NVIDIA COMPUTEX 2026: five tiers into the agentic era

NVIDIA's COMPUTEX 2026 keynote did something no prior NVIDIA event has done in a single sitting: it pushed every compute tier — personal PC, deskside workstation, standalone CPU, rack-scale server, and physical-AI robot — into the agentic-AI era at once. On May 31, 2026, GTC Taipei was less a product launch than a unified thesis about where computing is heading.

The headline products are RTX Spark, a superchip that runs large models locally on a thin laptop; Vera, NVIDIA's first CPU designed specifically for AI agents; and Vera Rubin, the rack-scale system NVIDIA says has entered full production. Around them sit DGX Station for Windows, the open Cosmos 3 physical-AI model, and the Nemotron 3 model family. Each one is a story; together they are a strategy.

This is a builder's first take, not a press recap. We cover what actually shipped versus what was merely announced, which numbers are independently confirmed versus vendor-stated, and the practical question every engineering team should be asking after a keynote like this: which of my workloads can move on-device, and which still need a rack? Everything below is sourced to NVIDIA's newsroom and corroborating trade press from launch day.

Most COMPUTEX coverage reads as a list of separate product launches. The more useful framing is that NVIDIA moved all five of its compute tiers into the agentic era on the same stage, on the same day. That simultaneity is the actual news: it signals that NVIDIA now treats "an agent" as the unit of work across the entire hardware ladder, from a laptop in your bag to a multi-rack AI factory.

The five tiers map cleanly onto where work happens. RTX Spark is the personal PC. DGX Station for Windows is the deskside workstation. Vera is the standalone CPU for agentic servers. Vera Rubin NVL72 is the rack-scale AI factory engine, and Vera Rubin POD knits five racks into one supercomputer. Cosmos 3 and the Isaac GR00T humanoid reference design extend the same stack out into physical AI.

That "you ask, the PC does the work" line is the through-line of the whole keynote. If the agent is the new unit of compute, then every tier needs to be re-architected around the agent's actual workload pattern — long context, lots of tool calls, branch-heavy reasoning, and unpredictable bursts of retrieval. NVIDIA's argument is that yesterday's silicon, optimized for dense matrix throughput, is not the right shape for that pattern. Whether the silicon delivers is the question the rest of this post pressure-tests.

RTX Spark is the most consequential consumer announcement of the keynote. It is a 70-billion-transistor superchip built on TSMC's 3nm process, pairing a 20-core NVIDIA Grace CPU (developed with MediaTek, codename N1X) with a Blackwell RTX GPU over NVLink-C2C at 600 GB/s. NVIDIA positions it as the first Windows laptop chip to run the full CUDA software stack natively — a meaningful detail, because it means the existing CUDA developer ecosystem transfers to the device without a porting effort.

The specifications that matter for builders are the memory and the local-model claim. RTX Spark carries up to 128 GB of unified LPDDR5X memory at 300 GB/s, 6,144 Blackwell CUDA cores with fifth-generation Tensor Cores, and what NVIDIA states is 1 petaflop of FP4 AI performance. NVIDIA says that combination runs 120B-parameter LLMs locally with a 1-million-token context window — all vendor-stated, but corroborated in part by Tom's Hardware and Microsoft's own Windows Experience blog on launch day.

One caution. Trade press (StorageReview) described RTX Spark as "roughly RTX 5070-class" for gaming, but that comparison is not in NVIDIA's official materials, so treat it as informed speculation rather than a spec. The same goes for the headline laptop dimensions — as thin as 14mm, as light as 3 lbs, single-digit watts up to roughly 80W — which are NVIDIA's targets, not measurements of a shipping product. RTX Spark is announced for fall 2026, with more than 30 laptop and 10-plus desktop configurations planned across the ecosystem; it is not sampling or in production today.

Vera is the most architecturally interesting announcement for anyone building agentic systems. NVIDIA's pitch is that the CPU — not the GPU — has become the bottleneck for agents, because an agent spends enormous time on branch-heavy, memory-sensitive work that GPUs are bad at: tool calls, sandboxed code execution, Python and JavaScript runs, data retrieval, and orchestration. Vera is the first CPU NVIDIA has designed explicitly around that pattern.

The chip carries 88 NVIDIA Olympus cores with up to 1.2 TB/s of LPDDR5X memory bandwidth. NVIDIA reports up to 50% higher IPC than the previous-generation Grace core and roughly 40% lower peak memory latency than x86, plus support for Spatial Multithreading and up to 1.8 TB/s of coherent GPU-to-CPU bandwidth via NVLink-C2C. The named early adopters read like a who's-who of frontier labs and clouds: Anthropic, OpenAI, ByteDance, CoreWeave, Oracle Cloud Infrastructure, Lambda, Cloudflare, Akamai, and Together AI among them.

The headline performance claim — more than 1.8x higher agentic sandbox performance than x86 CPUs across code compilation, code analysis, and Python execution — is where editorial discipline matters. NVIDIA does not identify the x86 SKU, the configuration, or the test harness behind that number, and there is no independent replication. We present it below decomposed by workload type so builders can reason about it, with every row explicitly marked as NVIDIA-reported.

Vera Rubin is the rack-scale story, and NVIDIA's framing is that it entered full production as of the keynote, with shipments beginning fall 2026. Both the Vera CPU and the Rubin GPU are built on TSMC's 3nm process — a generation ahead of the N4 used for Grace Blackwell. The flagship configuration, Vera Rubin NVL72, packs 36 Vera CPUs and 72 Rubin GPUs unified by NVLink Switch 6, delivering a stated 3.6 exaFLOPS of NVFP4 inference, 75 TB of fast memory, and 260 TB/s of scale-up bandwidth in a single rack.

Above the rack sits Vera Rubin POD: five purpose-built racks operating as one integrated AI supercomputer, unifying Vera Rubin NVL72, the Vera CPU, Groq 3 LPX, BlueField-4 storage, and Spectrum-6 Ethernet. Jensen Huang described the supply chain behind it as "twice as large as Grace Blackwell," involving 150 ecosystem partners in Taiwan and more than 350 factories across 30 countries. Cloud availability is slated for the second half of 2026 at providers including CoreWeave, Lambda, Oracle Cloud Infrastructure, Microsoft Azure, and IBM Cloud.

Every COMPUTEX recap we have seen is siloed by product. The table below is our attempt to put all the compute tiers side by side so a builder can read across them in one pass and see which tier fits a given inference budget. Figures are drawn from NVIDIA's newsroom releases, Tom's Hardware, and VideoCardz's NVL72 deep-dive; performance numbers marked as such are vendor-stated.

Reading across the matrix, the pattern NVIDIA wants you to see is a continuum rather than a catalog. The same agentic workload — a long-context reasoning loop with heavy tool use — can run on a laptop for privacy, a deskside box for a single power user, or a rack for production scale, with the open-model lineup and CUDA stack carrying across all of them. The practical decision is no longer "cloud or local" as a binary; it is choosing the tier that matches each workload's privacy, latency, and cost profile.

Hardware needs models to run, and NVIDIA used the keynote to extend two open-model lines. Cosmos 3 is billed as the world's first open physical-AI omnimodel, using a mixture-of-transformers architecture that pairs a reasoning transformer with an expert generation transformer. It natively handles text, images, video, ambient sound, and action trajectories in a single model, and ships in two sizes available immediately — Super at 32B parameters and Nano at 8B — via build.nvidia.com and Hugging Face, with an Edge size to follow.

NVIDIA reports that, among open models, Cosmos 3 ranks first on seven or more robotics benchmarks, including Physics-IQ, R-Bench, RoboArena, and VANTAGE-Bench. Keep the qualifier: this is a claim about ranking among open models, not against closed frontier models, and the benchmark results are vendor-reported with independent verification still pending. NVIDIA also says Cosmos 3 was trained on roughly 20 trillion multimodal tokens — itself a vendor-stated figure.

A keynote this dense is also a wall of benchmarks, and most of them share a property worth naming plainly: they are NVIDIA numbers, run on NVIDIA hardware, evaluated by NVIDIA. That does not make them wrong. It makes them a strong vendor signal rather than a verified design spec — and the difference matters when you are committing a budget. The chart below sorts the keynote's biggest performance claims by how confidently a builder should lean on them today.

The interpretive point is not skepticism for its own sake. NVIDIA has a long track record of shipping silicon that broadly delivers on its architectural promises, and the configuration-level facts — core counts, memory capacities, process nodes — are the kind of thing independent outlets like Tom's Hardware and VideoCardz confirmed on launch day. It is the comparative multipliers — 1.8x, 10x, 300+ tokens/second — that should carry an asterisk until a third party runs the same workload on the same hardware. Build your plan on the confirmed specs; treat the multipliers as upside.

None of this hardware ships before fall 2026, so the right move now is not procurement — it is workload classification. The single most valuable exercise after this keynote is to audit your agentic workloads and sort each one by where it should run: on-device for privacy, deskside for a power user, or rack-scale for production throughput. The decision matrix below is how we frame that for clients.

For most agencies and engineering teams, the honest near-term action is to run your own evaluations on the open models that are already downloadable, and to build a cost-and-privacy model for each agentic workload so you are ready to map them onto tiers the moment hardware ships. If you are weighing on-device versus rack-scale deployment for a specific agentic pipeline, our AI and digital transformation engagements start with exactly this kind of workload audit. For teams further along, our CRM automation work is a frequent first home for privacy-bound, on-device agents.

This keynote is best read alongside the longer NVIDIA arc. The production ramp announced here is the operational follow-through on NVIDIA's trillion-dollar AI infrastructure pipeline from the earlier GTC keynote, and the Nemotron 3 and OpenShell announcements extend the enterprise agentic-AI platform unveiled at GTC 2026. Nemotron 3 Ultra is the direct successor to Nemotron 3 Super, NVIDIA's open coding model.