Lambda exits GPU hardware sales
No side quests for Lambda
This marked Lambda becoming a pure compute utility, not a hardware seller. Once the same H100 or B200 could either be boxed up once for a fixed gross profit or rented repeatedly into large training clusters, the economics favored keeping every scarce GPU on balance sheet. That also fit Lambda’s shift from university labs and developers toward multi year contracts with Nvidia, Microsoft, and other large buyers that can keep clusters full around the clock.
-
Lambda’s hardware roots were real, with the workstation and server business reaching about $100M by 2024 and serving 97% of top U.S. research universities. Shutting it down was not trimming a failed side project, it was redeploying scarce chips from lower lifetime value sales into a much bigger cloud revenue engine.
-
The market was already moving toward reserved capacity. Earlier research showed CoreWeave winning enterprises that could commit to thousands of GPUs on yearly or multi year deals, while Lambda served more flexible growth stage teams. By late 2025, Lambda moved decisively toward the CoreWeave playbook, concentrating on predictable utilization and anchor contracts.
-
Selling hardware ties revenue to one shipment. Renting GPUs creates an annuity as long as demand stays tight. That helps explain why Lambda at $760M annualized revenue and a 7.8x multiple was being compared to CoreWeave at $5.13B revenue and roughly 12x, because public markets reward scaled, contract backed infrastructure more than transactional hardware resale.
The next step is a tighter split across the stack. Large neoclouds like Lambda and CoreWeave will keep pushing scarce GPUs into dense training and hyperscaler contracts, while self serve inference and developer workflows keep drifting to companies like RunPod, Modal, and Baseten that package smaller chunks of compute into easier to consume software products.