Scuttlebutt

Scuttlebutt

Chip compute grew 1M-fold in 25 yrs, memory bandwidth just 40-fold, Fractile CEO finds

11 arguments: why AI labs build their own chips, why land and power are now the bottleneck, and why he'd rather serve inference than run a neocloud

Scuttlebutt's avatar
Scuttlebutt
Nov 15, 2026
∙ Paid

20VC · Nov 14, 2026 · 1h 15m · 64m saved

With Walter Goodwin, Founder and CEO of Fractile

Chip compute grew 1M-fold in 25 yrs, memory bandwidth just 40-fold, Fractile CEO finds

Walter Goodwin, founder and CEO of Fractile, told Harry Stebbings on 20VC that chip compute has grown about a millionfold in 25 years while memory bandwidth has grown about 40-fold. That gap, he said, is what limits how fast AI models produce answers.

Most new chip entrants rely on Broadcom for physical design and manufacturing. Fractile keeps that work in-house and builds its memory directly on top of the processor.

"What we've been working on at Fractile over the past couple of years is stacking that memory on top of the logic — rather than have this limited beachfront, everybody gets to have a seaside property." — Walter Goodwin

Walter has silicon back from the foundry and has just closed a $720M Series C led by Thrive, Founders Fund, Lightspeed and Redpoint, which the show notes say values Fractile at about $7.5B.

The full interview is covered here so you can skip it. 75 minutes of audio, 11 minutes of reading.

Here are the 11 arguments that matter.


This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.


Key Takeaways

  • Chip compute grew about a millionfold in 25 years, memory bandwidth about 40-fold

  • Labs and hyperscalers build their own chips to avoid being squeezed by chip vendors' margins

  • Land and power have replaced GPUs and the supply chain as the binding constraint on AI data centers

  • Walter expects far more spending on running models in 10 years than today

  • Software-only AI companies face a coming wave of chip variety that could compress their value

  • Walter would rather be an inference specialist than a neocloud, and expects multi-hundred-billion-dollar winners there

  • Nvidia trades at a lower forward multiple than Cisco after the dot-com crash, Walter said

1. A Chip Company, Not a Lab

Walter describes Fractile as a chip company that earns its money selling chips and racks of compute. It keeps its own machine-learning researchers and systems engineers so it does not misjudge where AI models are heading.

Fractile will not sell its own AI models, so it never competes with the labs that buy its chips

"We never intend to compete with our customers. We will never monetize internal models — that's how we've chosen to build this company. We've chosen to compete in the chip layer." — Walter Goodwin

Walter reads the wave of in-house chip projects as a fight over margins. OpenAI's "Jalapeno" chip, Meta's MTIA and silicon efforts at DeepSeek and Anthropic all give labs and cloud giants a counterweight, so incumbent chip vendors' margins do not squeeze them "all the way down."

2. Frontier Models Keep Margins

User's avatar

Continue reading this post for free, courtesy of Scuttlebutt.

Or purchase a paid subscription.
© 2026 The Golden Age LLC · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.