Gimlet Labs' Valuation Jumps 7.5X to $3 Billion in Five Months

Gimlet Labs raised a $300 million Series B led by Andreessen Horowitz on Sept. 4, the company said in its funding announcement, which put the AI-inference startup's valuation at $3 billion. That is 7.5 times the $400 million valuation Gimlet Labs disclosed five months earlier, in March, when it raised an $80 million Series A. Gimlet Labs sells software that runs AI inference, the work of answering a request with an already-trained model rather than training one, across several types of computer chips inside one job.

The round included new backers Arm and Microsoft's venture arm M12, alongside returning investors Sapphire Ventures, Menlo Ventures, Samsung Ventures, Tiger Global Management and 645 Ventures, among more than a dozen participants in total, according to the company. Total funding since the 2025 seed round Factory led now stands at $392 million, Gimlet Labs said. Andreessen Horowitz, which had closed an additional $1.75 billion for its Growth Fund V four days earlier, did not disclose the size of its check.

A 7.5x jump on a $92 million base

Gimlet Labs' Series A, announced March 23, priced the two-year-old company at $400 million post-money, according to a funding-timeline analysis published by aifunding.me. TechCrunch, which covered the Series A the same week, described that figure as unusually high for a company five months out of stealth. Less than six months later, the same company was worth $3 billion, a jump the Series B announcement does not explain beyond citing new customer contracts.

1. Series A, March 2026400 $M
2. Series B, Sept. 20263000 $M

Source: Company funding announcements via aifunding.me and Yahoo Finance, accessed 2026-09-14

What the multi-silicon pitch claims

Gimlet Cloud, the company's product, splits an inference request into stages and assigns each to a different type of chip: GPUs from Nvidia or AMD for the compute-heavy prefill step, memory-optimized hardware for decode, and CPUs for network-bound tool calls. Gimlet Labs says the approach delivers 3 to 10 times the throughput of running the same job on a single chip architecture, at the same cost and power draw.

The company frames the opportunity around one number: TechCrunch reported in March that GPU utilization in production AI deployments runs between 15 and 30 percent, because most inference stacks are still built around one type of chip rather than a multi-stage pipeline. Gimlet Labs said it now works with Nvidia, AMD, Intel, Arm, Cerebras and d-Matrix hardware, a wider list than the five chipmakers it named in March. AWS took a related but separate swing at the same problem this month, when its SageMaker HyperPod cut inference cold starts from up to 30 minutes to seconds through model caching rather than chip-mixing.

What the statement doesn't say

Gimlet Labs said it added "one of the top three frontier labs" and "one of the top three hyperscalers" as customers and has signed "billions of dollars in contracted revenue" since March. It did not name either customer or disclose a specific revenue figure.

Quasa.io, an infrastructure-focused analysis site, wrote that the financing "does not establish that heterogeneous inference works economically at production scale," and that Gimlet Labs' performance claims leave out the model type, numeric precision and context length needed to judge them independently. Moving data between chip architectures can erase the efficiency gains the company advertises if the orchestration layer doesn't manage it well, the site wrote.

Founders built and sold Pixie in 2020

Gimlet Labs was founded by chief executive Zain Asgar, an adjunct professor at Stanford, along with Michelle Nguyen, Omid Azizi and Natalie Serrino. The four previously built Pixie, an observability startup New Relic acquired in 2020.

"We've reached a turning point where inference is the dominant AI workload and demand for tokens is explosive," Asgar said in the funding announcement.

The company said it is scaling toward "hundreds of megawatts" of managed compute capacity, part of a data center pipeline it measures in gigawatts. Gimlet Labs did not give a date for when that capacity would come online.