Gimlet Labs has raised $300 million in a Series B funding round that values the AI inference infrastructure startup at $3 billion, giving the company fresh capital as it pushes its multi-silicon approach deeper into the physical infrastructure behind artificial intelligence.
Andreessen Horowitz led the round, with Arm Holdings, Microsoft’s venture capital arm M12 and Samsung Ventures also participating. The financing takes Gimlet Labs’ total funding to $392 million, including $92 million previously raised through its seed and Series A rounds.
What makes the company’s approach unusual is that Gimlet Labs is not developing another AI chip. Instead, it is building software and infrastructure designed to make several different types of processors work together.
The company’s platform currently supports hardware from six semiconductor manufacturers: Nvidia, AMD, Intel, Arm, Cerebras and d-Matrix.
One AI workload, different types of silicon
Gimlet Labs describes its core product as a “multi-silicon inference cloud.”
Rather than assigning an entire AI inference workload to a single type of processor, its technology breaks the workload into stages. Individual stages can then be sent to a GPU, CPU or dedicated accelerator depending on which hardware is better suited to the task.
For developers, the idea is to write inference code once while the underlying system determines where that workload should run. Gimlet Labs says its software can route jobs according to factors such as cost and speed, reducing dependence on a single hardware vendor.
That distinction matters because different parts of AI inference do not necessarily place the same demands on hardware.
The prefill stage, where input information is processed, requires substantial compute capacity. The decode stage, where tokens are generated sequentially, is more dependent on memory bandwidth. AI agents can make that balance less predictable because they may repeatedly call models and external tools during a single task.
Gimlet Labs is betting that these differences create an opportunity to divide workloads across different processors instead of forcing every stage onto the same architecture.
The company claims this heterogeneous approach can improve the performance of frontier AI workloads by between three and 10 times under the same cost and power conditions. It also says throughput and interactive performance can improve by as much as 10 times within an equivalent power envelope. These are company claims and the supplied information does not provide independent benchmark results.
Gimlet Labs is moving beyond software
The larger shift now taking place at Gimlet Labs is its expansion into data center infrastructure.
Running different classes of silicon inside the same facility introduces another set of challenges. GPUs, CPUs and specialized accelerators can have different requirements around power density, networking and cooling. Existing data centers were not necessarily built around that combination.
That practical constraint has pushed Gimlet Labs beyond software orchestration and into infrastructure itself. The company is helping customers with data center buildouts while also expanding its own managed capacity.
Since March, Gimlet Labs has secured multi-billion-dollar contracts and increased its managed capacity to several hundred megawatts, according to the information provided. It is also working with Arm to operate its software on Arm-based semiconductors.
That combination of software orchestration and physical infrastructure is an important part of the company’s current pitch. Gimlet Labs is effectively trying to solve the same problem at two levels: choosing the right silicon for each inference task and creating infrastructure capable of housing those different chips together.
From Stanford research project to a $3 billion valuation
Gimlet Labs was founded in 2023 and grew out of a Stanford University research project. The San Francisco-headquartered company remained out of public view until October 2025, when it emerged from stealth and introduced its inference platform for AI agents.
CEO Zain Asgar As and members of the founding team had previously worked together at Pixie, a Kubernetes observability company that was acquired by New Relic in 2020.
The latest financing gives the two-year-old company unicorn status without requiring it to design or manufacture its own processor. Instead, its valuation rests on the proposition that future AI infrastructure will involve multiple chip architectures operating together rather than workloads being tied permanently to one supplier.
There are still important unknowns.
Gimlet Labs has not disclosed its revenue or customer count, making it difficult to assess the $3 billion valuation against its present commercial performance. The valuation therefore reflects expectations around future demand as much as disclosed operating results.
There is also a structural question surrounding the company’s position between chipmakers and customers. Semiconductor companies that currently work within the Gimlet Labs ecosystem could eventually develop more of their own orchestration capabilities. Restrictions on access to any of the supported hardware platforms could also weaken the promise of being able to move workloads freely across architectures.
For now, however, Gimlet Labs is building around a clear thesis: AI inference does not have to belong to one type of chip.
Its $300 million Series B, six-manufacturer hardware footprint, multi-billion-dollar contracts and expansion to several hundred megawatts of managed capacity suggest the company is attempting to turn that thesis into infrastructure at significant scale
Also Read: iPhone 18 Pro Prices Leak via Vodafone, but the iPhone Ultra Figure Steals the Show
