What should a government price before signing a sovereign AI datacenter deal?

What should a government price before signing a sovereign AI datacenter deal?

Five commercial dimensions, each of which can dwarf the headline compute price if left unresolved. A sovereign AI datacenter deal is not a standard infrastructure PPP or an IT procurement. It combines elements of both while adding technology-cycle risk that neither category normally contains. The five dimensions the government must price before commitment are:

1. Power supply: volume, price, duration, curtailment rights and the allocation of grid-upgrade costs
2. Land tenure: ownership or lease terms, permitted use, expansion rights and reversion on termination
3. Technology refresh: minimum compute thresholds, refresh cycles, upgrade pricing and obsolescence responsibility
4. Data residency and jurisdiction: governing law, access rights, sovereign data protection and data portability on exit
5. Exit rights: termination triggers, handover process, data extraction, asset valuation and staff transition

Each of these is a commercial exposure, not a technical specification. Treating them as part of the IT brief rather than the commercial negotiation is the single most common error in sovereign datacenter deals.

How is a sovereign AI datacenter different from a data embassy?

A data embassy addresses jurisdiction. A sovereign AI datacenter addresses commercial capability. The distinction matters because a government may solve the jurisdiction question—where its data sits and under whose law it is protected—through a data-embassy framework, and still face all five commercial exposures listed above. A data embassy grants sovereign continuity for state data held abroad. A sovereign AI datacenter provides the compute infrastructure on which AI workloads run. The two are complementary, not interchangeable. For the jurisdiction framework, see what is a data embassy. This article addresses the commercial terms that sit beside it.

• Data embassy = jurisdiction, immunity, inviolability of state data
• Sovereign AI datacenter = power, compute, storage, networking, refresh, pricing, exit
• A government needs both frameworks, not one or the other

Confusing the two leads to a datacenter that has sovereign branding but no jurisdictional protection, or a data embassy that has legal protection but no commercial discipline around cost and technology cycle.

Why is power the largest commercial exposure?

Because AI datacenters consume power at a scale that can strain a national grid, and the cost of that power over the facility's life can exceed the cost of the compute hardware. The IEA's analysis of datacenter energy demand shows that a single hyperscale AI facility can draw 100–300 MW, and next-generation clusters are projected to draw more. If the government guarantees power supply at a fixed price while the operator scales compute, the buyer is effectively underwriting an electricity subsidy that grows with every GPU generation. The power agreement must address:

• A defined volume cap: maximum power the government commits to supply at the agreed price
• A price-adjustment mechanism: how the power price changes if generation costs or grid charges increase
• Curtailment rights: the government's right to reduce supply during grid emergencies, and the operator's obligations during curtailment
• Grid-upgrade allocation: who pays for the transmission infrastructure needed to deliver power to the site

Without these provisions, the government has signed an open-ended power commitment whose cost it cannot forecast and whose volume it cannot control.

How should technology-refresh obligations be structured?

As a defined buyer obligation with specified cycles, minimum thresholds and a pre-agreed pricing mechanism. AI hardware—GPUs, networking fabric, storage systems—evolves on a 2–4 year cycle. An operator that controls refresh timing can run ageing hardware while charging current-generation service prices, or treat each upgrade as a separate commercial negotiation from a position of incumbency. The refresh framework should include:

• A minimum compute-capacity threshold, expressed in industry-standard metrics that can be independently verified
• A refresh cycle: typically 3–5 years for GPU clusters, with the first refresh window defined before signature
• A pricing mechanism for upgrades: benchmarked to independent hardware cost indices, not the operator's list price
• Obsolescence responsibility: who bears the cost of decommissioning and replacing hardware that falls below the minimum threshold

The OECD's sovereign cloud governance guidance flags technology lock-in as a specific risk in government cloud arrangements. A refresh clause that leaves timing and pricing to the operator is the commercial mechanism by which lock-in becomes lock-out—the government cannot leave because it cannot afford to rebuild elsewhere.

What do data residency and jurisdiction require beyond physical location?

Physical hosting in-country does not guarantee legal protection of state data. Data residency means the data sits on servers within the government's territory. Data jurisdiction means the law that applies to that data—who can access it, under what authority, with what remedies for violation. A datacenter built on sovereign soil but operated under a foreign law or subject to foreign legal process has not solved the jurisdiction problem. The EU's AI Act framework makes this distinction explicit. The commercial agreement must specify:

• Governing law for data stored and processed in the facility
• Access rights: who may access the data, under what conditions, with what audit trail
• Foreign legal process: whether the operator must comply with foreign court orders or disclosure demands, and what happens if it receives one
• Data portability: the government's right to extract its data in a usable format, at its own instance or on termination

These provisions sit alongside the jurisdiction framework of the data embassy. They are the commercial counterpart to the diplomatic instrument.

What exit rights should the government negotiate before signing?

Exit rights that are clear, enforceable and sufficient to permit the government to continue operating or to transfer to a successor operator without loss of capability. Too many sovereign IT agreements treat exit as a contingency to be addressed later. In a datacenter deal, exit must be defined before the first server rack is installed because the cost of extraction rises with every year of operation. The exit framework should include:

• Termination for cause: material breach, sustained underperformance below the minimum compute threshold, insolvency
• Termination for convenience: whether the government can exit without cause, on what notice, at what cost
• Handover process: data extraction in standard formats, system documentation, configuration records, staff transition plan
• Asset valuation and transfer: a mechanism for valuing the physical infrastructure if the government wishes to retain it, and a right of first refusal if the operator wishes to sell
• Data disposition: a clear, binding statement of what happens to all state data on termination—deletion, return, or continued access under an interim arrangement

A government that signs a datacenter deal without exit rights has not procured a service. It has accepted a permanent dependency that no future administration can unwind without losing capability.