Building on AI when one company finances the whole chain

The company that makes the chips has committed billions to the labs that buy them. On 18 November 2025 Anthropic committed to purchase $30 billion of Azure compute capacity, and in the same announcement Nvidia and Microsoft said they were committing to invest up to $10 billion and up to $5 billion respectively in Anthropic [7]. Two months earlier, Nvidia said it intended to invest up to $100 billion in OpenAI “progressively as each gigawatt is deployed”, against at least 10 gigawatts of Nvidia systems [6]. On 3 September 2026 Nvidia agreed to acquire Hugging Face, which its announcement calls “a vibrant home for the open model developer community”, for $12,930,300,000 [8]. Money leaves one balance sheet and comes back to it as revenue, and the loop is tight enough that a handful of companies now sit on several sides of your stack at once.

None of that is a scandal, and none of it is yours to fix. It is a supply chain, and you are at the thin end of it, paying a subscription or a few hundred dollars a month in API spend. This guide is for someone in that position who wants to know which parts of the arrangement can actually reach their week. It is not investment guidance, it contains no view on whether any of these companies is worth what the market says, and it is not for anyone with a procurement function that already runs vendor-risk reviews. The narrow question is what breaks on your desk when compute gets tight or a financing decision changes somebody’s roadmap, and how much warning you get before it does.

The circle is real, and it is not the part that reaches you

Look at the structure of the deals rather than the totals. Nvidia’s OpenAI investment is staged: the money goes in “progressively as each gigawatt is deployed”, the first phase is targeted to come online in the second half of 2026 on Nvidia’s Vera Rubin platform, and the announcement of 22 September 2025 described a letter of intent, with both companies saying they looked forward to finalising the details in the coming weeks [6]. Anthropic’s side of its deal runs the other way. It is a commitment to purchase $30 billion of Azure compute capacity and to contract additional compute capacity up to one gigawatt, initially with Nvidia Grace Blackwell and Vera Rubin systems [7]. One party promises capital contingent on buildout, the other promises purchases contingent on nothing.

That asymmetry is the useful part. Commitments of that shape are multi-year and hard to unwind, which is why the effects on customers arrive slowly and indirectly rather than as a sudden failure. A lab that has promised to spend $30 billion on cloud capacity has a strong reason to keep its paying customers, raise revenue per customer, and retire anything that costs more to serve than it earns. That is not a prediction of doom. It is a description of ordinary commercial pressure, and it shows up in three places you can actually see: the list of models you are allowed to call, the queue your requests sit in, and the price per token.

Everything else about the financing is unauditable from where you sit. Private companies do not publish their compute contracts, a letter of intent is not a binding schedule, and the reporting that fills the gap is mostly informed guesswork. Time spent tracking it is time not spent on the three visible things, which are documented, dated and in your control.

The shock arrives as a model id that stops working

Your real contract with a model provider is its deprecation policy, and the two big ones differ by a factor of three. Anthropic commits to at least 60 days’ notice before model retirement for publicly released models, moving each one through active, legacy, deprecated and retired, at which point requests to the model fail [2]. OpenAI commits to at least 6 months for generally available models, at least 3 months for specialised variants such as chat, codex and deep research, and warns that preview models “may be retired with much shorter notice, such as 2 weeks” [5]. OpenAI also carves out an exception: if safety or compliance concerns require an earlier retirement, it will give “as much notice as reasonably possible” [5]. Anthropic’s page states the 60-day commitment without publishing an equivalent exception [2].

Those are not abstractions. Anthropic’s list shows claude-opus-4-1-20250805 deprecated on 5 June 2026 and retired on 5 August 2026, and claude-sonnet-4-20250514 deprecated on 14 April 2026 and retired on 15 June 2026 [2]. OpenAI announced on 11 June 2026 that gpt-5-2025-08-07 and o3-2025-04-16 shut down on 11 December 2026, announced on 20 July 2026 that the older realtime and audio families shut down on 20 January 2027, and announced on 26 August 2026 that whisper-1 and the gpt-4o-transcribe variants shut down on 26 February 2027 [5]. If a pinned model id is sitting in an automation you built last year, one of those dates is your outage date.

Parameters go the same way, more quietly. Anthropic’s documentation lists temperature, top_p and top_k as deprecated on Claude Opus 4.7 and later, returning a 400 error when set to a non-default value on Claude 4.7 and later models [2]. A workflow that has always passed a temperature of 0.2 does not degrade when you upgrade the model. It stops. The fix is dull and takes an hour: list every model id and parameter your tools send, exactly as written, and check each one against the provider’s page. Anthropic points you at the Console usage page, where you click Export and review the downloaded CSV, which breaks usage down by API key and model [2].

Scarce compute is sold to you as queue position

When capacity is tight, providers do not switch you off. They sell the front of the queue, and the documentation shows the shape of it clearly. Anthropic’s standard tier prioritises requests “alongside all other requests with best-effort availability”, its priority tier targets 99.5% uptime with prioritised computational resources and minimises server overload errors, and its batch tier is for asynchronous work that “can wait or benefit from being outside your normal capacity” [3]. A priority commitment consisted of a number of input tokens per minute, a number of output tokens per minute, a duration of 1, 3, 6 or 12 months, and a specific model version [3].

The line worth reading twice is the one in the past tense: priority tier capacity commitments are no longer available for purchase, and organisations with an existing commitment can continue to use the tier through their contract end date [3]. That is a supplier deciding it has more demand for guaranteed capacity than capacity to guarantee. Requests beyond a committed amount fall back automatically to the standard tier, and the service_tier parameter accepts auto or standard_only, with the response usage object reporting which tier actually served the request [3].

The same logic runs through the token prices. Anthropic’s Batch API is 50% off both input and output tokens, a cache hit costs 10% of the standard input price on most models, a 5-minute cache write costs 1.25 times the base input price and a 1-hour cache write 2 times [1]. Fast mode on Claude Opus 5 and Claude Opus 4.8 is priced at $10 per million input tokens and $50 per million output, against $5 and $25 on the standard tier [1]. OpenAI prices the same way, with fast mode at 2 times standard, batch at a 50% discount, cached input at 10% of the standard input rate, and long context at double the short-context input rate [4]. Geography is metered too: Anthropic applies a 1.1 multiplier for US-only inference on Claude 4.6 and later models [1], OpenAI a 10% uplift for regional processing on models released on or after 5 March 2026, and OpenAI notes that “fast mode is unavailable for GPT-6 Astra with EU data residency” [4].

Read as a group, those are the dials a provider turns when compute is scarce, and every one of them is a choice you can make first. Work that can run overnight belongs in a batch tier. Work that repeats the same long preamble belongs behind a cache. Work that genuinely needs to answer while a customer waits is the only work that should be paying the fast-mode multiple.

A second provider is cheap to keep warm and expensive to build in a hurry

Prices at the working tier have converged enough that portability is a real option rather than a slogan. Claude Sonnet 5 lists at $2 per million input tokens and $10 per million output [1]. OpenAI’s gpt-5.6-terra lists at $2 and $12 on short context, with gpt-5.6-sol at $4 and $20 and gpt-5.6-luna at $0.20 and $1.20 [4]. At the top, Claude Opus 5 is $5 and $25 and Claude Fable 5.1 is $10 and $50 [1], while gpt-6-astra is $10 and $50 [4]. For most small-operator work the difference between two vendors at the same tier is smaller than the difference between a careless prompt and a tidy one.

Keeping a second provider warm means something specific and small. Once a quarter, take your highest-volume prompt, run it on the other vendor’s mid-tier model, and save the output next to the original. You are not benchmarking. You are checking that the prompt still makes sense without the first vendor’s defaults, that your code has somewhere to put a different model id, and that you know what the output looks like when it arrives in a slightly different voice. The cost is an hour and a few cents of tokens. The alternative is discovering, on the morning your model id starts returning errors, that the prompt was tuned to one model’s habits and nobody remembers why.

The same drill applies to subscriptions, not just APIs. If your writing, research or coding work lives inside one assistant, spend one working session a quarter doing the same job in another. What you are buying is the knowledge of how much of your process is portable, and the answer is usually more than you feared and less than you hoped.

Neutral infrastructure stops being neutral when someone buys it

Hugging Face was, for most of a decade, the thing everyone treated as plumbing. Nvidia’s announcement puts more than 18 million developers, researchers and creators on it, sharing more than 3 million models, 500,000 datasets and 1 million applications, with more than 200,000 companies using the platform [8]. On 3 September 2026 Nvidia agreed to acquire it for $12,930,300,000 [8]. The stated commitments are explicit. The announcement says “Hugging Face will remain an open platform for the entire AI ecosystem” and that “NVIDIA compute will not be required to build on or deploy through Hugging Face” [8]. Nvidia also describes itself as the largest contributor of open models and data to Hugging Face, with more than 500 models and more than 250 open datasets released there [8].

Take the commitment at face value and still do the boring thing. Make a list of the parts of your stack you have been treating as neutral ground rather than as a vendor: the model hub you pull open weights from, the registry your automation platform installs connectors from, the place your prompts and evals live. For each one, work out what you would lose if its terms changed in a year, and whether you hold a local copy of the specific artefacts you depend on. Downloading the two or three open-weight models and the dataset you actually use costs disk space and one afternoon. It is the cheapest insurance in this entire guide, and it is unaffected by whatever the owner does next.

What still goes wrong

The financing story is close to useless as a timing signal. What is public is a set of company announcements about company plans: a letter of intent whose details were still to be finalised [6], investments stated as up to a number rather than as a paid-in sum [6][7], and a purchase commitment with no published schedule [7]. None of that tells you what happens to your account next quarter. If you find yourself refreshing chip-company news instead of shipping, the habit has cost you more than the risk it was meant to manage.

Diversification is not free either, and the version of this advice that says “always use two providers for everything” is wrong. Running every workflow twice doubles your bill, doubles your prompt maintenance, and gives you two sets of quirks to remember instead of one. The proportionate version is one warm alternative for the one or two workflows that would actually hurt to lose, checked quarterly, and nothing at all for the rest. Notice windows do most of the work here: 6 months of warning on a generally available model [5] is enough time to migrate calmly, and the 2 weeks a preview model may get [5] is the real argument for not building a business process on a preview.

The last limit is the one this guide cannot solve. Notice periods, price tiers and export buttons protect you from disruption you can see coming. They do not protect you from a model that quietly gets better at the thing you sell, or from a supplier deciding your whole category is now a feature. That is a different problem, it is not addressed by keeping a second API key, and no amount of watching the compute layer will tell you when it arrives.