Dylan Patel on semiconductor manufacturing
From Dylan Patel (SemiAnalysis): The Datacenter in 2026: CPUs, RL Environments & Agent-Driven Workloads · · Daytona and SemiAnalysis
“AI is buying all the capacity on 3 nanometer and 2 nanometer in a couple years that people are having to turn to other directions right like one of the like you know people make all these nonsense reasons up why Nvidia acquired Grok part of it was because they want to have really fast inference but part of it is that Grok is manufactured on Samsung right because there's no three nanometer capacity for them at TSMC they need to chip somewhere else. And if you just if AI is as crazy as we believe it is and demand is as crazy as we believe it is it's going to be even crazier next year. And so just make any decent chip and it'll sell right.”
On , Dylan Patel, Founder, CEO, and Chief Analyst at SemiAnalysis, spoke about semiconductor manufacturing during Dylan Patel (SemiAnalysis): The Datacenter in 2026: CPUs, RL Environments & Agent-Driven Workloads on Daytona and SemiAnalysis.
About Dylan Patel
Dylan Patel, founder and CEO of SemiAnalysis, has been speaking at several industry events in early 2026 about AI infrastructure, benchmarking, and market dynamics. At an Aria Networks launch event in April, Patel stated that AI inference demand has grown so rapidly that the rental price of three-year-old H100 GPUs has risen from around $160-170 per hour to over $240 per hour in six months, with no spare capacity available. He also discussed the InferenceX project, which he described as a free and open-source benchmarking effort with over a thousand GPUs donated by companies including OpenAI, Microsoft, and Nvidia. In a March interview at the Daytona Compute Conference, Patel said that hyperscalers like Google, Amazon, and Microsoft were slow to move into AI, creating an opportunity for "NeoClouds" that could skip complex legacy software. He also noted that the entire cloud market had run out of CPUs, with Amazon's CPU server installations tripling year-over-year. In an April interview with Patrick O'Shaughnessy, Patel said his firm's AI token spend had skyrocketed from tens of thousands of dollars annually to $7 million, driven by non-technical staff using AI for coding. He stated that "ideas are cheap and plentiful but execution is very easy," and warned that people who do not use more tokens, generate value from them, and capture that value will "never escape the permanent underclass." Patel also predicted a "large scale protest against Anthropic and AI," citing a Pew survey that he said showed AI is less popular than politicians. In a panel at the Beyond Summit, Patel asserted that vendor benchmark claims are "lies, impossible to achieve," and that "if you're not pissing off people with your benchmark, then you're not testing something useful."