Anthropic Eyes UK Startup Fractile's Analog AI Chips

Anthropic, the maker of the Claude AI model, is reportedly in early-stage negotiations with British chip startup Fractile to integrate its novel inference processors into the company's AI infrastructure. The move would make Fractile's chips a fourth compute resource for Anthropic, joining NVIDIA GPUs, Amazon Trainium, and Google TPUs in powering one of the world's most prominent AI platforms.

According to a report from The Information, Anthropic aims to deploy Fractile's hardware by 2027, though discussions remain preliminary. The deal, if finalized, would represent a major endorsement of an unconventional chip architecture that promises dramatic improvements in speed and cost — and a significant signal that the AI industry is actively seeking alternatives to the dominant NVIDIA ecosystem.

Key Takeaways

• Anthropic is exploring Fractile's inference chips as a complement to NVIDIA, Amazon, and Google silicon
• Target deployment is 2027, with negotiations still in early stages
• Fractile uses 'analog in-memory computing', a fundamentally different approach from conventional digital AI accelerators
• The startup claims 25x speed improvements and 1/10 the cost when running leading AI models
• Former Intel CEO Pat Gelsinger is among Fractile's investors
• The deal would diversify Anthropic's compute supply chain, reducing reliance on any single hardware vendor

Why Anthropic Is Looking Beyond NVIDIA

The AI industry's dependence on NVIDIA has become one of its most discussed vulnerabilities. NVIDIA controls an estimated 80-90% of the AI accelerator market, and its H100 and B200 GPUs remain the gold standard for both training and inference workloads. But this dominance comes at a steep price — literally. NVIDIA's top-tier chips can cost $30,000-$40,000 each, and demand consistently outstrips supply.

For companies like Anthropic, which operate massive inference fleets to serve millions of Claude users, compute costs represent a significant portion of operating expenses. Inference — the process of running a trained model to generate responses — now accounts for the majority of AI compute spending industrywide, and that share is growing as consumer and enterprise adoption accelerates.

Anthropic has already taken steps to diversify. The company runs workloads on Amazon Web Services using Amazon's custom Trainium chips, and it also leverages Google Cloud's TPU infrastructure as part of its deep partnership with Google, which has invested billions in the startup. Adding Fractile would give Anthropic a fourth option, specifically optimized for inference efficiency.

Inside Fractile's Analog Computing Breakthrough

What makes Fractile stand out is its fundamentally different approach to chip design. While virtually all modern AI accelerators — from NVIDIA's GPUs to Google's TPUs to AMD's Instinct series — rely on digital computation, Fractile employs an analog in-memory computing architecture.

In traditional digital chips, data must be constantly shuttled between memory and processing units, creating a bottleneck known as the 'von Neumann bottleneck' or the 'memory wall.' This data movement consumes enormous energy and limits throughput. Analog in-memory computing sidesteps this problem by performing calculations directly where the data is stored, using the physical properties of memory cells themselves to execute mathematical operations.

The approach is particularly well-suited for matrix multiplication — the core mathematical operation that dominates neural network inference. By encoding model weights as analog values in memory cells and processing inputs in place, analog chips can theoretically achieve massive parallelism with far less energy consumption than their digital counterparts.

Fractile claims its chips deliver:

• 25x faster inference compared to current mainstream solutions
• 10x lower cost per inference operation
• Dramatically reduced power consumption due to eliminated data movement
• Optimized performance for large language models and transformer architectures

If these claims hold up at scale, the implications for AI economics would be profound. A 10x cost reduction in inference could fundamentally alter the pricing models for AI services and make advanced AI capabilities accessible to a much broader range of organizations.

Pat Gelsinger's Bet on Analog AI

Fractile's investor roster lends additional credibility to the startup's ambitions. Pat Gelsinger, the former CEO of Intel, has personally invested in the company. Gelsinger, who led Intel from 2021 until his departure in late 2024, is one of the semiconductor industry's most experienced executives, with decades of chip engineering and business leadership experience.

His involvement suggests that Fractile's technology has passed scrutiny from someone who deeply understands both the technical and commercial challenges of bringing new chip architectures to market. Gelsinger's tenure at Intel was marked by ambitious attempts to revitalize the company's manufacturing capabilities and compete more aggressively in the AI chip space — efforts that ultimately fell short but demonstrated his conviction that the AI hardware market needs more competition and innovation.

The startup is based in the United Kingdom, adding a geopolitical dimension to the story. The UK government has been actively trying to position Britain as a hub for AI and semiconductor innovation, and a high-profile deal with Anthropic would bolster those ambitions significantly.

The Broader Race to Dethrone NVIDIA

Anthropic's interest in Fractile fits into a much larger industry trend: the urgent search for NVIDIA alternatives. Virtually every major AI company is either building custom chips or investing in startups that promise to break NVIDIA's stranglehold on the market.

The competitive landscape includes:

• Google TPU v6 (Trillium): Google's latest custom AI accelerator, used internally and offered through Google Cloud
• Amazon Trainium 2: AWS's next-generation training and inference chip, designed to undercut NVIDIA on price-performance
• AMD Instinct MI300X: AMD's flagship AI GPU, which has gained traction with cloud providers and enterprises
• Microsoft Maia 100: Microsoft's custom AI chip, deployed in Azure data centers
• Groq LPU: A startup offering ultra-fast inference through its Language Processing Unit architecture
• Cerebras WSE-3: A wafer-scale chip designed for both training and inference at unprecedented scale

What distinguishes Fractile from most of these competitors is its use of analog computing. While the companies listed above all employ digital architectures (albeit with varying designs), Fractile is betting that analog's inherent efficiency advantages will prove decisive as inference workloads scale.

Technical Challenges and Skepticism

Despite the promising claims, analog computing faces well-documented challenges that explain why it has not yet achieved mainstream adoption. Precision is the most significant concern. Analog computations are inherently less precise than digital ones, as they are susceptible to noise, temperature variations, and manufacturing inconsistencies.

For AI inference, however, this limitation may be less critical than it first appears. Research has shown that many neural network operations can tolerate reduced numerical precision without meaningful degradation in output quality. Techniques like quantization — deliberately reducing the precision of model weights — have become standard practice in the industry, suggesting that the analog approach's precision trade-offs may be acceptable for many use cases.

Other challenges include:

• Manufacturing complexity: Analog chips require different fabrication processes and quality control standards
• Software ecosystem: Developers need new tools and frameworks to program analog hardware effectively
• Scalability: Demonstrating performance at lab scale is different from delivering it in production data centers
• Reliability: Analog components may degrade differently than digital ones over time

The 2027 target deployment date gives Fractile roughly 2 years to address these challenges — an ambitious but not impossible timeline, assuming the technology is already at an advanced prototype stage.

What This Means for the AI Industry

If Anthropic successfully integrates Fractile's chips, the ripple effects could extend far beyond a single company's infrastructure decisions. A validation of analog computing by one of the world's leading AI labs would likely trigger a wave of investment and interest in similar technologies.

For developers and businesses building on Claude and other AI platforms, cheaper and faster inference could translate into lower API prices, faster response times, and the ability to deploy more complex models in production. The economics of AI applications remain challenging for many startups, and a 10x cost reduction at the infrastructure level could make previously unviable business models suddenly attractive.

For NVIDIA, the news is a reminder that its dominance, while formidable, is not guaranteed. Every major customer is actively seeking alternatives, and novel architectures like analog computing represent a potential disruption vector that NVIDIA's traditional digital GPU roadmap may not easily counter.

For the broader semiconductor industry, Fractile's potential deal with Anthropic validates the thesis that the AI hardware market is large enough and growing fast enough to support fundamentally new approaches to chip design. The era of one-size-fits-all GPU computing for AI may be drawing to a close, replaced by a more diverse ecosystem of specialized accelerators.

Looking Ahead: The Road to 2027

Several key milestones will determine whether this reported partnership materializes and succeeds. First, the early-stage negotiations must progress to a formal agreement, which will likely involve extensive technical due diligence, benchmark testing, and contractual commitments on performance and delivery timelines.

Fractile will need to demonstrate that its chips can reliably run Anthropic's Claude models at production scale, meeting stringent requirements for accuracy, latency, and uptime. The startup will also need to scale its manufacturing capabilities, likely partnering with an established foundry to produce chips in the volumes Anthropic would require.

Meanwhile, the competitive landscape will continue to evolve rapidly. By 2027, NVIDIA will have released multiple new GPU generations, Amazon and Google will have advanced their custom silicon programs, and other startups may emerge with their own breakthrough architectures.

The stakes are enormous. The AI inference market is projected to reach tens of billions of dollars annually by the end of the decade. If Fractile can deliver on its promises of 25x speed and 10x cost reduction, even capturing a small slice of that market would make it one of the most valuable semiconductor companies in the world. For Anthropic, the partnership represents a strategic bet that the future of AI compute will be more diverse, more efficient, and far less dependent on any single chip vendor than it is today.