Fractile announced it has raised $220 million to accelerate the development of a new generation of inference hardware designed to dramatically increase the speed and efficiency of frontier AI systems.
Founded in 2022, Fractile was created on the belief that the long-term impact of advanced AI systems would ultimately be constrained not by model quality alone, but by the time required to generate useful outputs. The company said it has spent the past several years building chips and systems specifically engineered to overcome the performance bottlenecks limiting large-scale AI inference workloads.
According to Fractile, frontier AI systems are increasingly dependent on generating extremely long output sequences, often involving tens of millions of tokens, in order to solve complex reasoning tasks. As models become more capable, the compute required for inference has become both a technical and an economic bottleneck for the broader AI industry.
The company noted that modern reasoning systems are already demonstrating how additional inference-time compute can directly improve model performance. Fractile pointed to systems like DeepMind’s AlphaGo, which achieved superhuman gameplay performance by repeatedly running neural network inference over many potential future states rather than relying on a single prediction.
Fractile argued that similar dynamics are now emerging within large language models, especially as reasoning-focused AI systems tackle increasingly difficult intellectual workloads requiring long chains of sequential computation. The company said many of today’s most advanced AI applications already involve generating up to 100 million tokens for a single task.
At current inference speeds of approximately 40 tokens per second on existing hardware, Fractile said generating outputs of that size can take nearly a month to complete. The company believes future AI systems will require inference speeds closer to 1,200 tokens per second in order to compress those workloads into a single day while still supporting long-context reasoning and large-scale model execution.
Fractile said its architecture is specifically designed to address these challenges by overcoming memory bandwidth limitations and improving the economics of high-performance inference at scale.
The company believes faster inference infrastructure will enable entirely new classes of AI workloads beyond current use cases such as agentic coding. Fractile highlighted future applications across areas including drug discovery, software engineering, materials science, and broader scientific research, where AI systems may eventually perform extremely long and complex chains of intellectual reasoning.
Fractile said its engineering work spans the full hardware and AI stack, including foundational AI research, foundry process innovation, chip microarchitecture, and systems-level optimization. The company said its goal is to break traditional trade-offs between cost and latency in AI inference systems.
The $220 million financing round was led by Accel, Factorial Funds, and Founders Fund. Additional participants included Conviction, Gigascale, O1A, Felicis, Buckley Ventures, and 8VC, alongside existing investors.
Fractile said the new funding will help accelerate commercialization efforts and support deployment of its first chips and systems to customers. The company is also expanding hiring efforts across offices in London, Bristol, San Francisco, and Taipei.
