Bloomberg Tech

Gary Dickerson on why AI comes down to tokens per second per watt

Gary Dickerson· CEO of Applied Materials at Applied Materials
·~8 min·English·Bloomberg
GPUAI InfrastructureInferenceBusiness Strategy
TL;DR

Applied Materials' CEO reframes the AI race from chip design to materials physics: the real constraint is energy-efficient compute (tokens per second per watt), built from the transistor and wiring materials his company makes for over 60% of AI chips.

01Core Mental Model

The Race Is Tokens Per Second Per Watt

Dickerson collapses the AI contest into one metric, <strong>tokens per second per watt</strong>, set by the atomic-scale materials in the transistor and the wiring.

There's a technical term on tokens per second per watt. How fast you can process the data, those are innovations in the transistor, and the materials that enable that are created by applied materials.

— Gary Dickerson, Bloomberg Tech
Key Insight
Most AI coverage watches the chip designers and the fab. Dickerson points one layer deeper: the number that actually caps AI, energy per token, is decided at the atomic scale of the materials, where his company does the work the chip's logo gets credit for.
02Positioning

Over 60% of AI-Chip Materials

Because <strong>over 60% of the materials in an AI chip</strong> come from Applied, it sits upstream of the compute shortage, invisible to buyers but load-bearing for builders.

Over 60% of the materials that are in these AI chips are created by Applied Materials.

— Gary Dickerson, Bloomberg Tech
Key Insight
The 60% figure is a quiet claim to pricing power. If more than 60% of what makes an AI chip work ships from a single supplier, Applied sits above the whole compute shortage, upstream of the designers the market actually prices.
03Advanced Packaging

The Interconnect Is the New Bottleneck

Energy efficiency now turns on pushing data across <strong>thousands of miles of wiring</strong> inside a package with almost no resistance, not on the transistor alone.

In some of these advanced packages, you can have thousands of miles of wiring, and you're moving that data in incredibly high speeds with almost no resistance. That's what determines energy efficiency and power consumption.

— Gary Dickerson, Bloomberg Tech
Key Insight
For decades, progress meant a smaller transistor. Dickerson is describing a different frontier: moving data between stacked dies with almost no loss. The phrase thousands of miles of wiring says the hard problem is now the interconnect, not just the switch.
04AI on AI

AI Building the Machines That Build AI

Applied runs AI across its own R&D and operations, claiming <strong>50% faster on key technologies and 40 times on some tasks</strong>.

In some cases, 50% faster with key technologies that we're innovating and developing and applied materials. In some cases, we're improving task 40 times.

— Gary Dickerson, Bloomberg Tech
Key Insight
The striking part is where the gains land: not in a new product line but inside Applied's own R&D, operations, and supply chain. A hardware supplier using AI to compress its own cost base is the same flywheel its customers are buying, turned back on the toolmaker itself.
05The Forecast Question

Demand Too Fast to Forecast

Dickerson <strong>refused a specific forecast</strong> because demand is changing too fast to pin a point on, yet still committed to doubling output in two years.

we didn't give a specific forecast because it's changing so quickly.

— Gary Dickerson, Bloomberg Tech
Key Insight
Refusing a number is itself a signal. The stock fell on the missing guidance, which reads as the market wanting caution; Dickerson framed the gap the other way, as a demand curve bending too fast to quote, while still committing to double output.
06Bubble or Demand?

A Bet on Demand, Not Valuations

Asked about the casino charge against AI investors, Dickerson <strong>sidestepped valuations and pointed at unbuilt demand</strong> from agentic and physical AI.

I see so many extremely valuable applications that are in the pipeline that are going to consume more compute demand.

— Gary Dickerson, Bloomberg Tech
Key Insight
He does not argue with the casino charge, he reroutes around it. By pointing at agentic and physical AI as demand not yet built, he separates frothy valuations, which he will not defend, from compute consumption, which he says is real and growing. The bet is on usage, not multiples.