Your compute. Your account. Your bill.
🤫 Agent One runs on hardware a person owns and bursts to a cloud account that is also theirs. Not resold through us, not pooled, not anonymised into somebody else's tenancy. Their key, their quota, their invoice, and a workload we never see. This is our partnership plan for the clouds that would carry it.
Public information only. None of the providers named here is a partner, investor, or customer of ours.
They are not short of capacity. They are short of reach.
The instinct is that we need GPUs and they have GPUs. That is backwards as a pitch, because capacity is the one thing a neocloud is not short of. What none of them can reach today is an individual, and there are a great many individuals.
They are not short of capacity
They are short of demand they can reach. A neocloud sells reserved blocks to buyers with a procurement function. There is no route from that motion to one person running an agent at home, and that is a market none of them currently serves.
Individual demand is counter-cyclical to theirs
Enterprise training jobs are lumpy, scheduled, and fought over. One person's agent bursts for ninety seconds at nine at night. That load shape fills exactly the gaps a reserved-block business leaves behind.
We do not resell them, we route to them
On the free tier and on bring-your-own, the account is the person's and the bill is the person's. The neocloud gains a direct customer relationship rather than a reseller sitting on top of one.
A four-to-six times price spread on identical silicon
Public on-demand H200 rates run from about $2.29 to $13.78 an hour as of August 2026. An agent that can move a workload to whichever account is cheapest for that job, on the person's behalf, is worth real money to the person and real volume to whoever is cheapest that week.
The same silicon, at four to six times the spread.
Every figure below is published by the provider or by a public survey, and dated, because a GPU-hour price without a date is not a number. Rates move; check them before relying on one.
| Provider | H200 / hr | B200 / hr | Terms |
|---|---|---|---|
| Nebius | $4.50 | $5.50 | Per-hour billing. Up to ~35% off on longer commitments and larger blocks. |
| CoreWeave | $6.31 | ~$6.50 | Multi-year committed contracts at roughly 15 to 30% off on-demand. |
| Lambda | n/a | $6.99 | Widely quoted mid-market B200 rate. |
| Market range | $2.29 to $13.78 | $7 to $10 | The spread across all providers. A four to six times spread on the same silicon is the whole opportunity. |
The roadmap
Five steps. The first needs nothing from anyone.
Bring your own account
Ready nowBuilt. Shipping.
The person connects their own Nebius, CoreWeave, AWS, GCP, Azure or Alibaba account. Their key, their quota, their bill. Agent One bursts to it and we never hold the credential in a form we could use ourselves. This needs nothing from any provider, which is why it is first.
What we would ask for: Nothing. It works against a public API today.
A named reference integration
ProposalOne provider, one quarter.
One neocloud we build against properly and document properly, so the connect flow is two taps rather than a key paste. Published as a reference so every other provider can see exactly what integrating looks like.
What we would ask for: A technical contact and a test account.
Wholesale capacity behind the paid rungs
ProposalFollows the ladder shipping.
Plus through Ludicrous sell capacity. That capacity is bought wholesale, and committed volume is where a neocloud's 15 to 35% commitment discount becomes the margin that makes the rung work. This is the only part where we are a customer rather than a router.
What we would ask for: Committed-volume pricing, and honest terms on what happens if our volume misses.
Route to the cheapest account for the job
ProposalAfter 01, and only with consent.
When somebody has more than one account connected, the agent picks the one that is cheapest or fastest for that specific workload, and shows them why afterwards. Providers compete on price and performance for each job rather than for an annual contract, which is better for the person and honest for everyone.
What we would ask for: A published rate endpoint, or agreement that we may read the public price page.
The edge grid, meeting the cloud
ProposalLong horizon.
🤫 Puppy One owners host compute in garages and warehouses on Starlink. That is not a replacement for a data centre and we will not pretend it is. It is a different cost curve for inference close to the person, and the interesting question is which workloads genuinely belong at which end.
What we would ask for: A technical conversation about where the line actually falls.
Who we would work with, and what each is best at.
If the person does not own the account, it is not bring your own compute.
On the free tier and on every bring-your-own path, the cloud account belongs to the person. We hold their credential the way an agent holds a key, scoped and revocable, and we never see the workload. The moment we pool people into our own tenancy to save money, they stop owning their compute and we become one more company running their data on our hardware, which is the thing this whole company exists to argue against.
The paid rungs are the one exception and they are labelled as such: there, the capacity is ours and we are reselling it. A person who wants neither can stay on Free forever, which remains the whole agent.
Sources.
- Nebius published GPU pricing
- Nebius compute pricing documentation
- CoreWeave vs Nebius GPU cloud pricing comparison, 2026
- H200 provider pricing survey, 2026
- B200 cloud pricing survey, 2026
Rates as published in August 2026 and they move constantly. Nobody named on this page is a partner, investor, or customer of ours, and nothing here has been agreed with any of them.