Gimlet Labs, a San Francisco-based startup that helps companies run AI models more efficiently across different types of computer chips, just raised $300 million at a $3 billion valuation, more than tripling its value in six months. Andreessen Horowitz led the round on September 4, 2026, joined by Microsoft's venture arm and chip designer Arm, signaling serious confidence in the company's approach to a growing infrastructure challenge.

The funding reflects a fundamental shift in how enterprises spend on AI. For the first time in 2026, companies are spending more money running AI models in production than training new ones. Gartner forecasts that global spending on inference will reach $23.3 billion this year, surpassing training spending of $19 billion. That budget shift is exactly where Gimlet is positioned to capture value.

What Problem Is Gimlet Labs Actually Solving?

Here's the technical reality that Gimlet addresses: running an AI model involves two distinct phases with completely different hardware needs. The first phase, called "prefill," processes your prompt and is compute-heavy, benefiting from raw processing power. The second phase, "decode," generates output one token at a time and is memory-bandwidth-heavy, meaning it needs fast access to data rather than sheer computing speed.

Most companies today run both phases on the same hardware, typically Nvidia GPUs, even though this is inefficient. Gimlet's software disaggregates these phases and routes each to whichever chip architecture handles it best. That might mean Nvidia GPUs for prefill and custom chips like Google TPUs, AMD accelerators, or Amazon's Trainium chips for decode. The result is lower latency, better throughput, improved hardware utilization, and reduced power consumption.

Before Gimlet existed, only the largest AI labs and cloud companies had the engineering resources to build this routing layer themselves. Gimlet's bet is that turning in-house tooling into a standalone managed service is a large enough market to support an independent company, rather than a feature that hyperscalers will eventually ship for free.

How Does Gimlet's Growth Compare to Other AI Infrastructure Companies?

Gimlet's $3 billion valuation sits below several inference-serving competitors, but the comparison reveals an important gap in how the company is communicating its scale. Here's how the major players stack up:

• Fireworks AI: Valued at $17.5 billion with approximately $800 million in annualized revenue as of May 2026, focused on custom model serving and deep customization
• Baseten: Valued at approximately $600 million with nine-figure annualized revenue as of Q1 2026, emphasizing managed inference and serving reliability
• Together AI: Valuation not fully disclosed, focused on open-source model serving and price competition
• Gimlet Labs: Valued at $3 billion with "billions" in contracted revenue rather than annualized revenue, specializing in multi-silicon disaggregated inference for agentic AI workloads

The key difference is that Gimlet has disclosed "billions of dollars in contracted revenue," which represents signed multi-year commitments rather than revenue already collected and recognized. Every peer above it on valuation has published specific annualized revenue figures, which is why some analysts view Gimlet's story as incomplete.

Who Is Backing Gimlet and Why Does It Matter?

The Series B round included strategic investors beyond Andreessen Horowitz. Existing backers Sapphire Ventures, Menlo Ventures, and Factory all returned, while two new investors with significant industry weight joined: M12, Microsoft's corporate venture arm, and Arm, the chip design company whose instruction-set architecture powers much of the mobile and increasingly server chip market.

Microsoft's participation is particularly notable given that the company has its own cloud infrastructure and AI ambitions. Arm's involvement signals confidence that a multi-chip approach will become standard practice in AI infrastructure, rather than a niche optimization.

What's the Founding Story Behind Gimlet?

Gimlet Labs was founded in 2023 by Zain Asgar, who serves as CEO, alongside Michelle Nguyen, Omid Azizi, Natalie Serrino, and James Bartlett. The founding team previously worked together at Pixie Labs, and the company's core technology grew out of a Stanford University research project on splitting AI compute workloads across heterogeneous hardware.

The company stayed out of public view until October 2025, when it emerged from stealth with its Series A and disclosed eight-figure annualized revenue. Growth has been rapid: the company tripled its customer base between its October 2025 stealth exit and its March 2026 Series A, adding one of the world's top three frontier AI labs and one of the top three cloud hyperscalers as customers in that window. By the Series B announcement, Gimlet said it is scaling its managed infrastructure footprint toward several hundred megawatts of heterogeneous compute capacity.

Why Is a16z Doubling Down on AI Infrastructure?

Gimlet's funding comes as Andreessen Horowitz is significantly expanding its AI and hardware investment strategy. Just days before closing Gimlet's Series B, the firm closed a $1.1 billion Machine Age Fund dedicated solely to hardware investments, including chips, memory, networking, storage, data centers, and robotics. The firm also expanded its fifth growth fund to $8.5 billion in August 2026, up from $6.75 billion when it launched in January.

Hardware startups now account for more than 20 percent of a16z's deal flow, up from a small share just two years ago, according to PitchBook analyst Nick Rescigno. Semiconductor and autonomous-machine companies have raised roughly $100 billion over the past year, reflecting the industry's recognition that AI infrastructure is becoming as important as the models themselves.

What's the Open Question About Gimlet's Long-Term Value?

Despite the strong funding round and impressive customer base, one fundamental question remains unanswered: whether the multi-chip routing layer is worth a standalone $3 billion company or whether it will eventually become a standard feature that Nvidia, AMD, or the hyperscalers themselves ship natively.

Gimlet's pitch to customers is straightforward: because agentic AI workloads chain together many inference calls rather than one, small per-call efficiency gains compound quickly across an entire agent session. The company claims this approach improves latency, throughput, hardware utilization, and power efficiency versus single-architecture serving. Whether that efficiency advantage is defensible long-term, or whether it will eventually become commoditized, will likely determine whether Gimlet becomes a category-defining company or a feature that gets absorbed into larger platforms.