Fractile eyes $6.5B valuation following $250M Anthropic inference deal

The massive valuation jump for a company without a shipping product underscores the industry's desperate search for Nvidia alternatives specifically optimized for inference. As large language models move from training to deployment, the bottleneck has shifted to cost-effective execution, making Fractile's in-memory architecture a high-stakes bet for labs like Anthropic. Within the broader streaming and AI ecosystem, this reflects a move away from general-purpose GPUs toward specialized silicon that can handle massive data throughput more efficiently. The market is now pricing early-stage inference startups at multi-billion dollar premiums to secure future capacity. Watch for whether Fractile can maintain this momentum as competitors like Groq and the now-public Cerebras bring their own hardware to market sooner.

Fractile enters a crowded and well-funded inference accelerator landscape where several competitors have already reached commercial deployment or are nearing it. Etched raised $120 million in a Series A round co-led by Primary Venture Partners and Positive Sum Ventures to develop its Sohu chip, a transformer-specific ASIC manufactured on TSMC's 4nm process that the company claims delivers an order of magnitude better inference performance than Nvidia's GPUs. The Cupertino-based startup, founded by Harvard dropouts Gavin Uberti and Chris Zhu, argues that hardwiring the transformer architecture into silicon eliminates unnecessary hardware overhead and reduces latency for use cases like AI agents and real-time voice. Fractile's in-memory compute approach shares this philosophy of architectural specialization, though it targets memory bandwidth rather than compute density as the primary bottleneck.

The business dynamics around Fractile's Anthropic deal reflect a broader pattern of AI labs diversifying their inference supply chains away from Nvidia. Amazon's Trainium program has become the most prominent example of this shift. AWS announced the general availability of Trainium2 chips in December 2024, offering 4x the performance of their predecessors with up to 20.8 petaflops per instance, and the company is building a cluster of hundreds of thousands of Trainium2 chips for Anthropic that it describes as the world's largest AI compute cluster reported to date. By March 2026, Amazon had deployed 1.4 million Trainium chips across three generations, with Anthropic's Claude running on over 1 million Trainium2 chips. AWS also signed a deal to supply OpenAI with 2 gigawatts of Trainium computing capacity, signaling that even Nvidia's largest customers are hedging their compute strategies. Amazon formally launched Trainium3 UltraServer at its annual conference in December 2025, built on a 3-nanometer process with more than 4x the performance of Trainium2, and teased a Trainium4 roadmap that will support Nvidia's NVLink Fusion interconnect for hybrid deployments.

On the technical side, Fractile's valuation premium depends on proving that its in-memory architecture can outperform both Nvidia's latest GPUs and the growing field of specialized alternatives. Etched's CEO Gavin Uberti has claimed that one Sohu server replaces 160 H100 GPUs for transformer inference workloads, a benchmark that sets an aggressive bar for any challenger entering the market. Meanwhile, Nvidia continues to raise the performance ceiling with its Blackwell architecture, and AWS noted that Trainium2 already handles the majority of inference traffic on Amazon's Bedrock service. For Fractile, the challenge is delivering production chips in 2027 that can justify a $6.5 billion valuation against competitors who are already shipping or will ship sooner, in a market where the largest AI labs are simultaneously building their own custom silicon partnerships.