Manifesto · October 2026
By the end of 2028, AI inference is on course to become a bigger market than databases¹. We think it will define software the way the database has for decades.
Europe’s AI compute will grow tenfold by 2031 and still hold about 5% of the world’s total². In the largest hubs, a new grid connection takes seven to ten years³. Power is becoming the limit on AI everywhere. Europe hits it first. Every GPU already connected can do more in the meantime. That is our work, and Europe is where we start.
1 The workload has changed
GPUs were designed for throughput. They give up speed on any single task so that a large batch of work finishes sooner. The defaults in every serving stack work against this. They assume chat, where one person waits for the next word, so servers keep batches small and much of each GPU’s compute goes unused.
Coding agents, security scanning agents and overnight code review pipelines run for hours with nobody watching. Work like this can wait a little, so the GPU can run full batches, as it was designed to. On OpenRouter, the largest network that publishes its traffic, programming passed half of paid tokens by the end of 2025⁴. Open-weight models now let organizations run this work on their own hardware.
This work has deadlines in hours or days, so time to first token matters little. Throughput matters more, and we take it one step further: finished work per watt, meaning tasks that pass their checks inside their deadline, per watt of grid connection.
their deadline, per daywatts of grid connection
2 We know this workload because we generate it
We come from Fiberplane, where we run autonomous software factories and background coding agents in production. We have measured this traffic down to what each request does to the hardware. We know which steps are urgent and which can wait. We know how the load rises and falls over a day, and how much power the GPUs draw between tasks.
3 What we are building
Most serving stacks are assembled from parts built by different teams, each tuned against its own benchmark. We’re building a vertically integrated stack, tuned as a whole for the work it will serve.
Each GPU finishes more tasks, because the kernels and the engine are tuned for agent work. Work that can wait moves to the hours when the fleet would otherwise sit idle. It can also move to smaller sites outside the big hubs, including racks with poor uptime, because a task that fails can restart somewhere else. The result is more finished tasks per megawatt already connected, without raising the peak the grid connection allows.
The aim is to put every GPU that is short of power to work, wherever it sits.
If you own GPU capacity, run agents at scale, or fund the companies that do, talk to us.
¹ Tunguz, “Inference Is the Most Important Market in Software”, Oct 2026, citing Gartner: about $130B inference in 2026 and about $350B in 2027, against $190B for databases in 2027.
² Europe 2031 compute forecast: about 2 GW at end-2026 to almost 21 GW in 2031; share about 5% now and slightly above 5% in 2031.
³ IEA, grid connection queues in Frankfurt, London, Amsterdam, Paris and Dublin, Nov 2025.
⁴ OpenRouter, State of AI, Dec 2025.
