The @hyperbolic_labs paper runs analysis around 60,000 price paths. The assumption I would spend more time on is that 90% of capacity finds a buyer. This is where coffee feels like a useful comparison, I have been on the receiving end of coffee bean inflation ☕️. A benchmark does not make every coffee bean origin and grade equally saleable. In compute, networking, location and reliability still determine which workloads a cluster can serve. The year three numbers in their analysis show why that matters. At the modelled selling price of roughly $4.90 per billed hour, 90% utilisation earns about $4.41 for every available hour. The full contracted capacity cost is $4.30, including the prepaid deposit. At 85% utilisation, revenue becomes about $4.17 per available hour. That falls below the full capacity cost before paying for the hedge. The selling price has not changed. Just five percentage points of utilisation. My addition would be to let utilisation and customer mix move alongside price. Include a scenario where rates fall and fewer hours sell. Feed those assumptions with actual bookings, repeat usage and reasons customers leave. This can also explain the recent @CoreWeave debt news flow. Building inference products at @ionet has made me pay attention to how demand changes around model launches and new AI tech trends. The workload customers wanted last month may need different capacity or a different service today. That is product work with a direct effect on the economics: which workloads to support, how to package, extract peak performance from the capacity and how much demand actually returns. Coffee traders care about the specific micro lots and its buyers. Compute needs that same attention, with one extra constraint. Unroasted coffee beans can wait in a warehouse for another buyer. An unused GPU hour cannot. Thank you @hyperbolic_labs team on putting this paper out there.

io intelligence now works with four open source projects: Gajae-code (@bloatedaislop) Claude-mem (@Claude_Memory) Vellum-assistant (@vellum_ai) Wakil Pick provider and run on @ionet More integrations are on the way. Building in the open? Give us a shout and we'll help you add your project.

500M tokens a day in March 2026. 15B tokens a day yesterday. That is daily volume across @ionet hosted models on @OpenRouter, served on an almost same mix of H100, H200 and B300 GPUs since Day 1. Along the way, we became day zero launch partners for open source reasoning models including @Zai_org, @Kimi_Moonshot, @Alibaba_Qwen , Nemotron, @MiniMax_AI and GPT OSS. Nvidia has built something useful with the recent Nemotron 3.5 Lightning. A fast, capable and efficient open model, with weights, datasets and recipes that give teams like ours room to optimise. Yesterday, the OpenRouter provider token share showed @ionet at 64% token share for Nemotron 3.5 Lightning, the largest provider share. Effective prices per million tokens: Input: $0.0335 Output: $0.179 Our volumes and price reflect both usage and affordability. We are serving this workload profitably on an inference stack built and operated end to end in house. Our small team still runs two to three experiments every week. We track tool calling changes, frontier lab updates and what communities on vLLM and Sglang are discussing, then test what matters on our own workloads. Demand changes with time zones and days of the week. Launch excitement fades. New providers arrive and price wars start. You get used to revisiting your assumptions. With GPU rental costs rising, compute efficiency and reliable tool calling matter more. A low token price only helps if the task gets done correctly. Open models are becoming capable enough to build around and economical enough to serve at scale. There is a lot of engineering behind making both true at once. Thank you @NVIDIA and the Nemotron team, @OpenRouter and @alexatallah for allowing providers like ourselves to work seamlessly, and our engineers, who know how much work sits behind a neat row in a pricing table.