Blackwell set to dominate GPUs

Nvidia’s newest artificial intelligence chips are turning into a one-platform market: TrendForce now says the Blackwell family will make up 71% of Nvidia’s high-end graphics processing unit shipments in 2026, up from an earlier 61% forecast. Hopper falls to 7%, and Rubin drops to 22% as supply problems push buyers back toward the chips they can actually get. (trendforce.com) That concentration matters because a graphics processing unit is the engine inside an artificial intelligence server, and Nvidia sells those engines in families that data-center operators plan around years in advance. If one family becomes the default choice for most big buyers, delays in that one family ripple through cloud budgets, construction schedules, and model launches. (trendforce.com) Blackwell is not just a chip name but a full platform Nvidia unveiled in March 2024 for training and running large artificial intelligence models at scale. Nvidia says Blackwell systems are built around newer memory, faster links between chips, and rack-level designs meant to pack many processors together in one cabinet. (nvidia.com) Nvidia has been steering customers toward those full-rack systems, including Grace Blackwell and Vera Rubin rack designs, because selling the whole cabinet raises the chip count in each order. TrendForce says that push toward integrated rack products is one reason total high-end graphics processing unit shipments are still expected to grow even as the mix shifts heavily toward Blackwell. (trendforce.com) The surprise is on the chip that comes after Blackwell. TrendForce says Rubin’s expected 2026 share fell from 29% to 22%, and The Register reports the downgrade is tied to memory qualification, networking changes, higher power draw, and tougher liquid-cooling demands. (trendforce.com) (theregister.com) The memory issue sits at the center of this. Rubin is designed around High Bandwidth Memory 4, which is a stack of ultra-fast memory chips placed next to the processor like a pantry built into the stove, and TrendForce says validating that new memory is taking longer than expected. (theregister.com) The networking change adds another delay. TrendForce says Rubin systems also need a move to Nvidia’s ConnectX-9 network interface cards, which are the parts that let one server talk to the rest of the data center at very high speed, so a chip launch now depends on memory, networking, and system design lining up at the same time. (theregister.com) Then there is heat. TrendForce says Rubin’s higher overall power consumption means more advanced liquid cooling, and liquid cooling is not a small accessory but plumbing, pumps, manifolds, and facility work that have to be installed before a rack can go live. (theregister.com) That leaves Blackwell in the strongest position simply because it is the platform customers can standardize on now. Nvidia’s own description of Blackwell focuses on “artificial intelligence factories” built for inference, which is the stage where trained models answer questions and generate outputs, and that is exactly where cloud providers are racing to add capacity. (nvidia.com) The risk is that the market gets efficiency and fragility at the same time. If more than seven out of ten of Nvidia’s top-end shipments land on Blackwell, any packaging snag, memory shortage, or rack integration delay hits a much larger share of the industry than it would in a more balanced product cycle. (trendforce.com) For now, the roadmap is not collapsing so much as bunching up. Blackwell is absorbing demand that might have spread across Hopper, Blackwell, and Rubin, and the result is a 2026 market where Nvidia still dominates but more of that dominance sits on one very crowded platform. (trendforce.com)