Anthropic is in early talks to buy AI inference chips from Fractile, a London-based startup whose SRAM architecture sidesteps the expensive DRAM bottleneck that has become a critical constraint for AI companies. The discussions, reported by sources familiar with the matter, highlight a pivotal shift in how major AI developers are approaching infrastructure costs. With chips expected to arrive in 2027 or later, this is a long-term bet—but the timing reveals something urgent happening now.
Key Takeaways
- Anthropic is negotiating with UK startup Fractile to purchase specialized AI inference chips designed to reduce reliance on expensive DRAM memory.
- Fractile’s SRAM-based architecture addresses a critical supply and cost crunch in AI infrastructure.
- No deal terms, pricing, or purchase volumes have been disclosed; discussions remain preliminary.
- Chips are expected available in 2027 or later, positioning this as a strategic long-term investment.
- The move reflects Anthropic’s broader strategy to diversify suppliers beyond Google, Amazon, and Nvidia.
Why AI inference chips matter now
AI inference—running trained models to generate responses—has become a cost center that dwarfs everything else. Unlike training, which happens once, inference happens billions of times per day across every user query. The bottleneck is memory: moving data in and out of expensive DRAM chips burns power and slows computation. Fractile’s approach flips the architecture. By embedding computation directly into SRAM (static RAM, which is faster and more power-efficient but traditionally smaller), the startup reduces how much data needs to move through that expensive external memory pipeline. It’s not revolutionary in principle, but it’s a direct answer to a real problem that’s become impossible to ignore.
The semiconductor industry has been unable to supply enough DRAM to meet AI demand, and prices have stayed elevated as a result. Companies like Anthropic are paying premium rates for memory that was never designed for the scale of inference workloads they now run. Any architecture that reduces that dependency becomes strategically valuable—which is why a major AI developer is talking to a startup that won’t ship anything for at least two years.
Anthropic’s supplier diversification strategy
Anthropic currently relies on Google, Amazon, and Nvidia for its AI hardware infrastructure. That concentration is a vulnerability. Google and Amazon are also competing AI companies with their own inference demands. Nvidia, while dominant, controls the pricing power—and customers have limited alternatives. Fractile represents a third axis of independence: a specialized vendor focused purely on inference acceleration, with no competing AI service of its own. If the chips deliver, Anthropic gains leverage in negotiations with existing suppliers and reduces its exposure to any single vendor’s supply constraints.
This move is part of a wider industry pattern. OpenAI has invested in custom chip design. Meta is building its own AI accelerators. Microsoft is exploring partnerships with chip startups. The message is clear: the major AI labs no longer trust the traditional semiconductor supply chain to keep pace with their needs. They are hedging by sponsoring alternatives. Fractile’s SRAM approach is one such hedge.
The 2027 timeline and what it means
Two years is a long wait in AI. By 2027, the current DRAM shortage may have eased—or it may have worsened as even more AI workloads come online. The timing suggests Anthropic is not expecting relief from traditional memory suppliers anytime soon. This is a bet that specialized inference silicon will be necessary infrastructure by the time it arrives, not a luxury.
No details have been shared about pricing, purchase volume, or performance benchmarks. The secrecy is standard for early-stage supplier negotiations, but it also means the market cannot yet assess whether Fractile’s architecture actually delivers the promised cost savings. That verification will come only when chips ship and real-world inference workloads run on them.
What happens to the existing suppliers?
Anthropic’s diversification does not mean abandoning Google, Amazon, or Nvidia. All three will remain critical to the company’s infrastructure. What it does mean is that Anthropic is signaling it will not be held hostage by any single supplier’s capacity or pricing. For Nvidia, which has enjoyed near-monopoly pricing power in AI accelerators, this is a warning shot. For Google and Amazon, it reinforces that they need to keep their internal chip programs competitive or risk losing customers to alternative suppliers.
Is Anthropic definitely buying Fractile chips?
No. These are early talks, not a signed contract. Preliminary discussions can collapse if technical benchmarks disappoint, if costs remain uncompetitive, or if other suppliers solve the DRAM problem first. The reporting makes clear that nothing is finalized.
Why would Fractile’s SRAM design matter for AI inference?
SRAM is faster and more power-efficient than DRAM, but it has traditionally been too expensive and small-capacity for large-scale computing. By redesigning the inference chip to keep more computation local to SRAM and reduce trips to external memory, Fractile reduces both latency and the amount of expensive DRAM bandwidth required. This is particularly valuable for inference, where the same model runs millions of times and memory access patterns are predictable.
When will these chips actually be available?
Expected in 2027 or later in the decade, according to reports. This is not imminent. The timeline reflects the long lead time required to design, fabricate, and validate custom silicon at scale.
Anthropic’s move to explore AI inference chips from Fractile is less about an imminent product launch and more about positioning for a future where memory constraints force the industry to rethink chip architecture. Whether Fractile’s approach succeeds remains an open question, but the fact that a major AI company is betting on it signals that the current infrastructure model is unsustainable. The real story is not about one startup or one deal—it is about an industry finally acknowledging that inference costs are becoming the dominant problem, and custom silicon is the only way out.
Edited by the All Things Geek team.
Source: Tom's Hardware
