20VC · · 1h 16m
20VC: Taking on NVIDIA: Fractile CEO on the Future of Chips | What Everyone Gets Wrong About Capex Spending | The Winners and Losers in Neo Clouds with Walter Goodwin
With Walter Goodwin
Summary
Walter Goodwin, in this 20VC episode about Fractile and the future of AI chips, argues that inference demand will grow enormously over the next decade, driven by long-running reasoning and agent workloads, and that the AI build-out is justified unlike the 2000-2001 fiber overbuild. He contends that memory bandwidth, not FLOPs, is the binding constraint on token generation, and that Fractile's 3D stacking of SRAM, DRAM and compute aims to remove the tradeoff. He argues that Nvidia is an extraordinary business but that specialist inference chips can carve out space alongside it, and that custom chips built on a shared Broadcom and TSMC back end face a supply crunch in TSMC and HBM capacity. On the neocloud layer he says cloud compute is commoditizing, so providers are vertically integrating, and that FluidStack's move to Google TPUs shows the upside of breaking single-vendor dependence. He is bullish on TSMC and Nvidia's valuation, and says power and land, not just chips, are becoming the bottleneck for compute buildout.
Key takeaways
- Walter Goodwin (bullish on Fluidstack): Fluidstack's move off Nvidia onto Google TPUs was a bold bet that has paid off with an extraordinary year.
- Walter Goodwin (neutral on NVIDIA): Nvidia builds extraordinary chips, but its general-purpose design leaves room for specialist inference chips.
- Walter Goodwin (neutral on NVIDIA): Nvidia's Groq deal answered a fast-inference need, not insurance; SRAM chips can't scale LLM inference.
- Walter Goodwin (bullish on Taiwan Semiconductor Manufacturing Company): TSMC is cheap and underpins world chip production; he would buy the stock every day.
- Walter Goodwin (bullish on NVIDIA): Nvidia could reach a $10 trillion value in five years; it trades cheaply versus Cisco post-crash.
What was said, by company and theme
24 views from 1 speaker. Every quote is checked against the episode's transcript (how).
NVIDIA
Across the shows: 147 speakers in other episodes, 134 bullish and 106 bearish calls, past 30 days
Where everyone stands on NVIDIA
Walter Goodwin · FractileNeutral
Nvidia builds extraordinary chips, but its general-purpose design leaves room for specialist inference chips.
Goodwin calls Nvidia an extraordinary business that builds, by many metrics, the best AI chips and systems, and says it has done brilliant work on data center inference. He argues Nvidia excels at delivering high-quality products at very high volume that are unopinionated general-purpose processors. Specialist inference chips can be built for a particular focus, such as huge memory bandwidth, that trades away some generality. He argues the frontier now makes up a large share of the total market, and specialist workloads are no longer niche. This creates room for specialist chips alongside Nvidia.
“And I think if you look at NVIDIA's business, what they excel at is delivering high quality products at extraordinary volumes that are, in a sense, unopinionated.”at 14:27
Walter Goodwin · FractileNeutral
Nvidia's Groq deal answered a fast-inference need, not insurance; SRAM chips can't scale LLM inference.
Walter says the Groq team is strong and Jensen wants smart people, and that the deal filled a real gap for Nvidia in fast inference when it happened. He does not view it as an insurance policy against losing OpenAI as a customer, although he notes the timing alarm of a large customer signing a rival deal. He thinks that solution has since been superseded: SRAM-based chips lack the capacity, and the cost of speed is too high for long context and agents running for days. They give snappier chatbots but do not unlock the largest prize of frontier intelligence.
“Today, though, we've already seen ourselves, I think, move on from this. You know, you cannot solve LLM inference with these SRAM based chips. There isn't enough capacity.”at 1:00:11
Walter Goodwin · FractileBullish
Nvidia could reach a $10 trillion value in five years; it trades cheaply versus Cisco post-crash.
Walter says Nvidia is a very positive-sum business and believes it trades at a lower forward multiple than Cisco did after the dot-com crash. He attributes this to investors being traumatized by the dot-com bubble and so having priced in exuberance from the outset. He argues the market is not yet appreciating how sustained the build-out will be. Because megawatts are finite, chips must be replaced every few years with better ones, which keeps demand going without new data centers. He flags the comparison as a soundbite.
“It's a very positive sum game. NVIDIA today trades... I believe it trades at a lower forward multiple than Cisco after the crash, you know, after the .com bubble burst.”at 1:09:46
Fluidstack
Walter Goodwin · FractileBullish
Fluidstack's move off Nvidia onto Google TPUs was a bold bet that has paid off with an extraordinary year.
Goodwin says FluidStack was in some ways the first neocloud to decide to step away from Nvidia. It deployed a bunch of Google TPUs. He argues that crossing this Rubicon earned FluidStack the right to start deploying other chips. He characterizes the year it has had since as extraordinary. He uses this as evidence that a neocloud breaking its dependence on one GPU vendor can reap big benefits, even though the bet is risky.
“They were, in some ways, the first Neo Cloud to kind of decide that they were going to step bodily away from NVIDIA, and they deployed a bunch of Google TPUs. And they've had an extraordinary year.”at 12:56
Broadcom
Across the shows: 16 speakers in other episodes, 17 bullish and 6 bearish calls, past 30 days
Where everyone stands on Broadcom
Walter Goodwin · FractileNeutral
Broadcom-built custom AI chips are one shared platform with different names, limiting real differentiation.
Goodwin says many custom AI chips, including TPU, MTIA, Jalapeno and others, are rumored to use Broadcom and the same TSMC nodes and HBM. He cites a Broadcom presentation of a next-gen platform expected in production with six major customers in 2028. Each chip gets a different name, but it is essentially the same platform: the same HBM stack and the same number of chips. He says the back-end partner dictates many fundamental building blocks, so customers can only change the color of the paint. Fractile's in-house back-end work lets it build different building blocks.
“A lot of the sort of fundamental building blocks are dictated by the back end partner because it's a platform technology play. It's a kind of you get to change the color of the paint, basically.”at 30:55
Taiwan Semiconductor Manufacturing Company
Across the shows: 8 speakers in other episodes, 7 bullish and 0 bearish calls, past 30 days
Where everyone stands on Taiwan Semiconductor Manufacturing Company
Walter Goodwin · FractileBullish
TSMC is cheap and underpins world chip production; he would buy the stock every day.
Walter says that virtually everything in the AI chip business traces back to TSMC, since the majority of frontier chips are built there. He describes it as an extraordinary business with an extraordinary team, and says it is cheap at the price. Asked which supplier he would invest in, he chose TSMC.
“I think it's pretty well known that TSMC underpins all of the world's chip production. It is cheap at the price. I would buy that stock every day of the week.”at 1:06:49
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Walter Goodwin · Fractile
Inference spending will far exceed today's levels in ten years, needing thousands of times more compute.
Goodwin's opening claim is that the world will spend far more on running models in ten years than it does today. He says it feels like thousands of times more compute than today will be needed. He ties this to the growth of long-running reasoning and agent workloads, which need very long contexts and sequential token generation. He gives no detailed sizing for the estimate, so the magnitude rests on his intuition about how these workloads will scale.
“We will be spending far more on running models in ten years' time than we are today. It feels like we need thousands of times more compute than we have today.”at 0:06
Walter Goodwin · Fractile
Land and grid power, not chips, are the binding constraint on AI data center buildout.
Walter says the easy data-center sites, such as retrofitted industrial sites and US locations where gas turbines can be permitted in about 90 days, have been exhausted. Data-center capacity, meaning land and power, is now the fundamental bottleneck on how much compute can be built. Grid connection queues in the US and UK are an order of magnitude larger than the energy available on the grid. He adds that whether all those projects are legitimate is open to question, but the demand looks staggering.
“the queues for grid connection will be an order of magnitude more than there is energy in the grid.”at 39:14
Walter Goodwin · Fractile
Inference compute is the key axis for AI capability, so inference demand is enormous.
Walter argues that training is a fixed cost, while inference is the use of models to do useful things and is the part that matters economically. Test-time compute scaling, first shown by reasoning models, means capability rises as more compute is spent at inference. He cites a Navier-Stokes solution produced by about 10,000 agents running for 80 hours. Solving hard problems therefore means generating billions of tokens, so every frontier application of a hard problem will look like this. Demand for inference compute is therefore very large.
“Inference compute is the axis that we have, and the Navier Stokes example is kind of evidence of that... It's the knob that we have to turn to get greater capability.”at 45:13
Walter Goodwin · Fractile
AI GPU buildout is fully utilized and won't collapse like the 2000s fiber bust.
Walter argues the AI build-out is justified and will not collapse, unlike the fiber overbuild of 2000-2001. His evidence is that all GPUs powered on in data centers are fully utilized, with no dark GPU fleets. He links this to the large number of problems that still need solving and the tools now available to tackle them, plus the expanding amount of inference thinking that could be done with more compute.
“far from the fiber build out of 2000, 2001, we have complete full utilization of the GPUs that are powered on in data centers. We don't have dark GPU fleets”at 48:38
Walter Goodwin · Fractile
Specialist inference providers are a better bet than neoclouds, which lock up their own capacity.
Walter says inference providers are better placed than neoclouds because they have not yet locked up large amounts of their own capacity. That keeps them free to dodge and weave, bet on new chips and carve out new parts of the market. He argues this segment will not fully commoditize, since specialists can serve multimodal, text and fast inference. He would rather be an inference provider today than a neocloud, because it gives more scope to build an enterprise customer flywheel that consumes tokens.
“They haven't yet locked up huge amounts of their own capacity. They're not a Neo Cloud. ... I would rather be an inference provider today, a specialist in running these models... than I would a Neo Cloud, actually.”at 1:01:10
Walter Goodwin · Fractile
A couple of AI inference companies will be worth over $100 billion each.
Walter says the market will not be winner-take-all like Uber versus Lyft, and he doesn't expect only three players as in cloud. He expects a long tail of players, and with neoclouds added, tens of players maneuvering around billions of dollars. Within that field he expects a couple of companies that become worth over $100 billion.
“I think we will have a couple of multi $100,000,000,000 companies built in that space.”at 1:02:31
AI chips & semiconductors
Where everyone stands on AI chips & semiconductors
Walter Goodwin · Fractile
Frontier labs and hyperscalers need their own chips so incumbent chip margins don't squeeze them.
Goodwin says there is a fight to keep margins expanded across the AI stack. If you serve frontier models, it matters that you are not squeezed down by very high margins at incumbent chip players. A first-party or third-party chip can serve as a counterpoint to that margin pressure. He argues that a combination of forces means frontier labs and hyperscalers need their own retort to parts of this stack. This is why every layer is trying to vertically integrate.
“if you are serving frontier models, it is important that you're not squeezed all the way down by very high margins on incumbent chip players.”at 7:17
Walter Goodwin · Fractile
Memory bandwidth lags compute scaling by about 25,000x, making LLM inference increasingly bandwidth-bound.
Goodwin explains that over the last 25 years, chip FLOP counts have scaled about a million-fold, while memory bandwidth has scaled only about 40-fold. Bandwidth is more intrinsically limited by physics, so it lags FLOP scaling by roughly 25,000 times. This makes many operations bandwidth-limited. For LLM inference, every generated token requires loading the model weights and context from memory over that limited bandwidth. As a result, generation speed is capped by memory bandwidth rather than math capability, which is why a large frontier model might generate only around 80 tokens per second.
“Bandwidth, by contrast, is kind of much more intrinsically physics limited. We've scaled it about 40 fold in the same time. So, you know, it's lagged behind the FLOP scaling about 25,000 times.”at 17:19
Walter Goodwin · Fractile
New frontier chip entrants face a scarce-capacity race for TSMC and HBM, making scale hard to reach quickly.
Goodwin says there is a real challenge for new chip entrants at the frontier in accessing TSMC and HBM capacity. Many custom chips from TPU and MTIA to Jalapeno, Matx and Etched are rumored to use the same back-end partner, Broadcom, and the same TSMC process nodes and HBM technology. This creates a gold rush for a finite resource. He argues this makes it challenging for chip companies to reach meaningful scale in a short time. Fractile's in-house back-end capability helps it reach manufacturing partners directly.
“There is a real challenge, especially for new chip entrants at the frontier, in accessing, for instance, TSMC capacity, HBM capacity.”at 28:22
Walter Goodwin · Fractile
HBM memory and TSMC logic shortages will persist for years, constraining AI chip supply.
Walter describes a shift from an Nvidia GPU shortage three years ago to a broader silicon and supply-chain shortage. He points to acute HBM memory shortages that stretch years into the future and logic silicon shortages at TSMC. Land and power are a separate constraint, but the chip supply shortages run for years.
“we found ourselves in a[n] underlying silicon and supply chain shortage, you know, acute memory shortages today for HBM and so on, which stretch years into the future, logic silicon shortages with TSMC as well.”at 36:32
Walter Goodwin · Fractile
Value in AI will accrue to chips, since long design lags limit the race to the bottom.
Walter argues chips are where value accrues because of the intrinsic lag to design, build and deploy a chip, and because a chip must have a useful lifetime in a data center. This means the chip space does not face the same fully efficient, potential race to the bottom seen elsewhere. A better idea or sustained advantage can therefore earn lasting value, and he also expects value to accrue at the frontier of model construction.
“I am very long on chips in terms of where that value accrues or being one big place where that value accrues is actually because of the nature of the challenge of placing the right chip bet.”at 51:01
Walter Goodwin · Fractile
Mixed-chip inference orchestration is transient; heterogeneity likely collapses into fewer chip types.
Walter says the coming wave of mixed chips, such as Nvidia pairing Rubin GPUs with Groq chips for prefill-decode and attention-FFN disaggregation, creates orchestration value that may not last. Deployments mixing chips risk stranded capacity, because the right chip mix shifts as context lengths and model sizes change and data-center capex is fixed. SRAM chips lack the memory for attention and the FLOPS density for prefill, so a single chip that does everything would remove the need for heterogeneity. He argues Fractile's 3D stacking of an SRAM layer, a DRAM layer and a full compute die achieves this.
“I think a lot of that could collapse back down over time. You know? There isn't a fundamental existential reason why we desperately need to have seven different platforms to run a workload.”at 55:18
AI pricing & unit economics
Where everyone stands on AI pricing & unit economics
Walter Goodwin · Fractile
Algorithmic gains are cutting the cost to serve fixed intelligence sharply, with no hardware change.
Walter says the cost to serve a given level of intelligence is falling deeply, with no hardware change, because of model efficiency work. He points to innovations in compressing context and sparse attention as context lengths become a challenge. He frames this as more intelligence per chip at iso-compute, and says hardware improvements will layer on top.
“the cost assuming no hardware change to serve a given level of intelligence is coming down Deeply, eight deeply in efficiency with no hardware change.”at 47:44
Walter Goodwin · Fractile
Total spending on running AI models will be far higher in ten years than today.
Walter says a token in ten years will deliver far more capability because the models will be orders of magnitude larger. Inference margins will fluctuate and accrue in different places across chips, clouds, frontier labs and open-model providers. Even so, he says total spending on running models in ten years will clearly be far greater than today.
“We will be spending far more on running models in ten years' time than we are today. You know? That part, I think, is is clear.”at 50:40
Enterprise AI adoption & AI software
Where everyone stands on Enterprise AI adoption & AI software
Walter Goodwin · Fractile
Company-specific fine-tuned models lose their edge with each new frontier or open release.
Walter says fine-tuning a model for a company's use cases has a positive impact, but the advantage is transient. It disappears when the next release or next open model arrives. He sees a possible exception where an internal specialized variant is stacked on top of the best frontier capability. He says it may become the case that specialization is only possible with open models, and then the trade-off is whether to specialize an open model or rely on the frontier model, which keeps improving compute efficiency.
“there is positive impact in taking a model and fine tuning it for your use cases. But it's transient thing. Generally, the advantage of having something that is internally fine tuned will disappear as soon as the next release, if the next open model arrives.”at 52:44
Walter Goodwin · Fractile
Companies working purely in the AI software layer face danger from chip heterogeneity fading.
Walter says there is a danger for anybody working purely in the software layer. Some clouds and software providers have positioned themselves as the orchestration layer over a coming wave of heterogeneous chips. He thinks much of that heterogeneity could collapse back down over time because there is no existential or strong economic case for running seven platforms per workload, so the orchestration value may not last.
“I think there is a danger for anybody that is working, you know, purely in the software layer.”at 53:58
Power demand & the grid
Where everyone stands on Power demand & the grid
Walter Goodwin · Fractile
China leads the West on energy for AI; Western data centers should co-locate with renewables and hydro.
Walter says China has stolen a march on the West on energy. Mega hydro projects are ideal for AI compute because an AI data center does not need to be near population, so 15 gigawatts of compute can sit next to a hydro plant. He argues the US, UK and Europe can do much more of this, especially with renewables such as wind farms, and that governments should promote it more.
“I think this is one of the places where China has somewhat stolen the march on the West.”at 40:06
Frontier AI labs & model race
Where everyone stands on Frontier AI labs & model race
Walter Goodwin · Fractile
Verifiable tasks like maths and coding will be solved by frontier models within twelve months.
Walter says the pace of capability advance in models has been extraordinary, with this summer's maths results. People talk about shortened timelines, and he thinks anything verifiable, such as maths and coding, will be solved in the next twelve months. He expects frontier models to tackle any hard verifiable challenge within that period.
“I think what has become clear is anything that is verifiable, you know, maths and coding tends to fall into this, will be solved in the next twelve months.”at 1:08:48
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
- NVIDIA147 speakers · 134 bullish, 106 bearish calls
- Broadcom16 speakers · 17 bullish, 6 bearish calls
- Taiwan Semiconductor Manufacturing Company8 speakers · 7 bullish, 0 bearish calls
Topics
- 3D memory stacking
- AI buildout vs fiber bust
- AI margins
- AI orchestration software
- Groq acquisition rationale
- Nvidia valuation
- Nvidia's general-purpose design vs specialists
- TSMC as foundry
- chip replacement cycle
- custom AI chip platform
- custom silicon vs incumbent margins
- data center power and land constraints
- enterprise fine-tuned models
- foundry and HBM capacity constraints
- heterogeneous inference hardware
- inference compute demand
- inference cost per unit of intelligence
- inference provider market structure
- memory and logic chip shortage
- memory bandwidth bottleneck
- neocloud custom chip adoption
- neocloud differentiation
- neocloud vs inference providers
- power siting for AI compute
- specialist inference chips
- total inference spending
- value capture in AI chips
- verifiable task capability
More from 20VC
- 20VC: Jev Creator on Why Data Is More Important Than Compute | Will US Open-Source Models Catch Chinese | The Cost of Intelligence and Tokens: What No One Sees But Needs to Know with Diogo Almeida · November 16, 2026
- 20VC: Jensen Huang Investigation: What Happens Now | Sequoia Leave London, Close Scout Program and End Evergreen Fund | NASDAQ Loses $2.5TRN in 24 hours | Anthropic targets $1.4TRN IPO valuation · November 12, 2026
- 20VC: Mistral CEO on Why Energy Is Not the Bottleneck to AI | Why OpenAI and Anthropic Will Come for Your Business | The Future of the Inference Market | Are Chinese Open-Source Models Truly Safe? · November 9, 2026
- 20VC: From $8BN to $19BN: Inside WashU's Endowment Fund | SpaceX Made Up 20% of Our AUM | Why Europe and China Has So Much Opportunity | The Reckoning That is Coming for AI Companies with Scott Wilson · November 7, 2026
- 20VC: OpenAI Cuts Valuation to $1.25TRN | Public Market Meltdown: GOOG Loses $500BN in 12 Hours | DoorDash Buys Instacart for $17BN | Kalshi Raises $1BN and Cognition Eyes $80BN Valuation · November 5, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus