Special Analysis
For three years the strategic question in the business application of artificial intelligence has been which model to choose. This week's events point to something else. In four separate decisions, taken in four jurisdictions by four different kinds of actor, the variable that determined who is allowed to supply and deploy artificial intelligence was not the capability of the technology but permission: who may sell, who may buy, and through whose components, licences and models the transaction has to pass.
Taken individually, a chipmaker's quarterly results, a data centre tender in Cairo, a court order in San Francisco and a procurement decision in Seoul have little to do with one another. Taken together, they describe a market in which the supply of artificial intelligence is allocated by physical scarcity, filtered by export policy, constrained by government designation and shaped by rules on domestic origin. For an organisation that buys artificial intelligence rather than builds it, this moves the critical risk from the model layer to the procurement layer.
Components and export licences
NVIDIA reported revenue of 96.2 billion dollars in the second quarter, up 106 percent year on year, with data centre revenue of 89 billion dollars and guidance of approximately 108 billion dollars for the third quarter. The demand figure is not the most interesting part. The two figures beneath it are.
The first concerns cost. The results statement projects a gross margin of 74 percent in the third quarter, but on the investor call CFO Colette Kress said margins were expected to bottom out in the fourth quarter at 71 to 72 percent, attributing the pressure to memory prices and a memory shortage caused largely by the build-out of AI infrastructure itself. The most profitable supplier in the accelerator market is telling investors that its margin is now set by the component market.
The second concerns policy. The third-quarter guidance explicitly assumes no data centre compute revenue from China. Export control is no longer context around the forecast. It sits inside it, as a stated assumption.
The scale of the build-out, meanwhile, is not slowing. On the same day, AWS and NVIDIA announced plans to deploy two million additional NVIDIA GPUs across AWS's global infrastructure during 2027 and 2028, together with a separate capacity of 100,000 processors on secured infrastructure for United States federal and security purposes.
The composition of the risk has changed, not its size. AI infrastructure budgets are increasingly exposed to the economics of a component market that no buyer controls and no manufacturer's roadmap can cushion, while export policy has become material enough to be written into a revenue forecast as an assumption. Anyone planning multi-year AI capacity should treat the hardware line as commodity exposure rather than a settled technology decision, and should expect that exposure to arrive through the price of cloud even if they never buy a single server.
The same pattern is visible in procurement at state level. According to published reports, Huawei responded to an Egyptian government tender with an offer to export 1,408 of its top-end Ascend 950 series chips for a cloud intended for model training, together with a further 600 units, either the same chips or the older 910B model, for two inference clusters, with a twelve-month build plan. According to the same reports, the State Department approached NVIDIA, AMD and Microsoft regarding a possible competing consortium. None of those three companies has confirmed participation, the American bid has not been confirmed, and Egypt has not selected a winner. This is a reported offer and reported discussions, not a signed contract.
The structure is what makes the case worth attention even before the outcome is known. It is reported that this could be the first time the United States and China compete directly for the same government tender for an AI data centre, and that shipments of advanced American AI processors to Egypt have required an export licence since 2023. Sovereign procurement of artificial intelligence thereby acquires a visible geopolitical layer: the choice of hardware can simultaneously determine which export controls, licence restrictions and long-term supply chain dependencies the buyer takes on, and those terms are set by governments that are not party to the contract. Public sector and regulated buyers evaluating larger infrastructure projects should treat licence exposure, re-export conditions and substitution paths as first-order commercial terms, not as a legal appendix to be settled after the technical evaluation.
Suppliers and model provenance
On 27 August, a federal judge in San Francisco ordered the Pentagon to remove the designation under which it had labelled Anthropic a supply chain risk, finding the measure unlawful. In a 59-page order, District Judge Rita Lin found that the designation constituted unlawful retaliation and a violation of the First Amendment, and that the company had been denied the process required by the Fifth Amendment. The dispute followed Anthropic's refusal to permit military use of its models for mass domestic surveillance and for fully autonomous weapons.
The mechanism matters more than the participants. This is the first time an American company has been publicly designated a supply chain risk under procurement rules written to protect military systems from foreign sabotage. A supplier was removed by a designation, not by a contractual dispute or a failure to perform. And the reversal is partial: a separate proceeding in Washington, concerning a different designation, has not concluded.
Supplier continuity in artificial intelligence can therefore be interrupted by a decision of state policy rather than by commercial failure or technical fault, and restored slowly and incompletely, through litigation. The useful response is not to choose a different supplier. The response is to build portability into critical AI deployments: the right to substitute models, export of data and prompts, and an exit plan, so that a decision taken in one jurisdiction does not halt a production system in another.
Model provenance, meanwhile, is becoming a matter of procurement regulation. On 28 August, South Korea's Ministry of Science and ICT selected three consortia, led by SK Telecom, KT and Kakao, from six applicants, to operate an "AI for all" service: a general-purpose chatbot and public service agent available across the entire country without usage restrictions. The state will make a total of 512 NVIDIA B200 GPUs available to the three operators, and from 2027 intends to cover the cost of the public service from the national budget, with an obligation on operators to develop their own revenue models.
The binding condition is a rule on origin. Operators must run at least 50 percent of the service on domestic AI models meeting state criteria for independent foundation models, and at least a further 30 percent on domestic models developed by Korean companies other than the operator itself. That second provision prevents any single operator from serving the entire country exclusively on its own technology.
The government has thus used a consumer service to write model provenance into procurement as a hard percentage. It is simultaneously a distribution decision and an industrial policy decision: it determines what citizens will expect from a public digital service and it channels demand towards domestic development teams. Suppliers bidding for public sector work should expect provenance to become a criterion that is assessed rather than a claim made in positioning, which means being able to demonstrate where a model was trained, where it runs and who controls it.
Take away
None of these four events turned on what a model can do. Memory availability, export licences, a government designation and a percentage of domestic origin were decisive. Capability has become an assumption. Permission has become the variable.
For organisations in South East Europe that will buy artificial intelligence rather than build frontier models, this carries a concrete consequence. The competence that decides whether an AI programme survives the next two years is not model selection. It is contracting: knowing which licence regime the chosen stack sits inside, what happens if a supplier is designated, how a shock in component prices reaches you through your cloud bill, and what proof of provenance a public buyer will demand. These are procurement questions, and they can be answered in advance.
The question worth asking: if your primary model supplier became unavailable tomorrow for reasons having nothing to do with its technology, how long would it take you to move to another, and what would it cost?