Gimlet Labs, a three-year-old startup, just raised $300 million in Series B funding that values the company at $3 billion, more than tripling its valuation in just six months. The San Francisco-based company sells software that splits AI inference workloads across different types of chips,Nvidia GPUs, Google TPUs, AMD accelerators, and Amazon's custom silicon,rather than running everything on a single architecture. The funding round, led by Andreessen Horowitz and closed on September 4, 2026, signals that enterprises and AI labs are willing to pay for smarter ways to run AI models at scale.

What Problem Is Gimlet Actually Solving?

Running an AI model in production involves two distinct phases with very different computing needs. The first phase, called "prefill," processes the user's prompt and requires raw computational throughput. The second phase, called "decode," generates output tokens one at a time and is bottlenecked by memory bandwidth rather than processing power. Most AI inference platforms run both phases on the same GPU fleet, typically Nvidia's, even though this approach wastes resources.

Gimlet's insight is straightforward: different hardware excels at different tasks. Nvidia GPUs might be optimal for prefill, but cheaper chips with faster memory access might handle decode more efficiently. By disaggregating these two phases and routing each to whichever chip architecture handles it best, Gimlet claims to improve latency, throughput, hardware utilization, and power efficiency compared to single-architecture serving. The company specifically targets agentic AI workloads, which chain together many small model calls and tool uses rather than generating a single chatbot response. In these scenarios, small per-call efficiency gains compound quickly across an entire agent session.

How Does Gimlet's Multi-Silicon Strategy Work?

• Workload Routing: Gimlet's software analyzes each inference request and routes the prefill phase to whichever chip architecture offers the best throughput, then routes the decode phase to whichever chip offers the fastest memory access.
• Hardware Diversity: The company's six-chip strategy includes Nvidia GPUs, Google TPUs, AMD accelerators, Amazon Trainium chips, Amazon Inferentia chips, and other custom silicon, allowing customers to use whatever hardware they already own.
• Agentic Optimization: The platform is specifically designed for AI agents that make multiple sequential model calls, where efficiency gains in each individual call compound across the entire agent workflow.

The practical appeal is clear: before Gimlet existed, frontier AI labs and hyperscalers with enough engineering talent were already building similar routing tools in-house. Gimlet's bet is that turning that internal tooling into a standalone managed service is valuable enough to support an independent company, rather than a feature that hyperscalers will eventually ship for free.

Who Is Backing Gimlet, and What Does That Signal?

The Series B round attracted heavyweight investors beyond Andreessen Horowitz. Existing backers Sapphire Ventures, Menlo Ventures, and Factory all returned, joined by two strategically significant new investors: M12, Microsoft's corporate venture arm, and Arm, the chip design firm whose instruction-set architecture underpins much of the mobile and server chip market. This investor mix suggests confidence that multi-silicon inference is not a niche problem but a structural shift in how AI infrastructure will be built.

Gimlet's customer base reinforces this narrative. The company has locked in billions of dollars in contracted revenue from one of the world's top three frontier AI labs and one of the top three cloud hyperscalers, though neither has been named publicly. Between its October 2025 stealth exit and its March 2026 Series A, Gimlet tripled its customer base and added these marquee customers. By the Series B, the company says it is scaling its managed infrastructure footprint toward several hundred megawatts of heterogeneous compute capacity.

How Fast Did Gimlet Grow, and What Does That Tell Us?

Gimlet's valuation trajectory is striking even by 2026 AI funding standards. The company was founded in 2023 by Zain Asgar, 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. Gimlet remained in stealth until October 2025, when it emerged with its Series A and disclosed eight-figure annualized revenue. The March 2026 Series A valued the company at roughly $980 million, meaning the September 2026 Series B represents a better than 3x markup in approximately six months.

This compressed timeline reflects a real inflection in how enterprises spend on AI infrastructure. Gartner forecasts that worldwide AI-optimized infrastructure-as-a-service spending will grow 96 percent in 2026 to $42.3 billion. For the first time, global spending on inference, projected at $23.3 billion, will exceed spending on training, projected at $19 billion, in 2026. That shift from paying to build models to paying to run them at scale is exactly the budget line Gimlet, Baseten, Fireworks AI, and Together AI are all competing for.

How Does Gimlet Compare to Other AI Inference Startups?

Gimlet's $3 billion valuation sits well below several inference-serving peers that have disclosed harder revenue numbers. Fireworks AI has reached a $17.5 billion valuation on the back of roughly $800 million in annualized revenue as of May 2026. Baseten crossed a multibillion-dollar valuation on roughly $600 million in annualized revenue as of Q1 2026. Together AI and Groq have not fully disclosed their revenue figures, though Groq was valued at approximately $1.1 billion in 2025.

The comparison exposes a gap in Gimlet's public narrative: every peer above it on valuation has published a specific annualized revenue figure, while Gimlet's largest disclosed number is a contracted-revenue total with no breakdown of actual revenue recognized. Contracted revenue represents signed multi-year commitments rather than revenue already collected and recognized, making it harder to compare Gimlet's scale directly to competitors.

Gimlet's positioning differs from its peers in another way. Fireworks AI emphasizes custom model serving and deep customization. Baseten focuses on managed inference and serving reliability. Together AI competes on open-source model serving and price. Gimlet's unique angle is multi-silicon disaggregated inference specifically optimized for agentic AI workloads, a narrower but potentially higher-margin niche.

What's the Open Question About Gimlet's Future?

The practical problem Gimlet is pointing at is real: Nvidia GPUs remain the default choice for both prefill and decode phases of inference even though the two phases have different hardware requirements. Building a routing layer across chip architectures from Nvidia, AMD, Google, and Amazon is hard engineering that most AI labs and enterprises would rather buy than build in-house. Whether that routing layer is worth a standalone $3 billion company, versus a feature Nvidia, AMD, or the hyperscalers themselves eventually ship natively, is the open question the next funding round will have to answer.