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
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
Hosted by Harry Stebbings
- 1h 15m
- Transcript
Summary
Written by a model from the transcript; not yet reviewed.
Walter Goodwin, founder and CEO of Fractal, joins Harry Stebbings to discuss the future of AI chips, frontier compute demands, and the realities of semiconductor manufacturing. The guest explains how Fractal addresses the memory wall by stacking DRAM on top of logic die to achieve high bandwidth and high capacity simultaneously, avoiding the constraints of traditional HBM and SRAM chips. They explore why inference demand is accelerating due to test-time compute scaling and agent swarms, examine the commoditization pressures facing neo-clouds, and debate the strategic moves of industry giants like Nvidia and OpenAI.
The discussion provides a detailed look at the physical and supply-chain realities of building specialized processors, the challenges of data center power constraints, and why the guest remains intensely optimistic about the long-term value of the silicon layer. Listeners gain a rare, technical insider's perspective on how hardware architecture must evolve to match the advancing frontier of artificial intelligence.
- Frontier intelligence is not yet commoditizing, and specialized chip providers can still capture high margins by targeting the evolving frontier of AI workloads. 0:16
- Sticking memory directly on top of the logic die solves the memory wall by combining SRAM-level bandwidth with DRAM-level capacity. 20:35
- Long-running agent swarms require massive test-time compute capacity and speed, creating intense demand for specialized inference chips. 25:20
- Data center land and power limitations represent a more acute bottleneck for AI scaling than raw logic silicon availability. 36:46
Chapters
- 0:00Podcast Intro and Sponsor Segments
The episode opens with host Harry Stebbings introducing the show and the sponsors, including Eight Sleep, MongoDB, and Framer.
- 0:30Walter Goodwin Introduces Fractal
Walter Goodwin explains Fractal’s focus on frontier AI chips and the company’s recent $720 million Series C funding.
- 2:09Is Vertical Integration the Future of Chips?
Goodwin discusses whether the industry is moving toward vertical integration and the challenges of competing as a pure chip company.
- 7:58Commoditization vs. Frontier Capability
The conversation explores what layers of AI infrastructure are becoming commoditized and why frontier AI remains a high‑margin space.
- 9:00Fractal’s 3D Memory Stacking Innovation
Goodwin details Fractal’s proprietary 3D memory stacking technology that delivers SRAM‑like bandwidth with DRAM‑like capacity, giving a competitive edge.
- 18:14Memory Bandwidth Bottleneck in LLM Inference
Goodwin explains the physics of memory bandwidth limits, the difference between DRAM and SRAM, and how 3D stacking can overcome these constraints.
- 27:37Production Challenges and Tape‑Out Process
The host asks about the realities of chip production, tape‑out, and the hurdles of scaling to volume manufacturing.
- 33:59Inference as the Dominant AI Market
Goodwin argues that inference, not training, is the key driver of AI economics and explains the growing demand for inference hardware.
- 39:38Data Center Capacity and Energy Constraints
Goodwin discusses how data‑center land, power, and energy supply are becoming the new bottlenecks for AI compute, and compares U.S. and China’s approaches.
- 42:04Strategic Vision for Frontier AI and Long‑Term Impact
The episode concludes with Goodwin reflecting on the long‑term vision for frontier AI, the importance of inference compute, and the societal impact of solving hard problems.
- 47:16Fractal’s Vision for Continuous Innovation
The host and guest discuss the necessity of ongoing innovation across the AI stack, focusing on chip development to support future intelligence.
- 48:47Capitalizing on Existing Data Center Capacity
Fractal’s strategy to maximize current GPU utilization and avoid dark fleets by leveraging existing data center infrastructure.
Highlights
- 0:35
Walter Goodwin is the founder and CEO of Fractal, which just raised a $720 million Series C led by Thrive, Founders Fund, Lightspeed, and Redpoint at a $7 billion valuation.
- 9:42
DeepSeek Inference Efficiencystory
DeepSeek released their full inference stack, showing how going full stack allowed them to achieve higher throughput, better interactivity, and healthy margins despite running open models.
- 16:15
The Memory Bandwidth Wallinsight
While math capability has scaled a million-fold in 25 years via Moore's Law, memory bandwidth has only scaled 40-fold, creating a critical bottleneck for LLM inference generation speed.
- 21:18
3D Memory Stacking Solutionstrategy
Fractal overcomes the memory beachfront limitation by stacking the entire DRAM memory layer directly on top of the logic die, using vertical copper connections instead of lateral wires.
- 28:43
Avoiding the Broadcom Bottleneckstrategy
Unlike competitors who use Broadcom as an intermediary backend partner and share identical chip platforms, Fractal keeps physical design and foundry interaction in-house to build custom building blocks.
- 43:59
Navier-Stokes Millennium Problem Agent Swarmstory
Solving complex science problems like Navier-Stokes equations required 80 hours of execution from 10,000 agents in a swarm, generating millions of tokens sequentially.
- 56:56
Stranded Capacity in Heterogeneous Chipsinsight
Mixing fixed proportions of different chip vendors in a data center risks stranded capacity when context lengths or workload ratios shift unexpectedly over time.
Quotes
“We will be spending far more on running models in 10 years' time than we are today.”
Unnamed speaker · 0:07 “I think there's a verticalization in that as all of these different components come together, and especially as you kind of look into the future and you work out, you know, what are these models going to look like?”
Unnamed speaker · 5:10 “We monetize by selling chips. We monetize by selling racks of compute to our customers.”
Unnamed speaker · 5:35 “The way that I think about this, Nvidia is an extraordinary business and they build, by many metrics, in a sense, the best AI chips in the world, right?”
Unnamed speaker · 13:43 “We don't have dark GPU fleets because, as you've seen, you know, we have this enormous number of problems as a species that we still have to solve.”
Unnamed speaker · 48:47 “I think there's absolutely space today in just, if you look at the technology, there is positive impact in taking a model and fine tuning it for your use cases.”
Unnamed speaker · 52:39 “If I have finite numbers of megawatts, I need to replace my chips every few years with better chips.”
Unnamed speaker · 1:10:26 “I think it's, again, we're using these models now. What are we using them for? We're using them to solve problems we couldn't solve before.”
Unnamed speaker · 1:11:39
Sponsored segments
- Eight Sleep 0:57–1:45
- MongoDB 1:41–2:58
- Framer 2:54–4:00
- 8Sleep 1:12:40–1:13:32
- MongoDB 1:13:27–1:14:30
- Framer 1:14:26–1:15:30
About this episode
Walter Goodwin is the founder and CEO of Fractile, the British chipmaker taking on Nvidia in the race to make AI models answer faster. Fractile has raised $940M, including a $720M Series C that valued the company at $7.5BN. Thrive, Founders Fund, Lightspeed and Redpoint led the latest round. AGENDA: 0:00 Can the world build enough chips for the AI we'll want in ten years? 5:04 Can you still win by just building chips? 13:26 Where can a specialist chip beat Nvidia at its own game? 16:14 Why are the world's fastest AI chips still waiting on memory? 27:37 How much of the AI chip industry is bullshit? Who's actually shipping? 35:48 What happens when you have the chips but nowhere to plug them in? 50:13 What will a single AI token cost in ten years? 58:46 Was Jensen's Groq deal an insurance policy against losing OpenAI? 1:04:00 Fractile raised $720 million. Is that actually enough? 1:09:40 Will Nvidia be worth $10 trillion in five years? 20VC:
Transcript
14,431 words · 3 speakers
Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.
Topics
- Fractal
- DRAM
- NVIDIA
- SRAM
- HBM
- Grok
- GPU
- TSMC
- OpenAI
- DeepSeek
Page updated · transcript and summary are machine-generated