Two‑track inference reality
Investors and startups are increasingly accepting a two‑track market: GPUs remain the standard for training and experimentation, while specialized silicon and SRAM chips (Groq/Cerebras/others) get traction for low‑latency inference. Gimlet’s multi‑silicon positioning is the clearest commercial signal of that split. (menlovc.com)(techcrunch.com)
Menlo Ventures led an $80 million Series A in Gimlet Labs, a round announced March 23, 2026 and publicized on Menlo’s blog and in TechCrunch coverage. (menlovc.com) Gimlet bills its product as the industry’s “multi‑silicon inference cloud,” software that slices agentic workloads and runs segments across different vendors and architectures including NVIDIA, AMD, Intel, ARM, Cerebras and d‑Matrix. (techcrunch.com) The company claims the orchestration can deliver roughly 3×–10× faster inference for the same cost and power envelope and says it launched publicly in October with eight‑figure revenue out of the gate. (techcrunch.com) Gimlet announced a partnership with d‑Matrix on March 12, 2026 to deploy Corsair low‑latency, memory‑optimized accelerators alongside GPUs; the joint release cites “order‑of‑magnitude” improvements and up to 10× gains in latency and throughput per watt versus GPU‑only deployments. (prnewswire.com) Menlo’s lead investor Tim Tully frames the market shift as one where GPUs remain the compute workhorse while SRAM‑centric accelerators from players like Groq, Cerebras and d‑Matrix deliver step‑function latency advantages, with Gimlet providing the missing software layer to tie the heterogeneous fleet together. (menlovc.com) Gimlet says its stack can import existing PyTorch or Hugging Face pipelines and is offered as software, an API, or via Gimlet Cloud, positioning the product for frontier model labs and large data‑center operators rather than generalist app developers. (techcrunch.com)
