Gimlet Labs raises $300M to scale a multi-silicon inference cloud for AI agents

Gimlet Labs has raised a $300 million Series B to expand its AI inference platform, six months after announcing an $80 million Series A. The round is led by Andreessen Horowitz with participation from a long list of infrastructure and strategic investors including Arm, Samsung Ventures and Microsoft’s M12. Independent enterprise-tech coverage reports a roughly $3 billion valuation.\n\nThe product behind the financing is more relevant to AiToolMap than the round itself. Gimlet Cloud is an inference platform designed for agentic workloads, where one user request can trigger multiple model calls, retrieval, tools and non-model stages. Rather than running the entire workload on one homogeneous GPU fleet, Gimlet says its software traces and decomposes workloads and schedules different stages onto the hardware architecture best suited to each stage.\n\nA common example is prefill/decode disaggregation. Prefill is compute-heavy, while token-by-token decode is often memory-bandwidth constrained. Gimlet’s architecture can place those phases on different accelerators and can dynamically rebalance capacity as the workload changes. The company also describes broader splits such as speculative-decode or attention/feed-forward disaggregation and says it operates across GPUs, SRAM-centric accelerators, CPUs and other silicon.\n\nThe provider says this heterogeneous approach can deliver 3–10× faster performance for frontier workloads, 5–10× speedups at equivalent power in some configurations, or similar throughput gains at a target latency. Those are meaningful technical claims, but they are still Gimlet-run measurements and architectural models. Earlier company research discloses concrete examples, including multivendor prefill/decode work and speculative decoding on d-Matrix hardware, yet AiToolMap has not verified a neutral benchmark that generalizes those gains across models and customer workloads.\n\nWhat is independently clear is that inference economics have become a major financing category. SiliconANGLE reports the $300 million round and the company’s focus on dividing model phases across chip types; Crunchbase’s weekly funding roundup also lists Gimlet among the largest U.S. rounds of the week. Gimlet says it is already working with frontier labs and large inference consumers while ramping capacity for a broader audience.\n\nFor the directory, Gimlet Cloud is a credible infrastructure-tool candidate rather than merely a datacenter company. Canonical review should verify access model and pricing, supported models/accelerators, SLAs and observability, and separate customer-reproducible performance from the vendor’s own benchmark claims.