AI Performance 100 Times Higher and Costs 10 Times Lower: Anthropic Could Revolutionize AI

Anthropic

Anthropic, a company known for the development of Claude AI, is reportedly about to make a huge leap forward in the sector through the development of proprietary chips that would complement collaborations with NVIDIA, Google, and Amazon. According to rumors reported by The Information, the company has initiated preliminary talks with the British startup Fractile, an emerging entity that promises to redefine performance in the field of AI inference.

The central issue revolves around the technology developed by Fractile, called Memory Compute Fusion Architecture, a solution that concentrates processing directly within the chip and significantly reduces data transfer to DRAM. This approach limits dependence on external memory, one of the most costly aspects in terms of latency, consumption, and especially expenses in infrastructures dedicated to artificial intelligence.

Fractile bases its project on a new implementation of high-performance SRAM technology, with a philosophy similar to that of the LPU adopted by Groq and integrated by NVIDIA into its Vera Rubin ecosystem. In this context, NVIDIA defines the Groq 3 as an inference accelerator, with notable specifications such as 500 MB of SRAM, 150 TB/s of SRAM bandwidth, and 2.5 TB/s of bandwidth scale-up, while the complete LPX rack integrates 256 LPU and 128 GB of SRAM.

Fractile's proposal, at least on paper, pushes these ambitions even further: the startup claims that its architecture could ensure an increase of up to 100 times in AI inference performance and a cost reduction of up to 10 times compared to current NVIDIA/Groq technologies. These are extremely aggressive figures, especially considering that the company has not yet produced a prototype.

The project thus remains purely theoretical for the moment, but it takes on particular relevance in the current competitive landscape, where major AI players are increasingly seeking independence from traditional semiconductor manufacturers. Anthropic, while maintaining important agreements such as the multi-gigawatt contract with Broadcom and the possible future integration of AMD into its hardware portfolio, is evaluating options that could offer greater control over performance, operational costs, and scalability.

It should also be emphasized that the Fractile team consists of seasoned professionals from NVIDIA, Graphcore, and Imagination Technologies, a detail that strengthens the technical credibility of the project despite the current absence of a validated chip.

With the resources available to Anthropic, we could see the first fully customized chip in action within a few years, which would provide Anthropic with even greater autonomy over the supply chain and the development of internally optimized solutions specifically for its model.