The team behind continuous batching says your idle GPUs should be running inference, not sitting dark (venturebeat.com)
<p>Every GPU cluster has dead time. Training jobs finish, workloads shift and hardware sits dark while power and cooling costs keep running. For neocloud operators, those empty cycles are lost margin.</p><p>The obvious workaround is spot GPU markets — renting spare capacity to whoever needs it. But spot instances mean the cloud vendor is still the one doing the renting, and engineers buying that capacity are still paying for raw compute with no inference stack attached. </p><p>FriendliAI's answer is different: run inference directly on the unused hardware, optimize for token throughput, and split the revenue with the operator. FriendliAI was founded by Byung-Gon Chun, the researcher whose paper on continuous batching became foundational to vLLM, the open source inference engine used across most production deployments today.</p><p>Chun spent over a decade as a professor at Seoul National University studying efficient execution of machine learning models at scale. That research produced a paper called<a href="https://www.usenix.org/conference/osdi22/presentation/yu"> <u>Orca</u></a>, which introduced continuous batching. The technique processes inference requests dynamically rather than waiting to fill a fixed batch before executing. It is now industry standard and is the core mechanism inside vLLM.</p><p>This week, FriendliAI is launching a new platform called InferenceSense. Just as publishers use Google AdSense to monetize unsold ad inventory, neocloud operators can use InferenceSense to fill unused GPU cycles with paid AI inference workloads and collect a share of the token revenue. The operator's own jobs always take priority — the moment a scheduler reclaims a GPU, InferenceSense yields.</p><p>"What we are providing is that instead of letting GPUs be idle, by running inferences they can monetize those idle GPUs," Chun told VentureBeat.</p><h2>How a Seoul National University lab built the engine inside vLLM</h2><p>Chun founded FriendliAI in 2021, before most of the industry had shifted attention from training to inference. The company's primary product is a dedicated inference endpoint service for AI startups and enterprises running open-weight models. FriendliAI also appears as a deployment option on Hugging Face alongside Azure, AWS and GCP, and currently supports more than 500,000 open-weight models from the platform.</p><p>InferenceSense now extends that inference engine to the capacity problem GPU operators face between workloads.</p><h2>How it works</h2><p>InferenceSense runs on top of Kubernetes, which most neocloud operators are already using for resource orchestration. An operator allocates a pool of GPUs to a Kubernetes cluster managed by FriendliAI — declaring which nodes are available and under what conditions they can be reclaimed. Idle detection runs through Kubernetes itself.</p><p>"We have our own orchestrator that runs on the GPUs of these neocloud — or just cloud — vendors," Chun said. "We definitely take advantage of Kubernetes, but the software running on top is a really highly optimized inference stack."</p><p>When GPUs are unused, InferenceSense spins up isolated containers serving paid inference workloads on open-weight models including DeepSeek, Qwen, Kimi, GLM and MiniMax. When the operator's scheduler needs hardware back, the inference workloads are preempted and GPUs are returned. FriendliAI says the handoff happens within seconds.</p><p>Demand is aggregated through FriendliAI's direct clients and through inference aggregators like OpenRouter. The operator supplies the capacity; FriendliAI handles the demand pipeline, model optimization and serving stack. There are no upfront fees and no minimum commitments. A real-time dashboard shows operators which models are running, tokens being processed and revenue accrued.</p><h2>Why token throughput beats raw capacity rental</h2><p>Spot GPU markets from providers like CoreWeave, Lambda Labs and RunPod involve the cloud vendor renting out its own hardware to a third party. InferenceSense runs on hardware the neocloud operator already owns, with the operator defining which nodes participate and setting scheduling agreements with FriendliAI in advance. The distinction matters: spot markets monetize capacity, InferenceSense monetizes tokens.</p><p>Token throughput per GPU-hour determines how much InferenceSense can actually earn during unused windows. FriendliAI claims its engine delivers two to three times the throughput of a standard vLLM deployment, though Chun notes the figure varies by workload type.
Most competing inference stacks are built on Python-based open source frameworks. FriendliAI's engine is written in C++ and uses custom GPU kernels rather than Nvidia's cuDNN library. The company has built its own model representation layer for partitioning and executing models across hardware, with its own implementations of speculative decoding, quantization and KV-cache management.</p><p>Since FriendliAI's engine processes more tokens per GPU-hour than a standard vLLM stack, operators should generate more revenue per unused cycle than they could by standing up their own inference service. </p><h2>What AI engineers evaluating inference costs should watch</h2><p>For AI engineers evaluating where to run inference workloads, the neocloud versus hyperscaler decision has typically come down to price and availability.</p><p>InferenceSense adds a new consideration: if neoclouds can monetize idle capacity through inference, they have more economic incentive to keep token prices competitive.</p><p>That is not a reason to change infrastructure decisions today — it is still early. But engineers tracking total inference cost should watch whether neocloud adoption of platforms like InferenceSense puts downward pressure on API pricing for models like DeepSeek and Qwen over the next 12 months.
"When we have more efficient suppliers, the overall cost will go down," Chun said. "With InferenceSense we can contribute to making those models cheaper."</p>
Most competing inference stacks are built on Python-based open source frameworks. FriendliAI's engine is written in C++ and uses custom GPU kernels rather than Nvidia's cuDNN library. The company has built its own model representation layer for partitioning and executing models across hardware, with its own implementations of speculative decoding, quantization and KV-cache management.</p><p>Since FriendliAI's engine processes more tokens per GPU-hour than a standard vLLM stack, operators should generate more revenue per unused cycle than they could by standing up their own inference service. </p><h2>What AI engineers evaluating inference costs should watch</h2><p>For AI engineers evaluating where to run inference workloads, the neocloud versus hyperscaler decision has typically come down to price and availability.</p><p>InferenceSense adds a new consideration: if neoclouds can monetize idle capacity through inference, they have more economic incentive to keep token prices competitive.</p><p>That is not a reason to change infrastructure decisions today — it is still early. But engineers tracking total inference cost should watch whether neocloud adoption of platforms like InferenceSense puts downward pressure on API pricing for models like DeepSeek and Qwen over the next 12 months.
"When we have more efficient suppliers, the overall cost will go down," Chun said. "With InferenceSense we can contribute to making those models cheaper."</p>
Comments