Gimlet Labs Raises $300M Series B at $3B Valuation

The hardest part of heterogeneous computing begins after the benchmark. A mixed fleet can post impressive speed gains, but customers still need it to behave like infrastructure when models, traffic, chips, power, and cooling change at once. Gimlet Labs now has a $300M Series B led by Andreessen Horowitz to build for that operating reality.

The September 4, 2026 financing values the San Francisco AI infrastructure company at $3B and brings its announced total funding to $392M. It arrives only 5 months after Gimlet's $80M Series A and less than a year after the company emerged from stealth.

Gimlet says the capital will expand its team and build operations for a multi-silicon inference cloud designed to run AI workloads across GPUs, CPUs, and purpose-built accelerators. The bet is that the next unit of AI capacity can come from matching each phase to the right hardware, rather than simply buying more of one kind of chip.

A $300M Bet on Multi-Silicon Inference

Andreessen Horowitz led the Series B, with Sapphire Ventures, Menlo Ventures, 645 Ventures, Arm, Eclipse, Emergence, Factory, Hudson River Trading, M12, OnePrime Capital, Prosperity7, QuantumLight, Samsung Ventures, Tiger Global Management, Triatomic, Wing Ventures, and XTX Markets participating. Bloomberg and Reuters independently corroborated the amount, $3B valuation, and Andreessen Horowitz's role as lead investor.

The capital history moved quickly. Gimlet announced a $12M Seed led by Factory when it launched publicly in October 2025, followed by an $80M Series A led by Menlo Ventures on March 23, 2026. The new Series B increases the company's disclosed funding more than fourfold in a single round.

Gimlet says it has added billions of dollars in contracted revenue since March, built a gigawatt-scale datacenter pipeline, and is scaling toward hundreds of megawatts in managed capacity. Those figures are company-reported, as are the unnamed frontier-lab and hyperscaler customers described in earlier announcements, but they help explain why the financing moved from venture-software scale toward physical-infrastructure scale.

Why AI Inference Is Becoming a Hardware-Matching Problem

Training a large model is demanding, but inference introduces a different pattern of work. A single agent may handle prompts, retrieve data, call tools, run code, and move among models, with each step placing different pressure on compute, memory, networking, and latency.

Gimlet's software traces and decomposes those workloads, then schedules pieces onto the hardware suited to them. Compute-heavy prefill can remain on GPUs, memory-bound decode can use another architecture, and CPUs or specialized accelerators can handle other parts of the graph. A compiler, runtime, and orchestration layer coordinate the result behind one inference interface.

Andreessen Horowitz's investment note describes the problem in physical terms as well as software terms. Different processors can require different networking topologies, rack densities, power profiles, cooling systems, and even water temperatures. The developer may see one API, while Gimlet has to make incompatible machinery behave like a dependable shared fleet.

The Team Has Built Infrastructure Together Before

Gimlet's leadership team includes 5 co-founders. Zain Asgar is Co-Founder and CEO; Michelle Nguyen is Co-Founder and Head of Engineering; Omid Azizi is Co-Founder and Head of Hardware Platforms; Natalie Serrino is a co-founder leading KForge; and James Bartlett is Co-Founder and System Architect.

The team's shared history matters because the product crosses so many layers. Menlo Ventures says Michelle Nguyen architected Pixie's core data plane and developer experience, Omid Azizi built its data collector and kernel-instrumentation system, Natalie Serrino was a founding engineer at Pixie and Observe, and James Bartlett was a founding engineer at Pixie. Zain Asgar previously co-founded and led Pixie Labs before New Relic acquired it in 2020, after earlier work as a GPU architect at NVIDIA and an engineering lead at Google AI.

That continuity gives investors evidence that the founders have already shipped complex distributed systems together. It does not remove the risk of running a capital-intensive cloud, but it explains why the syndicate is willing to finance a jump from software research into datacenter operations.

Company-Reported Performance Meets Physical Scale

Gimlet reports 5-10X speedups for the same power footprint, or comparable throughput gains at the same latency, by breaking models across different accelerators. Andreessen Horowitz separately describes up to 10X gains in throughput and interactivity on frontier models within the same power envelope. These are company and investor claims rather than independent benchmark results, so the useful takeaway is the architecture being tested, not a universal performance guarantee.

The new capital is meant to expand that architecture into a larger managed service. Gimlet says it works with hardware from NVIDIA, AMD, Intel, Arm, Cerebras, and d-Matrix, and it offers both its own cloud and managed deployments in customer environments. Every added hardware family creates another opportunity to improve utilization and another integration surface that must stay reliable.

Power makes the commercial case sharper. Gimlet argues that AI demand is increasing faster than data centers and power infrastructure can expand, which makes throughput per kilowatt as important as raw accelerator count. If the software can route work effectively, older or specialized hardware can become useful capacity instead of stranded inventory.

The $300M Round Changes the Operating Obligation

Gimlet is no longer asking the market to judge only a compiler, scheduler, or clever systems paper. The Series B gives the company the resources and obligation to procure capacity, build datacenter operations, hire across the stack, and support customers whose AI products depend on latency and uptime.

That changes the proof investors and buyers will require. Customers need to know that performance survives different models, traffic patterns, chips, and facilities, while Gimlet must make the added complexity less visible rather than more visible. A heterogeneous cloud earns trust when developers stop thinking about the handoffs that make it work.

For the wider AI infrastructure market, the financing puts more value on orchestration across silicon instead of loyalty to a single architecture. Gimlet now has $300M to turn that thesis into operating capacity, and the next stage will be measured in customer workloads, reliable megawatts, and whether mixed hardware can feel ordinary enough to become infrastructure.