Blackwell scarcity keeps A100 relevant

Nvidia’s newest Blackwell systems are selling into a market where demand still exceeds supply, and that is helping keep the older A100 relevant well beyond a normal chip cycle. Nvidia said on May 20 that fiscal first-quarter revenue rose to $81.6 billion and data center revenue reached $75.2 billion, with Blackwell demand driving results. Separate reports published on May 24 said hyperscalers are taking available Blackwell inventory almost immediately, while A100s are still being used across enterprise AI and academic inference workloads. That split matters because it shows two markets operating at once. Nvidia chief executive Jensen Huang has described Grace Blackwell as delivering lower cost per token for inference, but customers that cannot get enough current-generation systems are still deploying older hardware where the economics work. ### Why are A100s still showing up in 2026 workloads? (investor.nvidia.com) The A100 remains useful because many inference jobs do not require the newest chip if operators can control memory use, batch efficiently and keep utilization high. The May 24 24/7 Wall St. report said enterprise AI customers, academic labs and inference users are still relying on A100s even as Blackwell systems are prioritized by the largest cloud buyers. (nvidianews.nvidia.com) Nvidia made a similar point earlier this year when it said A100 GPUs were still running at full utilization years after launch. That suggests older Ampere systems are not simply stranded assets; they are still part of active capacity for customers whose workloads fit within their limits. ### Why can’t Nvidia just make enough Blackwell chips? (247wallst.com) TSMC remains the key constraint in the supply chain, according to Michael Parekh, a former Goldman Sachs executive interviewed by Stocktwits on May 24. Parekh said Huang has been flying to Taiwan “almost every month” to ask for more production capacity and described TSMC as the central bottleneck in Nvidia’s long-term AI growth. (stocktwits.com) Nvidia’s own filings point to the same imbalance from the demand side. The company said on May 20 that Blackwell helped produce the fastest product ramp in its history, and Huang has said computing demand is growing exponentially as customers build agentic AI systems. ### Why does that make software efficiency more important? Blackwell scarcity raises the value of software that can extract more throughput from whatever GPUs are available. (stocktwits.com) When operators are serving models on A100s or mixed fleets, techniques such as tighter KV-cache management, better schedulers and higher batch efficiency can lower cost per token without waiting for new hardware allocations. This is an inference drawn from the market conditions described in the cited reports and Nvidia’s own emphasis on token economics. (investor.nvidia.com) Huang’s public framing has moved in the same direction. In Nvidia’s fiscal 2026 fourth-quarter release, he said Grace Blackwell with NVLink is “the king of inference today” and delivers an order-of-magnitude lower cost per token, underscoring that the competition is not only about raw performance but also serving economics. ### Who gets the newest chips first? (247wallst.com) Hyperscalers are at the front of the line for Blackwell capacity, according to the May 24 24/7 Wall St. report. That leaves enterprises, universities and smaller model developers more likely to keep running on older Nvidia hardware or to buy whatever capacity becomes available in secondary channels. (nvidianews.nvidia.com) The result is a market where A100s remain commercially relevant even as Nvidia pushes customers toward newer systems. Nvidia’s next disclosed checkpoints on supply and demand will come through future fiscal 2027 filings and earnings updates, while TSMC capacity remains the named variable investors and customers are watching. (investor.nvidia.com) (247wallst.com)