Jensen Huang took the stage at GTC Taipei on Monday to reaffirm what the AI industry has anticipated since January. NVIDIA’s Vera Rubin platform, a seven-chip architecture that redefines rack-scale supercomputing, is in full production and on track to reach customers beginning in the third quarter of 2026. The timeline, first locked in at CES in January and expanded at GTC in March, now hinges on execution across a global supply chain that spans more than 350 factories in 30 countries. For a market starved of compute, the message was clear: the silicon is real, the racks are building, and shipments are only weeks away.
The Vera Rubin platform is not a single processor but a tightly integrated suite of seven new chips now rolling off production lines. The lineup includes the Vera CPU built on Olympus Arm cores, the Rubin GPU, the NVLink 6 Switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU, the Spectrum-6 Ethernet switch, and the Groq 3 LPU. Together they form the NVL72 rack, a configuration that pairs 72 Rubin GPUs with 36 Vera CPUs and binds them via NVLink 6. NVIDIA positions the NVL72 as a unified AI factory designed to handle pretraining, post-training, test-time scaling, and agentic inference without cobbling together disparate hardware from multiple vendors.
Early performance claims suggest a substantial leap over the prior Grace Blackwell generation. According to NVIDIA, training large mixture-of-experts models now requires roughly one-quarter the GPU count previously needed, while inference throughput per watt jumps by up to a factor of ten. Those figures have circulated since Huang’s January keynote, but the Taipei audience focused on what they mean for economics. Lower cost per token and higher agent throughput translate directly to margins for cloud providers running thousands of racks, and that is precisely the bet NVIDIA is asking its customers to make.
Hardware is already leaving factories. Dell Technologies recently shipped the first production rack of the Vera Rubin NVL72 to CoreWeave, the specialized AI cloud provider. That single system delivers 3.6 exaFLOPS of NVFP4 inference compute. The delivery marks a concrete milestone in a year otherwise defined by roadmaps and press releases. It also signals that partner qualification is moving faster than many supply chain observers expected.
Independent benchmarks of the Vera CPU are also surfacing. In early testing, the chip’s 88 Olympus Arm cores outpaced competing x86 offerings from AMD and Intel in data center workloads. Initial results position the Vera processor as a legitimate orchestration engine for agentic systems rather than a mere companion to the GPU. NVIDIA reportedly hand-delivered the first CPU racks to OpenAI, Anthropic, SpaceX, and Oracle ahead of broader availability, suggesting that major labs are already rewriting software stacks to exploit the CPU’s memory bandwidth and core density.
The production ramp has become a case study in supply-chain concentration, and Taiwan is firmly at the center of it. More than 150 local partners contribute to assembly, testing, and packaging. Social media during the keynote highlighted the sheer scale of the operation.
Vera Rubin is ramping into full production to power Agentic AI Factories Worldwide
— Kawz (@KawzInvests) June 1, 2026
Market observers note that the Rubin generation arrived well ahead of some internal expectations. Reports from earlier this spring described the move into full production as faster than anticipated, giving NVIDIA breathing room before volume shipments accelerate in late 2026. For now, Blackwell continues to satisfy immediate demand, but cloud providers and enterprise customers are already allocating capital for the Rubin transition. The overlap is intentional; NVIDIA typically manages two generations in the channel to avoid the kind of supply vacuum that hampered rivals during previous node shifts.
The shift toward rack-scale systems reflects a broader industry pivot. Individual chips matter less than the orchestrated stack inside a data center. NVIDIA’s Spectrum-X Ethernet Photonics, also now in production, adds high-bandwidth networking fabric that ties these AI factories together. The company has opened its MGX server reference designs to partners, allowing faster customization while keeping the core architecture consistent. NVIDIA has spent years building consumer familiarity with its brand through gaming and streaming devices, but Vera Rubin marks a decisive return to its data center roots.
Agentic AI is also driving new consumer form factors; Samsung and Google recently unveiled smart glasses built around on-device intelligence. Yet the heavy lifting for those services will happen inside Rubin-powered data centers, not on the lenses themselves. That dependency only strengthens NVIDIA’s position as the pick-and-shovel play for a generation of software agents.
As June progresses, attention will turn to qualification timelines and partner announcements ahead of the Q3 initial shipments. If the current production pace holds, the volume ramp projected for Q4 2026 should position Vera Rubin as the dominant training and inference infrastructure heading into 2027. For an industry racing to build agentic AI at planetary scale, the machinery is finally leaving the assembly line.
