What chip export news actually changes about your AI bill

Every few weeks a chip story lands hard enough that someone forwards it to you with a one-line note: does this change anything for us. Export licences loosen, a country group gets reclassified, a memory supplier announces a new plant. The story is always framed as a supply shock, and the implied question is always whether you should be buying differently, budgeting differently, or hedging something.

Almost always the answer is no, and the reason is worth understanding once so you can stop re-deriving it. You are not buying chips. You are buying tokens from three or four companies that publish their prices on a web page and change them on their own schedule. This guide is about the actual chain between a hardware headline and your invoice, how long that chain takes to move, and the much shorter list of vendor notices that will genuinely cost you money or break your code. It is not for people who buy GPUs, run their own data centres, or hold semiconductor stock. Those readers have different questions, and this is not their guide.

The lag between a policy headline and your invoice is measured in quarters

The clearest recent example runs on a public timeline you can check. On 8 December 2025 the US administration said it would allow Nvidia H200 sales to approved Chinese customers, and the Bureau of Industry and Security issued the implementing rule on 13 January 2026 [1]. The rule moved these chips to case-by-case licence review rather than a presumption of denial, and attached conditions: the exporter has to show the sale will not reduce global semiconductor production capacity currently available to US customers, the Chinese purchaser has to have adopted export compliance procedures including customer screening, and the product has to have undergone independent third-party testing in the United States to verify its performance and security [1]. The Federal Register version, published 15 January 2026, sets the technical boundary of what qualifies: a total processing performance under 21,000 and total DRAM bandwidth under 6,500 GB/s when destined for end users in China or Macau [2].

That is about as clear a policy signal as this area produces. Now look at what it produced commercially. Nvidia’s second quarter of fiscal 2027 ended 26 July 2026, and the results, published 26 August 2026, show revenue of $96.2 billion with data centre revenue of $89.0 billion, up 117 percent year over year [3]. The first H200 shipments to China happened inside that quarter, and they came to less than 1 percent of that $89 billion [8]. Beijing had restricted purchases while it promoted home-grown alternatives, and only began allowing selected Chinese AI companies to buy limited quantities in July 2026 [8]. For the following quarter, Nvidia’s own outlook language is flat: “NVIDIA is not assuming any Data Center compute revenue from China in its outlook” [3].

So the sequence took about seven months from the announcement to the first physical shipments, and the company at the centre of it declined to forecast a single dollar from the market that opened. Nothing in that chain reached the price of a token. The 117 percent growth in data centre revenue was produced almost entirely outside the market the rule opened, because the China H200 sales inside it were under 1 percent of the segment [3][8]. If in December 2025 you had rearranged your stack in anticipation, you would have spent the following seven months being wrong about a market that produced under 1 percent of one quarter’s data centre revenue for one vendor.

What actually sets the price you pay

Model prices are posted numbers, set by vendors, and they do not track chip availability in any way you can observe from outside. As of today, Anthropic lists Claude Opus 5 at $5 per million input tokens and $25 per million output, Claude Sonnet 5 at $2 and $10, and Claude Haiku 4.5 at $1 and $5 [4]. OpenAI lists GPT-5.6 Sol at $5 and $30, GPT-5.6 Terra at $2 and $12, and GPT-5.6 Luna at $0.20 and $1.20 [5]. Google lists Gemini 3.8 Flash at $0.75 and $3.75, and Gemini 3.1 Pro Preview at $2 and $12 for requests up to 200,000 tokens [6].

The spread inside a single vendor’s own line-up is larger than any plausible effect of a supply story. Moving a workload from Opus 5 to Haiku 4.5 divides the input price by five and the output price by five [4]. Moving from GPT-5.6 Sol to GPT-5.6 Luna divides input by 25 [5]. No export licence is going to do that to your bill in either direction.

The discounts are also posted, and they are larger than most people’s optimisation efforts. Anthropic’s batch API takes 50 percent off both input and output tokens, and a prompt cache hit is charged at 0.1 times the base input rate [4]. If you send the same 30,000-token brief with every request, caching it is a real reduction on a real line item, available today, requiring no forecast about anything.

The price change worth diarising is on the vendor’s own calendar

Here is the kind of thing that does hit you, and it is sitting in plain sight. Google’s pricing page shows Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 output through 31 December 2026, and $1.50 and $7.50 starting 1 January 2027 [6]. That is a scheduled doubling, published in advance, on a model many small teams use for exactly the high-volume work where a doubling matters.

Nobody will forward you that as a headline. It is not news, it is a footnote on a pricing table. But if your monthly spend on that model is meaningful, 1 January 2027 is a date that belongs in your calendar in a way that no export rule does. The general form of this rule: introductory or promotional pricing has an end date, and the end date is usually printed next to the price. Read the small print on the pricing page of every model you depend on, once a quarter, and you will catch these while you still have time to test an alternative.

The other disruption that reaches you has a 60-day clock

Model retirement is the second thing that actually breaks something. Anthropic’s deprecation policy commits to notifying customers with active deployments and “providing at least 60 days’ notice before model retirement for publicly released models” [7]. The page also carries the live list. Over the past year it has retired specific pinned model IDs on published dates, including claude-3-7-sonnet-20250219 on 19 February 2026, claude-sonnet-4-20250514 on 15 June 2026, and claude-opus-4-1-20250805 on 5 August 2026, each with a named recommended replacement [7]. A retired model does not degrade gracefully. Requests to it fail [7].

Two details on that page matter more than they look. First, deprecated-but-not-yet-retired models are described as likely to be less reliable than active ones [7], so the quiet period before a retirement is not a free ride. Second, the policy applies to Anthropic’s own platforms, and partner-operated platforms such as Amazon Bedrock and Google Cloud set their own retirement schedules [7]. If you reach a model through a cloud marketplace, the calendar you need is the marketplace’s, not the model vendor’s.

Sixty days is enough time to swap a model and re-test if you have a test to run, and not enough if you have to invent one under deadline. That asymmetry is the whole argument for the next section.

Portability is the hedge you can actually buy

You cannot hedge chip supply. You can make yourself cheap to move, and that hedge covers price rises, retirements, quality regressions and outages all at once, which is more than any forecast covers.

Three things do most of the work. Keep model IDs and prompts in configuration rather than scattered through your code, so switching is an edit in one place rather than a search across a repository. Keep a fixed set of about 20 real inputs from your own work, with the outputs you accepted, saved somewhere plain; that file is your switching test, and it is the difference between evaluating a replacement in an afternoon and guessing. And price the same workload against a second vendor at current list prices once a quarter, using the calculator below and the numbers on the pricing pages, so you know what the alternative costs before you need it.

Doing this has a cost, and it is honest to name it. Prompts tuned hard against one model do lose a little when moved. Vendor-specific features, caching behaviour and tool-calling formats among them, do not port cleanly. Portability is not free, and the correct amount of it is proportional to how much you spend, not to how alarming the news is.

What to watch instead of chip headlines

The useful watchlist is short and boring. Each vendor’s pricing page, checked quarterly for changes and for end dates on promotional rates [4][5][6]. Each vendor’s deprecation or model-lifecycle page, checked for anything with a retirement date inside your planning horizon [7]. Your own usage dashboard, because a cost surprise is far more likely to come from your own traffic pattern than from anyone’s supply chain. That is the entire list.

Hardware and policy news earns a place on it only when it changes one of those three. The H200 story is a fair test case: seven months of process, a licence regime with published technical thresholds [2], first shipments under 1 percent of one quarter’s data centre revenue [8], and a vendor outlook that assumes zero from the market going forward [3]. Interesting, well documented, and not something you can act on.

What still goes wrong

The main failure of this approach is that it tells you to ignore something that will occasionally matter. Supply and policy do reach end users eventually, usually not as a price change but as availability: a new model that launches in some regions and not others, a rate limit that stays low for months, a queue for a capacity tier. When that happens the signal shows up in your own logs and in the vendor’s status and rate-limit pages, several weeks after the headline that supposedly predicted it. Watching the headline earlier would not have helped you, but pretending the physical layer does not exist is also wrong. The honest position is that you cannot time it, and you can notice it quickly.

The second problem is that posted prices are not the whole bill. Enterprise agreements, committed-spend discounts and cloud marketplace rates differ from the public pages, and if you buy through a reseller or a cloud provider your effective price and your retirement dates are theirs, not the model vendor’s [7]. The calculator above gives you list-price arithmetic, which is the right starting point and the wrong final answer if you have a contract.

Finally, this guide assumes your dependency really is just tokens. If you have committed to hardware, signed a multi-year reserved-capacity deal, or built a product whose margins only work at one vendor’s current price, the exposure is real and the advice here is too light for you. The test is simple: if a model you use were retired with 60 days’ notice [7], and a second model you use doubled in price on a published date [6], could you absorb both without renegotiating anything. If yes, chip news is reading material. If no, the work to do is contractual, and no amount of headline-watching substitutes for it.