A chip orchestration layer just became more valuable than most chip companies.
The Summary
- Gimlet raised $300 million at a $3 billion valuation, led by Andreessen Horowitz, to build software that routes AI workloads across different chip architectures
- The company doesn't make chips. It makes the intelligence layer that decides which chip does what work, when
- This validates the thesis that the real money in AI infrastructure isn't in silicon, it's in the orchestration software that makes silicon work together
The Signal
Gimlet's $3 billion valuation marks a shift in where venture capital thinks the AI infrastructure stack creates value. The company builds software that divides computational tasks across GPUs, TPUs, custom AI accelerators, and even CPUs based on workload requirements, cost, and availability. Think of it as a smart load balancer for the heterogeneous chip world we're entering.
The timing matters. Nvidia's H100 shortage forced companies to get creative about compute. Instead of waiting months for the "right" chip, teams started using whatever silicon they could get their hands on. That created a coordination problem: how do you split a training run across Nvidia, AMD, and Google chips without rewriting your entire stack?
"The real bottleneck in AI isn't chip supply anymore. It's chip utilization across a fragmented landscape."
Gimlet solves this by sitting between the model and the metal. Developers write to Gimlet's API. Gimlet figures out which parts of the job run best on which architecture, handles the data movement, and manages the orchestration. The result: companies can use 40-60% less compute for the same job by routing tasks to the cheapest available silicon that can handle them.
Here's why a $3 billion valuation makes sense:
- Cloud providers spend billions on chips that sit idle 30-40% of the time because workloads are optimized for specific architectures
- Every AI company is now hedging chip vendor risk after watching Nvidia's pricing power
- The semiconductor market is fragmenting fast, with new AI accelerators from Amazon, Microsoft, Meta, and a dozen startups
The $300 million round brings Gimlet's total raised to $475 million. Andreessen Horowitz led with participation from Sequoia and Lightspeed. The company declined to share revenue numbers, but sources familiar with the deal say annualized run rate crossed $100 million in Q2 2026, primarily from contracts with three hyperscalers and a handful of large AI labs.
What Gimlet is really selling is infrastructure independence. If your AI stack is locked to one chip vendor, you have no negotiating leverage and no fallback if supply tightens. Gimlet makes multi-vendor strategies practical without requiring teams to become experts in six different chip architectures.
The Implication
Watch for the orchestration layer to become the new chokepoint in AI infrastructure. Whoever controls how workloads route across chips controls pricing power, vendor relationships, and ultimately which chip companies survive. Gimlet's valuation suggests VCs believe software will capture more value than hardware in the next phase of the AI build-out.
If you're building AI products, this matters immediately. Chip-agnostic infrastructure isn't a nice-to-have anymore. It's table stakes for managing costs and avoiding vendor lock-in as the compute landscape fractures.