The speaker below sells the thing he is arguing for: Walter Goodwin is founder and CEO of Fractile, which builds inference chips designed around memory bandwidth.His figures are spoken on a podcast with no source given, and the quotations come from a transcription of the recording rather than a published transcript.
Walter Goodwin, founder and CEO of a memory-bandwidth chip company, argues the AI industry is measuring progress on the wrong axis. Here is what he claims, what he does not prove, and why the structure of the argument matters anyway.
Two numbers carry the whole argument, and both are his. Goodwin says FLOPs have scaled about a million-fold over the last 20 years, while memory bandwidth has risen about 40 times across the same period. He offers no source for either figure; they are spoken in conversation, not cited.
Walter Goodwin is the founder and chief executive of Fractile, a company building inference chips designed around memory bandwidth rather than raw compute. He appeared on the No Priorspodcast, hosted by Sarah Guo, in an episode titled The Future of Frontier Model Architectures. He is not a neutral analyst. His central argument, if accepted, describes the exact market that Fractile exists to serve. The reader should hold that in mind for everything that follows.
On the podcast, Goodwin claimed that compute — measured in FLOPs, or floating-point operations per second — has scaled by roughly a million-fold over the last twenty years, while memory bandwidth has grown by only about forty times over the same period. He offered no source for either figure. They are spoken numbers in a conversation, not citations from a published study. This publication verified the episode title, the speaker, the company, and the host. It could not obtain an independent transcript; the quotations here come from a transcription of the recording.
The gap between those two figures — a million-fold versus forty-fold — is the foundation of his case. He argues that the AI industry’s standard vocabulary for measuring progress, denominated almost entirely in FLOPs, reflects that historical emphasis rather than any law of nature.
The key insight: The interesting part of Goodwin’s argument is not the specific numbers, which are unverified. It is the structural claim underneath them: that measuring AI progress in FLOPs is a choice of denominator, not a physical constant. Whether that claim holds up to scrutiny is a separate question.
The gap between those two numbers is the whole of his argument. It is also a claim by someone whose company is built on it being true.
The Structural Read
The argument Goodwin is making has a specific logical shape. He is not simply saying that memory bandwidth matters. He is claiming that compute and memory bandwidth are, to some meaningful degree, substitutable inputs for reaching a given level of model capability.
In his words, as transcribed from the recording: “as we elevate this other quality, this property of memory bandwidth, which hasn’t really been scaled so much on chips recently… So as you scale that frontier, you get to conserve more of the other thing — we get to reduce the number of FLOPs that we’re using on these models to get to a certain level of intelligence.” He calls that a “multiplying factor on global throughput.”
He also states that “the scaling laws traditionally, we think of them as like FLOP scaling laws” and that “if you look at just like the known landscape of ideas today, we already see some scaling laws for bandwidth.” That second claim is his reading of existing research, offered without naming any specific paper or author. It must be understood as his interpretation, not a settled finding.
Goodwin provides no benchmark, no measurement, and no third-party result for the substitution claim. A reader who walks away believing bandwidth can replace FLOPs as a demonstrated, verified fact has been misled by the framing. What Goodwin argues is that the trade-off is possible. That is a meaningfully different statement.
FDE Framework — Enabler Play
Redefining the Denominator Is an Enabler Strategy
In the FDE framework — Founders, Distributors, Enablers — infrastructure companies win by making a constraint disappear. Fractile is positioning as an Enabler that replaces one scarce input (FLOPs) with a different input (bandwidth) that it happens to specialize in. The commercial incentive to frame the argument this way is obvious. That does not automatically make the argument wrong. It does make independent verification essential before acting on it.
Two Claims, Both 25x, Entirely Different Things
There are two separate figures in the available material that both happen to be twenty-five times something. They must not be read as the same claim or as one confirming the other.
On the podcast, Goodwin refers to Fractile’s chip design as having “like 25 times more bandwidth per chip than an HBM-based chip.” HBM — high bandwidth memory — is the memory standard used on current AI accelerators. That figure is a claim about memory bandwidth per chip, and it describes his own company’s design.
Separately, Fractile’s own website advertises running “the most advanced models up to 25x faster and at 1/10th the cost.” This publication read the Fractile site directly to establish that distinction. That figure is a claim about inference speed and operating cost. It is a different quantity entirely.
Neither figure is accompanied by a published benchmark, a shipped product, or any third-party verification in the material reviewed here. There is no known volume production of a Fractile chip, no independent customer confirmation, and no external audit of either the bandwidth or the speed and cost figures.
Walter Goodwin — No Priors Podcast, transcribed
“as we elevate this other quality, this property of memory bandwidth, which hasn’t really been scaled so much on chips recently… So as you scale that frontier, you get to conserve more of the other thing — we get to reduce the number of FLOPs that we’re using on these models to get to a certain level of intelligence.”
Three Implications
For the AI Infrastructure Narrative
If the substitution argument has any validity at all, the current consensus that more FLOPs equals more capability is a convention rather than a ceiling. The industry’s entire benchmarking vocabulary — training compute, inference compute, scaling curves — assumes FLOPs as the primary input. A credible challenge to that assumption would matter structurally, regardless of who is making it.
For Chip Company Strategy
Fractile is not the only company betting on bandwidth as a bottleneck. Whether other companies share that reading is not something this episode establishes. Goodwin is making that case explicitly and publicly. Nothing here establishes how widely that framing is held.
For How You Read Founder Podcasts
A founder describing the market in terms that precisely justify his own product is a normal thing. It is not automatically wrong, and it is not automatically right. The Goodwin appearance is a clean example of an argument that has structural interest but zero verification. The correct response is not to dismiss it and not to repeat it as fact. It is to locate the independently published research he gestures at and read that instead.
Business Engineer Framework
The FDE Framework — Where Fractile Is Placing Its Bet
In the FDE model, Enablers win by removing a constraint that Founders and Distributors cannot remove themselves. Fractile is claiming the bandwidth constraint is real, large, and addressable by its chip design. That is a classic Enabler thesis. The Map of AI shows exactly where that layer sits in the current stack — and which companies are already occupying it.
Walter Goodwin’s argument — that FLOPs are the wrong axis and that memory bandwidth is a substitutable input for reaching a given level of AI capability — is structurally interesting precisely because it challenges a convention rather than a law. But it is an argument made on a podcast by a founder whose company is built on that argument being correct, supported by figures he did not source and a trade-off he did not benchmark. Hold the structure. Suspend the conclusion. Wait for the paper.
Walter Goodwin is the founder and chief executive of Fractile, a company building inference chips designed around memory bandwidth. The argument reported above describes the market his own product is built to serve, and that interest should be read alongside every figure in it.This publication verified that the No Priors episode exists with the title, speaker, company and host described. It could not retrieve an independent transcript, so the quotations come from a transcription of the recording, with timestamps at roughly 29 and 31 minutes.
The claim that FLOPs have scaled about a million-fold in 20 years while memory bandwidth has risen about 40 times is Goodwin’s, spoken in conversation with no source given. Neither figure has been independently confirmed here.Two separate claims in this story both involve the number 25 and are not the same claim. On the podcast Goodwin refers to 25 times more bandwidth per chip than an HBM-based chip.
Fractile’s website separately advertises running models up to 25 times faster at one tenth the cost, which is a speed and cost claim. Neither corroborates the other.Nothing above establishes that memory bandwidth can be substituted for FLOPs, that scaling laws for bandwidth exist, that any Fractile chip exists in volume, or that any benchmark supports any figure here. Nothing above predicts anything and nothing here is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
