A $300 Million Bet Against Single-Vendor AI Chips
Gimlet Labs raised a $300 million Series B, announced September 4, 2026, that values the AI infrastructure startup at $3 billion, roughly EUR 2.7 billion or GBP 2.35 billion. Andreessen Horowitz led the round, joined by Arm, Samsung Ventures, M12, Tiger Global, Sapphire Ventures, Menlo Ventures, and more than a dozen other investors, according to Gimlet's own announcement. The round comes just over five months after Gimlet's Series A, and the company says its customer base has tripled since then, with contracts for its Gimlet Cloud platform now worth billions of dollars combined.
What Gimlet Cloud Actually Does
Gimlet Cloud is built to run a single AI inference workload across several different kinds of chips at once, rather than binding it to one accelerator vendor's hardware. The platform disaggregates the workload across GPUs, CPUs, near-memory compute, and dataflow architectures, and automatically routes each part of the job to whichever chip handles it most efficiently. Gimlet says this delivers five to ten times the inference throughput for the same power footprint compared with infrastructure locked to a single accelerator type, though that figure comes from the company itself and has not yet been independently benchmarked by a third party.
The Same Week, a Mirror Bet on Software
Gimlet's funding lands in the same week Servola reported on Wonderful, an Amsterdam-based startup that raised $550 million at a $5 billion valuation to sell software that coordinates AI agents across a company's existing tools regardless of which model vendor powers them. The two rounds point at the same anxiety from opposite ends of the AI stack: Wonderful sells freedom from picking one model vendor at the software layer, while Gimlet sells freedom from picking one chip vendor at the hardware layer. Investors backed both bets in the same seven days, which is itself a signal that vendor lock-in has become a priced-in risk across the entire AI infrastructure market, not a concern confined to one layer of it.
The Risk Underneath the Anti-Lock-In Pitch
An abstraction layer that promises freedom from any one chip or model vendor still has to be built, priced, and contracted by somebody, and that somebody becomes the new dependency. If Gimlet's routing software becomes the only thing that knows how to split a workload efficiently across mixed hardware, switching away from Gimlet itself becomes exactly the kind of switching cost the company markets itself as eliminating. Nothing in Gimlet's own announcement addresses contract terms, data portability, or what happens to a customer's inference pipeline if the company is acquired or shuts down, questions that matter more now that billions of dollars in contracted revenue are already running through the platform.
What This Means for EU and UK Buyers
Neither Gimlet's throughput claims nor Wonderful's coordination promises have been tested at scale by an independent buyer yet, so an EU or UK organization evaluating either category should ask for its own benchmark before trusting a vendor's own numbers, and should read the exit terms as carefully as the performance claims. Chip supply concentration is already a live procurement risk in Europe, and a genuine multi-silicon option is worth watching, but watching is not the same as committing a production workload to an eighteen-month-old company's routing layer without first confirming what happens on the way out.
Read next: AI's Bottleneck Just Moved From the Chip to the Socket | Nvidia Turns Hugging Face Into a Chokepoint
