Gimlet Labs co-founder and CEO Zain Asghar announced over the weekend that the company had secured $300 million in fresh funding, placing its post-money valuation at around $3 billion. The round was led by new investor Andreessen Horowitz (a16z). This publication has obtained undisclosed details indicating that, compared to the valuation recorded by PitchBook in the prior round, this company's valuation has surged 16-fold, with additional intricacies hidden within the fundraising process.

According to insiders familiar with the company, this three-year-old startup focuses on enabling AI developers to simultaneously utilize multiple AI chip types for running large language models. During the fundraising phase, executives informed certain potential investors that, based on signed contracts, OpenAI could spend over $100 million annually on their services. However, Gimlet only exited stealth mode this October, and if this revenue projection materializes, it would represent a substantial income stream. Yet, this high revenue figure remains a forward-looking estimate. An OpenAI spokesperson stated, "We have explored collaboration possibilities with Gimlet, but they are not currently a paying customer."

OpenAI primarily relies on Nvidia chips at present. Nevertheless, like Anthropic and other major model developers, OpenAI faces pressure to adopt at least a portion of GPU solutions beyond Nvidia, which enhances the appeal of services like those offered by Gimlet. Last year, OpenAI entered into an agreement with a supplier, committing to purchase its chips while also acquiring equity in that enterprise. Recently, OpenAI also unveiled its proprietary AI chip, Jalapeno. According to a16z's disclosure, Gimlet is concurrently collaborating with an unnamed hyperscale cloud provider, a category typically encompassing companies like Google or Amazon.

Another notable aspect of Gimlet's current round is how the publicly announced valuation was achieved. Bloomberg first reported the round's headline valuation at $3 billion, but that figure corresponds only to the valuation at which the majority of the round's capital was invested. Two individuals familiar with the investment revealed that the round actually involved three distinct valuation tiers for share purchases. The three capital tranches were invested at valuations of $2.5 billion, $3 billion, and a "significantly higher" tier, respectively. Arm and Microsoft's venture capital arm M12 also participated in this round. It remains unclear at which valuation tier these two entities invested. Arm declined to comment, and M12 did not respond to interview requests.

Tiered valuation financing has become increasingly common in recent months, particularly among red-hot AI startups; however, the market typically sees two tiers, making a three-tier structure relatively rare. When demand for a round becomes overheated, some investors bid higher prices to secure allocations. Other investors may accept a higher publicized valuation in exchange for a larger equity stake. Gimlet Labs CEO Zain Asghar and a16z both declined to explain the financing structure. Nonetheless, it is evident that capital markets are fervently pursuing startups that enhance the inference efficiency of large language models.

The surge in inference demand has spawned numerous large transactions, such as Nvidia's $20 billion licensing agreement with inference chip company Groq, and specialized AI chip startups like Etched achieving extremely high valuations. a16z has just established an $1.1 billion hardware-focused fund, and combined with an additional $8.45 billion earmarked for AI infrastructure and growth-stage projects, the firm possesses ample capital and rationale to secure a substantial stake in Gimlet. This investment also positions a16z within the cloud services sector, an area previously missing from its portfolio. Asghar told Bloomberg that Gimlet aims to "build the world's first multi-chip heterogeneous cloud," having originated as a software business that now assists clients in building data centers while also constructing its own computing infrastructure.

Martin Casado, a general partner at a16z, explained the investment thesis in a recent interview: "Cost, performance, and latency are constraints everywhere, with limitations even extending down to the power level. Nearly every node in the supply chain demands transformation from us." Casado, who led Gimlet's growth-stage investment, further wrote in a Friday blog post: "AI inference is one of the fastest-growing markets in capitalist history, and we are on the verge of undersupplying nearly all the physical resources needed to support it."