- Gimlet Labs raised $300M at a $3B valuation, led by Andreessen Horowitz
- The startup’s software splits AI workloads across different types of chips
- Arm and Microsoft’s M12 joined as new backers as Gimlet scales into data centres
AI infrastructure is getting harder to build around a single chip. Gimlet Labs, a startup developing software that allows AI workloads to run across different types of chips, has raised $300 million at a $3 billion valuation. The round was led by Andreessen Horowitz, with new backing from Arm Holdings and M12, Microsoft’s venture fund.
It first raised a $12 million seed round backed by Factory and angels including Intel CEO Lip-Bu Tan, Figma CEO Dylan Field and a16z general partner Raghu Raghuram. Six months later came the $80 million Series A, led by Menlo Ventures and joined by Eclipse, Factory, Prosperity7 and Triatomic. The latest $300 million round takes the company’s valuation to $3 billion.
The problem Gimlet is trying to solve
Gimlet Labs was co-founded by Zain Asgar, Michelle Nguyen, Omid Azizi, Natalie Serrinoc and James Bartlett. The founding team previously worked together at Pixie, an open-source Kubernetes observability company. Asgar is the CEO; the rest of the team leads engineering, hardware platforms and systems research. Asgar previously founded Pixie Labs, which was acquired by New Relic in 2020.
Gimlet started with a relatively simple premise: AI workloads do not necessarily need to run on the same type of chip from start to finish. Its software can divide AI tasks and route different portions to the hardware best suited to them — GPUs for some workloads, CPUs or specialised accelerators for others.
The company calls this a multi-silicon inference cloud. Its platform is designed to orchestrate AI workloads across heterogeneous hardware, with Gimlet reporting three-to-ten-times inference speed improvements for the same cost and power envelope.
But deploying that software exposed another problem: data centres themselves were largely designed around more homogeneous infrastructure. That has pushed Gimlet beyond software. Asgar said the company is now helping customers configure data centres and is also working on data centres of its own. Different chips can require different cooling systems, temperatures and hardware configurations, turning a software problem into a physical infrastructure problem.
From inference software to the data centre
Gimlet is concentrating particularly heavily on AI inference, the stage where trained models generate responses, because Asgar says combining different chips is producing compelling performance gains.
That puts the startup in a rapidly expanding market alongside specialised inference players such as Groq and Cerebras. Tech Funding News has covered how
Groq’s $650 million raise and subsequent $350 million round at a $3.5 billion valuation reflect the growing importance of inference infrastructure, especially after Nvidia’s licensing deal reshaped Groq’s business.
Cerebras, meanwhile, has gone from private AI-chip challenger to public company, raising $1 billion at a $23 billion valuation in February 2026 and later reaching a market valuation above $56 billion at its Nasdaq debut.
The difference is that Gimlet is not trying to win by designing one superior chip. It is building the software layer that allows multiple architectures to work together.
Why a16z is leaning into multi-silicon
For Andreessen Horowitz, Gimlet fits neatly into a much broader shift toward the physical infrastructure underneath AI. On August 31, a16z expanded its Growth fund to $8.5 billion, just days after launching its $1.1 billion Machine Age Fund, which targets chips, memory, networking, storage, data centres and robotics. Its portfolio already includes OpenAI, Databricks, xAI, Mistral, Anduril and SpaceX, alongside newer infrastructure bets.
The bigger opportunity for Gimlet may therefore be less about picking the next Nvidia challenger and more about making sure companies do not have to pick just one. If AI infrastructure becomes genuinely multi-silicon, the software coordinating those chips could become infrastructure in its own right.
