Data Centres Are Not AI Sovereignty: Why Control Depends on the Business Model
Around the world, from Switzerland to Abu Dhabi, governments and economists answer the AI question with the same demand: build more data centres. This post argues that a data centre on your own soil gives you no control over AI. Control is decided by the business model, above all by who owns the model you run and the chips you run it on. I take two recent interviews, with Danny Bürkli and Monika Schnitzer, as the case and run the data centre through the four elements of my business model framework. I then look at who pays for the power and who carries the risk of stranded assets, and ask why this is bad strategy in Richard Rumelt’s sense. China’s leap from the combustion engine to the electric car shows the alternative: start a new race. The post ends with four questions for any AI strategy and five lessons for every industry.
The AI Data Centre Business Model Nobody Examines
There are many interviews like this one. In an October 2026 interview with the Swiss daily Tages-Anzeiger, Danny Bürkli, a Swiss political scientist who advises governments on AI for the Windfall Trust, lays out the stakes with care. AI could automate large parts of every digital service. Labour markets, tax bases and social systems would feel it. His remedy for Switzerland: build far more AI data centres, because computing power is becoming the critical input of every economy, much like oil. Europe, he says, hosts only about five percent of global capacity. The limiting factor is electricity.
A few months earlier, Monika Schnitzer, chair of Germany’s Council of Economic Experts, made a similar case in an interview with DER SPIEGEL in June 2026. What she misses most in the German debate, she said, is a vision of what the economy will look like in the future. In the next sentence she named what Germany needs: the infrastructure to take part in AI at least to some degree. Her concrete proposals were an AI safety institute, more skilled workers, and luring American AI companies to build data centres in Europe, following France’s “Choose France” campaign.
I do not single out either of them. Both argue more carefully than most. But read the two interviews side by side and a pattern appears. Asked for a vision, the debate answers with a shopping list of inputs: compute, power, access, talent. Customers do not appear. Business models do not appear. Nobody asks who will earn money with all this capacity, or who will control it.
The two interviews are local voices of a global pattern. In May 2025, OpenAI announced Stargate UAE, a one-gigawatt cluster in Abu Dhabi built with G42, Oracle, Nvidia, Cisco and SoftBank. It is the first partnership under “OpenAI for Countries”, an initiative that, in OpenAI’s own words, helps governments build sovereign AI capability in coordination with the U.S. government. Saudi Arabia’s Humain has signed similar deals with Nvidia, AMD and Amazon, and the UK is courting data centres with “AI Growth Zones”. Sovereignty has become a sales promise, and the owners of the bottlenecks are the ones making it.
This is the same pattern I described in the powertrain post. There I argued that Europe is on course to become a customer of the battery business model rather than its author. With AI the same pattern is repeating, only faster. And it begins with a mistake about what a data centre actually is.
What We Are Talking About: A Short Glossary of the AI Stack
The debate mixes up terms that describe very different things. Before the analysis, here is how I use them in this post.
| Term | What it is | Why it matters here |
|---|---|---|
| Model | The trained software that answers a request. It consists of billions of learned parameters, the so-called weights. | The core component of the stack. Whoever holds the weights decides who may use the model. |
| Frontier model | The most capable models at any given time, built by a handful of labs such as OpenAI, Anthropic and Google. | Usually delivered as a service and not handed over. |
| Open-weight model | A model whose weights are published, so that anyone can run it on their own hardware. | The only form in which a model runs fully under the operator’s control. |
| Training | The process that creates a model. The lab feeds it vast amounts of data and adjusts its weights over weeks or months on very large clusters of chips. It is repeated for every new model generation. | Training produces the asset. Whoever trains the model owns its weights, and with them the right to decide who may use it. |
| Inference | Running a finished model to answer a request: every question, every generated text, every agent step. It happens continuously, wherever the model is served. | Inference is where the model meets the customer. It happens with every single request, and that is where access is granted or refused. |
| Access (API, licence) | The interface and the contract through which a customer uses a model that someone else owns. | The login. This is where control sits, wherever the servers stand. |
| Data centre | A building with power supply, cooling and network connections that houses servers. | The most interchangeable layer of the stack. |
| Servers and AI chips | The computers inside the data centre and their specialised processors, such as Nvidia’s GPUs or Google’s TPUs. Replaced every three to six years. | Almost 70 percent of the upfront capital, and a recurring payment to a few suppliers. |
| Hyperscaler | The largest cloud operators, such as Amazon, Microsoft and Google, which build and run data centres at scale. | Usually the tenant or owner. The host country supplies them with land and power. |
| Colocation | A data centre that rents space, power and cooling to tenants who bring their own servers. | The tenant may have no local operations at all. |
| Export controls | State rules on who may receive chips, chipmaking machines or models. The US foreign direct product rule extends them to foreign products that contain US technology. | The lever of power politics in the AI stack. |
| Large-load tariff | A utility contract for very large electricity users, with minimum payments, long terms and collateral. | Decides who carries the risk if the data centre leaves. |
| Stranded asset | An investment that loses its value before the end of its life because the demand it was built for disappears. | The risk host countries take on when they build power and grid for a data centre. |
Factory Logic: Why Owning the Building Is Not Owning the AI
Both interviews think in factory logic. In the industrial world, whoever holds the factory can produce. If the owner sits in another country and one day decides the plant should stop, the machines are still on your soil, the workers still know how to run them, and in an emergency the state can take over. The factory is the means of production, and it stands where it stands.
An AI data centre does not work that way. The value does not sit in the building. It sits in the model and in the chips, and control over the model is exercised continuously, with every single request. Whoever grants access through authentication, licences and remote management decides whether the model answers. Owning the server it runs on gives you no say in that.
The distinction between training and inference makes this concrete. Training creates the model and with it the asset; it happens where the lab chooses to do it. Inference is the daily use of that asset. A data centre that runs a foreign frontier model for local customers hosts the inference of a model trained, owned and controlled elsewhere. The host gets the electricity bill of every request. The owner keeps the asset that makes the request possible.
The point is no longer hypothetical. On June 12, 2026 the US government issued an export-control directive that suspended all access to Anthropic’s Fable 5 and Mythos 5 models by any foreign national, inside or outside the United States. To comply, Anthropic had to disable both models for all of its customers. The company disputed the technical basis of the order, the controls were lifted on June 30 and access returned on July 1. For the argument here, one detail decides everything: the directive was tied to the nationality of the users, not to the location of the servers. A data centre in Zurich or Frankfurt would have changed nothing.
There is a second reason the factory analogy fails. Data centres are among the most interchangeable assets in the AI stack. They can be built, rented and relocated, and many companies offer them. The frontier models come from a handful of labs. The chips come from even fewer suppliers. The bottleneck is not the physical asset. A strategy that secures the building secures the one part of the system that was never scarce.
The Value Proposition Sounds Great. The Architecture and the Revenue Model Tell a Different Story
To see why, run the data centre through the four elements I use in my book Das Richtige gründen: value proposition, value-creation architecture, revenue model and company spirit. The framework is useful here because it forces a question the interviews skip. Whose business model are we looking at?
Start with the customer. In a data-centre deal, the country is not the customer. Its offer is addressed to the hyperscaler: cheap and reliable power, land, fast permits, a grid connection, often tax breaks. The host country becomes a supplier in someone else’s business model. Once you see that, the four elements read very differently from the press release.
| Element | What is promised | What the analysis shows |
|---|---|---|
| Value proposition | Sovereignty, access to AI, prosperity, jobs | A location. Access is granted with every login, and the model owner holds the login. |
| Value-creation architecture | Our own infrastructure, our own capacity | The country owns the most interchangeable layer: building, power, grid. Model and chips belong to others. |
| Revenue model | Investment, tax revenue, value added at home | Most of the money flows to the owners of the bottleneck assets. The cost of the grid and the risk of stranded assets stay at home. |
| Company spirit | “We have to be part of it” | A defensive, imitative stance: a race run by rules others have written. |
The value proposition: written for the voter
The value proposition of a data centre is attractive because it is visible. It has a groundbreaking ceremony, an investment figure and a job count. But a value proposition is the benefit you promise a real customer. The benefit the country actually receives is a tenant on its land. The capability that matters, a model that answers, is delivered as a service by its owner and can be withdrawn.
Sovereignty is the promise. Access under licence is what you get
The same gap runs through the offers the AI companies make to governments, and it helps to keep two things apart. The value proposition is the benefit that is promised. The product is what you actually receive. Promises cost nothing; what decides is what you get.
In the powertrain post I wrote that what the customer buys is never the product but the benefit the product delivers. That remains true for what the customer wants. A government that signs a sovereign-AI deal wants sovereignty, and so would every one of us. The tension appears when the promised benefit and the delivered product come apart: the customer pays for the promise and receives the product, which is delivered by a whole value-creation architecture that cannot support the value the customer wanted.
“OpenAI for Countries” promises sovereign AI capability. Read the promise on its own terms and the contradiction is already in the sentence: sovereignty, built in coordination with the U.S. government. What the country actually receives is capacity on its soil and access to a model under licence, with the model, the chips and the right to switch it off remaining with someone else. Sovereignty is promised. It is not part of what is delivered. This is not a point about one company. It applies to every provider that sells sovereignty while keeping the core components, OpenAI’s competitors included.
I made the same argument about renewable diesel in the HVO100 post: a value proposition the customer loves still fails if the architecture cannot deliver it. With sovereign AI the customer loves the promise, and the architecture is built so that it cannot be delivered. Before signing, every government should ask one plain question: what exactly do we get, and who can take it away?
The value-creation architecture: one borrowed block on someone else’s land
In the post on online marketing I described what makes a business model defensible: the uniqueness of the business model, or UBM, which arises from several interdependent components you own yourself. A data centre running a foreign frontier model on foreign chips contains no such component. The model is borrowed, the chips are bought, and the building could stand anywhere with enough power.
The revenue model: where the money actually goes
Here the numbers are unambiguous. Epoch AI has modelled a stylised one-gigawatt AI data centre owned and operated by a US hyperscaler and equipped entirely with Nvidia GB200 systems, at US average costs. Of its roughly USD 38 billion in upfront capital, the building and its technical systems account for about 30 percent. Servers and networking account for almost 70 percent. Land and grid connection together come to about one percent.
That split understates the asymmetry, because the two parts age at different speeds. Epoch assumes 14 years for the facility and five years for the IT equipment, and critics argue that the real economic life of AI chips is closer to three. The building is paid for once. The servers are bought again and again. On my own rough arithmetic, at constant prices per generation, close to nine out of ten dollars invested over the life of the building go to hardware: 86 percent if the servers are replaced every five years, 91 percent if every three. Over the same period the money flows to a very small group of owners: Bernstein estimates that Nvidia’s gross profit alone equals about 29 percent of the total cost of an AI data centre, and Nvidia’s gross margin stood at 75 percent in the quarter to July 2026.
What stays local is modest. In Epoch’s annual cost breakdown, labour comes to about USD 40 million out of roughly USD 8.5 billion, about half a percent. The best causal study so far, by Bahar and Wright for Brookings, finds that data centres do create local jobs, though fewer than the industry claims, and that colocation facilities often serve tenants with no local operations at all. Even the operator in the middle earns little. Oracle has reportedly struggled to reach a gross margin above 25 percent in its AI infrastructure business, a point I made in the Oracle post. The data centre sits sandwiched between two owners of bottlenecks: the chips below it and the model above it.
The company spirit: running someone else’s race
The fourth element is the one I care about most, because it is about people and values. Here the stance is defensive. “We have to be part of it” is a fear of being left out, and it produces imitation. The people who build the components that matter, the model researchers and chip designers, work elsewhere. The data centre gives the country something to show. It does not give it anything to build on.
The Hidden Bill: Power, Stranded Assets and Who Carries the Risk
Bürkli is right that power is the limiting factor. What his argument leaves open is who pays for it, and who carries the risk if the demand does not last.
Start with the lifetimes, because they do not match. Server and GPU fleets are depreciated over three to six years. The data-centre building is modelled at about 14. The best large-load contracts that US utilities now demand run for 12 to 15 years. The grid assets built to serve them last far longer: FirstEnergy plans with 40 to 50 years for transmission transformers and 50 to 60 years for line conductors. A country that builds power for a data centre commits for decades to a customer whose own equipment lives a few years.
Whether other customers already subsidise data centres is contested, and both sides deserve a hearing. The Harvard Electricity Law Initiative documents how special contracts and regional cost allocation can shift the costs of serving data centres onto households. In the PJM grid, the independent market monitor estimated that data centres caused 63 percent of the price increase in the 2025/26 capacity auction, USD 9.3 billion recovered from all customers. On the other side, E3’s review of the few quantitative studies found no evidence of a cost shift so far, including in Virginia, the world’s largest data-centre market. The honest reading is that the past is disputed and the future risk is widely acknowledged. Many US states have reacted with large-load tariffs. Virginia’s Dominion GS-5 class, effective January 2027, requires 14-year contracts, payment for at least 85 percent of contracted transmission and distribution demand and 60 percent of generation demand regardless of use, and upfront collateral to cover grid upgrades.
Then there is the demand itself. The cost of AI inference at a given level of performance has been falling by a factor of between 9 and 900 per year, depending on the task according to Epoch AI, and chips keep getting more efficient. That can cut both ways. Cheaper AI may simply be used more. But it has happened before that efficiency outran demand: between 2010 and 2018, global data-centre computing grew by 550 percent while electricity use rose by only 6 percent (Masanet et al., 2020). Jonathan Koomey, a co-author of that study, recently put the state of knowledge bluntly: nobody knows what AI electricity use will be even a few years ahead, and planners should include scenarios in which it declines.
The data centres themselves age as well. Uptime Institute’s 2025 survey found that 82 percent of operators had a highest-density rack below 30 kW, and its research director estimates that fewer than four percent of US data centres can currently support the newest AI racks. The industry tends to build new rather than retrofit.
Put this together and the risk structure becomes clear. The general grid and generation behave like optical fibre: if the data centre leaves, they still serve electrification, heat pumps and industry. Everything built specifically for one campus behaves like the data centre itself, and loses its value when the tenant goes. The more interchangeable the data centre, the easier it is for the tenant to leave once the contract ends. The host country writes an option, and the owner of the bottleneck holds it.
Bad Strategy: Why the Data Centre Misses the Crux
Why, then, do so many AI strategies start with a data centre? I see two reasons, and they reinforce each other. The first is a lack of understanding of the technology, which leads people to reach for the logic they know: the factory. The second is the wish to be seen at the front. In AI there is little to show. A data centre is the one thing you can open, photograph and quantify.
Richard Rumelt has a name for this. In Good Strategy Bad Strategy, two of the hallmarks of bad strategy are a failure to face the challenge and the mistake of treating a goal as a strategy. “Being part of AI” is a goal. Without a diagnosis of where the bottleneck sits, no action can address it. I applied the same test to renewable diesel in the HVO100 post.
In The Crux (2022), Rumelt sharpens the idea. The crux is the most important part of a set of challenges that is also addressable, meaning it has a good chance of yielding to coherent action. Tellingly, Rumelt himself names dependence on a peer competitor for critical supplies as the crux of today’s security challenge. Apply his two filters and the options sort themselves.
| Critical? | Addressable? | |
|---|---|---|
| Building data centres | No: the most interchangeable layer | Yes |
| Owning today’s frontier models and chips | Yes: the bottleneck | Hardly: that race is decided |
| Starting a new race | Yes | Yes, where you already have capabilities |
A data-centre strategy picks the part of the challenge that is addressable but not critical. It is the political equivalent of polishing the car while the engine is somewhere else. The wish to show something tangible is understandable. It simply points at the wrong layer.
Contracts Do Not Protect in a World of Power Politics
The usual answer to all this is contractual: we will secure access through agreements. Bürkli suggests that Switzerland’s foreign and economics ministers should be given an explicit mandate to secure access to US frontier models. In a world of rules, that would be a reasonable plan. We no longer live in one.
The political scientists Henry Farrell and Abraham Newman call the mechanism “weaponized interdependence”. Global networks grow asymmetrically, with a few central nodes. States with jurisdiction over those nodes can use them for coercion, most directly through the “chokepoint effect”: denying others access to the network. The AI stack is a textbook case, and the chokepoints have already been used at every level. At the level of the model, the June 2026 directive. At the level of the chip, US export controls. At the level of the machine that makes the chip, the Netherlands’ restrictions on ASML, introduced step by step under American pressure. When the Dutch rules went less far than Washington wanted, the US used its foreign direct product rule to claim jurisdiction over ASML machines containing US technology. Europe’s strongest component is itself under foreign control.
The United Arab Emirates show how far this goes even for a buyer with almost unlimited capital. In November 2025 the US Commerce Department authorised exports of up to 35,000 Nvidia Blackwell chips to G42 and Saudi Arabia’s Humain, with security and reporting conditions attached. In July 2026 the US moved the UAE into its most favourable export tier, which allows approved companies to buy without individual licences. That access is time-limited. According to Bloomberg, G42 may buy licence-free only until around April 2027 and needs to change its corporate structure for continued access. Its executives have held exploratory talks about selling a majority stake to American companies or setting up a new vehicle in the US; no final decision has been made. To keep access to the chips, the flagship of Abu Dhabi’s AI strategy is considering giving up control of itself. No contract could have expressed the balance of power more clearly.
A balance of deterrence does exist in this world, but it runs between the two powers that own core components. In April 2025 China placed export controls on seven rare-earth elements. In October 2025 it extended them beyond its borders for the first time, to products made abroad with Chinese rare-earth materials or technology, its own version of the American rule. It kept the controls as a bargaining chip even during the trade truce, and the effect is real: Japan has received only four percent of its previous dysprosium imports since the restrictions began. Deterrence is available only to those who own something the other side cannot do without. A data centre running someone else’s model on someone else’s chips is not such a thing.
Do Not Chase a Race Whose Rules Others Have Set: What China Learned from the Combustion Engine
If the current race is decided, the strategic question changes. You do not catch up in a race whose rules others have written and in which they have already raised their walls and moats. You start a new one.
China did exactly that with the car. Wan Gang, a mechanical engineer who had spent a decade at Audi, became China’s minister of science and technology in 2007. His argument, as documented in a study for Singapore’s RSIS, was that China could not overtake the incumbents in combustion engines because they were too entrenched. It had to leap to electric cars instead, which would also address its energy-security problem. Twenty years later China is the world’s largest car exporter.
Two details make the case useful rather than merely impressive. First, China did not only enter the new race. It took control of its core component, the battery, and the materials upstream of it, which is exactly the leverage it now uses in the rare-earth conflict. Second, the new race was won where China already sat on a learning curve. As I described in the powertrain post, BYD started in 1995 as a maker of handset batteries, and CATL is a 2011 spin-off of ATL, which made cells for laptops and phones. A new race is best started where you already have capabilities.
Life is not a pony farm, and this was not cheap. The Center for Strategic and International Studies estimates Chinese government support for the EV industry at USD 230.9 billion between 2009 and 2023, by its own account a conservative figure, and very few Chinese EV and battery makers are profitable. A new race takes time and capital far beyond Horizon 1. But it is a different kind of spending from paying, year after year, for data centres in a race already lost, where your only role in value creation is to pay.
Where New Races in AI Could Start
So where could new races begin? None of the following is a recommendation. They are options, each with the bottleneck you would have to own and the new dependency it would create. Christensen and Raynor’s law of conservation of attractive profits is the guide: when one stage of a value chain is commoditised, attractive profits migrate to an adjacent stage. The bottleneck does not disappear. It moves, and the question is whether you stand where it is moving to.
Open-weight models. On some benchmarks the gap between the best open and the best closed models shrank from eight percent to 1.7 percent within a year, according to Stanford’s AI Index. With open weights, the model runs under the operator’s control. The new dependency: most open frontier models come from Chinese labs or from Meta, and giving weights away needs a patron.
Specialised chips. Custom accelerators such as Google’s TPU and Amazon’s Trainium are already breaking Nvidia’s position. For a country without its own chips, control merely moves from one American company to another, and manufacturing stays at TSMC.
Efficiency and small, specialised models. If smaller models suffice for most tasks, the value of the frontier falls and value moves to applications, data and customer access. This is Bürkli’s own scenario in which Switzerland would be “excellently positioned”. It is also the scenario in which today’s data-centre investments are most likely to strand.
AI on the device and in the machine. If inference moves from the data centre into phones, cars and machines, control moves to the makers of those devices and their chips. Here Europe starts from an unusual position: Infineon led the global microcontroller market in 2025 with 23.2 percent, and three of the five largest microcontroller makers are European, according to Omdia data reported by heise. This position is entangled too: Infineon and STMicroelectronics run fabs and packaging in Asia, and Infineon and NXP are partners in TSMC’s new fab in Dresden.
New architectures. Approaches beyond today’s transformers and GPUs could reshuffle everything. This is Horizon 3, and nobody can say whether or when it arrives.
Three of these five options move value towards fields where Europe has real capabilities: domain knowledge, industrial customers, production data and embedded systems. That is where a new race could start, on ground where the learning curve already runs in Europe’s favour.
Four Questions for Any AI Strategy
The four elements translate into four questions that any country or company can put to an AI project before the first spade goes into the ground.
- Value proposition: Who is our customer, and what benefit does that customer get that it could not get without us?
- Value-creation architecture: Which components do we own ourselves, and is there one among them that others cannot do without?
- Revenue model: Where does the money flow over the entire lifetime, and who carries the risk if our assumptions do not hold?
- Company spirit: Are we chasing a race someone else defined, or starting our own where we already have capabilities?
A data centre for a foreign frontier model fails all four. China’s electric-car strategy passed all four, at a high price and over twenty years.
Five Business Model Lessons for Every Industry
Strip away the chips and the power lines, and this case leaves tools that work in any industry.
1. Location is not control. Where an asset stands says little about who controls the value it creates. In digital business models, control is exercised through access, continuously. Always ask who can switch it off.
2. Read a business model from its revenue model, not from its value proposition. Value propositions are written to persuade. The architecture and the flow of money over the full lifetime show who actually wins. If you only read the brochure, you will miss that you are the supplier.
3. Find the bottleneck before you invest. The visible asset is rarely the scarce one. Profits and power collect where scarcity sits, and they move when scarcity moves. A strategy that does not name the bottleneck is a wish.
4. Long-lived investments for short-lived customers are an option you write. When your commitment lasts decades and your partner’s lasts a few years, you carry the risk of their departure. Price that option, or let the other side carry it.
5. Do not chase a race whose rules others have set. Catching up in a decided contest means paying for the privilege of losing slowly. Start a new race, and start it where you already sit on a learning curve.
A data centre is a building. It can be a good building, well powered and well cooled, and still leave the country that hosts it without any control over the intelligence that runs inside. The question for every AI strategy is not where the servers stand. It is which business model you are part of, which component of it you own, and whether that component gives you a say when it matters. Understand. Imagine bigger. Act.
Annotated References
Bahar, D., & Wright, G. (2026). New evidence on data center employment effects. Brookings Institution. Compares labour markets with built data centres to those with announced but cancelled projects, which controls for site selection. Finds that data centres do create local jobs, fewer than industry claims, and that colocation facilities often serve tenants without local operations. Link.
Bernstein (2025), as reported by Investing.com. How much does a GW of data center capacity actually cost. Estimates about USD 35 billion per gigawatt of AI data-centre capacity, with GPUs at 39 percent of capex and Nvidia’s gross profit alone equal to roughly 29 percent of total cost. Link.
Bloomberg, as reported by The Edge Malaysia. (2026, September 4). Abu Dhabi’s G42 weighs US ownership to safeguard AI chip access. G42 may buy advanced US chips licence-free only until around April 2027 and needs to change its corporate structure for continued access; executives have held exploratory talks about a US majority shareholder or a new US vehicle. The sharpest illustration that even unlimited capital buys access only on the bottleneck owner’s terms. Link.
Christensen, C. M., & Raynor, M. E. (2003). The Innovator’s Solution. Harvard Business School Press. Chapter 6 formulates the law of conservation of attractive profits: when modularisation and commoditisation make profits disappear at one stage of the value chain, the opportunity to earn them appears at an adjacent stage with proprietary products. Used here to read disruption as a migration of the bottleneck. Overview.
CSIS, Kennedy, S. (2024). The Chinese EV Dilemma: Subsidized Yet Striking. A conservative estimate of USD 230.9 billion in Chinese government support for the EV sector from 2009 to 2023, with subsidies falling from over 40 percent of sales in the early years to 11.4 percent in 2023, and very few EV and battery makers profitable. Link.
E3. (2026). Understanding the Drivers of Rising Electricity Rates and the Role of Data Centers. Reviews the quantitative studies on cost shifting and finds no evidence so far that other customers subsidise data centres, including in Virginia. The counter-position to the Harvard paper. Link.
Epoch AI: Michael, A., & Cottier, B. (2026). Servers account for 60% of the total cost of ownership of a one-gigawatt AI data center. A stylised cost model of a 1 GW data centre owned by a US hyperscaler with Nvidia GB200 systems: USD 38 billion upfront capital, about 30 percent facility, almost 70 percent servers and networking; labour about half a percent of annual cost. Open spreadsheet with sources. Link.
Epoch AI. (2025). LLM inference prices have fallen rapidly but unequally across tasks. Prices for a fixed level of model performance fell by a factor of 9 to 900 per year depending on the benchmark. Evidence for the efficiency side of the demand uncertainty. Link.
Farrell, H., & Newman, A. L. (2019). Weaponized Interdependence: How Global Economic Networks Shape State Coercion. International Security, 44(1), 42–79. Shows how asymmetric global networks allow states with jurisdiction over central nodes to coerce others through the “panopticon” and “chokepoint” effects. The theoretical frame for why contracts do not protect in the AI stack. Link.
FirstEnergy. (2025). End of Life Project Methodology (PJM informational filing). Expected service lives of 40 to 50 years for transmission transformers and 50 to 60 years for line conductors. Basis for the lifetime mismatch between grid assets and data-centre equipment. Link.
Koomey, J. (2026). Electricity Demand Growth and Data Centers (talk, Ohio Manufacturers’ Association). Argues that efficiency and service demand for AI are deeply uncertain, that utility forecasts often adopt industry projections, and that scenarios should include declining AI electricity use. Link.
LeVine, S. (2010). The Great Battery Race. Foreign Policy. Documents how China’s government recruited Wan Gang, after ten years at Audi, to lead its electric-vehicle research. Link.
Martin, E., & Peskoe, A. (2025). Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power. Harvard Electricity Law Initiative. Identifies the mechanisms, from confidential special contracts to regional cost allocation, through which data-centre costs can shift to other ratepayers. Link.
Masanet, E., Shehabi, A., Lei, N., Smith, S., & Koomey, J. (2020). Recalibrating global data center energy-use estimates. Science, 367(6481), 984–986. Bottom-up estimate showing that global data-centre energy use rose only about 6 percent from 2010 to 2018 while compute instances grew 550 percent. The precedent for efficiency outrunning demand. Link.
Monitoring Analytics, as reported by IEEFA. (2025). Projected data center growth spurs PJM capacity prices by factor of 10. The independent market monitor attributes 63 percent of the 2025/26 capacity-price increase, USD 9.3 billion, to data-centre load. Link.
OpenAI. (2025). Introducing Stargate UAE. Announces a 1 GW cluster in Abu Dhabi with G42, Oracle, Nvidia, Cisco and SoftBank as the first partnership under “OpenAI for Countries”, described as helping governments build sovereign AI capability in coordination with the U.S. government. The offer’s own wording contains the contradiction this post analyses. Link.
Rumelt, R. P. (2011). Good Strategy Bad Strategy. Crown Business; and Rumelt, R. P. (2022). The Crux: How Leaders Become Strategists. PublicAffairs. The first book names the hallmarks of bad strategy, among them the failure to face the challenge and the confusion of goals with strategy. The second defines the crux as the most important part of a challenge that is also addressable. Used here to show why a data-centre strategy picks the addressable but non-critical part. The Crux.
Stanford HAI. (2025). AI Index Report 2025. Reports that the performance gap between leading open-weight and closed models shrank from 8 to 1.7 percent on some benchmarks within a year. Link.
Stähler, P. (2001). Geschäftsmodelle in der digitalen Ökonomie. Josef Eul Verlag. My dissertation, among the first works to treat the business model as a distinct unit of analysis. ResearchGate.
Stähler, P. (2021). Das Richtige gründen: Werkzeugkasten für Unternehmer (5th ed.). Murmann. Source of the four-element framework used as the diagnostic grid: value proposition, value-creation architecture, revenue model and company spirit.
Tan, K. Y. (2026). The Untold Story of Wan Gang and the Success of China’s Electric Vehicle Industry. RSIS, Singapore. Describes Wan Gang’s argument that China could not overtake entrenched combustion-engine incumbents and had to leap to electric cars, with energy security as a second motive. Link.
Uptime Institute. (2025). Global Data Center Survey 2025, as reported by Data Center Knowledge. 82 percent of operators report a highest-density rack below 30 kW; fewer than four percent of US data centres can currently support the newest AI racks. Evidence that data centres themselves age with each chip generation. Link.
US Government export-control directive on Fable 5 and Mythos 5, as described by Anthropic. (2026, June 12). The directive suspended access for all foreign nationals inside or outside the US, forcing a shutdown for all customers; controls were lifted on June 30 and access restored on July 1. The case showing that access, not location, determines control. Link.
