Let's stack the numbers

Alphabet shares rose 1.51% Monday on a report from The Information that the company is building a new server chip, "Frozen v2," which embeds Gemini's architecture directly into the silicon. Google's own engineers say that's worth six to ten times more tokens per watt than current TPUs. The same day, AMD began shipping Helios, its first rack-scale AI system, adding Microsoft to a customer list that already includes Meta, OpenAI and Oracle. AMD's data center revenue rose 57% year over year in the first quarter, and the company expects most of that future growth, from 2027 on, to come from Helios.[1], [2]

And TSMC: CFO Wendell Huang said the company is putting another $100 billion into Arizona to meet a "multi-year demand mega trend" — bringing the total investment pipeline to $265 billion. 2-nanometer production is becoming the company's newest revenue engine as it enters the third quarter. These aren't bubble numbers; this is capacity with contracts and shipping schedules attached.[3]

There is no cloud, there's a factory

The common thread across all three: none of them claims the model got "smarter." Google's chip aims to run the existing architecture with less computation. AMD is selling a system that can actually ship against Nvidia's Grace Blackwell and Vera Rubin. TSMC is building capacity that's agnostic to who's producing what. We'll keep reading the bill in megawatts, because that's where the real competition now sits.[1], [2], [3]

One line from Huang stuck with me: building a fab in the US costs four to five times more than in Taiwan. So that $100 billion isn't just capacity — it's also a geopolitical risk premium. I'll keep tracking the Taiwan concentration story; this week's news doesn't reduce that risk, it just prices it.[3]

Last week's model score, this week's factory

Last week I wrote that as the Nasdaq dropped 1%, we were reading the bill by model score, not megawatts — the market swings between watching the model and watching the factory, but they land on the same balance sheet.[4]

My expectation: TSMC's 2-nanometer revenue will show up concretely in the third-quarter report.[3]

AMD's Helios shipments will turn into real numbers by year-end.[2]

Google's chip likely won't reach production before 2028.[1]

All three are saying the same sentence in different words: the constraint is no longer silicon, it's watts. We will go on reading the bill in megawatts.[1], [2], [3]