Prologue
Europe is obsessed with owning intelligence.
Building its own models. Funding research. Pursuing AI sovereignty. Yet intelligence doesn’t live in white papers. It runs in datacenters. We’re moving from asking AI about the weather to asking it to discover drugs, engineer new materials, and build software autonomously.
These aren’t questions that take seconds to answer. They’re missions that can take days to complete.
This is an industrial computing problem, with consequences for productivity and economic growth, constrained by electricity, compute capacity, and the economics of running intelligence at scale.
Owning the models means little without the power and compute to put them to work.
Subnivean is building the infrastructure for industrial-scale intelligence, unlocking more productive capacity from every megawatt of AI compute.
Thesis · November 2026
Intelligence is becoming a production output
The cost of running intelligence is becoming part of the cost of making things. It will shape how much software a team can build and test, which new medicines researchers can pursue, and how much work factories can hand over to robots. How efficiently we run intelligence will shape how much our economies produce and how fast they grow.
ii
Europe’s compute shortage risks becoming a productivity gap.
Europe hosts an estimated 5% of global AI compute1. In its major hubs, new datacenters can wait 7–10 years for a grid connection2. GPU rental prices are rising at major providers, adding to the cost of putting AI to work3. That shortage risks limiting what European businesses can build and produce. We need more capacity, and better utilization of every megawatt already connected.
iii
The workload has changed
The cloud was built for eight billion people. Every layer of it assumes a person is waiting at the other end. A request takes a few hundred milliseconds. Datacenters are placed close to where people live. Traffic peaks at nine in the morning and eight at night, and half the servers sit idle in between. Billing is by the machine hour, because the only expensive thing in the loop is a person’s attention.
In February 2026 the cloud got a second kind of user. An agent’s job runs for minutes to hours, in parallel with hundreds of others, with no one watching it. It has a deadline, not a cursor. A security scan works through every repository overnight. A deep-research run on a drug target reads hundreds of papers and comes back with a shortlist. A personal agent, the kind Muse or Instinct run, keeps a task open for days, waits for replies and finishes when the job is done. Its context is long and branches: a code review forks into the files, the hard cases and the tests, and comes back together. Each turn feeds the next, so most of what the model reads it has read before.
Agents change the scale and shape of compute. Growing demand, long-horizon tasks, branching context and parallel work have opened a new scaling axis for inference, and created new opportunities to organize compute.
Agents passed humans in inference in February 2026 and were doing seven times as many tokens by September. Token demand is expected to grow more than twenty-fold by 2030, most of it agents running in the background. At this scale, the cost of each completed job matters. Building infrastructure around agent workloads will shape how much of the coming demand we can afford to serve.
iv
Building compute around the workload
Agent tasks often have flexible deadlines and reuse the same context across many steps. That gives us room to group work into larger batches and make better use of each GPU. We’re building Subnivean as one integrated stack, from the API and scheduler to the inference engine, GPU kernels and fleet.
v
Make every watt more productive
Building new hardware capacity takes years. Tuning our existing hardware around new workloads can increase what we produce today, tomorrow, and for years to come. That work deserves a place alongside investment in new datacenters and grid connections.
On Fiberplane’s background-agent workloads, which are based on a mix of internal traces and public benchmark tasks, our stack can fit 4x as many of them concurrently (56 vs 224 running agents), running at 3x throughput of the reference architecture (218 tasks per hour, up from 58) on the same Hopper-class GPUs (2022 generation hardware).
This means that with full utilization, a user could run twice as many security scans, even on old hardware, and still end up with a lower bill.
On newer, Blackwell-generation hardware, our stack is seeing even higher throughput gains.
These metrics have far-reaching implications for AI capacity, and we see them as the floor of what’s possible. Every workload we process is a chance to optimize further. The stack will only get better.
vi
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.
Before Fiberplane we built Wercker, a container-native pipeline that ran other companies’ builds and tests on shared machines, by the thousands, with nobody waiting. Agents are that workload with a model in the loop.
vii
Europe needs an industrial base for intelligence.
To meet the challenges of this century, Europe needs an industrial base for intelligence. We cannot ration our ambition because intelligence is too scarce or too expensive.
Europe needs power, compute, and the engineering to make both productive. We need to connect new capacity and get more work from every megawatt already available.
Making intelligence economical and abundant is integral to Europe’s ability to compete and grow.
The payoff will be felt in the EU and around the world. People will test more ideas, build more transformative products, and take on more impactful work. This is only the beginning.
viii
Talk to us
If you own GPU capacity, run agents at scale, or fund the companies that do, talk to us.