Who Sets the Terms of Your AI?
Organisations are building real capabilities on technological foundations whose terms are set elsewhere. What matters when those terms change is how much room to respond a company has kept for itself.
In June of this year, a number of organisations saw how far their authority over their own tools extends. They had built real capabilities on two of Anthropic's most advanced AI models: workflows ran on them, staff had learned them, output depended on them. Then Anthropic suspended access to the two models worldwide to comply with United States export controls, restoring it on the first of July once the controls were lifted. The models did not cease to exist and the rest of the industry was untouched; no contract was breached, the provider was obeying its government, and no escalation available to a customer could reverse an export control. The capabilities those organisations had built were genuine. But for roughly three weeks, an important condition governing their continued use was determined outside the organisations themselves. How much that mattered would have turned on preparation, since for a firm with a tested alternative the suspension meant a configuration change, while for a firm without one it meant an interruption lasting as long as the controls did, and that variation is the subject of this essay.
Few companies have ever owned their payment networks, their logistics or their software, and the gains from specialisation are among the oldest arguments in economics. The capabilities organisations are building on AI are real in the only sense that matters commercially: work gets done that could not be done before. The distinction that deserves closer scrutiny is between the capability itself and the conditions that sustain it. The capability an organisation develops may be embedded within its own operations, but its sustaining conditions, the model's availability, the cloud's pricing, the chips' supply, the financing behind all three and the jurisdictions above them, are partly determined elsewhere. Control over those conditions is out of reach for almost every buyer, and chasing it would be a poor use of capital. Agency is a different matter. An organisation keeps agency when it understands its dependencies well enough to see change coming and has arranged, in advance, the room to respond: an alternative tested, a contract with teeth, an exit it has priced. The question worth asking about November's market turmoil, as about June's suspension, is how much of that room organisations have kept.
What November changed, and what it did not
Consider what the first week of November changed for a buyer. The index carrying the industry's largest companies fell into correction territory in the days after midterm elections whose results investors read as a sign of closer scrutiny for data-centre construction. Lenders were demanding yields above 16% on bonds of CoreWeave, a company that rents out AI computing, against under 6% for Meta paper of similar maturity. The four largest platforms have guided to roughly $730 billion of capital spending for 2026, about 78% more than last year on comparable definitions, and two of the four, Amazon and Alphabet, spent more in the second quarter than their operations produced in cash.
None of these facts changes any tool today, and none of them guarantees anything about next year. The reason they belong in a buyer's assessment is a chain with two distinct links. The first link is an observation: dearer financing and closer political scrutiny can change the economics of infrastructure investment. The second link is a possibility rather than a consequence: a provider whose economics have changed may review its prices, its contract terms, its product priorities or its capacity, or it may instead absorb the higher costs, reduce its capital expenditure, renegotiate its financing, improve its efficiency or delay its expansion. Which path any given provider takes cannot be known from outside, and we predict none of them. The point is more modest and more useful: the financial condition of the companies beneath an organisation's tools is one of the conditions that can move, and examining it is what gives a buyer the chance to see movement coming. Whether a particular customer would feel any such review depends on bargaining power and alternatives, which is to say on agency, and terms are hardest to negotiate at the moment they are being tightened.
How dependence has been governed before
As electricity networks expanded a century ago, many businesses found it increasingly economical to rely on centralised power generation rather than maintain their own generating capacity. The transition was gradual, it varied by country, industry and period, and it was not simply technological: it created new questions about the obligations of utilities, the regulation of prices and the responsibilities owed to users who had become dependent on a shared infrastructure. The societies that made that transition did not leave the relationship to commerce alone; they built institutions around it, because infrastructure that everyone depends on was understood to need governance beyond the bilateral contract. The analogy to AI has clear limits. Electricity was typically a local monopoly with no substitute, while AI computing is supplied by several competing providers, and a model, unlike a kilowatt, can often be replaced by a rival's at the cost of testing and migration. The comparison is useful for the question it raises rather than any answer it supplies: when organisations depend on infrastructure they do not control, what combination of public rules, contractual rights and private preparation preserves their room to act?
For now, the honest answer is that the burden falls mostly on private preparation, and the figures usually cited in the European debate show why, with each figure measuring something specific. Three American providers hold around 70% of Europe's cloud market, by Synergy Research Group's count, which is a measure of concentration in cloud services generally. Governments have begun to respond, with worldwide spending on sovereign cloud infrastructure projected by Gartner to reach $80 billion this year and European spending on it growing by 83% in 2026, though European alternatives still trail the American platforms in capability. Alongside concentration sits a question of legal jurisdiction: the US CLOUD Act establishes obligations for covered American providers to produce data in their possession, custody or control regardless of where it is stored, subject to legal process and the statute's conditions. It does not give authorities unrestricted access to everything these companies hold, but it does mean that the law applicable to data can follow the provider rather than the territory. Infrastructure concentration, legal jurisdiction and the broader ambition of technological sovereignty are related problems, but they are not the same problem, and an organisation mapping its dependencies needs to know which of the three it is actually exposed to. Public governance of this dependence may yet develop, and an organisation planning the next three years cannot wait for it.
Where the risk settles
Parts of the producing side have begun to redistribute risk among themselves by agreement. Nvidia has taken on payment obligations capped at $105 billion, disclosed in an SEC filing, backstopping lease and power commitments at a data-centre campus OpenAI is leasing in Ohio; the guarantee pays out only if specified defaults occur, and OpenAI has agreed to reimburse Nvidia for anything it pays. Critics describe the arrangement as circular, since a supplier is underwriting a customer whose spending returns to the supplier, and Nvidia's chief executive defends it as the ordinary use of a strong balance sheet to get necessary infrastructure built. One arrangement cannot stand for an industry; what it does reveal is how risk is being allocated at the top of the chain: suppliers, infrastructure developers and the largest customers have the visibility and the leverage to negotiate among themselves who bears which losses, and they are using both. Ordinary buyers sit outside those negotiations, though a guarantee that gets a data centre built may indirectly benefit every customer who later runs workloads on it. The difference is positional: the parties at the top shape the allocation of risk, while the ordinary buyer receives whatever allocation results, and the only risk decisions still open to that buyer are the ones inside its own walls.
Buyers hold a set of protections that deserve a precise accounting. Service-level agreements refund fees when a service fails, though they typically exclude interruptions required by law and rarely cover the value of the work that never happened. Larger organisations negotiate stronger terms, carry continuity arrangements and sometimes insurance, and maintain alternative providers and backup systems. The distinction that matters is that contractual protection and operational resilience are different things: a contract allocates compensation after a disruption, while resilience determines whether the disruption interrupts the work at all, and June rewarded only the second. Beyond the paperwork, exposure turns on variables that differ enormously from one organisation to the next: how substitutable the tool is, what switching would cost, how much negotiating weight the contract carried, and whether the vendor can pass its own rising costs through. Read this way, the familiar procurement variables are really measurements of agency: an organisation with a substitutable tool, a priced exit and a credible alternative negotiates as a party with options, while one without them accepts what it is offered.
The capability still has to be built at home
Access, meanwhile, does not by itself create capability, and the research record shows how persistent the gap between the two has been. MIT's NANDA initiative found that about 95% of generative AI pilots produced no measurable effect on profit and loss, a 2025 snapshot of early pilots rather than of mature programmes, and BCG found 60% of companies seeing minimal revenue or profit from the technology in the same year. The newest figures are better, with 28% of businesses reporting revenue growth from AI in Deloitte's latest survey, up from 20% earlier in 2026, while only a quarter of organisations have moved even 40% of their pilots into production and budgets concentrate on visible, front-office tools even as the clearest measured returns appear in finance, operations and process automation. These studies measure different populations with different methods, and none of them passes judgment on any particular organisation. Read together, they describe a persistent gap between obtaining access to a technology and building the internal capacity, the redesigned processes, the named owners, the agreed measures of success, that turns access into results.
External conditions and internal capability are separate dimensions of the same investment, and they can fail independently: an organisation can govern its vendors scrupulously while never redesigning its own work, and another can be superbly productive with tools whose foundations it has never examined. The two weaknesses can share a common origin when acquiring access is treated as the completion of adoption, a framing that discourages both the examination of external dependencies and the internal redesign of work required to realise the technology's value. June's suspension shows what preparation is worth. No public reporting names which organisations fared how, but the logic is straightforward: an organisation that had already tested an alternative model would have faced a configuration change, while one that had not would have lost the use of the capability for as long as the suspension lasted, and the difference between those two positions was created months earlier, when testing was cheap. Preparation of that kind confers no control over the chain and guarantees nothing, since resilience is a matter of odds and costs; what it reliably changes is the price of surprise.
Organisations already practise this discipline toward other conditions they cannot command. No company sets exchange rates, energy prices or shipping times, and the mature ones carry treasuries, hedges and second suppliers, an apparatus whose entire purpose is to preserve decisions in circumstances set elsewhere. Around AI, the equivalent apparatus is forming unevenly: regulated industries with vendor-risk programmes have a head start, while across much of the economy adoption still runs ahead of the mapping, and the questions specific to AI, questions of jurisdiction, model substitution and the finances of the ground layer, are new even to sophisticated buyers. Preparation does not convert uncertainty into certainty, since a well-prepared organisation can still be interrupted and an unprepared one can be lucky; what it changes is the exposure. The question the past six months have been quietly testing is whether each organisation keeps enough room to respond when the terms beneath its capabilities move. That is the question Cerndaia was set up to help organisations answer.
Sources
- Anthropic: Fable and Mythos access
- InvestmentNews: Nasdaq enters correction as AI stocks sink after midterm elections
- TIKR: CoreWeave bonds now yield over 17%
- FourWeekMBA: AI daily roundup, 7 November 2026
- ValueAddVC: AI spending, from Q3 2026 earnings releases
- Amazon Investor Relations: Second quarter 2026 results
- Tom's Hardware: Alphabet goes cash-flow negative as AI capex doubles
- The Register, reporting Synergy Research Group: European cloud market share
- Gartner press release: Worldwide sovereign cloud IaaS spending to total $80 billion in 2026
- The Register: European sovereign cloud spending in 2026
- Digital Chiefs: Digital sovereignty 2026
- US Department of Justice: CLOUD Act white paper
- Fortune: OpenAI data centre deal with Nvidia
- Data Center Knowledge: Nvidia backs OpenAI's Ohio data centre buildout with $105B guarantee
- Axios: Nvidia, OpenAI and AI financing
- Benzinga: Nvidia and OpenAI deal and circular financing
- Fortune: MIT report finds 95% of generative AI pilots failing
- AI for CFO: The AI ROI gap
- Forbes Tech Council: The productivity mirage
- Yahoo Finance: Deloitte report on AI progress
- Deloitte press release: State of AI in the Enterprise 2026