Special Analysis
Between 19 and 25 August, several important things happened, and none of them concerned an improvement in models. Credit analysis showed that borrowing tied to AI infrastructure had reached roughly eighteen times the volume of a year earlier, while investors are demanding a higher premium to take on that risk. Two American regulators opened proceedings that will affect how AI is priced and monetised, although neither is being pursued under any regulation written specifically for artificial intelligence. Autonomous rides became bookable through Uber in Zagreb, the platform's first such service in Europe. At the same time, internal data emerged showing that the company behind one of the most widely used AI coding agents uses that tool internally almost six times more often than its business customers do.
Taken together, these events describe a week in which every decisive variable sat outside the model layer itself. In none of the four cases was model capability the binding constraint. What mattered were financing costs, permission to deploy, and the degree of adoption inside organisations, and each of those is becoming a regular determinant of business outcomes rather than merely background circumstance.
That is an uncomfortable fact for anyone whose AI strategy still rests, predominantly, on assessing model capability, because these constraints cannot be discovered by testing models. They appear in bond prospectuses, regulatory proceedings, local permits and adoption data, and all of them can be analysed this quarter, without waiting for the next generation of models.
Financing AI infrastructure and the new regulatory frontier
The AI infrastructure build-out has leaned further on borrowed capital, and the bond market has begun to price that risk differently. A Reuters analysis published on 21 August showed that US corporate bond issuance connected to AI infrastructure reached 220 billion dollars from the start of 2026 to 10 August, against 12.5 billion dollars in the same period of 2025. Investors are charging for the privilege: Amazon's 25 billion dollar issue was priced at a yield roughly 120 basis points above US Treasuries, where the company had previously paid around 60 basis points.
The next pressure point is larger and considerably less certain. On 20 August Bloomberg reported that Broadcom is negotiating a senior secured financing package of more than 60 billion dollars which, together with a subordinated tranche, could approach 100 billion dollars. The financing would be carried out through a special purpose vehicle and, according to those reports, would serve capacity for Anthropic. Broadcom, Anthropic, Apollo and Blackstone declined to comment, so this remains negotiations reported in the media rather than a concluded transaction.
This shift matters because of who the cost may ultimately be passed to. Until recently, much of the build-out was financed from the cash flows of the large cloud providers, which kept financing costs out of buyers' sight. Debt, however, carries interest that providers may attempt to recover through future prices, contractual terms or smaller discounts. Companies negotiating cloud or inference capacity for 2027 should test their assumptions against a higher cost of capital rather than simply rolling forward today's price lists.
In this region, many subsidiaries buy services under group framework agreements negotiated in other markets, so new and more expensive terms may only surface at renewal, when local negotiating room is limited.
A second, less direct channel runs through savings. European Central Bank economists estimated, in a staff blog post of 17 August that does not represent the position of the ECB Governing Council, that European households hold around 440 billion euros in the shares of the largest American technology companies, with similar exposure held by pension funds and insurers. The effect on any individual portfolio, however, depends on its specific structure.
Two American regulators addressed the economics of AI at the same time, and neither invoked a law written specifically for artificial intelligence. On 19 August the Federal Trade Commission voted 2:0 to open a public consultation on a proposed enforcement policy under which personalised pricing without clear notice to the consumer could be treated as an unfair or deceptive practice under Section 5 of the FTC Act. The practice is defined as the use of personal data to set a price according to the amount a company believes an individual consumer is willing to pay. The document does not mention artificial intelligence, which is precisely the point: algorithmic pricing may be a highly valuable business application of AI in retail and travel.
On the same day, the Commodity Futures Trading Commission requested comment on the listing of derivative contracts tied to compute capacity. The request was published in the Federal Register on 21 August, with a 60-day comment period. CFTC Chairman Michael Selig suggested that a developed derivatives market for compute could support US competitiveness in artificial intelligence.
Together, these two proceedings sketch the economic boundaries of AI from both sides: one regulator is considering what companies must disclose when personal data affects the price a consumer pays, while the other is considering a market in which the price of compute would be set. Neither needed a dedicated AI statute in order to act, which is an important lesson for European boards as well.
Under the EU's revised timeline, the key obligations for high-risk systems under Annex III now begin to apply in December 2027 rather than August 2026, while some national supervisory mechanisms are still being established. On 24 July Romania's ANCOM stated that it and other competent authorities will only be able to verify compliance and sanction breaches once national implementing legislation enters into force. It would be risky to read this as a period without obligations: a regional retailer, airline or insurer applying personalised pricing today may already be exposed to consumer protection, competition and data protection rules. The dedicated AI regime has been postponed, but the general legal framework never stopped applying.
From pilot to actual adoption
The first autonomous rides available through Uber in Europe became bookable in Zagreb. On 19 August Uber announced that autonomous rides, delivered with its Croatian partner Verne and the Chinese company Pony.ai, can be booked through its app in Zagreb. These are the first autonomous rides available through Uber anywhere in Europe. The service uses Pony.ai's seventh-generation system in Arcfox Alpha T5 vehicles, within the UberX and Comfort categories, with a licensed safety operator behind the wheel. Fully driverless operation remains a goal for the future. Fleet size has not been disclosed. This represents an important step in platform integration and commercialisation, but not the start of the service in Zagreb itself, which began operating earlier in 2026.
The technology is not the only significant part of the story, since the same system already operates commercially in Guangzhou and Shenzhen, where Pony.ai reports having reached break-even unit economics. The more instructive aspect is that Zagreb offered an early European path from supervised commercial operation to distribution through Uber. Smaller jurisdictions can sometimes move faster and have stronger incentives to attract investment, although a single example is not enough for a general conclusion. The Croatian operator now sits between an American platform and a Chinese autonomous driving system, which is a more valuable position than that of an end market.
Internal data shows that Codex is used by 98 percent of OpenAI employees, while among business customers the usage rate is 17 percent. On 24 August TechCrunch published internal OpenAI data indicating that 98 percent of employees used its Codex coding agent in June 2026, against 17 percent of business plan users and fewer than one percent of individual subscribers. ChatGPT Work and Codex together have approximately 20 million users, while ChatGPT overall has more than a billion. On 19 August, in a new update to its labour market indicator, Yale's Budget Lab again concluded that there is still no clear trace of an AI effect on the structure of occupations.
The gap between 98 and 17 percent is one of the most useful figures of the week. It suggests that capability claims often originate inside organisations with unusually good access to the technology, strong incentives and support for adopting it. Most companies face legacy systems, procurement controls, training needs and uneven demand, and a better model does not remove those constraints.
On 20 August Axios noted a shift in corporate language: after two years in which executives linked AI to efficiency and headcount, Etsy said that laying off 220 people was not driven by artificial intelligence, while Microsoft stated that 4,800 eliminated roles were not replaced by AI. This does not prove that AI played no indirect role, but it does make public explanations for layoffs a less reliable measure of labour substitution. On 12 August the Stanford Digital Economy Lab reported that employment among people aged 22 to 25 in occupations highly exposed to AI was around 19 percent below a comparison path based on peers in less exposed occupations, against 15 percent in the previous estimate. The authors describe this as an early descriptive indicator rather than a causal estimate. For an economy whose IT services sector relies on hiring large numbers of entry-level staff, even a signal framed this cautiously deserves attention.
Take away
The speed of regulatory action is worth treating as a property of a location with measurable business value. If a particular application has stalled in a large jurisdiction, it is worth assessing whether a smaller neighbouring market offers a lawful path to a pilot.
Before buying on the basis of capability claims, the likely degree of adoption should also be assessed. If a supplier records a 98 percent internal usage rate while the comparable rate among its customers is 17 percent, that gap shows how much implementation conditions matter, and a company has more influence over those than over any model roadmap.
Nvidia reports results on 26 August, with revenue guidance of 91 billion dollars, which will be an early indication of whether caution in the credit market reflects the trajectory of demand or merely the volume of debt issuance. From which follows one question for the next management meeting: among the commitments due to be signed in the next two quarters, which rest on a technical assumption, and which on an assumption about financing, permissions or adoption that nobody has yet verified?
Assumptions about cloud and inference pricing for 2027 need to be adjusted to the cost of capital as it changed during August, rather than simply extending this year's prices. Where capacity is procured under a group framework agreement negotiated in another market, renewal scenarios should be requested now. Every pricing, eligibility or scoring system that uses AI and is already in production should be reviewed against existing consumer protection, competition and data protection rules, not only against the AI Act implementation timeline, because those legal regimes apply already today.