The Latent State

Who Controls the Compute?

Wed Nov 18 2026

On 9 November 2026, Deutsche Telekom announced that it would expand its AI factory in Munich. The initial phase had gone live in February and was fully booked seven months later, the company said. Siemens and Daimler Truck were already using the infrastructure. The next phase would take the site to 20 megawatts and add Nvidia B300 systems to its existing B200 machines. Germany had gained computing capacity that industrial customers could use close to their operations. For the expansion, Telekom went back to the same chip supplier. Deutsche Telekom, November 2026

Those two facts describe much of what governments now call sovereign AI. A domestic operator can decide who gets access to a cluster, employ the people who run it and host sensitive work nearby. The decisions that determine whether the cluster can grow, be repaired or move to a different stack are spread among equipment makers, software vendors, regulators and financiers.

The practical test is how much room a buyer has to act when one of those parties changes the terms. An inventory of servers tells you what is installed. A replacement plan tells you how long the operator can keep its options open.

Table of Contents

The capacity a country actually gets

France’s February 2025 AI summit produced announcements of more than €109 billion in infrastructure investment. Britain’s September 2025 state-visit headline was £150 billion in US investment commitments across multiple sectors. The British government’s notes show how much the categories matter. The total included £90 billion of forward commitments from Blackstone alone. Blackstone also reaffirmed an earlier £10 billion data-centre investment as part of its own £100 billion UK pledge. For a ministry trying to reserve computing time, the relevant figures are the capacity due to enter service, the date it becomes available and the share the ministry can book. Élysée, February 2025 · UK government announcement and notes to editors

Munich shows what can be gained. An automotive company can reserve capacity in Germany, negotiate directly with a domestic operator and run simulations near the teams that use the results. Telekom can build a business on those advantages. Its customers also have a commercial reason to buy the system that runs their existing code and comes with mature libraries and tools. Once engineers have tuned a production workload for Nvidia’s CUDA stack, changing accelerators can mean replacing libraries, adapting custom code and repeating performance and reliability tests.

A hosted model API leaves the model and its serving environment with the provider. Running a model on rented infrastructure gives the customer more control over software and scheduling, subject to the rental terms. Running it on equipment operated locally can add control over access, maintenance and capacity allocation. For public procurement, the first decision is which of those rights the service actually needs.

flowchart TD
    accTitle: Who can interrupt a locally operated AI service
    accDescr: The local operator schedules workloads. External dependencies include replacement hardware, software, licences for future shipments, site financing and electricity.
    Operator["Local operator"] -->|Schedules and runs| Service["AI service"]
    Vendor["Hardware and software suppliers"] -->|Parts, drivers and support| Operator
    Authority["Export authorities"] -->|Licences for covered shipments| Vendor
    Site["Utility and financiers"] -->|Power and terms of expansion| Operator
    Customer["Customer"] -->|Data, model and workload choices| Service

A licensing decision may first be felt when an operator orders the next batch of cards. A software decision may surface when a newly purchased accelerator needs a different driver or compiler. A power constraint can postpone a whole expansion even when the equipment is available. In its 2025 Energy and AI report, the International Energy Agency projected global data-centre electricity use of around 945 TWh in 2030, more than double its level at the time of the study. At a particular site, the useful number is the megawatts the utility has agreed to deliver by a particular date. International Energy Agency, Energy and AI · The Power Bill

Where Washington has leverage

In April 2025, the United States told Nvidia it would need a licence to export its H20 chips to China. Nvidia reported a $4.5 billion charge for excess inventory and purchase obligations in the quarter that ended on 27 April. In January 2026, the Commerce Department announced that it would review applications for the more powerful H200, AMD’s MI325X and similar chips case by case, subject to requirements including third-party testing and customer screening. The decisions show how quickly the government can change the economics of future supply, including for an American vendor. A licence review policy still leaves each exporter waiting on the outcome of its application. Nvidia earnings release · US Bureau of Industry and Security, January 2026

The country that designs a processor, the country that fabricates it and the country that licenses its export can all be different. Under the US Foreign Direct Product Rules, specified foreign-made items can fall under US export regulations based on the technology or software used to produce them, or on the equipment in the plant that makes them. Each rule has its own conditions for the product and the transaction. For a buyer, the immediate question is whether a specific shipment is covered. The producing country alone cannot answer it. 15 CFR § 734.9

ASML shows the reach and the limits of American leverage. The Dutch company supplies the extreme-ultraviolet lithography tools used to pattern many leading-edge chips. Controls also cover certain advanced machines that use deep ultraviolet light. Dutch licensing decisions matter in their own right, while the US can exercise separate leverage where its own rules apply. Washington depends on the continued cooperation of countries with equipment makers and foundries that it cannot readily replace. The Choke Point

Congress has also considered extending oversight beyond the point of sale. The Chip Security Act, introduced in 2025, proposed location verification for covered advanced chips and reporting in certain cases of possible diversion. That would give regulators an additional source of information after shipment. It would also give legitimate operators reason to ask who could receive the resulting information and what would happen if their equipment failed to verify its location. The Senate bill’s listed actions as of November 2026 were introduction and referral to committee. S. 1705 bill text · Legislative status record

A local server can still answer to a foreign court

The US CLOUD Act addresses a different form of control. Under 18 USC § 2713, a covered provider must comply with applicable preservation or disclosure obligations under that chapter for data within its possession, custody or control, regardless of where the data is stored. A server in France can therefore be relevant to a US legal process. The law provides a route for providers to contest certain orders when disclosure risks violating the law of a qualifying foreign government and other conditions are met. Which entity operates the service, what it can access and which legal procedure applies matter alongside the location of the facility. 18 USC § 2713 · 18 USC § 2703(h)

Who can grant administrative access? Can support staff outside the country reach the production environment? Which components can see plaintext while the workload runs? Customer-managed keys can reduce a provider’s access in some designs. A key held outside the provider’s control still leaves questions about live processing, logs, backups and the control plane through which administrators operate the service. The answers depend on the architecture and contract for the specific product.

The EU Data Act, applicable since September 2025, addresses a second part of the problem. It introduces obligations to make switching between data-processing services easier and provisions concerning unlawful third-country government access to non-personal data. The switching rules give customers a better basis for asking what can be exported and on what timetable. The work of reconstructing infrastructure permissions, moving a large dataset and validating the restored service still needs an operator, a destination and a rehearsal. European Commission, Data Act

What a supplier can ask in return

In 2024, Microsoft announced a $1.5 billion investment in Abu Dhabi’s G42, including a board seat and plans to run G42 applications and services on Azure. Bloomberg reported that the Emirati company had agreed with the US government to reduce its exposure to Chinese technology before the deal. In May 2025, OpenAI and its partners announced Stargate UAE, a planned one-gigawatt computing cluster. The UAE was offering land, power and capital on a scale most countries could struggle to match. For G42, access to American technology also came with constraints on which other suppliers it could use. Microsoft’s announcement · Bloomberg, via Business Standard · CNBC, Stargate UAE announcement

Dependence also runs through less celebrated parts. In autumn 2025, the Dutch government intervened at Nexperia, a Netherlands-based chipmaker owned by China’s Wingtech, citing governance concerns. China then restricted exports of Nexperia chips, disrupting a supply chain in which chips made in Europe were sent to China for packaging. Carmakers were exposed to shortages of inexpensive components. The Dutch government suspended its order in November. The dispute exposed an industrial risk well beyond frontier AI. Even an inexpensive qualified component can depend on a single acceptable route from wafer to finished part. The Guardian, November 2025

Who finances the next upgrade

In September 2025, Nvidia and OpenAI announced a letter of intent covering at least 10 gigawatts of Nvidia systems. Nvidia said it intended to invest up to $100 billion in stages as the systems were deployed. In February 2026, OpenAI announced a funding round that included $30 billion from Nvidia, along with plans to use five gigawatts of Nvidia capacity for training and inference. The announcements describe different arrangements. What they share is a supplier on both sides of a relationship. Nvidia stands to benefit from the customer’s demand for its hardware and has a stake in whether the customer can raise the money to pay for it. Nvidia and OpenAI, September 2025 · OpenAI, February 2026

SoftBank’s accounts give a different view of the flows. The Japanese group said it put $30 billion into OpenAI through Vision Fund 2 across two closings in 2025. Other investors contributed $11 billion, bringing the funded total to $41 billion. SoftBank gained an equity stake in a US company. A Japanese customer seeking guaranteed computing capacity would need a separate arrangement. The distinction matters when a profitable overseas investment is counted as progress towards a country’s industrial goals. SoftBank Group, December 2025

The Bank for International Settlements projected in 2026 that the five largest hyperscalers would spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026. Its annual report describes supplier investments and customers’ purchase commitments linking the fortunes of companies that appear independent on paper. Private credit adds another layer. An investment fund can finance a specialised data-centre developer against revenue promised by a large tenant. If demand falls or the tenant runs into trouble, the lender’s protection depends on the contract, who guarantees it and how easily somebody else can use the site. A country renting critical capacity from the project has its own version of that problem. It needs to know who can keep the service running if the operator must renegotiate with lenders. BIS, 2026 Annual Economic Report · BIS Bulletin 120 · The Capex Confession

The financial precedents

Early twentieth-century US dollar diplomacy tied loans and financial supervision to strategic interests. In Nicaragua, US involvement in the country’s finances went hand in hand with military intervention. After the US ended dollar convertibility into gold in 1971, a different relationship took shape. Major oil exporters earned dollars and placed some of the proceeds in US financial assets. The US–Saudi arrangements of the 1970s helped organise those flows. Dollar oil pricing spread through markets and commercial practice across producers. US State Department, Office of the Historian · Middle East Institute, US–Saudi joint commission

Dollar stablecoins show how such arrangements can survive a change in technology. Someone can use Tether’s USDT to hold or transfer a dollar-denominated asset outside a traditional bank account. Tether, as issuer, places part of its reserves in US government debt. At the end of 2025 it reported about $141 billion in direct and indirect US Treasury exposure. The US GENIUS Act, enacted that year, established a framework for payment stablecoins and allows several types of eligible reserve assets, including certain short-dated Treasuries. Demand for digital dollars can therefore support demand for familiar dollar assets. Tether’s year-end disclosure · GENIUS Act, Public Law 119-27

Computing has some of the same features. Overseas demand supports American suppliers and investors. Access can carry political terms. There is a difference in how switching works. A barrel of the right grade can often be sourced from another producer. A replacement accelerator can require a different compiler, interconnect and set of libraries, and weeks or months of testing. Sellers gain leverage from how much of that work the customer has already done on their platform.

What China has paid to build another option

China’s third national semiconductor investment fund was registered in 2024 with 344 billion yuan in capital, roughly $47.5 billion at the time. The number describes the registered capital of that phase. The policy behind it gives Chinese firms a chance to develop manufacturing tools, materials and other capabilities that customers would otherwise have little financial reason to buy early. Investors can wait longer for returns. They also risk funding projects that fail to reach competitive yields or acquire customers. Reuters, May 2024

The technical evidence is more interesting than a count of national suppliers. DeepSeek reported that the full training of V3 took 2.788 million Nvidia H800 GPU-hours, including pre-training, context extension and post-training. Its roughly $5.6 million calculation used an assumed rental rate of $2 per GPU-hour and excluded prior research and architecture experiments. The model showed what a team could gain from using its available hardware more efficiently. The training run was still carried out on Nvidia chips. DeepSeek-V3 technical report, Table 1

At the hardware level, Reuters reported in October 2024 that a TechInsights teardown found a chip made by TSMC in a Huawei product. That finding illustrates why the nationality of the brand reveals so little about the entire system. A domestic accelerator project still needs fabrication, advanced packaging, high-bandwidth memory, networking and a software stack that other people can use. The commercial test is whether customers keep using it when they can buy a faster or easier-to-program alternative. The strategic test is whether it can carry enough essential workloads if that alternative goes away. Reuters, October 2024

Europe has a less centralised way of building options. The EuroHPC AI Factories programme pools public computing capacity across countries and gives eligible researchers and companies a route to apply for access. A publicly governed allocation can solve a real problem for firms that cannot afford their own cluster. A second route to capacity is more valuable if it can withstand the disruption that took out the first. Two operators may use the same chips, rely on the same vendor for support or buy power from the same constrained grid. EuroHPC Joint Undertaking, AI Factories

The test is what keeps running

A hospital deploying a document-processing model, a carmaker running simulations and a lab training a frontier model have different tolerance for an interruption. A hospital may need to keep inference running on an older machine for a month. A carmaker may need capacity on another platform by the next design cycle. A research lab may need several thousand tightly connected accelerators for its next training run. A credible contingency plan begins with those differences.

flowchart TD
    accTitle: How to test a critical workload for supplier dependence
    accDescr: Choose a minimum service level and test whether the workload can continue on current equipment and recover on a second platform within the required time.
    Work["Set a minimum service level and recovery deadline"] --> Installed["Run on existing equipment without vendor support"]
    Work --> Second["Restore the workload on a second platform"]
    Installed --> Record["Measure capacity, recovery time and operating cost"]
    Second --> Record
    Record --> Contract["Reserve the capacity and rights the service needs"]
If this changes What the operator needs to demonstrate
A new shipment requires a licence How long existing capacity and spare parts can sustain the service
A cloud contract expires How quickly the team can restore the workload elsewhere, including data, model weights where available, permissions and logs
A foreign authority seeks disclosure Which entity can produce which data, under what process, including backups and live access
The site operator loses financing or power Who has contractual rights to keep the service running and where replacement capacity is reserved
The incumbent offers a substantially better price What it costs to keep an alternative tested and available

Some countries will pay to build manufacturing capacity. Others can get significant protection by maintaining an independent operator, qualifying a second architecture or pooling spare capacity with allies. These choices have recurring costs. A second system requires engineers and work to run on it. A backup cluster booked to full capacity has little room for anyone else during a disruption. And a written right to move a workload has limited value when the other site needs six months to connect to the grid.

The Munich expansion is a real gain for German industry. The next measure of its value is whether the workloads developed there can keep running through a supply interruption and move when customers need them to. Governments will learn more from the time and money it takes to do that than from the size of the next investment announcement.

Sources