Blackwell to dominate 2026 GPUs
Analysts now expect Nvidia’s Blackwell chips to account for more than 70% of the company's high‑end AI GPU shipments in 2026, a shift driven partly by delays to the next generation. Supply tightness is already prompting Nvidia to offer rack-scale scheduling software and pushing startups to sell forecasting tools so CFOs can model future GPU costs. The result is that compute scarcity is becoming both a technical and financial planning issue for enterprises rolling out AI at scale. (cloudnews.tech) (blockchain.news) (siliconangle.com)
Nvidia’s newest problem is not selling artificial intelligence chips. It is helping customers decide which jobs get scarce chips first, because analysts now think Blackwell will make up 71% of Nvidia’s high-end graphics processing unit shipments in 2026. (trendforce.com) That forecast changed because the next family, Rubin, now faces delay risk from supply-chain tuning and high-bandwidth memory 4 certification, while older Hopper chips are losing share. TrendForce said Blackwell’s 2026 mix rose from an earlier 61% estimate to 71%. (trendforce.com) A graphics processing unit is the part of a server that does the heavy lifting for training and running artificial intelligence models. When companies cannot get enough of them, they do not just wait longer for hardware; they start rationing computing time inside the data center. (nvidia.com) Blackwell is Nvidia’s current architecture for those jobs, and the company is packaging it into rack-scale systems like the GB200 NVL72 and GB300 NVL72. Nvidia says each of those systems links 72 graphics processing units across 18 compute trays with NVLink so they can act more like one giant machine. (developer.nvidia.com) Once a rack holds 72 chips, the bottleneck shifts from buying hardware to assigning work. Nvidia’s Mission Control software maps which chips are tightly connected inside the rack and feeds that information into schedulers such as Slurm and Nvidia Run:ai so jobs land on the right part of the system. (developer.nvidia.com) Nvidia has been pushing that software angle for more than a year. In March 2025, the company said Mission Control could raise graphics processing unit utilization by 5 times and improve training and inference efficiency across Blackwell infrastructure. (blogs.nvidia.com) The shortage is now spilling into finance teams. Silicon Data, a startup covered on April 8, says companies need forward-looking price data for renting graphics processing units because chief financial officers cannot plan large artificial intelligence budgets around opaque spot prices. (siliconangle.com) Silicon Data’s pitch is simple: treat compute like a commodity market instead of a mystery bill. Its service tracks rental pricing, benchmarks, and market data so buyers can estimate what a future training run may cost before they sign cloud or colocation contracts. (silicondata.com) That is a new kind of constraint for enterprise artificial intelligence. A few years ago, the question was whether a company had a model worth deploying; in 2026, the question is also whether it can lock in enough Blackwell capacity, schedule it efficiently, and explain the bill months in advance. (trendforce.com)
