Gimlet Labs, an AI infrastructure company focused on inference performance, has raised a $300 million Series B round at a $3 billion valuation. The financing was led by Andreessen Horowitz and included new investments from Arm and Microsoft's M12, alongside Menlo Ventures, Sapphire Ventures, Tiger Global Management, and other partners. The company said the capital will scale its multi-silicon cloud and datacenter capacity as demand for faster AI inference accelerates.
The Growing Need for Throughput and Speed
Monthly token generation has increased sixfold in twelve months, and some projections point to another twentyfold rise by 2030. AI data centers consumed roughly 18 gigawatts of capacity in 2025, a figure expected to triple by the end of the decade. Gimlet argues that power is now the most critical bottleneck, making throughput per kilowatt a central measure of infrastructure value.
Fixing Inference with Software and Heterogeneous Hardware
Gimlet was founded on the view that AI inference has become the majority software workload but remains orders of magnitude less efficient than it should be. Its approach breaks models into compute-bound and memory-bandwidth bound phases and schedules each phase on the most suitable accelerator, including GPUs, CPUs, dataflow architectures, and near-memory compute. By disaggregating workloads across different chips, Gimlet reports speedups of three to ten times for frontier inference workloads.
From Pixie Team to Rapid Funding Momentum
The Series B follows a $12 million seed round and an $80 million Series A that closed only about six months earlier, bringing the company's valuation to $3 billion. Gimlet was co-founded by Zain Asgar, Michelle Nguyen, Omid Azizi, Natalie Serrinoc, and James Bartlett, a team that previously built the Kubernetes observability company Pixie, which New Relic acquired in 2020. Return investors in the new round include Menlo Ventures, Eclipse, Factory, Prosperity7, and Triatomic.
Expanding Beyond Software into Data Centers
The startup initially focused on software that routes different parts of an AI workload to different silicon, but deployment quickly exposed a physical infrastructure gap. Many data centers were built around homogeneous chip environments, while heterogeneous architectures can require different cooling, temperature, and hardware configurations. Since March, it says it has added billions in contracted revenue, gigawatts of datacenter pipeline, and is scaling to hundreds of megawatts in managed capacity.
Strategic Investor Backing and Market Position
The round aligns with Andreessen Horowitz's broader push into AI infrastructure, including its Machine Age Fund and an expanded $8.5 billion Growth fund. Strategic new investors Arm and M12 signal support from both chip design and cloud ecosystems as the industry diversifies beyond a single dominant processor. Rather than betting on one accelerator, Gimlet is positioning itself as the coordination layer that allows multiple architectures to operate together.
Gimlet Labs enters its next phase with substantial capital, strategic partners, and a clear focus on making AI inference faster and more power efficient. The company is working with frontier labs and large-scale inference consumers while preparing to broaden access to additional capacity. As model sizes, context windows, and agentic workloads continue to grow, its heterogeneous approach may become a critical layer for the next wave of AI software.