more than 1 million GPUs, 1.7 gigawatts), and what share of the new capacity comes from sovereign AI and regional operators — this ratio determines whether the incremental source of compute demand is genuinely spreading beyond the hyperscalers.

The next GTC is set for March 15–18, 2027, again in San Jose.

How to use this: This mapping is best suited to two uses — locating suppliers' and customers' relative positions along the industry chain, and identifying which links' constraints are shifting (such as the move from chip supply to power and construction). It is not suited to assessing the competitive architecture landscape, since direct competitors are structurally absent; nor is it suited to judging technical leadership, which requires going back to papers, open-source repositories and third-party benchmarks. Any judgment on a specific name should still be grounded in revenue structure, customer concentration, gross margin and cash flow.