Anthropic Could Speed Up Artificial Intelligence by 100 Times
Anthropic is in talks with British startup Fractile, which promises to accelerate AI inference by 100x. Details about the company’s new hardware strategy are in our news.
Anthropic, the developer of the Claude AI model, is in early-stage talks with UK-based startup Fractile, which is claimed to be able to speed up AI inference processes by 100 times and reduce costs by 10 times. While the company is trying to diversify the chip infrastructure it currently supplies from giants such as NVIDIA, Google and Amazon, it has embarked on new searches to develop its own special artificial intelligence chips.
As increasing computational demands in the artificial intelligence industry force Anthropic to find more efficient and cost-effective hardware solutions, the innovative memory architecture offered by Fractile becomes a strategic target.
- Fractile is working on a new memory architecture that promises to increase AI inference speed by 100 times.
- Anthropic is conducting early-stage talks with the startup to strengthen its hardware infrastructure and reduce costs.
- The technology in question uses a special SRAM-based structure that reduces the need to move data off-chip.
Transformation is Aimed with Fractile Memory Architecture
The technology developed by Fractile is called “Memory Compute Fusion Architecture”. This architecture aims to perform all data processing processes directly within the chip by minimizing the need to move data onto DRAM. This technology, which works with a similar logic to NVIDIA’s Groq LPU solutions, offers high bandwidth and optimized SRAM capacity.
The new architecture has the potential to revolutionize AI processes compared to traditional methods.
A New Era Begins in NVIDIA and Groq Competition
With NVIDIA’s acquisition of Groq, Groq 3 LPU technology has become the centerpiece of the Vera Rubin platform. Offering 500 MB SRAM capacity and 150 TB bandwidth per second, these chips play a critical role for low-latency operations. Fractile, on the other hand, promises to offer similar functionality at a cost 10 times lower than NVIDIA’s existing solutions. These ambitious targets are enough to attract the attention of major players in the industry.
Anthropic Creates Its Own Chip Strategy
Anthropic has already signed a multi-gigawatt agreement with Broadcom and is preparing to add AMD’s Instinct series chips to its portfolio.
But conversations with Fractile make clear the company’s desire to develop customized hardware in-house. Although there is no chip in the testing phase yet, experts from NVIDIA, Graphcore and Imagination Technologies in the Fractile team prove the seriousness of the project. The upcoming period will clarify whether Anthropic will make a radical change in its hardware strategy.
How do you think the fact that large artificial intelligence companies such as Anthropic produce their own chips will affect the market dominance of giants such as NVIDIA? Share your thoughts with us in the comments section.
