Cheap AI compute is financed, not earned

You compare two prices. Claude Sonnet 5 costs $10 per million output tokens [5]. DeepSeek V4 Flash, served on Together AI by a company whose hardware you will never see, costs $0.28 per million output tokens [4]. Both numbers are real and both are published. What the invoice does not show is the machinery behind the cheaper one: hardware bought with borrowed money by companies losing money on the quarter [1], increasingly backed by credit support from the chip vendor itself [2].

This guide is for people who buy AI capacity as a line item: a few hundred dollars a month of API tokens, sometimes a rented GPU for a fine-tune or an overnight batch job. It is not for anyone training a frontier model, and it is not procurement advice for a company with a legal team and a vendor risk questionnaire. The question here is narrower. When a compute price looks too good, what should you check, and what should you actually do about the answer.

The chain between someone else’s debt and your monthly bill

A neocloud is a company that borrows money, buys graphics cards, and rents them out by the hour. That is the whole business. The economics only work if the rental contracts outlast the depreciation on the hardware, which is why these companies chase long commitments so hard.

The scale is easier to see with published figures. CoreWeave reported revenue of $2,078 million for the first quarter of 2026, a net loss of $740 million, capital expenditure of $7,695 million in that single quarter, and total debt of $24,859 million as of 31 March 2026 [1]. Against that it reported a revenue backlog of $99.4 billion, which it defines as “remaining performance obligations, plus other amounts we estimate will be recognized as revenue in future periods under committed customer contracts” [1]. That is a company spending roughly four times its quarterly revenue on hardware in the same quarter, funded by debt, against contracts that have not been performed yet.

None of this is hidden or unusual for infrastructure. It matters to you for one reason: a company in that position has a strong structural preference for long, prepaid, committed contracts, and a strong structural need for the price it quotes you today to be low enough to win the contract. Those two forces do not both survive contact with a bad quarter.

The July 2026 program showed how fast the terms move

The money to buy the cards increasingly comes from, or is guaranteed by, the company selling them. In July 2026 Nvidia introduced a program under which it acts as a financial backstop for cloud operators buying its chips, agreeing to rent back unused GPUs at a fixed rate, in return for a share of the operator’s cloud revenue [2]. Nvidia described it as a way for AI clouds to “procure Nvidia infrastructure for AI-native, enterprise, and ISV customers through economic alignment with a revenue-sharing and credit-support model” [2]. Two named early adopters were Firmus, deploying 170,000 GPUs in Batam, Indonesia, and Sharon AI, deploying 40,000 GB300 GPUs [2]. The announcement did not disclose the revenue-share percentages [2].

Eight weeks later, on 28 August 2026, the program was reported paused. A quarterly filing put the arrangement at $36 billion in total commitments, tied to agreements of a typical six-year duration [3]. The sticking point was not the money. Nvidia insisted on vetting its partners’ end customers and mandated that capacity be distributed across multiple smaller AI firms rather than concentrated with single large buyers, and several cloud providers objected that client selection should remain their own decision [3]. Nvidia employees had raised concerns with clients that the tight terms could draw regulatory scrutiny [3]. Nvidia’s response was that “the new business model we introduced in July that opens up compute access to the fast-growing AI ecosystem is still in place and continues to evolve due to high demand” [3].

Whether the program returns in a different shape is not the point for a buyer. The point is the timescale. A financing structure worth $36 billion over six-year terms was announced and paused inside two months, and one of the terms under dispute was who is allowed to rent the capacity. If you had signed a two-year contract with one of those providers in July on the strength of its expansion plans, the ground under that contract moved before your first renewal conversation.

Four things worth checking before a price convinces you

The first is the gap between the on-demand price and the committed price. CoreWeave publishes on-demand rates alongside a statement that it offers “up to 60% discounts over our On-Demand prices for committed usage” [6]. A discount that large is not generosity. It is the price of certainty, and the size of it tells you how much the provider needs certainty.

The second is who paid for the building. Public filings, where they exist, are the fastest read: backlog against debt, and capex against revenue [1]. A private provider gives you nothing to read, which is itself information about how much weight to put on its roadmap.

The third is concentration. CoreWeave’s first-quarter announcement highlighted a $21 billion commitment signed with Meta in March and a multi-year agreement with Anthropic [1]. Contracts of that size are good news for the provider’s solvency and bad news for your position in the queue when capacity is tight, because your workload is the one that is cheap to disappoint.

The fourth is whether the price has already moved once. It has. CoreWeave raised prices by more than 20% in late 2025 and began requiring three-year contracts from smaller customers [7]. Over two months in early 2026, spot pricing for Nvidia Blackwell chips rose 48%, from $2.75 to $4.08 per hour [7]. A provider that has repriced before will reprice again, and the terms it added last time tell you what it will add next time.

A committed contract is where you take on the provider’s risk

Renting by the hour is expensive and safe. Committing for a year or three is cheap and is, functionally, you lending the provider certainty in exchange for a discount. That trade is fine when your usage is steady and terrible when it is not, so the usage figure you sign against has to come off an invoice rather than out of a plan.

The published numbers make the arithmetic concrete. Lambda lists an NVIDIA B200 SXM6 at $6.69 per GPU-hour on demand [8]. CoreWeave lists an 8-GPU HGX B200 node at $68.80 an hour on demand and $34.11 on spot [6], which at 730 hours is roughly $50,000 for a month of continuous running. Against a committed discount of up to 60% [6], the break-even is not a matter of opinion. Run it at the hours you actually used last month, taken from the invoice, not the hours you intend to use.

If the result is close to zero, do not sign. A commitment that barely breaks even at today’s usage is a bet that your usage grows, made with a counterparty whose own terms are unstable.

For most readers the risk arrives as rate limits, not as an invoice

Renting hardware is the wrong shape for most solo operators and small teams. A single 8-GPU node at $68.80 an hour [6] is not a line item you want next to a business that bills clients monthly. Buying tokens is the sane default: Claude Sonnet 5 at $2 per million input and $10 per million output tokens [5], or an open model like gpt-oss-120B served at $0.15 and $0.60 [4].

That does not put you outside the financing story. It changes how the story reaches you. When capacity tightens, the people renting hardware pay more and the people buying tokens get metered. Through March and April 2026, GitHub Copilot introduced new usage limits citing rapid growth and intensive usage, Windsurf replaced its credit system with daily and weekly quotas plus add-on capacity at API prices, OpenAI shifted enterprise Codex billing to token-based metering, and Anthropic adjusted session limits while temporarily offering doubled usage off-peak [7]. Reliability moves too: the Claude API’s uptime over the 90 days ending 8 April 2026 was 98.95%, below the 99.99% standard that established cloud providers typically maintain [7].

So the practical exposure for a small operator is not that a provider goes bankrupt and takes your data. It is that the plan you built a client deliverable on quietly acquires a quota, a queue, or a 20% higher price at renewal, on a schedule set by somebody else’s debt covenants.

The only hedge you fully control is portability

You cannot audit a private provider’s financing and you cannot negotiate its terms. You can make leaving cheap. That means keeping a second provider with a live key and a small monthly job running through it, so that switching is a config change you have already tested rather than a project you start on a bad day. It means keeping prompts, evaluation cases and any fine-tuning data in your own repository rather than only in a vendor console. And it means limiting prepaid balances, because prepaid credit is an unsecured loan to a company that is, by the numbers above, spending far more than it earns [1].

Portability has a real cost and it is worth being honest about it. Moving from a frontier model to a cheap open model is not a swap of equal parts. The prices differ by more than an order of magnitude [4][5] because the outputs differ, and prompts tuned for one behave differently on the other. The hedge only works if you have already run your own tasks on both and know what you would lose. A backup provider you have never tested is not a backup.

What still goes wrong

You cannot see the terms that matter. The revenue-share percentages in the Nvidia program were not disclosed when it was announced [2], and most neocloud financing is private. Press reporting is the only signal available to a small buyer, it arrives late, and it is sometimes contradicted by the companies involved within days [3]. Treat any read of a provider’s capital structure as a rough prior, not a finding.

The subsidy may also be working in your favour, and this guide is not an argument for avoiding cheaper providers. Capacity funded by vendor credit is genuinely cheap capacity for as long as it lasts, and taking it on month-to-month terms costs you nothing when it ends. The mistake is not using it. The mistake is building a client commitment on top of a price you have no reason to believe will hold, without having tested the alternative.

And portability costs more than it looks. Two providers means two bills, two sets of keys, two rate-limit regimes and two evaluation runs every time you change a prompt. For a one-person business, that maintenance can plausibly exceed the money at risk. If your total AI spend is under a few hundred dollars a month, the honest answer is to skip the second provider, keep the prepaid balance near zero, avoid multi-year terms, and accept that a reprice would be an annoyance rather than a threat.