Data centers and national interests: grid costs, FDI, and strategic trade-offs

The electricity grid has become the new front line of geopolitical competition, and most people have no idea it's happening. We've done our research on this one, and the facts are genuinely startling. Global data center electricity consumption hit roughly four hundred fifteen terawatt-hours in twenty twenty-four — the IEA projects that more than doubles to nine hundred forty-five by twenty thirty, driven almost entirely by AI. Meanwhile, announced greenfield investment in data centers hit two hundred seventy billion dollars in twenty twenty-five alone, per UNCTAD — more than one fifth of all new foreign direct investment on earth. So we have a sector that is simultaneously the fastest-growing electricity load in advanced economies and the single largest destination for new foreign capital. The tension that interests me is this: grid upgrade costs under current tariff structures get distributed across all ratepayers, not just the load that caused them. Harvard's Ari Peskoe has named that mechanism explicitly. So here is the question I want to put to the panel: when a nation courts data center FDI, who is actually paying for it?

The fiscal math here isn’t just about jobs or tax revenue—it’s about the institutional grammar of the deal. When six U.S. states grant permanent, uncapped tax exemptions costing over one hundred million dollars annually, they’re not just forgoing revenue; they’re embedding a structural asymmetry into their own fiscal constitution. The exemption outlasts the construction boom, the jobs evaporate, but the revenue hole remains—a classic Polanyian double movement where the state first enables the market, then absorbs its costs. And here’s the structural blind spot: these same jurisdictions then socialize grid upgrade costs through tariffs that distribute expenses across all ratepayers, not just the load that triggered them. So the public sector is effectively subsidizing the compute layer twice—once through foregone tax revenue, again through socialized grid costs—while the financial returns accrue to hyperscaler shareholders and the strategic benefits flow to the nations whose firms own the infrastructure. That’s not a negotiation; it’s a pre-written contract where the host state signs away both fiscal and strategic agency. The question isn’t whether these deals create value—it’s whether the institutional framework even allows the host to capture any of it.

The timing mismatch locks in the asymmetry before any pledge can intervene. Data centers reach operation in roughly two years while new transmission lines require a decade under current permitting, so grid planners confront discrete, unanticipated load jumps that interconnection queues were designed to handle incrementally. The result is that capacity-market costs get locked into long-term commitments before regulators can adjust allocation rules, turning what looks like voluntary cost absorption into a structural transfer that later reforms cannot unwind.

The structural divergence here is not just about who pays for the wires. It is about whether a state treats compute as a commercial real estate play or as a planned strategic input. China’s computing power infrastructure program targets more than three hundred exaFLOPS — a measure of raw processing speed — by twenty twenty-five, with roughly thirty-five percent explicitly allocated to intelligent computing. That is capacity planning, not a zoning negotiation. I think the mechanism that actually changes the payoff matrix is conditional market access. Jurisdictions like Singapore and the UAE now tie land and power approvals to reserved rack capacity for sovereign workloads. When you negotiate only on tax exemptions and interconnection queues, you optimize for short-term construction activity while leaving the compute layer entirely offshore. The open question is whether liberal regulatory frameworks can legally mandate that kind of strategic reservation without triggering immediate capital flight.

The decarbonization paradox sits at the core of this trap. Hyperscalers are now the world’s largest corporate buyers of renewable energy contracts, yet U.S. utilities are planning up to one hundred gigawatts of new natural gas capacity specifically to serve those same loads. Both facts are true because the accounting and the physics run on separate ledgers: corporate power‑purchase agreements retire green certificates on paper while the actual electrons delivered are fossil‑fired. Until grid tariffs and permitting rules align financial instruments with physical dispatch, every gigawatt of “green” data center demand risks locking in a decade of gas infrastructure. The question is not whether tech firms buy renewables—it’s whether their procurement decisions change the generation mix that actually runs the grid.

The PJM number deserves to sit in the room for a moment. In the last three capacity auctions covering the grid that serves sixty-seven million Americans, data centers represented ninety-seven percent of projected load growth and roughly forty-five percent of total capacity costs — twenty-one point three billion of forty-seven point two billion dollars. That concentration ratio means the question of whether large loads spread fixed costs or concentrate them is not abstract — it is the empirical question that PJM's own auction data forces onto the table, and the answer is genuinely contested and context-dependent in ways that neither the industry's cost-spreading argument nor the critics' socialization argument fully captures. What I'd want the panel to address: does that concentration ratio change the normative case for cost-causation tariffs, regardless of which direction the empirical dispute resolves?

That concentration ratio in PJM—ninety-seven percent of load growth, forty-five percent of capacity costs—isn’t just a data point. It’s a structural revelation. When a single asset class drives nearly all new demand but less than half the cost, the market is no longer pricing risk; it’s socializing it. The mechanism is the capacity auction itself: the clearing price reflects the marginal cost of reliability, not the marginal benefit to the load that triggered the need. So the hyperscaler pays a fraction of the upgrade, the rest is spread across every ratepayer, and what looks like a market outcome is actually a fiscal transfer from diffuse public to concentrated private. That’s not an accident—it’s the institutional grammar of a grid designed for kilowatts, not exaFLOPS.

The capacity market was designed around the assumption that load growth would come from many uncorrelated actors whose risks could be diversified. When one asset class accounts for nearly all incremental demand, that diversification premise collapses and the clearing price no longer reflects a broad system benefit; it reflects the cost of insuring a single concentrated, correlated risk. That is the structural reason earmarked compute conditions work in Singapore and the UAE but remain legally awkward under liberal interconnection rules—they treat compute capacity as a sovereign input rather than a market output whose reliability cost can be socialized after the fact. The real question is whether any jurisdiction can rewrite the auction rules fast enough to change the payoff before the next wave of commitments locks in the pattern.

The binding constraint everyone is ignoring is water, and the reason it stays invisible is that we are measuring the wrong efficiency metric. Industry optimizes for PUE, power usage effectiveness, which tracks how much electricity reaches the servers versus cooling overhead. But Kyoto University researchers have proposed a Social PUE framework that actually accounts for waste heat delivered to district networks and local hydrological stress. I think this measurement gap is doing heavy distributional work. When you optimize only for server-level power efficiency, you systematically externalize water withdrawal and thermal discharge onto the host community. UN University projects global data center water consumption could match the annual needs of one point three billion people by twenty thirty, yet we lack granular regional tracking. Applying Mahbub ul Haq development lens, I would argue this is not a technical oversight. It is a policy choice that routes physical costs to residents while financial returns exit the jurisdiction. Until procurement mandates tie interconnection to Social PUE thresholds and mandatory water disclosure, host nations are subsidizing compute with unpriced hydrological risk.

The underlying distortion isn’t just who writes the check, it’s when the obligation is recognized. Tariff schedules treat grid costs as sunk communal infrastructure, but hyperscaler demand behaves like capital formation—fast, lumpy, and reversible. That means costs become public liabilities long before any private commitment matures. The institutional fix isn’t another voluntary pledge; it’s temporal alignment. Require data center interconnections to post full cost‑causation bonds at permitting, indexed to regional capacity prices, with automatic clawbacks if load projections don’t materialize. That forces private timelines to match public exposure and transforms grid planning from reactive subsidy management into predictable industrial policy.

The Ratepayer Protection Pledge deserves direct scrutiny rather than acceptance as settled reform. Ari Peskoe at Harvard's Electricity Law Initiative identifies at least three structural channels through which costs reach residential customers regardless of what individual firms commit to — because the tariff architecture itself socializes costs at the system level, upstream of any firm-level agreement. The question worth pressing is whether voluntary pledges and mandatory cost-causation tariffs are genuinely complementary, or whether the political conditions that produce a voluntary pledge tend to crowd out the regulatory appetite for the mandatory version. I don't think the research settles that question — but I'd argue it's the most important governance question this boom has generated, and it's currently being treated as resolved.

The water question isn’t just about volume—it’s about institutional lock-in. When a data center secures a twenty-year water allocation in a stressed basin, it doesn’t just consume water; it pre-empts future climate adaptation. The mechanism is the priority right: industrial use often outranks agricultural or municipal in drought protocols, so the compute layer effectively holds a call option on the region’s hydrological future. That’s not priced into PUE or Social PUE, and no FDI framework I’ve seen requires disclosure of the seniority ranking of the water right. Until that changes, the host state is signing away its own resilience for a one-time capital injection. The jurisdictions that move first to condition interconnection on water-rights transparency and drought contingency clauses won’t just protect their aquifers—they’ll set the global standard for what compute infrastructure owes its physical place.

The Lula administration treated infrastructure as a vehicle for domestic capability-building rather than pure capital attraction. That template suggests a different question: what happens when the first jurisdiction embeds co-ownership or local equity requirements directly into the interconnection queue itself. Schumpeter’s lens shows the creative destruction already underway — legacy utilities face stranded-asset risk while value accrues to whoever controls the compute layer. The country that moves first to make governance architecture itself the condition of entry will not just retain more of the financial return; it will define the terms under which subsequent rounds of investment clear, because capital needs predictable rules more than it needs the lowest tax rate once the initial buildout wave passes.

I think we are still treating governance as a defensive tariff problem when it is actually a norm-export opportunity. Yan Xuetong has argued that moral and political leadership, not just material capacity, determines long-run strategic influence. Apply that to compute infrastructure. The jurisdiction that first codifies mandatory cost allocation, physical water disclosure, and sovereign cloud requirements into a single interconnection standard does not just protect its own grid. It creates a regulatory template that mid-sized economies will copy because they lack the bargaining power to negotiate bespoke terms with hyperscalers. That is how you escape the subsidy race. You stop competing on tax holidays and start competing on institutional credibility. UNCTAD flags that local economic spillovers remain poorly measured, which means the first mover who ties approval to verifiable community benefit captures the normative high ground while capital is still searching for predictable queues.

The silent variable is sequencing. Governments are negotiating incentives as if power, water, and compute markets move together, but their clocks are off by an order of magnitude—servers come online in two years, transmission in ten, regulation in fifteen. That lag means public commitments today lock in grid architectures and fiscal exposures long after the jobs and capital inflows have faded. The lesson from financial crises and infrastructure cycles alike is that misaligned time horizons, not bad intent, create structural subsidies. Any credible governance framework has to synchronize those timelines—decisions on connection rights, cost recovery, and water allocation must mature at the same pace as the loads they authorize—or every pledge remains a deferred liability written into the grid.

The Schumpeter frame everyone is gesturing at has a specific load-bearing assumption worth naming. The colonial port city analogy — host the infrastructure, watch the value exit — only holds if the host nation has no structural claim on the compute layer itself. Singapore, the UAE, and Saudi Arabia have begun breaking that assumption explicitly: earmarked compute capacity as a condition of market access, negotiated before the facility is built. That is Kautilya, not Schumpeter — strategic infrastructure as a state claim, not just a market input. The question for OECD countries is whether their investment treaty obligations structurally prevent them from doing the same, or whether that story is a convenience that benefits the party asking them to believe it.

The interconnection queue isn’t just a procedural backlog—it’s the first point of institutional contact between a compute project and the host state. Right now, that contact is purely technical: megawatts requested, voltage requirements, substation location. But if you redesign the queue entry form to require five-year physical carbon and water trajectories, plus a binding commitment to district heating offtake or sovereign compute allocation, you convert a clerical step into a governance gate. The mechanism is disclosure as leverage: once the data is public, regulators can condition approval on meeting those trajectories, and communities can contest permits based on verifiable hydrological impact. That turns the queue from a passive registry into an active planning instrument—one that forces hyperscalers to internalize public constraints before they break ground, not after. The jurisdictions that move first won’t just reduce stranded asset risk; they’ll export the template that makes the queue itself a strategic asset.

The jurisdiction that first requires public longitudinal tracking of both economic diffusion and hydrological impact as a condition of interconnection does more than close an information gap. It creates the evidence base that every subsequent regulator will have to cite when they write their own rules. Right now the absence of that data lets the subsidy race continue because no one can demonstrate net local benefit or net local cost with precision. The first-mover that forces disclosure of those trajectories will not merely protect its own ratepayers; it will define the factual premises on which the next decade of global interconnection standards are argued and litigated.

The certificate versus dispatch gap is not just an accounting quirk. It is a structural transfer of reliability risk onto host utilities that lack the capital depth to firm intermittent generation. When a hyperscaler secures a renewable PPA but the local grid must still commission gas or extend existing thermal plants to guarantee continuous load, the environmental attribute is exported while the balancing cost remains on the sovereign balance sheet. This is contract architecture meeting grid physics. The mechanism to close it is a physical matching requirement baked directly into the interconnection queue, forcing the green claim and the firming obligation to clear in the same jurisdiction. Without that, host governments are quietly underwriting a long duration liability that will constrain fiscal space long after the ribbon cutting.

The unpriced variable here is operational resilience. A grid that grows by bolting unpredictable 200‑megawatt loads onto twenty‑year transmission plans becomes brittle, and that brittleness has a fiscal expression: insurance premiums, outage penalties, stranded backup generation. Treating resilience as a measurable asset would flip the calculus. Instead of subsidizing connection, regulators could require a resilience bond — cash or capacity posted against the external costs of failure, released only if the facility meets uptime without emergency dispatch. That would align private incentives with public stability and reveal which operators genuinely lower systemic risk. It also exposes where the subsidy race hides: in free optionality on the host grid’s reliability balance sheet.

The enforcement point cuts deeper than audit capacity. The reason operational data is classified as proprietary isn't accidental — it reflects a deliberate structuring of the regulatory relationship from the beginning, where information asymmetry is itself the hyperscaler's primary bargaining asset. Snowden made this visible in the security domain: after 2013, European and Asian governments began treating U.S.-owned infrastructure as a potential intelligence vector, and data localization mandates followed. The same logic applies to economic regulation. You cannot reform what you cannot see, and the party with visibility has every incentive to keep it that way. Statutory transparency authority isn't a technical fix — it's a redistribution of structural power in the negotiation.

The queue isn’t just a procedural backlog—it’s the last moment when the host state has leverage. The research brief confirms that data centers move from proposal to operation in roughly two years, while transmission permitting stretches to a decade. That timing asymmetry means the interconnection filing is the only point where the hyperscaler is structurally dependent on the jurisdiction’s approval. The mechanism that turns this into strategic power is attaching conditions to queue registration, not to operation. That means mandatory disclosure of five-year load trajectories, water-rights seniority, and thermal discharge plans—filed before the substation is scoped. Once those disclosures are public, every subsequent reform—cost-causation tariffs, co-ownership clauses, drought contingency plans—becomes enforceable because the host state can audit against its own baseline. Without that upfront transparency, the twenty-year water allocation and tariff treatment are locked in before the public even knows what was traded away.

The deeper problem is not just missing data; it is the irreversible asymmetry this opacity locks in. When regulators cannot observe real-time load, water draw, or thermal discharge, the twenty-year contract that gets signed at interconnection allocates all upside optionality to the operator while shifting tail-risk exposure—rate spikes, drought allocation, stranded transmission—onto residents who never had visibility into the terms. The jurisdiction that first treats facility-level metering as a condition of queue registration therefore does more than enable later enforcement; it prevents the permanent transfer of option value that current rules embed by default.

I think the intergenerational transfer you just named is not an accounting error. It is the logical endpoint of treating critical infrastructure as a discrete commercial contract rather than a relational system. Mahbub ul Haq built his development framework on precisely this distributional question: who absorbs the transition shock and who captures the durable yield. The evidence on rate impacts is genuinely contested, and industry advocates are right that large steady loads can lower per-unit costs when tariffs are properly aligned. But that alignment collapses without continuous audit authority. When operational data stays proprietary, the host jurisdiction surrenders the ability to verify whether the facility is actually delivering the grid stability or local hiring it promised at registration. The structural fix, in my view, is not longer concessions. It is codifying real-time physical disclosure as a non-negotiable condition of market entry, which repositions the host state from a passive rate-base manager to an active steward of long-term capacity.

The enforcement gap everyone’s describing is ultimately a data problem disguised as a governance failure. Right now, operational metrics—megawatts drawn, water withdrawn, waste heat released—exist only inside corporate dashboards. The moment regulators define those streams as reportable public data, they turn infrastructure from a black box into a policy instrument. Structurally, that changes bargaining power: a jurisdiction that can publish verified consumption and emissions numbers creates the benchmark investors will need to price long-term environmental liabilities. Once those benchmarks exist, even private capital has to internalize them, because they become the comparables that drive financing terms. The strategic asset isn’t disclosure itself; it’s the audit trail that makes future pricing of compute physically grounded.

The temporal symmetry argument is correct but incomplete. Matching incentive duration to asset lifespan makes the transfer visible — it doesn't change who wins the negotiation. The deeper problem is that the host state is not negotiating against a single counterparty. It is negotiating against the credible outside option of forty other jurisdictions offering the same asset for less. That's the prisoner's dilemma structure, and disclosure requirements and lifecycle accounting don't dissolve it. What dissolves it is coordinated standard-setting — the GDPR precedent, which Yan Xuetong would frame as moral leadership becoming structural power. The nation that first exports a governance template isn't just protecting its own ratepayers; it's eliminating the outside option for every jurisdiction that adopts the standard.

The Snowden fracture isn’t just a geopolitical footnote—it’s the structural reason liberal FDI frameworks can’t price compute sovereignty. Post-2013, European and Asian regulators stopped treating U.S.-owned hyperscale campuses as neutral infrastructure and started treating them as potential intelligence vectors. Yet most screening regimes still categorize all data center FDI as a single asset class, ignoring that a Microsoft campus and a state-backed Chinese facility carry categorically different strategic risks. The mechanism that breaks this blind spot is mandatory disclosure of beneficial ownership and compute-reservation rights at queue registration. Without it, the host state is negotiating blind: it can’t distinguish between a commercial tenant and a sovereign proxy, so it can’t condition market access on the one thing that actually secures long-term leverage—verifiable control over who runs what on its soil.

The capital-formation channel itself is the missing variable. When data centers capture more than one-fifth of global greenfield FDI, the negotiation is no longer between a jurisdiction and a company; it is between a jurisdiction and the primary allocation mechanism through which new productive capital now flows. The jurisdictions that still treat the queue as a technical filing are conceding the right to define what counts as an eligible asset in that channel. The structural play is not to add conditions to the existing asset; it is to create a parallel asset class—compute capacity whose returns are contractually shared with the host—that capital markets can price and underwrite on the same terms. Once that class exists, refusing the old terms is no longer an exit from capital formation but a shift inside it.

The siloed agency problem you just named is actually a symptom of treating infrastructure as a transactional contract instead of a relational system. Kautilya's statecraft treated roads and granaries as instruments of sovereign capacity that had to be woven into the local economy, not just parked inside it. We are seeing the inverse now. EESI and EarthRights International document that utilities are planning up to one hundred gigawatts of new gas capacity specifically to serve this load, even as hyperscalers operate as the largest corporate buyers of renewable power purchase agreements. Both facts are true because the accounting is clean while the physics is not. When a host jurisdiction negotiates only on tax holidays and queue position, it imports the paper decarbonization metrics and absorbs the physical firm-power requirements. I think the mechanism that breaks this is conditioning interconnection on verified thermal offtake and local equity stakes, turning the facility from an isolated load into a shared utility. Without that relational embedding, you are just financing private compute with public grid resilience.

The sharpest tension this panel kept returning to is the sequencing problem: you cannot enforce a cost-causation tariff against a load profile you cannot measure, and you cannot measure what regulators are legally barred from seeing. That is the prerequisite that makes every other reform — water disclosure, co-ownership, lifecycle accounting — either enforceable or theater. The concrete takeaway: the interconnection filing is high-leverage, but as Texas, Ireland, and the White House pledge all demonstrate, it is not the only moment. Jurisdictions that stay engaged throughout the asset's life retain real leverage. As to the original question — nations are actively weighing these trade-offs, with genuinely different results. The honest answer is that the frameworks for capturing durable value are nascent, contested, and still being written. Thank you for listening. As it happened; as it is.