The hardware secret that gave OpenAI the edge on Anthropic

This article summarizes an episode of Dwarkesh Patel’s video series featuring Dylan Patel, CEO of SemiAnalysis.

Dylan Patel, CEO of SemiAnalysis/ Photo credit: Dwarkesh Patel

Dylan Patel, CEO of SemiAnalysis, warns that the technology industry is underestimating a fundamental physical limit on AI. At a time when companies assume growth only requires more money and software, Patel argues the real limit lies in manufacturing capacity.

If the industry ignores this manufacturing slowdown, traditional strategic plans will no longer work. Companies will need to completely alter how they buy and manage computer hardware.

In his view, surviving this transition requires accepting severe supply constraints. Companies that secure computing power early and maximize their existing equipment may win market dominance that hesitant firms can no longer reach.

Facing the manufacturing bottleneck

Many industry leaders think securing more funding is the best way to expand AI capacity. Patel sees the reality as a physical problem that will halt companies that do not plan for equipment shortages.

He points to a fragile supply chain that simply cannot build server farms fast enough to meet the current demand.

At the center of the challenge is a strict limit on highly specialized manufacturing equipment. To illustrate the scale of the problem, Patel breaks down the exact silicon required to power a single large AI data center:

• 55,000 advanced processor wafers (3-nanometer)
• 6,000 supporting processor wafers (5-nanometer)
• 170,000 memory wafers (DRAM)

Because these wafer layers can only be made by extreme ultraviolet (EUV) light machines, this volume creates a massive factory bottleneck just to supply one major facility.

“You’re at roughly two million EUV passes for a single gigawatt,” Patel notes.

This mathematical reality is a warning not to ignore physical production limits. Since only a single company, ASML, manufactures these specific machines, the entire industry must wait on its production schedule.

The problem with production speed
Software companies often assume hardware capacity will appear when needed. Patel argues this is a bad plan that leads to massive delays because these essential machines operate at a fixed speed.

“An EUV tool can do roughly 75 wafers per hour,” he says. “In the end, you need about three and a half EUV tools to do the two million EUV wafer passes for the gigawatt.”

The new hardware reality

The cost of early hesitation

The spot market penalty