Picture a data center floor at 3 a.m. Rows of accelerators are running at near-full utilization, not because a research team queued up a training job before going home, but because somewhere a few thousand agent loops are grinding through tool calls, retries, and self-checks on behalf of users who are asleep. Nobody is watching. The meter is running anyway.
That image is the quiet economic engine behind the news out of Nscale. The AI compute provider is in talks to raise $3.5 billion in pre-IPO financing, per TechCrunch, including as much as $1.5 billion in convertible notes sold to a group of investors. The raise lands weeks after Nscale announced a $45 billion compute deal with Anthropic. The sequencing matters more than either number in isolation.
Why the contract comes before the capital
The traditional order of operations in infrastructure was capital first, customers later. You raised money on a thesis, built capacity, then went hunting for demand. Nscale is doing the reverse: secure an enormous multi-year offtake commitment, then go raise against it. A $45 billion contract with a frontier lab is not a sales pipeline, it is closer to collateral.
Convertible notes fit that shape well. They let a company borrow against near-certain future revenue without pricing equity at a moment when the private market and the public market may disagree sharply about what an AI compute provider is worth. If the IPO goes well, the notes convert and everyone is happy. If it does not, the structure has already limited the damage to the existing cap table. It is a financing instrument designed for a company that believes its revenue is more predictable than its valuation.
The architectural reason demand looks predictable
This is where my interest as a researcher rather than a market watcher kicks in. Compute demand from agent systems has a fundamentally different shape than compute demand from chatbots, and I think that difference is what makes contracts of this size underwritable at all.
A single-turn assistant is bursty and human-paced. One prompt, one response, then a pause while a person reads. Utilization tracks waking hours in whatever time zones you serve. Agentic workloads break that pattern in several ways at once:
- Token amplification. A single user request can expand into dozens or hundreds of model calls once you add planning, tool invocation, observation parsing, and verification steps.
- Long horizons. Agents that operate over hours or days keep consuming inference well after the human has walked away.
- Retry and self-correction overhead. Reflection loops, critic passes, and failed tool calls all cost tokens. Reliability improvements often make workloads more expensive, not cheaper.
- Context growth. Long-running agents accumulate state. Larger contexts mean more memory pressure and more attention compute per step.
- Flatter duty cycles. Scheduled and background agents fill the overnight troughs that used to sit idle.
For a provider selling capacity, that last point is the quiet prize. Idle capacity is the enemy of infrastructure economics. Workloads that run continuously and can tolerate scheduling flexibility are worth more per installed watt than workloads that spike for eight hours and vanish.
What this implies about the buyer’s roadmap
I want to be careful here: the only facts on the table are the raise, the note structure, and the Anthropic deal. Nobody has published a workload breakdown. But committing to compute at the scale reported is a statement about time horizons. You do not sign a contract of that magnitude to serve interactive chat traffic that you could scale elastically on someone else’s cloud. You sign it when you expect sustained, growing, largely non-interactive consumption, and when you want price certainty on the input that dominates your cost structure.
The industry has spent two years arguing about whether inference or training would dominate spending. Agent architectures resolve that argument in an awkward way. They make inference behave like training: long-running, batchable, scheduler-friendly, and enormous.
The risk nobody has priced yet
Long-dated compute contracts assume that today’s efficiency curve holds. That is the exposure. Better routing between small and large models, aggressive caching of intermediate reasoning, distilled specialist models handling routine sub-tasks, and smarter early stopping in reflection loops could each cut tokens per completed task substantially. Stack a few of those together and the compute required per unit of useful agent work drops faster than demand grows.
My honest read is that efficiency gains have so far been absorbed by ambition rather than banked as savings. Every time inference gets cheaper, we spend the surplus on longer horizons and more verification. If that pattern continues, a $45 billion commitment looks conservative. If it breaks, the same commitment becomes a very expensive floor.
Nscale is raising $3.5 billion on the assumption that it continues. That is a reasonable bet on how agent systems are actually built today, and a genuine open question about how they will be built in three years. Anyone designing agent architectures right now is, whether they intend to be or not, a participant in settling it.
🕒 Published: