# Saturn Cloud adds Lilac's idle GPU network for per-token inference > Source: > Published: 2026-07-15 18:41:22+00:00 [Saturn Cloud](https://saturncloud.io/?ref=runtimewire) founder Sebastian Metti [announced a partnership](https://www.prnewswire.com/news-releases/saturn-cloud-partners-with-lilac-to-bring-enterprise-gpu-capacity-into-its-token-factory-platform-302826592.html?ref=runtimewire) on July 15 with [Lilac](https://getlilac.com/?ref=runtimewire), adding the YC-backed GPU network as a capacity provider for Saturn's token factory platform. The deal gives Saturn Cloud customers a way to route inference workloads onto GPU capacity Lilac says is already powered on inside enterprises and GPU clusters. The product pitch is direct: model serving through an OpenAI-compatible API, per-token pricing, and no reserved GPU commitment for customers whose inference demand moves up and down through the day. Metti framed the move around an operator's least forgiving metric: paid hardware sitting unused. "Operators want every usable GPU hour working, and a lot of capacity sits idle inside enterprises that never reaches the teams who need it," Metti said in the announcement. Selecting Lilac, he said, lets Saturn route token factory workloads to that capacity so customers can get serving and per-token inference without waiting for new infrastructure. That is the strategic reason this partnership matters for Saturn. Metti has been positioning Saturn Cloud as the control plane for GPU clouds, a layer above bare metal, Kubernetes and Slurm that lets operators sell higher-level AI services instead of raw GPU access. Saturn's own site tells operators to "turn your GPU fleet into a Token Factory" and lists managed inference, multi-tenant isolation, day-2 support, integrated billing, and white-label deployment as core pieces of the stack. The company says more than 100,000 developers use Saturn Cloud in production, a company-published figure that is not independently broken down by customer type or paid usage. ### Saturn wants to make GPU resale look like software Saturn Cloud's platform is built for GPU cloud operators, AI factory operators, and enterprises that want to expose compute as an internal or external product. The company describes the token factory as a productized inference business: customers hit OpenAI-compatible endpoints, models run on an operator's GPUs, and billing happens per token. Its platform also covers training, fine-tuning, GPU dev workspaces, metering, chargeback, security and governance. That matters because GPU operators face a harder business than renting out servers. Raw GPU-hours are increasingly commoditized, while inference buyers want endpoints, model catalogs, autoscaling, logs, predictable billing, isolation, and support. Saturn Cloud is trying to sit between the infrastructure owner and the application team, giving the owner a way to charge for model output rather than for bare capacity. Lilac extends that play by adding a supply source Saturn does not have to own. Lilac describes itself as a network of idle GPUs where customers can run inference, reserve clusters, fine-tune models, and process batch jobs on unused capacity. On its site, Lilac says users can point the OpenAI SDK at Lilac for pay-per-token open model inference with no contracts or minimums. It also says suppliers can install a Kubernetes operator, keep their own jobs first, and earn from idle GPUs that stay inside their infrastructure. The practical fit is bursty inference. A Saturn customer can route a serving workload to Lilac capacity when traffic rises, then release that capacity when demand falls. For a neocloud, that means offering inference to tenants without holding more reserved GPUs on its own balance sheet. For an enterprise AI team, it means testing or serving models without asking procurement for a standing hardware block. ### Lilac is young, and its supply depth is the proof point Lilac is still early. [Y Combinator lists Lilac](https://www.ycombinator.com/companies/lilac?ref=runtimewire) as founded in 2025, active, and a four-person team, and names Ryan Ewing and Lucas Ewing as founders. The brothers describe Lilac as a way to tap idle GPUs from cloud providers and enterprises, giving startups and researchers cheaper inference while helping companies monetize unused capacity. That founder story matches the product. Ryan and Lucas Ewing are building a market around a familiar infrastructure gap: engineering teams buy or reserve high-end GPUs, then struggle to keep them busy across changing workloads. The harder part is proving enough trusted capacity exists, across enough regions and service levels, for production inference buyers to rely on it. The announcement does not disclose how many GPUs Lilac can route, how many enterprise suppliers are live, which regions are available, what uptime commitments apply, or how pricing compares inside Saturn Cloud's customer contracts. Lilac's public site lists sample model and batch pricing, including pay-per-token model listings, but Saturn did not publish Saturn-specific Lilac prices in the partnership announcement. That omission is important because [inference infrastructure](/article/megaport-s-594-million-raise-puts-a-price-on-ai-inference-infrastructure) is becoming a capital race and a trust race at the same time. [Baseten said](https://www.baseten.co/blog/announcing-our-series-f/?ref=runtimewire) in June that it raised a $1.5 billion Series F. [Together AI announced](https://www.together.ai/blog/announcing-our-series-c?ref=runtimewire) an $800 million Series C in July. [Runpod said](https://www.prnewswire.com/news-releases/runpod-raises-100m-led-by-summit-partners-to-accelerate-the-ai-developer-cloud-302808689.html?ref=runtimewire) on June 24 that it raised $100 million in a Series A led by Summit Partners. Saturn and Lilac are taking a different path from those balance-sheet-heavy models. Saturn is selling the control plane and billing layer. Lilac is aggregating idle supply. The benefit is lower capital intensity. The risk is that buyers of production inference still care about availability, latency, data controls, and support when traffic spikes. ### The timing follows the utilization problem The timing is not random. AI infrastructure companies have spent the last year turning inference from a feature into a business line, and the economics favor software that squeezes more output from GPUs already in racks. [Cast AI's 2026 Kubernetes optimization report](https://cast.ai/press-release/2026-state-of-kubernetes-optimization-report/?ref=runtimewire) said average GPU utilization across non-optimized Kubernetes clusters was 5%. Cast is an optimization vendor with its own commercial interest in the finding, but the direction of the problem is visible across the market: GPU scarcity and GPU waste can exist at the same time. NVIDIA has been making the same economic argument from the software layer. In July, [NVIDIA said](https://nvidianews.nvidia.com/news/dynamo-1-0?ref=runtimewire) Dynamo 1.0 had entered production as an inference operating system for AI factories, aimed at orchestrating GPU and memory resources across clusters. Saturn's Lilac integration sits downstream of that same shift. The value is no longer just having H100s, H200s, B200s, or B300s available. The value is routing the right request to the right capacity, measuring the output, and billing for it cleanly. Saturn Cloud's bet is that GPU owners will need a commercial layer that looks less like cloud hosting and more like a software marketplace for tokens. Lilac's bet is that the cheapest next GPU for many workloads is the one an enterprise already bought and left idle. Together, the companies are asking operators to believe a fragmented pool of unused enterprise capacity can be turned into a product customers trust for inference. That is a founder-friendly bet because it attacks waste before asking for more hardware. It is also a difficult operational bet. If Saturn can make Lilac capacity feel like a normal part of its token factory, Metti gets a bigger supply pool without becoming a GPU landlord. If Lilac can keep enough idle GPUs warm, reliable, and compliant, Ryan and Lucas Ewing get distribution into the operators Saturn already wants to serve.