One billion against 725: the day renting a frontier model became a deliberate strategy

On September 1, 2026, John Ternus took over at Apple, succeeding Tim Cook after fifteen years. He inherits a market capitalisation of roughly $4.6 trillion and a singularity: Apple is the only very large technology company that does not own a frontier model. It rents one. Since the deal announced on January 12, 2026, it has been paying Google about a billion dollars a year for a custom version of Gemini with 1.2 trillion parameters, which runs the new Siri on Apple’s servers, not on Google’s. Meanwhile, the four big hyperscalers are committing capex of around $700 billion in 2026. The same day, another event did not take place: the 50% increase announced on Claude Sonnet 5 for September 1 was cancelled. We wrote on August 18 that the price of a token had stopped being a constant. One thing must be added: the question is no longer just what a token costs, but what it costs not to own the model that produces it. And that answer is now public.

On September 1, Apple confirmed it would not build its own model

The transition had long been prepared, but it comes at a specific point in the cycle. According to Fortune’s analysis published on August 31, Apple has “largely ceded the market for frontier models” to Google and OpenAI, to concentrate its resources on distributing the technology through hardware. Where Microsoft, Google, Meta, Amazon and Oracle build data centres at tens of billions a time, Apple has stayed out of the infrastructure race. That restraint has protected its margins; it leaves Ternus with internal training capacity and AI infrastructure that are smaller than those of its competitors.

The architecture of the deal with Google deserves a close look, because it is exactly the compromise most European organisations are looking for today. The model is customised for Apple, built on a mixture-of-experts architecture: only a fraction of the 1.2 trillion parameters is activated on each request, which lowers the unit cost of inference. Above all, it runs on Apple’s Private Cloud Compute servers, not in Google’s infrastructure: Apple rents the cognitive capacity without handing over the data. That is the operational definition of a sovereign core that does not train its own models.

Ternus’s first public test is set for September 9, at the company’s media event. The question put to him will not be whether he can sell hardware, but whether he will publicly own up to renting the intelligence that drives it.

The arithmetic of renting: 0.14% of hyperscaler capex

The figure to set against Apple’s billion a year is the 2026 capex of the four big hyperscalers. The tallies differ, and that should be said plainly, because these are sums of guidance, not audited accounts. Three independent surveys put the combined 2026 total for Amazon, Google, Microsoft and Meta between $630 and $725 billion. The most detailed count retains $725 billion, up 77% from the $410 billion of 2025, and breaks down as follows: Amazon about $200 billion (against 125 in 2025), Google $175 to 185 billion (against 91), Meta $115 to 135 billion (against 72), Microsoft $110 to 120 billion (against 90). We keep the range rather than the point: the order of magnitude is solid, the decimal is not.

The ratio, for its part, is robust whichever assumption is used. Apple gets access to a custom frontier model for about 0.14% of the combined capex of these four companies. Set against the supplier alone (Google, which is committing 175 to 185 billion), the rental comes to about 0.55% of what the builder spends in a year.

This ratio should not be read as a cost comparison: hyperscaler capex also funds their cloud, their own products and the sale of capacity to third parties; Apple’s billion buys only a right of use over a defined perimeter. But the strategic conclusion holds: building a frontier model and accessing one are not two points on the same continuum of spending, they are two different orders of magnitude. The best-capitalised company in the world has just decided, on a day of succession, in favour of the second. No Swiss organisation will ever have to ask the question in Apple’s terms; all of them ask it in the same economic terms.

The increase that did not happen

For everyone who had built an inference budget on Claude Sonnet 5, September 1, 2026 was to be the date of a 50% increase. The model had been launched at a price presented as introductory ($2 per million input tokens, $10 output), valid until August 31, 2026, with a scheduled move to $3 / $15 the next day.

The move did not happen. We consulted Anthropic’s official pricing page on September 2, 2026, after the fateful date: it carries an explicit note saying that the $2 / $10 rate, announced at launch as introductory until August 31, is now the standard rate, and that the increase planned for September 1 will not take place. The decision had been made public around August 10-11; its confirmation in the rate card after the deadline makes it an established fact, not an intention.

This is exactly the opposite movement to the one we documented two weeks ago. On August 16, DeepSeek replaced its single rate with an hourly grid that takes V4-Pro’s million output tokens from $1.10 to $3.96 at peak hours. On August 13, Google launched Gemini 3.7 Flash at a price whose documentation announced expiry on December 31. In fifteen days, the market will thus have produced a sharp increase, a promotion with an expiry date, and a promotion converted into a permanent rate.

The operational lesson is not that prices go up, nor that they go down: it is that an expiry date written in documentation is no longer a reliable budget input, in either direction. An organisation that had provisioned +50% on its Sonnet 5 line for September has tied up cash for nothing; one that built its 2027 plan on Gemini 3.7 Flash’s current rate is exposed to the symmetrical risk. The only tenable position is to measure the cost actually incurred, model by model, month by month, and to keep the technical ability to change supplier, the one variable the supplier does not control.

Sovereignty now has a posted price: +10%

The fact most directly usable by a Swiss decision-maker appears in no press release. It is written into Anthropic’s rate card, and we noted it there on September 2: for Claude 4.6 models and later, requesting inference run exclusively in the United States via the inference_geo parameter applies a multiplier of 1.1 across all token categories: input, output, cache writes, cache reads. Global routing, which is the default behaviour, stays at the standard rate.

The same logic applies at cloud resellers. On Amazon Bedrock and Google Cloud, from Claude Sonnet 4.5, Haiku 4.5 and Opus 4.5 onward, regional and multi-regional endpoints (those that guarantee the data transits through a specified geographic zone) carry a 10% premium over global endpoints.

It is worth measuring what this means. For two years, the question of where inference runs was treated as a contractual clause, negotiated case by case, invisible in the rate cards. It has become a public, priced, standardised line item. Geographic control of data is no longer a sales argument or a legal grey area: it is a product, and it costs 10% more.

For a private bank, a fiduciary or a Swiss firm, this is good news disguised as a surcharge. An explicit budget line can be arbitrated, audited and defended before a risk committee, which a confidentiality clause never will be. The corollary is just as clear: an organisation that has not instrumented its token consumption cannot know what its own sovereignty requirement actually costs, nor on which workloads it is paying for it to good purpose.

The tokenizer trap: −33% announced, −13% real

The same pricing page carries a second note, a technical one, with no fanfare, whose budget consequences exceed those of the first. Claude 4.7 and later models use a different tokenizer from their predecessors. This tokenizer contributes to their better performance, and it produces (Anthropic documents this explicitly) about 30% more tokens for the same text. Claude Sonnet 4.6 and earlier models use the old tokenizer. The exact scale, the note specifies, depends on the content and the shape of the workload.

Let us do the arithmetic, because it is rarely done. Sonnet 4.6 is billed $3 / $15 with the old tokenizer; Sonnet 5 is billed $2 / $10 with the new one. The headline announces a 33% cut. But for the same document submitted to both models, Sonnet 5 bills about 1.3 times more tokens. The real input cost thus goes from 3.00 to 2.60 units, and output from 15.00 to 13.00: the effective reduction is about 13%, not 33%. Two thirds of the advertised discount is absorbed by the change in unit of measurement.

We do not present this as a manoeuvre: a better tokenizer genuinely improves quality per unit of compute, and Anthropic publishes the information of its own accord, which is not the case for every supplier. But the methodological consequence is severe and general: the price per million tokens has ceased to be comparable from one model generation to the next, even within a single supplier. A price comparison table (including the one we publish below) no longer compares prices: it compares labels stuck on units that do not measure the same thing. The only honest figure is the cost observed on a real workload, measured on your own traffic.

Let us note in passing the reverse pricing innovation, discreet and very real. On Claude Fable 5.1 and Mythos 5.1, cache reads are billed at 0.025 times the base input price, instead of the 0.1 multiplier applied to all other models: that is $0.25 per million tokens on a model whose input is $10. It is not the price that moves, but the multiplier, and for an agentic workload with a long, stable context, it is a more powerful lever than a 20% discount on the list price.

Beneath the prices: compute, energy and the lawsuit

Three facts from the past fortnight shed light on what happens below the waterline of the rate cards.

Compute remains expensive and scattered. A survey carried out on August 20, 2026 on suppliers’ public APIs gives, for an H200 per hour, a minimum of $4.45 on the Akash marketplace and a maximum of $10.60 on Azure (ND96isr H200 v5, East US 2), a gap of 2.4 times for identical hardware. Other suppliers list $3.99 an hour on demand and $1.99 spot. The B200 generation trades around $7 to 10 an hour, a premium of 75 to 150% over the H200, which analyses judge hard to amortise outside workloads limited by memory bandwidth, very long contexts or FP4-quantised serving.

Energy has become the limiting factor, and capital has understood. At the end of August, Databricks raised $5 billion at a valuation of $190 billion, and Firmus $2 billion from Blackstone, Coatue and Nvidia at a valuation of $10.5 billion for its Asia-Pacific expansion. But the most instructive round is the smallest: Emerald AI raised $150 million at a valuation of $1.05 billion with a single purpose: making energy-hungry data centres able to modulate their consumption to take part in balancing the electricity grid. When a company whose product consists of switching off compute at the right moment crosses a billion in valuation, it means the constraint has shifted from silicon to the power socket. Bloomberg also reported on August 12 that Cognition was discussing a round at a valuation of at least $40 billion.

The cost of training data enters the balance sheet. On August 28, 2026, Sony Music Publishing and Warner Chappell Music sued Anthropic in the federal court for the Northern District of California, naming co-founders Dario Amodei and Benjamin Mann personally. The complaint concerns “tens of thousands” of compositions allegedly acquired and used without authorisation, and seeks up to $150’000 per work for wilful infringement, plus up to $25’000 per alleged removal of rights-management information. The text characterises the facts as among the largest thefts of intellectual property in history, and argues that the company treats this risk as a mere operating cost.

We report these amounts as claims, not convictions: no decision has been made, and a calculation multiplying a statutory ceiling by an unsettled number of works is not an exposure estimate. What matters for an AI buyer lies elsewhere: the legal risk attached to a model’s training data appears in no rate card, and it cannot be contracted at the token level. It is managed by the ability to change model without rewriting your system, that is, by the same architecture that protects against price risk.

API price comparison table

Model	Input ($/M tokens)	Output ($/M tokens)	Note
GPT-5.6 Sol	5.00	30.00	OpenAI flagship
GPT-5.6 Terra	2.00	12.00	−20% on July 30, 2026
GPT-5.6 Luna	0.20	1.20	−80% on July 30; floor of general-purpose offers
Claude Fable 5.1	10.00	50.00	Cache reads at 0.025× ($0.25/M)
Claude Opus 5	5.00	25.00	Fast mode: 10.00 / 50.00
Claude Sonnet 5	2.00	10.00	September 1 increase cancelled; new tokenizer
Claude Sonnet 4.6	3.00	15.00	Old tokenizer, not comparable line by line
Claude Haiku 4.5	1.00	5.00
Gemini 3.1 Pro	2.00	12.00	4.00 / 18.00 beyond 200k tokens
Gemini 3.7 Flash	0.75	3.75	Announced rate expiring December 31, 2026, not cross-checked, see note
Grok 4.6	2.00	6.00	4.00 / 12.00 from 200k tokens (whole request)
DeepSeek V4-Pro	0.66 to 1.32	1.98 to 3.96	Hourly grid since August 16: off-peak / peak
DeepSeek V4-Flash	0.22 to 0.44	0.66 to 1.32	Hourly grid since August 16
Qwen3.7 Flash	0.03	0.13	Under 32k tokens; rises to 0.20 / 0.80 beyond 256k

Claude prices verified on September 2, 2026 on Anthropic’s official pricing page. OpenAI, DeepSeek and Qwen prices come from a survey dated August 5, 2026; Google and xAI prices from a survey dated September 1, 2026. Two explicit caveats: (1) the Google survey of September 1 does not show Gemini 3.7 Flash, launched on August 13: we use the $0.75 / 3.75 rate announced at launch without having been able to cross-check it against a September source; (2) several aggregators present GPT-5.6 Luna as the market floor at $0.20, which is only accurate if Chinese suppliers are excluded: Qwen3.7 Flash is nearly seven times cheaper on short prompts. Claude 4.7 and later rates apply to a tokenizer producing about 30% more tokens: the lines in this table are not comparable with each other without correction.

What a Swiss decision-maker should take away

The past fortnight has delivered four lessons that all converge on the same practical conclusion.

First, renting a frontier model is a first-rank strategy, not a stopgap for an under-capitalised organisation: the best-endowed company in the world has just renewed it on a day of succession, paying about 0.55% of what its supplier spends annually. Next, a pricing expiry date is not a commitment: Sonnet 5’s was cancelled at maturity, Gemini 3.7 Flash’s is still running, DeepSeek’s turned into an hourly grid. Then, geographic sovereignty has become a 10% budget line: arbitrable for those who measure, invisible for those who do not. Finally, the billing unit itself has moved: 30% more tokens for the same text turn a 33% discount into a 13% saving.

These four facts do not call for four answers. They call for one: having an orchestration layer that measures the real cost per workload, applies the location constraint where it is required and nowhere else, and makes switching models painless. Without it, an organisation does not merely endure prices: it is unable to know which ones it is paying.