Gimlet Labs’ $300M Raise and OpenAI’s Unpaid Forecast

In September 2026, Bloomberg reported that Gimlet Labs had raised $300 million at a $3 billion valuation. Five days later, The Information reported that Gimlet had told investors OpenAI would likely spend more than $100 million a year on its cloud; OpenAI said it was “not yet” a paying customer.

In 2018, AWS introduced Amazon Elastic Inference, which let customers add GPU-powered acceleration to EC2 instances. AWS said the service could reduce deep-learning costs by as much as 75%. In 2020, Grid AI raised $18.6 million to help researchers scale models toward enterprise workloads. Both companies treated inference as something the existing cloud needed to fit more efficiently. By 2026, AI infrastructure providers were raising capital to assemble and orchestrate a separate compute layer before agent-driven workloads had settled into durable purchasing patterns.

Neocloud underwriters increasingly demand contracted, paid, durable utilization and enough realized margin to service hardware debt after inference costs change.

A forecast can support valuation before revenue

Gimlet Labs isolates the distinction. A16z led the round only six months after the company raised an $80 million Series A. The reported OpenAI forecast gave investors a possible anchor workload without establishing paid OpenAI usage.

OpenAI’s qualifier matters in both directions. The words “not yet” leave room for a commercial relationship or later spending, and a frontier lab’s anticipated demand can be genuine. Still, Gimlet’s investor forecast, a signed customer obligation and revenue collected from recurring workloads remain three different documents. A forecast can support an equity valuation; a contract can support a capacity decision; only a paid workload reveals how much of the fleet customers consume and what margin survives.

The four reports established investor appetite but disclosed no paid usage, realized utilization or gross margin.

Orchestration makes utilization the product

Gimlet describes its service as a multi-silicon inference cloud that divides AI tasks among chip types. To make heterogeneous AI compute pay, Gimlet must place each workload on hardware that meets its performance requirement at a cost the customer accepts, then keep enough of that hardware occupied to recover the capital committed to it.

Agent software complicates the placement decision. In interactive inference, a person notices delay immediately. In agentic inference, where a human may not be waiting for each result, a provider can weigh speed differently and optimize more aggressively for cost. The workload mix therefore determines the value of a multi-silicon fleet: one job may reward the fastest available chip, while another may tolerate delay if cheaper hardware completes it economically.

Documents reported by the Wall Street Journal put inference costs at more than half of OpenAI and Anthropic revenue. Months later, OpenAI engineers reportedly found a way to more than halve inference cost.

If Gimlet retains a technical saving, its margin widens; if buyers capture it through lower prices, their spending falls. Gimlet’s orchestration layer creates value only when customers redesign workloads around those trade-offs and continue buying the completed work. An idle accelerator remains idle even when software can describe it elegantly.

Contracts put a customer on the other side of the build

Crusoe supplies stronger demand evidence through confirmed contracts to provide computing power to Meta Platforms and Oracle. Meta reportedly contracted for roughly 1.6 gigawatts across two data centers in Texas and Missouri. The Meta agreement names a buyer, a quantity and physical sites. Those terms give Crusoe more usable evidence for a build decision than strategic interest alone.

OpenAI has also shown how a supplier can formalize anticipated demand. Its Guaranteed Capacity offering gives customers discounted compute access in exchange for one- to three-year commitments based on spend levels. When customers sign those commitments, OpenAI converts an expectation into an obligation and reduces the risk that reserved capacity waits for demand that arrives late.

A contract still does not establish the provider’s margin, and a concentrated contract can make one buyer disproportionately important. But lenders can assess its duration, committed spend, capacity and counterparty. Those terms are the machinery behind the contracted megawatt, where a legal commitment makes a future workload financeable before the building begins serving it.

Debt gives demand timing a due date

AI infrastructure companies borrowed more than $100 billion in 2025, while smaller companies faced higher interest rates because investors remained wary of unproven AI businesses. Lenders now have to finance, insure and underwrite data centers and specialized hardware as a novel asset class. Equity investors face no scheduled principal payments. A loan’s payment calendar does not move merely because a customer deployment does.

General Compute made the recovery question explicit when Upper90 provided a $400 million loan using inference-specific chips as collateral. The deal was reportedly the first such arrangement. Fluidstack reportedly raised $5.7 billion through junk bonds. Once lenders accept hardware as collateral, they must underwrite both the cash generated while the chips operate and the value they can recover if the expected workloads do not arrive on schedule.

Lenders can recover their principal only if another operator can use the chips and the hardware remains economical after software lowers inference costs. Customers must still pay enough to cover the financing. A provider may allocate every task correctly and still miss debt service if utilization arrives after the payment. That timing risk explains why AI data-center financing increasingly relies on backstops, long-duration agreements and other structures that put named counterparties behind projected demand.

Gimlet’s next test is paid utilization

To judge how Gimlet converts capital into customer cash, an underwriter needs the share of capacity covered by signed commitments, the share producing paid recurring usage, the concentration of that usage, and the time required to shift a workload when one chip becomes less economical than another.

If Gimlet lowers the cost of a task and retains the saving, its margin improves. If the customer receives the saving through lower prices, the same technical achievement can reduce revenue per task. If a financed chip cannot move into another profitable workload, its nominal capacity says little about its ability to service the attached financing.

Gimlet can fill a rack with several kinds of silicon and a data room with OpenAI’s $100 million forecast; an underwriter can count only the workloads customers actually buy.