Fractile raises $720M to build rack-scale AI inference systems
The startup closed a Series C at a $7 billion valuation to develop 3D-stacked memory chips. Founder Walter Goodwin argues that energy and land, not silicon, are now the primary bottlenecks for AI expansion.
Fractile, a frontier AI chip company founded by Walter Goodwin, closed a $720 million Series C round co-led by Thrive, Founders Fund, Lightspeed, and Redpoint, valuing the company at $7 billion. The firm develops rack-scale systems targeting large language model inference using a 3D-stacked memory architecture that combines the bandwidth of SRAM with the capacity of DRAM. Goodwin stated that the capital will fund multiple technology generations, expand systems and networking layers, and secure long-term supply chain agreements. Fractile maintains in-house backend design capabilities to avoid reliance on intermediaries like Broadcom, enabling direct relationships with foundries such as TSMC. The company recently hired Dan Rabinovich from Meta to lead systems build-out. Goodwin emphasized that the firm focuses on 'task zero,' the day a reference workload runs on an entire rack of Fractile chips, rather than just tape-out milestones. He believes the architecture can achieve a 10x advantage in cost and speed for specific workloads compared to conventional GPUs. Goodwin identified a shift in primary constraints for AI compute, moving from GPU shortages to acute shortages in HBM memory, logic silicon, and ultimately land and power. He noted that grid connection queues in the US and UK are an order of magnitude larger than available energy capacity. Goodwin contrasted this with China’s regulatory environment, which he described as capable of moving millions of people for energy sites, arguing that China has leveraged mega hydro projects to deliver 15 gigawatts of power to single sites. He stated that Western NIMBYism and environmental regulations delay projects, creating intense pressure to deploy chips that maximize intelligence per megawatt.