Nvidia confirms Vera Rubin roadmap, but doubts persist over Kyber

What happened?

Nvidia CEO Jensen Huang stated on July 15, 2026, that rumors about delays in the Vera Rubin platform are false, asserting that 'Vera Rubin is already in production. Gigantic amounts of production are coming.' However, he did not address the SemiAnalysis report indicating that the Kyber NVL144 system, designed to connect 144 Rubin Ultra GPUs via copper NVLink 7 interconnects, would be delayed until 2028 due to the complexity of its PCB backplane. According to Tom's Hardware, Nvidia had already confirmed Vera Rubin production in January 2026 and sampling in February. Huang's statement reinforces confidence in the base platform but leaves questions about the high-performance variant unanswered. Notably, this is not Nvidia's first technical setback: during the Blackwell launch, packaging issues delayed mass shipments by several months, causing stock volatility. At that time, Huang also opted for optimistic statements without detailing the hurdles, a pattern now repeating.

Why is it important?

Vera Rubin represents Nvidia's next-generation AI infrastructure, succeeding Blackwell. The platform combines Vera CPUs (with 88 custom cores) and Rubin GPUs (with 288 GB of HBM4), and is key to maintaining Nvidia's leadership in the AI accelerator market, valued at over $100 billion by 2027 according to IDC estimates. The Kyber NVL144 system is critical for massive model training applications, where NVLink scalability is essential. If the Kyber delay is confirmed, hyperscale customers (such as Microsoft, Meta, or Google) could see limited capacity to deploy 144-GPU clusters in a single NVLink domain, impacting performance in generative AI workloads and scientific simulations. For example, training models like GPT-5 or Gemini 2 requires low-latency interconnection that only NVLink 7 can provide; without Kyber, hyperscalers would have to resort to 72-GPU configurations or combine racks with slower network interconnects, increasing training time by up to 30%, according to SemiAnalysis analysis. This contrasts with AMD's strategy, whose MI400 platform promises native scalability of 256 GPUs via Infinity Fabric, and startups like Cerebras, which offer monolithic wafer-scale solutions without external interconnection.

What consequences will it have?

In the short term, confirmation of mass production of Vera Rubin NVL72 (72 GPUs) should sustain Nvidia's revenue growth in the coming quarters. In the first fiscal quarter of 2026, Nvidia reported record revenue of $39 billion, driven by Blackwell demand; Vera Rubin could add another $10 billion in the second half. However, uncertainty about Kyber could cool investor enthusiasm, as large-scale rack solutions generate the highest value per unit. According to Bloomberg, Nvidia shares fell 2% after the SemiAnalysis report, but partially recovered with Huang's statements. The potential delay opens a window for competitors like AMD (with its MI400 platform) or startups like Cerebras, seeking to capture customers needing immediate scalability. Nvidia could also accelerate development of alternatives such as the NVL72x2 design (two 72-GPU racks interconnected via optical cables) or silicon photonics (CPO) solutions to close the gap. However, these alternatives involve higher costs and development timelines, which could pressure Nvidia's margins, currently around 70% gross. In the past, Nvidia faced similar challenges with the DGX-2 system (2018), which used a complex backplane and had initial delays but ultimately dominated the AI supercomputing market.

What should readers know?

• Vera Rubin NVL72 is in production: systems with 72 Rubin GPUs are already being manufactured and shipped to select customers. First units will arrive at Microsoft and Meta data centers in Q3 2026.
• Kyber NVL144 faces delays: the copper backplane for 144 GPUs is complex to manufacture, with over 20 PCB layers and high-speed signals requiring nanometer tolerances. This could delay its launch from 2027 to 2028.
• Nvidia has not denied the Kyber delay: its statement of 'roadmap intact' is ambiguous and does not specify dates. A Nvidia spokesperson told Tom's Hardware that 'the roadmap is intact' but avoided answering specific questions about Kyber.
• The AI market remains volatile: any change in Nvidia's roadmap impacts data center investment decisions. Companies like Google have already begun diversifying suppliers, ordering TPU v6 and AMD accelerators to reduce dependence on Nvidia.
• Alternatives in development: Nvidia is exploring dual-rack and optical designs for future generations. The NVL576 system based on CPO (co-packaged optics) could offer 576 GPUs in a single NVLink domain, but availability is estimated for 2029 or later.

'Our roadmap is intact' — Nvidia spokesperson to Tom's Hardware. However, this statement does not clarify whether original timelines are maintained. Speculation: since Huang did not mention Kyber, the delay is likely real, but Nvidia prefers not to confirm it until an alternative solution is ready.