Fractile Raises $720M, Aiming for $7B Valuation, Product Expected in 2027

Fractile reaches a $7 billion valuation after raising $720 million, four years after its founding. No chips have yet been deployed in data centers, with the first deliveries expected in 2027. The company promises response times reduced by tenfold thanks to stacked memory, a claim that awaits independent verification.

A 2027 Timeline and Evidence Needed Before Deployment

Fractile's valuation comes ahead of the commercialization of its first chips, which are expected to be available in 2027. The company will rely on Asian foundries for manufacturing. To date, no Fractile chips are operational in a data center. The systems currently being evaluated by verified benchmarks like MLPerf Inference from MLCommons involve Nvidia offerings already in commercial production. It is therefore necessary to distinguish announcements from measurements conducted under real-world conditions. In the meantime, IT teams must maintain their existing servers and adhere to Fractile's timeline, while the adaptation of software tools to these alternative chips within the next two years remains uncertain.

A Funding Round That Propels Valuation from $1.1 to $7 Billion

On November 11, Fractile finalized a $720 million funding round, bringing its valuation to $7 billion. In May, the company was valued at $1.1 billion. Based in London and Bristol, Fractile brought together CapitalG, Thrive Capital, Founders Fund, Lightspeed Venture Partners, Conviction, and British public funds for this funding round. This progression comes four years after the company's establishment.

The Stacked Memory Bet and the Promise of Tenfold Response Times

Fractile announces an architecture that stacks memory directly on the computing silicon and promises to reduce response times by tenfold, a claim that is vendor-specific and remains to be verified, particularly regarding thermal stability at scale. Executing large language models on current servers requires continuously transferring billions of parameters between computing units and memory, a process illustrated in Nvidia's Blackwell documentation and costly in terms of energy. The Nvidia GPUs currently available on the market rely on external HBM memory and are designed to handle both training and inference, whereas Fractile is developing a chip specifically dedicated to LLM inference. Walter Goodwin, founder of Fractile, emphasizes that the floating-point computing power of the most advanced chips has increased a millionfold over twenty-five years, while memory bandwidth has only increased by a factor of forty, resulting in a differential of 25,000 times. This technological choice comes at a time when each inference request incurs energy and hardware costs, and reducing expenditure per response would help lower infrastructure costs over several years. However, the profitability of a specialized chip imposes strict conditions: models must be used continuously at over 80% capacity, and hardware must be amortized over three to five years, with limited financial interest for traditional machine learning tasks on already amortized servers.