newsfilter
← Interviews
20VC with Harry Stebbings1h 16m

The $7B Startup Building AI Chips to Challenge Nvidia | Fractile Founder & CEO, Walter Goodwin

Walter Goodwin, Harry Stebbings

Interview, Fireside Chat
  • Fractal AI, a frontier chip company founded by Walter Goodwin, recently closed a $720 million Series C funding round, bringing its valuation to $7 billion.
    • The round was led by Thrive Capital, Founders Fund, Lightspeed, and Redpoint.
  • Goodwin argues that the AI industry is currently traumatized by the dot-com bubble, resulting in market pricing that assumes immediate exuberance rather than sustained, long-term demand.
    • He projects that spending on running AI models in 10 years will significantly exceed current levels despite a potential plateau in physical data center growth.
  • Fractal distinguishes itself by vertically integrating ML research, system engineering, and silicon design to optimize for "test time compute scaling" and long-context inference.
    • The company explicitly refuses to monetize its own internal models, choosing instead to compete strictly in the hardware layer to avoid conflict with customers.
  • The core technological bottleneck for current AI inference is the "memory wall," where the gap between processor floating-point speed (scaled ~1 million-fold in 25 years) and memory bandwidth (scaled only ~40-fold) has widened 25,000 times.
    • Current reliance on external HBM (High Bandwidth Memory) creates latency and physical limits on the "beachfront" of copper wires connecting DRAM to the processor.
  • Fractal's primary innovation involves stacking DRAM vertically on top of the logic die (3D integration) to combine the high bandwidth of SRAM with the high capacity of DRAM.
    • This architecture aims to deliver a "holy trinity" of performance: maximum compute density, SRAM-level bandwidth, and DRAM-level capacity in a single chip.
  • Goodwin predicts the emergence of multiple multi-hundred-billion-dollar companies in the frontier AI infrastructure space, distinct from the commoditization of generic cloud compute.
    • He identifies "NeoClouds" and specialized inference providers as key winners, noting that heterogeneity in chip architecture (mixing GPUs with specialized chips) is a transient, inefficient solution to current capacity constraints.
  • The company is currently in production with silicon that has passed fab trials, with full product tape-outs and volume deployment expected within months.
    • Fractal retains in-house expertise for backend design and supply chain management, bypassing intermediaries like Broadcom to secure access to TSMC capacity and custom process nodes.
  • In a "Quickfire" segment, Goodwin identified several key market shifts and strategic beliefs:
    • Timeline Compression: He revised his view that math and coding verification tasks will be solvable by frontier models within the next 12 months, citing recent "truncation of timelines."
    • Nvidia's Valuation: He remains bullish on Nvidia reaching a $10 trillion valuation in five years, citing undervaluation relative to the dot-com era and the necessity of frequent hardware refresh cycles.
    • Ideal Board Member: He expressed a desire for Demis Hassabis (DeepMind CEO) to join Fractal's board.
    • Industry Misconception: He challenges the view of chip design as purely technical, emphasizing that success is fundamentally driven by human collaboration across the entire stack.
  • Energy supply has become the primary constraint on AI expansion, with grid connection queues in the US and UK estimated to be an order of magnitude larger than available capacity.
    • Goodwin notes that China is gaining a strategic advantage by bypassing Western regulatory hurdles (e.g., NIMBYism) to secure massive power sites, such as hydro plants, for data centers.
    • He advocates for the UK and Europe to accelerate renewable and nuclear deployment to avoid falling behind in AI infrastructure.
  • Goodwin predicts that while the cost of inference per token will drop due to hardware and algorithmic efficiency, total global spend on model execution will rise drastically.
    • Value will accrue to specialized hardware providers and frontier labs that can solve high-complexity problems (e.g., scientific discovery, drug discovery) via "test time compute," rather than to generic orchestration layers.
  • Future workloads will increasingly rely on "agent swarms" that run for extended periods (days) to solve problems, necessitating chips capable of handling massive context lengths without the efficiency penalties of current SRAM-only or GPU-only architectures.