What AI chip news means for your AI bill

Nvidia’s second quarter of fiscal 2027, ended 26 July 2026, brought in $96.2 billion, up 106% year over year, with data-center revenue alone at $89.0 billion and guidance of $108.0 billion for the following quarter [1]. The coverage arrives in one of two shapes. Either compute is running out and your costs are about to spike, or the whole thing is a bubble and prices are about to collapse. Neither shape is a plan you can act on.

You run a small business with a Claude subscription, an API key, and maybe an n8n or Zapier workflow that calls a model a few thousand times a month. The useful question is not what happens to the AI industry. It is whether anything in that headline should change what you do on Monday. The answer is almost always no, and it is worth understanding the reason once so you stop re-litigating it every earnings season. This guide is not for anyone buying GPUs, reserving cloud capacity, or negotiating a multi-year compute contract. Those readers have a genuinely different problem and none of the advice below applies to them.

A supply ceiling at the frontier is not a price rise on your invoice

Nvidia’s chief financial officer, Colette Kress, gave a preliminary expectation of a 70% jump in revenue for fiscal 2028 and said supply would remain a bottleneck at least until the end of that year [2]. Chief executive Jensen Huang said demand is “much greater” than that 70%, and that supply “allows us to confidently deliver 70%” [2]. Read plainly, that is a manufacturer saying its forecast is a production limit rather than a demand estimate.

The constraint behind it is physical, and physical constraints do not bend to willingness to pay. The International Energy Agency projects data-centre electricity consumption more than doubling to around 945 TWh by 2030, and its base case sees roughly 1,200 TWh by 2035 [3]. It estimates that unless the risks are addressed, around 20% of planned data-centre projects could be at risk of delay [3]. Transmission lines take four to eight years to build in advanced economies, and wait times for critical grid components such as transformers and cables have doubled in the past three years [3].

That is a real bottleneck. It is also not your bottleneck. The parties competing for the scarce thing are hyperscalers and frontier labs signing multi-year commitments for capacity that does not exist yet. You are buying a finished, metered service several layers downstream, at a price published on a page the vendor controls and changes on its own schedule. The distance between those two positions is the whole point. A shortage of fab capacity in 2028 reaches you, if it reaches you at all, as a model that ships later than announced or a rate limit that holds firm for another quarter. It does not reach you as a surprise line item.

The price of a fixed capability keeps falling while the frontier stays scarce

Both things are true at once, and the second one is the one that shows up in your account. Anthropic lists Claude Sonnet 4.5 and Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens, and Claude Sonnet 5 at $2 and $10 [4]. Same family, newer model, lower price. The Opus line shows the same shape on the same page: the retired Opus 4.1 is listed at $15 and $75, while Opus 4.5 through Opus 5 are all $5 and $25 [4]. Two families, two step-downs, one price list.

The spread inside a single vendor’s catalogue matters more than the year-over-year drift. Claude Haiku 4.5 is $1 input and $5 output, Claude Opus 5 is $5 and $25, Claude Fable 5.1 is $10 and $50 [4]. OpenAI’s range is wider still: gpt-5.6-luna at $0.20 input and $1.20 output against gpt-6-astra at $10 and $50, a 50x difference on input inside one price list [5]. Google lists Gemini 3.8 Flash at $0.75 input and $3.75 output, and Gemini 3.5 Flash-Lite at $0.30 and $2.50 [6]. Whatever the chip market does, the gap between the model you are using and the cheapest model that would do your job is usually a bigger number than anything in the news.

Published prices do move upward, so treat the current page as current rather than permanent. Google’s Gemini 3.8 Flash and 3.7 Flash rates are promotional through 31 December 2026, after which the listed price doubles to $1.50 input and $7.50 output [6]. That is a documented, scheduled increase you can plan around today, which is more than any chip forecast gives you.

Scarcity reaches you as a rate limit, not a bill

When capacity is tight, vendors ration throughput before they ration by price. Raising list prices annoys every customer at once; tightening limits quietly constrains the heaviest users and leaves everyone else alone. So the honest way to prepare for a tight compute market is to make your workflows survive being throttled, not to hedge against a price shock.

Anthropic publishes the mechanism. Organisations sit in usage tiers, and the throughput limits are set per model rather than per account. On the Start tier, Opus 5, Sonnet 5 and Haiku 4.5 each get 1,000 requests per minute, 2 million input tokens per minute and 400,000 output tokens per minute, while Claude Fable 5.x on that same tier gets 500,000 input and 100,000 output tokens per minute [7]. The Start tier carries a $500 monthly spend cap, Build raises it to $1,000, and Scale to $200,000 [7]. Exceed a rate limit and you get an HTTP 429 with a retry-after header telling you how long to wait. Exceed the monthly spend cap and you get a 429 with no retry-after at all, and access returns at 00:00 UTC on the first day of the next month [7].

That is the shape of the risk for a small operation. Your failure mode is an automation that stops at 4pm on the 27th, not an invoice that tripled. The fix is unglamorous. Retry on 429 using the retry-after value rather than a fixed sleep, watch the anthropic-ratelimit-input-tokens-remaining header and its request and output-token counterparts on the responses you get back [7], and know in advance which of your automations can tolerate an hour’s delay and which cannot. Do that once and a genuine capacity crunch becomes a slower Tuesday instead of an outage.

Four levers set your bill, and you control all four

Tier the model to the job. Most work in a small business is extraction, classification, reformatting and drafting, and the cheap tiers handle it. Reserve the expensive model for the small share of tasks where a wrong answer costs you real money. Moving a high-volume classification step from Claude Opus 5 at $5 and $25 to Claude Haiku 4.5 at $1 and $5 is an 80% cut on the same workload [4], and it takes an afternoon of testing on your own examples to find out whether quality holds.

Cache the stable part. If every call sends the same system prompt, style guide or reference document, you are paying full input price for identical tokens over and over. Cache reads cost 0.1x the base input price, and 0.025x on Claude Fable 5.1 and Mythos 5.1, which is $0.25 per million against $10 [4][8]. The mechanics have edges worth knowing. The minimum cacheable prefix is 1,024 tokens for Sonnet 5, 512 for Opus 5 and 4,096 for Haiku 4.5, and prompts below the minimum are silently not cached [8]. The default cache lives five minutes, measured from the start of the request that writes or reads it rather than from the end of the response [8]. A five-minute cache costs 1.25x on write and pays for itself after one read; a one-hour cache costs 2x and needs two [4]. Place the breakpoint on the last block whose prefix is identical across requests. Anthropic names the common failure explicitly: setting it on a per-request block carrying a timestamp and the user message, which caches nothing [8]. Changing a tool definition invalidates the entire cache [8].

Batch what can wait. Anthropic gives a 50% discount on both input and output through its Batch API across all models [4], and Google gives 50% off standard rates for batch on its current models [6]. Overnight report generation, bulk tagging, backfilling a spreadsheet, and anything else nobody is sitting and waiting for belongs here.

Cut the context. Output tokens cost roughly five times input tokens on most current models [4][5], so asking for a table instead of an essay is a direct saving. On the input side, sending a whole 40-page PDF when three pages are relevant is the most common form of quiet overspend, and the fix is retrieval or a first pass by a cheap model that pulls out what matters.

Run your real numbers through that before you run anyone’s forecast through it. A workload at 200 calls a day with 12,000 input and 2,000 output tokens each comes to roughly $194 a month at Sonnet 5 list prices. The same workload on a cheap tier, with a decent cache hit rate on a fixed system prompt and the batchable half moved to batch, lands in a different order of magnitude. No plausible chip-supply scenario moves your bill by as much as those three changes do.

What is worth committing to and what is not

The industry response to scarcity is long commitments. Hyperscalers and labs are locking in capacity years ahead because the alternative is queuing. That logic is correct for them and wrong for you, because the thing they are buying is scarce and the thing you are buying is not.

So resist the instinct to mirror it. Annual prepayment for API credits, long minimum terms with a reseller, or a wrapper product that only bills yearly all buy you a discount in exchange for freezing a decision in the fastest-moving part of your cost base. The Sonnet and Opus lines on Anthropic’s current price list both show the newer model priced below the older one [4]. Committing twelve months ahead means paying today’s price for next year’s commodity.

What is worth committing to is portability. Keep prompts in files rather than pasted into a vendor’s web interface. Keep the model name in one configuration value rather than scattered through the code. Write an evaluation set of twenty to fifty real examples with the answers you want, so that switching models is a half-day test rather than a leap of faith. Those three habits cost nothing and convert every future price change, on any vendor’s page, from a problem into a routine swap.

What still goes wrong

The biggest gap between this advice and reality is that cheap models are cheap until they are not. A tier swap that passes twenty examples can still fail on the twenty-first in a way that costs more to clean up than the saving was worth. That risk is real and it scales with how consequential the output is. Classification and drafting tolerate it. Anything a customer reads unedited, or anything that triggers a payment, does not. The evaluation set is the control, and an evaluation set of twenty examples is genuinely small.

Caching is fragile in ways the pricing table does not show. A prefix one token under the minimum caches nothing and reports no error [8]. A tool definition edited on Tuesday quietly invalidates everything [8]. The default five-minute window is short enough that a workflow running hourly gets no hits at all, and the one-hour cache costs 2x on write [8], which only pays back after two reads [4]. Check the cache_creation_input_tokens and cache_read_input_tokens fields in the response usage rather than assuming the configuration is working; if both are 0, nothing was cached [8].

And the honest limit on the whole argument: this holds for metered inference at small volume. If your product’s margin depends on inference at scale, or you are training rather than calling, you are closer to the frontier than this guide assumes, and the supply picture in [1], [2] and [3] becomes a real input to your planning rather than background noise. Even then, note who is speaking. The 70% figure came from Nvidia’s own CFO describing Nvidia’s own expected revenue, and Huang tied it to what supply “allows us to confidently deliver” [2]. Treat it as one company’s statement about its own production, not as a forecast of anything you will be charged.