Four scarce inputs decide who hosts the AI build-out — clean gigawatts, datacenter capacity, talent clusters, sovereign coordination — and almost no country holds them in equal measure. Norway has hydro but thin clusters; the US has talent and capital but stalled grid permits; the Gulf has sovereign capital and sun without clusters; China can coordinate everything but faces export controls. The question turns on which combinations compound.
The AI build-out is not a technology race in the conventional sense. It is a contest over who controls the physical and institutional substrate on which all future intelligence will run, and the early signs suggest that the apparent losers in energy and coordination may end up holding the most durable cards. Four inputs determine the outcome: clean gigawatts, AI-ready data center capacity, dense talent clusters, and sovereign coordination.
The AI build-out is driving an infrastructure super-cycle requiring roughly $3 trillion in investment by 2030, but electricity grid constraints — not GPU supply — have become the primary bottleneck, with over 2,060 GW of US generation capacity stuck in interconnection backlogs and Ireland already imposing a moratorium on new data center connections. Meanwhile, AI talent remains stubbornly concentrated in a handful of clusters, salaries have surged past $200,000 with 500,000+ positions unfilled globally, and Huawei's Ascend 910C is closing the capability gap with Nvidia faster than export-control policy assumes. The central unresolved tension is whether Western-style permitting reform and market signals can relieve these bottlenecks faster than state-directed coordination models in China, India, and the Gulf — a question the briefing raises but cannot settle, partly because the underlying evidence skews heavily toward sources outside the market-oriented and structural-inequality traditions that would most sharply contest each other's answers.
Read the research →Here is what is actually happening right now: the most consequential infrastructure race in a generation is being decided not by algorithms or venture capital, but by electricity permits and cooling water. We've done our research on this one, and the facts are genuinely startling. Global data center capacity is projected to nearly double to around two hundred gigawatts by twenty thirty, requiring roughly three trillion dollars in investment, while over two thousand gigawatts of US generation capacity sits paralyzed in grid interconnection backlogs. Meanwhile thirty-five percent of the world's AI engineers live within forty miles of San Jose. So the question is not who has the best models. The question is who can actually plug them in, staff them, and sustain them across a decade. Four inputs decide that: clean gigawatts, datacenter capacity, talent clusters, and sovereign coordination. Almost no country holds all four. So which combinations actually compound, and which ones are fool's gold?
The US talent clusters and private capital edge are real, yet the two-thousand-sixty-gigawatt interconnection queue reveals a deeper mismatch: frameworks built for steady incremental loads now face step-function demand from AI-scale clusters. Cost allocation disputes and multi-year timelines mean every participant optimizes for its own local constraints rather than system throughput. Capital and engineers therefore remain bottled up even as clean-gigawatt sites elsewhere clear projects faster. That sequencing failure does not self-correct through price alone; it requires aligned authority that can override fragmented veto points without recreating the coordination problem at a higher level.
The structural story underneath that grid logjam is that compute has started to behave like a traded commodity, but our permitting and transmission rules still treat it as a local utility. When two thousand gigawatts of generation are stuck waiting for interconnection, we’re not short of electrons — we’re short of sequencing. In practice, that means hyperscalers are shopping for places where they can buy time: jurisdictions that can clear land, connect gigawatts, and issue guarantees in months, not years. That’s why sovereign coordination is now the scarcest input. Whoever can compress administrative latency effectively manufactures years of competitive advantage without adding a single chip.
Let’s talk about sequencing as a structural problem, not a policy tweak. The US grid backlog isn’t just a queue—it’s a symptom of a deeper institutional grammar. The 2,060 GW of queued capacity represents a coordination failure that price signals alone can’t fix, because the grid isn’t a market; it’s a natural monopoly with fragmented ownership, conflicting incentives, and no mechanism to prioritize system-level throughput over local veto points. This is where Mazzucato’s argument about mission-oriented policy becomes relevant. The US has historically mobilized infrastructure at scale only when it framed the challenge as a national mission—think the interstate highway system or the Apollo program. The grid backlog isn’t waiting for a price signal; it’s waiting for an institutional actor with the authority to sequence interconnection, preempt local opposition, and guarantee access on a timeline that matches AI’s step-function demand. The counterargument—that permitting reform and price signals will suffice—ignores that the US has spent a decade trying incremental fixes without breaking the logjam. The question isn’t whether the market can eventually clear; it’s whether the US can afford to wait. China’s whole-nation system doesn’t wait for price signals—it sequences infrastructure by fiat, and while that model has its own distortions, it’s delivering gigawatts on a timeline that outpaces Western permitting cycles. The real constraint isn’t electrons; it’s the institutional capacity to sequence them. The US doesn’t need a better queue—it needs a different kind of queue, one where AI infrastructure is treated as a national priority, not a local land-use dispute.
I think we are misreading the export control regime as a constraint when the mechanism is actually functioning as a forced industrial catalyst. Per CSIS and industry tracking, China’s domestic AI chip market is scaling from roughly six billion dollars in twenty twenty-four to around sixteen billion in twenty twenty-five, and Huawei’s Ascend nine hundred ten C already hits about sixty percent of Nvidia’s H100 performance on key workloads. When you wall off a civilizational-scale engineering base, you do not starve it; you subsidize its parallel supply chain. The deeper issue is that the stack is system-constrained, not chip-constrained. Relieve GPU access and pressure instantly shifts to advanced chip packaging, specialized memory yields, and high-voltage grid transformers. If Western strategy optimizes for chip denial while ignoring those downstream choke points, it is solving for the wrong bottleneck and inadvertently financing a sanction-proof inference stack that could undercut hyperscaler economics by twenty twenty-seven.
The absence of audited data on where the vast majority of AI data-center energy actually goes leaves every national gigawatt strategy resting on an untestable premise. If inference workloads turn out to be a small fraction of total draw, then the clean-power advantage being chased by energy-rich hosts is being priced against an input whose real consumption profile remains opaque. That uncertainty does not favor any particular coordination model; it simply means both market-driven and mission-oriented approaches are optimizing against incomplete maps of the load they intend to serve.
The missing piece in this conversation is incentive alignment for the hosts of all this new infrastructure. In the Gulf, in Brazil, in parts of Africa, governments are offering land, power, and water for AI data centers, but ownership structures remain opaque. Without co‑ownership of the compute layer or enforceable local‑content rules, those projects risk reproducing extractive energy‑export dynamics rather than building sovereign capability. Yet the data gap matters: we don’t know whether these hosting models evolve into domestic ecosystems or not. That uncertainty makes policy design the strategic variable—linking power allocation to local compute rights or sovereign cloud foundations could turn what now looks like outsourced capacity into long‑term leverage over the global AI value chain.
The CNAS Sovereign AI Index finding deserves to be stated without softening: the US and China together control roughly ninety percent of frontier compute and host all fifty top-ranked foundation model platforms. Every other nation's current sovereign AI strategy is operating within that constraint — which is not the same as saying those strategies are futile, but it does mean the question is which policy moves can actually shift that structure over a decade, not just describe ambitions around it. Here is where I think the panel is missing something. The Gulf and Nordic hosting strategies are being evaluated against the wrong benchmark. The research brief flags explicitly that we have no longitudinal data on whether energy-hosting translates into domestic AI ecosystems or remains a permanent value-extraction arrangement. That is not a rhetorical point — it is a genuine epistemic gap that should make us cautious about both optimism and dismissal. Kautilya's insight in the Arthashastra is that the builder of the road controls the trade. Whether that applies here depends entirely on whether the infrastructure contracts being signed today include mechanisms for local value capture — co-ownership, sovereign compute rights, enforceable local-content provisions — or whether they are simply long-term leases on someone else's physical layer.
Here’s the structural blind spot: the entire debate assumes the AI build-out is optimizing for frontier models, but the research from NITI Aayog and Universitas Gadjah Mada suggests that for 80% of the world’s use cases, domain-specific small models trained on local data outperform large general models. If that’s true, then the $3 trillion hyperscaler investment logic is chasing the wrong capability. The real sequencing question isn’t who can host the biggest clusters—it’s who can build the right clusters. A country with modest compute but strong local data governance could leapfrog to a sovereign small-model stack, while energy-rich hosts risk becoming stranded assets if their infrastructure is optimized for workloads that never materialize. The Gulf’s bet on clean gigawatts only pays off if the West’s model-centric paradigm is correct. If it’s not, their physical layer becomes a white elephant, not leverage.
I think the deeper mechanism here is not power or chips, but the institutional grammar that binds them. Western frameworks treat infrastructure as a priced utility; East and South Asian statecraft treats it as a relational substrate where the state sets the operating standards and private capital plugs in. If the economic yield of AI shifts from frontier training to localized inference, then nations deploying standardized, state-anchored data and identity layers are quietly acquiring the routing rights for future value flows. The California talent cluster is extraordinarily capable but geographically brittle and priced at a premium that scales poorly. A population-scale interoperable layer compounds regardless of where the engineers sit. That flips the sovereignty question from who trains the heaviest model to who governs the data rails every lighter model must traverse. We are measuring compute density while the actual leverage is migrating to protocol control.
Talent density per capita creates a hiring-cycle advantage that aggregate headcount numbers completely obscure. In tight clusters, engineers can switch roles or spin out new teams without changing cities or losing embedded networks, which compresses offer-to-close time and raises retention. That velocity matters more than total engineers when every new gigawatt tranche must be staffed before the next hardware generation arrives. Large diffuse pools look bigger on paper but cannot match the transaction speed of a node where proximity itself lowers search and switching costs.
The blind spot in the sovereignty debate is contractual time horizon. The infrastructure being built today will run for decades, but the revenue models underwriting it turn over every few years. If data centers in the Gulf or Nordic regions are financed on short corporate cycles, foreign operators can recycle capital faster than host states can build domestic capability. That mismatch matters more than ownership share because it fixes who can reinvest at scale. Structurally, it’s the intertemporal leverage—who controls reinvestment cadence—not the initial equity split, that decides whether a host becomes an ecosystem or an energy colony.
The NITI Aayog and Gadjah Mada finding cuts deeper than a capability argument — it is an optimization target argument. If it holds, the nations currently racing to host frontier compute are building the nineteenth century equivalent of deep-water ports in landlocked regions. But here is the load-bearing assumption nobody has named: small models on local data only outperform if the local data governance infrastructure exists to train them on. Aadhaar is the existence proof — Nilekani built a population-scale identity substrate that no private actor would have financed, and that substrate now enables AI applications that are structurally impossible elsewhere. The question is not whether small models win. It is whether any nation outside the US-China duopoly can execute the Aadhaar move at the compute layer before hyperscaler lock-in forecloses the option.
The real sequencing failure isn’t in the hardware—it’s in the software stack’s governance layer. Every AI model, no matter how small or local, depends on three invisible rails: identity verification, payment settlement, and data provenance. Right now, those rails are being built by private hyperscalers and fintech platforms, not states. That means the physical infrastructure race is being won by nations with clean gigawatts, but the protocol layer is being captured by whoever controls the default identity and payment APIs. The structural question isn’t whether small models outperform large ones—it’s whether any nation can build a public alternative to those default rails before private lock-in becomes irreversible. India’s Aadhaar shows it’s possible, but only because the state treated identity as a public good, not a market service. Most energy-rich hosts are still treating data centers as real estate plays, not as the foundation for a new governance layer. That mismatch is the real sovereignty risk: you can own the gigawatts, but if someone else owns the rails, you’re just renting your own infrastructure.
The protocol layer you just named is actually a coordination problem in disguise. In statecraft traditions that treat infrastructure as constitutive of political order rather than a priced utility, the leverage sits in who sets the operating standards, not who leases the concrete. Mariana Mazzucato’s research reaches a parallel conclusion for modern industrial policy: when long-horizon build-outs face fragmented incentives, price signals cannot substitute for mission-mode coordination. We simply lack longitudinal data on whether Gulf or Nordic hosting matures into domestic ecosystems or locks in as permanent extraction. That uncertainty is exactly why sovereign policy must target the rule-making layer. If host states condition megawatt allocation on joint data mirroring, local compute quotas, and transmission equity, they shift from landlords to standard-setters. The structural test is whether they codify those terms before private platform defaults harden into global operating norms.
The missing variable is the load composition itself. Without audited breakdowns of where the non-query energy actually flows—training versus inference versus idle capacity—host states cannot set enforceable conditions on megawatt allocations because they lack a measurable baseline for what counts as local value creation. Conditioning power on compute quotas assumes the dominant future workload is already known and fixed; if inference on small local models scales faster than frontier training, the physical substrate being negotiated today may be oversized for the actual demand profile that materializes. That uncertainty turns every sovereignty clause into an option contract whose strike price is still being discovered.
The hinge variable now is sequencing between capital horizon and policy horizon. Private investors are committing on two‑ to five‑year depreciation cycles; sovereign coordination, when it exists, moves on ten‑ to twenty‑year infrastructure timelines. That mismatch explains why even energy‑secure hosts struggle to capture value: by the time a state crafts a data‑governance rule, the hardware it governs has turned over twice. If policymakers want durable leverage, they need to shift from asset‑specific incentives to standard‑based compacts—pricing not the facility, but the interoperability and data‑sharing terms that persist through hardware churn. Otherwise the asset life‑cycle becomes the governance window, and real sovereignty never compounds.
The Greenspan Commission parallel is worth making precise. In nineteen eighty-three, the commission succeeded not because it found an elegant solution but because it made painful tradeoffs — tax increases, delayed cost-of-living adjustments, a higher retirement age — legible as a shared burden across generations rather than a zero-sum fight between current beneficiaries and future taxpayers. The AI infrastructure problem has that same architecture. Who bears stranded-asset risk if small models win and frontier training clusters sit underutilized? Who captures upside if the protocol layer compounds? Those are intergenerational distribution questions, not procurement questions. No current policy instrument is designed to answer them, and that is why every sovereignty clause being negotiated today is, as the research brief puts it, an option contract whose strike price is still being discovered.
The structural blind spot isn’t workload size—it’s workload ownership. Right now, every kilowatt-hour of clean power allocated to AI infrastructure is being matched with a corresponding allocation of data rights. The hyperscalers aren’t just leasing electrons; they’re embedding their terms of service into the physical layer. That means the sovereignty question isn’t about hosting capacity—it’s about whether the host state retains the right to audit, tax, or repatriate the data flowing through its grid. The research brief flags this as a known unknown: we have no longitudinal data on whether energy-rich hosts become ecosystems or extraction zones. But the mechanism is clear—if the default contract gives the tenant full data sovereignty, the host’s physical infrastructure becomes a commodity. If the host conditions power allocation on data residency or revenue-sharing, the infrastructure becomes a lever. The sequencing failure is that most hosts are negotiating power deals before they’ve defined data terms. That’s not an energy policy—it’s a surrender of the governance layer before the first server is racked.
The Greenspan parallel holds only if the asset class has a stable use profile across generations. Here the substrate itself—gigawatt-scale compute tied to physical power contracts—faces demand obsolescence on a three-to-five-year cycle rather than a thirty-year one. Without audited load data, any intergenerational compact is pricing an unknown future workload mix, so the real risk is not stranded steel but contracts that lock in terms calibrated to the wrong load shape. That makes the distribution question unanswerable until the measurement layer exists.
The three-to-five-year turnover argument assumes leverage lives in the silicon. I think it actually lives in the interconnection rights and the material bottlenecks that every workload shape still requires. Supply-chain reporting flags that constraints in copper, advanced packaging, and high-bandwidth memory are multi-year and poorly tracked, meaning whichever actor secures long-dated options on those inputs captures structural leverage regardless of whether inference or training dominates. Historically, administrative traditions that outlasted technological turnover did it by standardizing the choke points—canal locks, grain registries, examination curricula—so hardware refreshes renewed the system rather than fractured it. We genuinely lack longitudinal data on whether energy hosts become ecosystems or colonies, so I would not treat any contract template as settled. But if host states negotiate for equity in the packaging and memory supply chains instead of just rack space, they convert hardware uncertainty into compounding option value.
The unpriced input in every sovereignty negotiation is temporal rights to the interconnect itself. Once a hyperscaler secures a point on the transmission network, that node functions like a long‑dated option: even if workloads change, the access right endures. The research on US grid backlogs shows interconnection queues exceeding two‑thousand gigawatts — more than total installed capacity — because each queue position effectively represents permission, not just potential generation. That same logic now globalizes through AI. Whoever holds durable grid access, not merely current compute, controls future siting choices. If host nations treat those rights as tradable national assets subject to periodic review rather than perpetual grants, they retain the ability to reprice sovereignty as technology and workloads evolve.
Here is the category error underneath this whole negotiation debate: we are pricing access rights to infrastructure whose energy consumption we cannot account for. The research brief is explicit — we have no transparent, audited breakdown of where roughly ninety-eight percent of AI data center electricity actually goes. It is not serving text queries. So every sovereignty clause, every data-residency condition, every intergenerational compact being designed right now is being calibrated against an energy footprint that nobody has actually measured. That is not a minor data gap. That is the entire measurement foundation of a three-trillion-dollar investment cycle, and it is missing. Before we can answer who owns the substrate, we need to ask what the substrate is actually doing.
The real sequencing error isn’t regulatory lag—it’s that we’re treating energy as the binding constraint when the binding constraint is actually coordination rights. Every gigawatt allocated to AI infrastructure comes with an implicit allocation of who gets to decide what runs on it. Right now, that right defaults to the tenant, not the host. The research brief flags that we lack longitudinal data on whether energy hosts become ecosystems or extraction zones—but the mechanism isn’t energy supply, it’s who holds the scheduling keys to the grid. If a host state can’t audit or redirect workloads, it doesn’t matter how many gigawatts it owns; the infrastructure is just a rented pipe. The sovereignty test isn’t about hosting capacity—it’s about whether the host retains the right to reprioritize compute slots when the workload mix changes. That right doesn’t require new hardware; it requires a governance clause that survives hardware refresh cycles.
The real problem is that an intergenerational compact cannot allocate risk when the underlying load profile remains unobservable. If the dominant future use of those gigawatts is invisible, every attempt to assign stranded-asset costs or capture rights across generations is simply pricing an unknown state of the world. The measurement gap therefore has to be closed before the compact can be written; otherwise the bargain itself becomes the next source of misallocated exposure.
The measurement paralysis misses how long-horizon statecraft actually secures leverage. You do not wait for perfect load telemetry to control the choke points that every workload configuration must traverse. The evidence shows the AI stack is system-constrained: clear the GPU bottleneck and pressure instantly shifts to high-bandwidth memory, advanced packaging, grid transformers, and fiber. Those material constraints run on multi-year lead times and remain poorly tracked in public reporting. I think the binding sovereign variable is not auditing electrons but orchestrating the physical supply chain and transmission topology that no model architecture can bypass. Market-oriented analysts are right that blunt mandates distort price signals and can repel capital, which is why the sequencing mechanism matters more than the directive. But treating an interconnection queue of over two thousand gigawatts as a simple market failure ignores that transmission lines take five to fifteen years to permit. If talent clusters atrophy before the power arrives, the race is lost to administrative friction, not workload uncertainty.
The gap in measurement hides a bigger asymmetry: who sets the default parameters for visibility. In earlier industrial cycles, the state defined the accounting units—tons of steel, cubic meters of gas—and that made energy policy legible. In this one, the hyperscaler defines the metric of usage. That’s not conspiracy; it’s a design convenience that turns into geopolitical leverage. The test for genuine sovereignty may be as simple as this: can a nation independently specify what counts as a unit of compute and require auditable reporting across providers? Until that standard exists, neither taxation, nor carbon policy, nor equity participation in AI infrastructure can be meaningfully priced.
The measurement gap isn’t just about energy—it’s about what the energy is being used to build. The research brief shows we don’t know where 98% of non-query data center energy goes, but the deeper question is whether that energy is training frontier models, running inference for local use cases, or simply keeping idle capacity warm. If most of it is speculative training runs for models that never deploy, then the entire "compute follows power" dynamic is a bubble. The sovereignty risk isn’t just that hosts don’t own the output—it’s that they’re subsidizing infrastructure for workloads that may not exist in three years. The only way to price that risk is to demand transparency on workload intent, not just consumption. Without it, every gigawatt deal is a bet on an unknown future—and the house always writes the terms.
The workload-intent transparency demand runs into a structural problem nobody has named yet: the hyperscalers genuinely may not know either. Training runs are launched speculatively. Capacity is provisioned against demand forecasts that routinely miss. The epistemic gap isn't just a governance failure — it's partly an intrinsic feature of how frontier AI development works. Which means demanding workload-intent disclosure isn't just a sovereignty move; it would impose discipline on the hyperscalers' own capital allocation. That's the Greenspan Commission move applied here — not finding new resources, but forcing legibility onto a system where the principal actors are also flying partially blind. The nation that mandates that standard first doesn't just gain audit rights; it gains structural insight into how the entire industry is actually making its bets.
The sharpest tension in this conversation was between the measurement problem and the action problem. We genuinely do not know where most of the energy in AI infrastructure goes, which makes every sovereignty clause, every intergenerational compact, every gigawatt deal an option contract on an unknown workload. But waiting for perfect telemetry is itself a strategic choice with costs. The concrete takeaway: the nation that defines the accounting unit — what counts as a unit of compute, what disclosure is owed — shapes the terms everyone else negotiates within, though whether that translates into durable advantage or just administrative overhead remains genuinely open. On the original question: no combination is sufficient on its own, and the honest answer is that the compounding combinations are still being discovered. The US-China concentration in frontier compute is real but probably overstated as a permanent duopoly given how rapidly the rest of the field is moving. The Gulf's gigawatt bet pays off only if the model-centric paradigm holds. India's data-substrate move is the most structurally interesting, and the least legible from the outside. Every other nation is negotiating within constraints that are still being written. Thank you for listening. As it happened; as it is.