An airline can tell you exactly what an empty seat costs: everything. The plane flies either way, the fuel burns either way, and the moment the doors close, the unsold inventory evaporates. Data center capacity behaves the same way, except the plane costs billions, the fleet depreciates on a schedule set by Nvidia’s product roadmap, and nobody is quite sure how many passengers are showing up in 2028.
That is the uncomfortable thought sitting underneath this week’s news. Bloomberg reported on October 8 that demand for the IPO of Firmus Grid Ltd., an Nvidia-backed data center company, cratered — enough to raise real questions about how long the AI funding boom can keep refinancing itself. It landed alongside the numbers from Nscale, the British AI cloud provider that filed for a U.S. listing on September 18 with revenue up 1,252% in the first half of 2026 and a net loss of $1.02 billion over the same six months, against $368.9 million a year earlier.
Both numbers are real. Both are growing fast. One is growing faster.
Growth and losses are not in tension here, which is the problem
The reflexive read is that Nscale is a bad business dressed up in a good narrative. I do not think that is right, and I think the easy dismissal obscures something more interesting about how AI infrastructure actually gets financed.
A 1,252% revenue increase alongside a billion-dollar loss is not a contradiction. It is what happens when you buy capacity years before the revenue arrives. You take on the capital expense, the power contracts, the networking build-out, and the depreciation schedule up front, and you sell against it later. The losses are the build. The revenue is the early tenants.
Which makes the entire structure a bet on one specific claim: that demand for AI compute in 2028 will be larger, more durable, and less price-sensitive than the demand you can see today. Public market investors are being asked to underwrite that claim with their own money, and Firmus Grid’s reception suggests some of them have started asking harder questions about the assumption rather than the operator.
Where agent architecture enters the picture
This is the part that interests me most, and it is the part the financial coverage tends to skip. The demand curve for AI compute is not a single thing. It is at least two things with very different shapes.
- Training is schedulable, bulk, and tolerant of latency. You can queue it, move it to cheaper power, run it overnight. It fills capacity predictably, which is exactly what a landlord wants.
- Agentic inference is none of those things. A multi-step agent holds context across dozens of tool calls, waits on external systems, retries, backtracks, and needs low latency throughout. It is spiky, stateful, and expensive to serve well.
The bull case for infrastructure leans heavily on agents, because agents consume dramatically more tokens per completed task than a single chat turn ever did. That part is true. But the same shift introduces a risk that pure training demand never had: agent workloads are enormously sensitive to architecture, and architecture improves fast.
Aggressive context caching, better retrieval so the model reads less, model routing that sends easy sub-steps to small distilled models, planners that stop an agent from re-deriving the same conclusion four times. None of these are speculative. They are the normal work of building agent systems, and each one cuts compute per completed task. Stack them and the reduction is not a few percent.
So you can have task volume growing explosively and compute per task falling just as fast. Total demand still rises, probably a lot. But it rises less predictably than a spreadsheet built on today’s efficiency assumes, and the financing structure has no tolerance for that kind of surprise. Depreciation does not wait for your utilization forecast to be correct.
Supply constraints cut both ways
The counterweight, and it is a serious one, is that supply remains genuinely tight. Bloomberg’s coverage has flagged networking as a potential bottleneck that limits how much AI compute can actually be delivered, regardless of how many chips exist on paper. An operator who controls power, interconnect, and floor space in the right places has something scarce.
Scarcity is a fine position to hold. It is a dangerous one to borrow against at high multiples, because scarcity resolves. Bottlenecks get engineered around. The firms that locked in capacity at boom-era costs will be competing against firms that build later and cheaper.
What the signal actually means
A single weak IPO is not a verdict on AI. Public markets are a lagging, noisy instrument, and they are pricing financing risk, not technical capability. The models keep improving. The agent systems being built on them are getting genuinely more capable.
What changed is that capital has started distinguishing between the technology and the balance sheets wrapped around it. For those of us building agent architectures, that is a useful discipline rather than a warning. Efficiency per task was always the right engineering target. It turns out it is also the variable that decides which infrastructure bets survive.
đź•’ Published: