The AI Compute Crisis: Why We Need Thousands of Times More Computing Power | Fractile Founder

Summary

Fractile’s Business Model

Founder and CEO Walter Goodwin describes Fractile as a chip and rack-systems company focused on AI compute. The company plans to earn revenue by selling hardware, rather than competing with customers through its own commercial models. It maintains machine-learning research expertise so its hardware decisions reflect changes in models and workloads.

AI Stack and Inference

Goodwin sees some open models and cloud-computing services becoming more interchangeable, while frontier intelligence and high-performance inference remain differentiated. He points to DeepSeek’s inference software as an example of how optimizing across the stack can improve throughput and speed, even when serving an open model. Specialized chips could also help cloud providers reduce reliance on a single GPU supplier.

Inference is often memory-bound, according to Goodwin. AI chips’ floating-point compute capacity has improved much faster than their memory bandwidth. During LLM generation, systems must repeatedly move model weights—and increasingly large contexts—between memory and compute. This data movement can limit token-generation speed even when the processor has unused mathematical capacity.

Fractile’s Memory Architecture

SRAM is fast and close to the processor, but expensive and limited in capacity. DRAM is denser and cheaper, but harder to access quickly. Fractile’s central architectural bet is 3D memory integration: placing memory above the logic layer to increase the number of connections between memory and compute. The goal is to combine high bandwidth with greater capacity, potentially supporting larger models and longer contexts without relying on multiple chip types.

New Workloads and Specialized Chips

Goodwin argues that reasoning models and agent systems can use far more inference compute than conventional question-and-answer interactions. Tasks may involve many parallel agents, long periods of work, and very large numbers of generated tokens. Fast, SRAM-based chips can struggle with these workloads’ capacity needs, while conventional GPU systems can face bandwidth and efficiency constraints.

Goodwin praises Nvidia’s general-purpose AI systems but argues that specialized silicon could be substantially faster or cheaper for particular workloads. He says new platforms could offer roughly a tenfold advantage on some tasks, though adoption is still early and carries risk. In his view, specialization is becoming more important as inference accounts for a much larger share of AI compute.

Manufacturing and Infrastructure

The discussion distinguishes tape-out—sending completed chip-design files to a foundry—from the longer process of receiving, testing, debugging, integrating, and producing reliable rack-scale systems. Goodwin says Fractile had received silicon from foundries and expected further test runs and product test batches in the following months.

Access to foundry capacity, HBM, and manufacturing partners presents major scaling challenges for chip startups. Fractile’s claimed advantage is that it handles more of its own physical-design and manufacturing work rather than relying on a backend intermediary such as Broadcom.

Land, electricity, and grid connections are also increasingly significant constraints alongside chip supply. Goodwin argues that improving hardware and model efficiency can deliver much more intelligence per megawatt and rack without requiring data-center capacity to grow a thousandfold. He sees China as having an advantage in rapidly developing large power projects, while renewable and nuclear power could support data-center growth in the West.

Demand, Economics, and Where Value May Accrue

Goodwin expects inference demand to grow substantially because deploying trained models enables practical use. More inference-time computation—such as generating and testing many candidate solutions—could help tackle difficult scientific or medical problems. He expects the cost per token to change over time, but says future models may provide far more capability per token and that total spending on model launches could rise.

Goodwin sees durable value in advanced chips and frontier-model development. He warns that software-only orchestration layers could be vulnerable if today’s fragmented hardware landscape consolidates. Multi-chip orchestration may address current limitations, but it can also create difficult capacity-planning and utilization problems.

Funding and Outlook

Fractile announced a $720 million Series C. Goodwin says the funding will support multiple technology generations, systems and networking work, supply-chain planning, research into future workloads, and international expansion. He is bullish on AI infrastructure investment and says he would invest in TSMC as a key supplier.

Speakers

• Harry Stebbings, host of 20VC
• Walter Goodwin, founder and CEO of Fractile

Content Type

This is an interview and analysis of AI chips, inference, manufacturing, and compute infrastructure. It is not a product review, guide, or tutorial.