The Electron Race

What DataCloud 2026 in Cannes revealed about the real contest beneath the AI boom — power, capital, and the sovereignty of the supply chain

The conference was billed as a data-centre event. Three days on the Croisette, a procession of hyperscalers, operators, turbine-makers, financiers and — newly (and conspicuously) nuclear physicists. The lanyards said digital infrastructure. But anyone who sat through all of it left with a different conviction: this was an energy-security summit that happened to be wearing a technology badge.

The tell was in the arc of the three days. Day One opened in the register of demand triumphalism — gigawatt campuses, SoftBank committing seventy-five billion dollars on top of a previously announced hundred-and-ten, the comfortable assertion that AI load is simply additive to a cloud business still compounding at twenty-five to thirty per cent a year. By Day Three the same industry was sitting in rapt attention as the room turned to the United States building a reactor at 2.5x budget, and to how some governments are quietly underwriting the risk that private capital will not touch.

In between, the conversation circled a buried fact: the West is trying to win the artificial-intelligence race on infrastructure it cannot yet reliably power, cannot yet cleanly finance, and cannot yet independently fuel — without leaning on the rivals it is racing.

I. The constraint inversion: power is the new geography

For thirty years, siting a data centre was an exercise in real estate, water-proximity and connectivity. You found land near a fibre route, near a metro, near customers, and you asked the utility for power more or less as an afterthought. That logic is dead.

One large operator now calls its method a “power-first strategy” — power is the first question in business planning, not the last, and it has built a dedicated power team to manage utility relationships and chase alternative routes to electrons. A global colocation operator described grid connection as having become a campaign rather than a request: a sustained, multi-party negotiation with transmission and distribution operators that can determine whether a project lives or dies. The radius of acceptable sites around major metros is expanding, not because anyone wants to build further out, but because power availability has overtaken connectivity as the binding constraint.

The numbers behind the problem are not subtle. Global IT power sits around 100 GW today, perhaps sixty per cent of it hyperscale; the moderator of the opening keynote framed AI as climbing toward roughly thirty per cent of all global IT power by the end of 2027. The standard deployment has swollen from 50 MW to campuses of 300 to 700 MW, with multiple operators now openly planning gigawatt-scale single sites. One large US operator described delivering four one-gigawatt projects and a European pipeline running from 130 MW to 700 MW per site. A pan-European developer, historically a 250 MW builder, said it is now working on one-gigawatt builds and that around eighty per cent of its demand is AI-related.

The independent data corroborates the mood. The International Energy Agency’s assessment puts data-centre electricity consumption roughly doubling to around 945 TWh by 2030 — close to Japan’s entire national consumption — with AI-specific demand tripling and the United States and China together accounting for nearly eighty per cent of the growth.

But the genuinely important insight from Cannes was not the size of demand. It was the timing mismatch underneath it, and a power-equipment executive stated it most cleanly: a data centre can be built in eighteen months to three years; the grid connection to feed it can take five to fifteen. A typical interconnection runs to seven years. That gap — anywhere from two years to a decade between when the load wants to exist and when the wires can carry it — is the single most consequential fact in the industry, and no amount of capital makes it close faster.

This is where the conference stopped being about technology and became about the State. Because the only actors who can compress a grid-connection timeline are governments and regulators, the question of who can build AI capacity has quietly become a question of which states can deliver electrons fast enough. Compute is no longer downstream of chips alone. It is downstream of national infrastructure capacity — and that makes it statecraft.

The regional scorecard that emerged was unsentimental. The United States enjoys structurally cheap energy (it is, as one panellist put it, one of the most energy-rich countries on earth, with abundant domestic oil and gas), deep capital markets, hyperscaler balance sheets and federal momentum behind nuclear. China brings state-backed execution, supply-chain scale and ruthless standardisation. Europe has superb engineering, a serious safety culture and ambitious climate policy — and is losing, because its permitting is slow, its regulation fragmented across jurisdictions, and its grid expansion glacial. The US holds roughly 50 GW of capacity to Europe’s 10–15 GW; one European operator called the fact that so much European data still lives on American soil “nonsense,” and argued that European training and inference belong in Europe. The subtext was sovereignty, even when the word went unspoken.

One detail deserves more attention than it received on the day. Chinese hyperscalers, the opening panel noted, are increasingly treating Europe as their route to becoming global brands — precisely because expanding in the United States has become commercially and politically fraught. Europe, in other words, is becoming the contested middle ground of the AI infrastructure contest: too slow to lead, too open to refuse, and courted by both poles of a bipolar compute world.

II. The bankability gap: capital cannot yet price what the State is asking it to build

If power is the physical bottleneck, capital is the financial one — and the most intellectually serious session of the conference was the one asking whether off-balance-sheet financing of data centres is “too good to be true.”

The structural migration the panel described is clear. Data-centre financing has moved from corporate balance sheets and real-estate lease structures toward project finance and PPP arrangements, and is now reaching for the bond and asset-backed securities markets — driven by a single forcing function. The capital required has simply outgrown everything else. Banks can comfortably underwrite a one-to-two-billion-euro European project; a five-gigawatt campus of the kind now contemplated in the Gulf requires pooling the entire capital stack, institutional bond buyers included. One private-credit investor was blunt: the amounts are now so large that the bond market must be tapped, and Europe — though it lags the US and the UAE on scale today — will outgrow its bank markets very soon.

The scale of the off-balance-sheet shift is now quantified beyond the conference room: technology groups moved more than $120bn of AI data-centre debt off balance sheet in under two years, via special-purpose vehicles, private credit and corporate bonds. The capex dwarfs the revenue — roughly $400bn of capex against perhaps $60bn of AI revenue in 2025, with a multi-trillion-dollar build-out projected by the end of the decade — which is exactly what is pushing issuers into the capital markets in the first place.

Here is the problem the markets have not solved; to sell this risk to institutional investors, you have to make development-stage risk legible — permitting, construction, energisation, cost overrun, pre-commercial revenue — and investors do not yet know how to digest it. The signature of that discomfort is a hard rating ceiling: the rating agencies have so far kept triple-A away from data-centre ABS, capping co-location collateral at single-A. Moody’s, having published a dedicated data-centre ABS methodology in early 2025, went on to warn that AI data-centre leases may not fully capture the underlying commitments — the market, in other words, is still learning to read the paper.

This rhymes with financial history. The last generation of project bonds was routinely monoline-wrapped: a third-party guarantor converted project risk into a AAA bond — until that market over-extended into structured credit and collapsed in 2008. The function never fully died (at least one guarantor survives and still wraps niche infrastructure), but the wrap shrank to almost nothing.

What is striking is that a replacement for the wrap is now emerging for data centres — and it is not, yet, an insurer. It is the hyperscaler balance sheet. The clearest example is the roughly $3.2bn debt transaction in which Google agreed to backstop a data-centre tenant’s lease obligations — a strong third-party covenant that converts otherwise unrateable tenant risk into investable paper. That is functionally a modern credit wrap.

And it carries the same flaw that destroyed the monolines. The backstop activates only once the lease commences; if construction is delayed beyond a set window the tenant can walk, and during the build-out itself bondholders bear execution risk with limited support. The new wrap, in other words, covers the operating leg and leaves the development leg — permitting, energisation, cost overrun, the very risks the whole conference kept circling — largely naked. That is the failure mode of a guarantee that fails when you most need it.

The open question, then, is whether a genuine credit-enhancement market re-forms to wrap construction and energisation risk on data-centre project bonds. The economics demand it: that unwrapped leg is exactly what caps ratings at single-A and locks out the deepest institutional pools, the insurers and pension funds that can only buy stabilised, highly-rated paper. Whoever productises a credible construction-phase enhancement — a surviving financial guarantor, a private-credit-aligned insurer, or a synthetic multi-party structure of tranching, residual-value insurance and EPC completion guarantees — unlocks the investment-grade bid and reprices the entire sector. The first generation will probably repeat the monoline mistake. The durable product is the one that wraps the riskiest leg, not the safest.

The deeper point, for a geopolitical reader, is this: capital has not yet worked out how to price national-strategic infrastructure built at venture speed. The bottleneck is migrating again — from balance-sheet capacity to risk-allocation engineering.

III. The contracting front line: where the strain shows first

Lawyers and procurement teams felt the same pressure from a different angle, and the contracting panels were where the abstractions became concrete.

Speed of delivery is now the dominant commercial driver, and it is breaking traditional contracting. Customers want capacity inside eighteen to twenty-four months — which happens to be roughly the average delay now reported on large projects — and cannot sit through eighteen-month negotiations to get it. Power readiness has migrated into the service-level agreement itself: if the power is not enabled, the customer may simply walk. Force majeure, several speakers warned, is being stretched well past its purpose, with parties trying to reclassify foreseeable risks — utility delays, supply-chain disruption, even geopolitical tension — as acts beyond reasonable control. The wiser counsel was to allocate known risks expressly and stop hiding them in boilerplate that will only be litigated when the relationship has already broken down.

Two threads here are pure geopolitics. First, GPUs remain the most supply-constrained component in the chain, and an Australian operator made the point that exposure to Indian and Chinese supply routes can be an advantage over relying solely on European chains — a quiet admission that supply-chain alignment is now a competitive variable, not a compliance footnote. Second, the financing of the chips themselves now turns on the quality of the offtake: lenders scrutinise termination and walk-away rights because a contract full of exits cannot support debt. The numbers are not trivial — one example put a six-month delay at roughly $80m in SLA penalties or equivalent renegotiation, dwarfing the liability caps inherited from a smaller era.

The financing-structure panels mapped the plumbing. One GPU-cloud provider’s delayed-draw term loan (a bankruptcy-remote, non-recourse vehicle holding customer contracts, leases and the GPUs themselves, self-amortising over five-year take-or-pay contracts) was held up as a genuine innovation, with one panellist comparing its potential trajectory to the way CMBS once became standard. A power developer described a market failure it is exploiting directly: capital wants riskless execution and execution waits on capital, a “chicken and egg” stand-off that it breaks by deploying its own equity first, in West Texas, and unlocking cheaper financing once the project is de-risked. The same panel floated West Texas as a potential 25 GW market — a reminder that in the US, the constraint really is capital scaling to meet demand, whereas in Europe the constraint is whether politics and permitting can catch up at all.

IV. The supply chain as a weapon

The hyperscaler keynote on procurement was, on its surface, an operations talk. Read properly, it was a briefing on economic statecraft.

The risks its procurement leader named were almost entirely geopolitical: instability, tariffs, sanctions, semiconductor constraints, single-country dependencies, long lead times, raw-material and rare-earth shortages. The prescription — “look forward and walk backward,” planning two to three years ahead, reverse-engineering requirements, standardising components and insisting on multi-tier visibility down to tier-three suppliers — is the language of supply-chain security, not merely efficiency. The concrete result cited was telling: lead times for critical power-conversion equipment that had blown out to sixty or sixty-five weeks during the pandemic were compressed back to six to ten through standardisation and dual-sourcing.

The principle that complexity is the enemy of speed, and that standardisation is therefore the first imperative, recurs throughout. It is also, not coincidentally, exactly the principle on which China’s industrial machine is built. When the dominant AI chipmaker spoke of helping partners extract thirty to forty per cent more compute from the same physical footprint, it was describing the other half of the same equation: if you cannot build power and supply chain fast enough, the remaining lever is to do more with the silicon you can actually energise.

V. The nuclear faith — and the fuel trap beneath it

Day Three belonged to nuclear. And the mood had shifted decisively from prior years. The industry has stopped asking whether nuclear is an option and started asking when, and how much.

The case made on stage was coherent. Hyperscalers have balance sheets larger than some countries — which makes them the natural first-mover market that risk-averse utilities have never been willing to be. Small modular reactors suit data-centre economics because their modularity matches the building-by-building rollout of a campus: phase the reactors with the racks, ring-fence each building’s power risk, and you get both “speed to revenue” and N+1 redundancy that a single thousand-megawatt turbine can never offer. One advanced-reactor programme is targeting a first commercial plant for 2028, with a 1.2 GW offtake agreement phasing in from 2030 and twenty-year power purchase agreements pitched around $120–140/MWh — competitive, it was argued, with new gas in major US markets. Nuclear’s capacity factor, the highest of any source, was offered as the clincher.

The external data confirms this is not mere conference enthusiasm. The IEA’s modelling has the conditional offtake pipeline between data-centre operators and SMR projects growing from 25 GW at the end of 2024 to roughly 45 GW, technology companies signing around forty per cent of all corporate renewable PPAs in 2025, and SMRs entering the data-centre supply mix as a baseload backbone after 2030.

But three hard truths sat under the optimism.

Truth one: cost realism. The most bracing talk of the event used Vogtle — the only large reactors the US has completed in a generation — as both proof and warning. The two new units came in around $37bn against an original estimate near $15bn. The distinction drawn between “cost overrun” and “cost realism” was sharp: a first-of-a-kind nuclear project in a country that has not built one for decades will always cost more than the optimistic estimate, and the honest fix is better estimating, not blame. But that creates a financing paradox — would it have been built if the costs had been known in advance? Perhaps not. The learning curve is real (the second of the two units came in twenty to thirty per cent cheaper than the first, and modelling suggests SMRs reach the bottom of the cost curve after roughly eight identical units), but somebody has to finance the expensive early units before the cheap later ones exist.

Truth two: nuclear cannot follow an AI load. The most technically arresting point of the day, drawing on national-laboratory analysis, was that AI training workloads swing violently — up to fifty per cent of maximum demand, changing at ten to twenty per cent per minute and, in bursts, ten to thirty per cent per second. A reactor is a magnificent provider of steady baseload and a hopeless follower of that kind of volatility. The pairing only works with batteries, inverters and serious power-electronics integration sitting between the reactor and the racks. The clean-firm dream, in other words, has a storage problem buried inside it.

Truth three — the one the room kept circling without quite naming — is fuel. Across multiple Day Three sessions, “fuel availability” and “the fuel cycle” were listed among the top blockers to nuclear deployment, with speakers carefully noting that the problem is more complex than mining or buying uranium. What went largely unsaid is why. Roughly two-thirds of SMR designs in development — including the high-temperature gas reactors and many of the advanced designs paraded at Cannes — require High-Assay Low-Enriched Uranium (HALEU), enriched between about five and twenty per cent. And HALEU is, today, a near-monopoly held by the West’s principal strategic rival: Russia controls roughly forty per cent of global enrichment capacity and has been, until very recently, effectively the sole commercial supplier, with only Russia and China producing it at scale.

This is the trap. The West’s chosen answer to the AI power crisis — advanced nuclear — depends on a fuel that the West cannot yet make in commercial volume. The United States banned Russian uranium imports in May 2024 (with waivers to 2028) and produced under a tonne of HALEU domestically in 2024 against a projected need exceeding fifty tonnes a year by 2035, and is now racing to build a sovereign fuel spine through domestic and allied enrichers — capacity that does not meaningfully arrive until 2027 and beyond (although perhaps that is soon enough). Europe, whose enrichers could in principle produce HALEU with existing centrifuge technology, says it still needs to see the business case before committing the capital.

So the geopolitical reading is hard to avoid. An industry racing to power AI with nuclear, in the name of energy security and sovereignty, is building toward dependence on an enrichment supply chain dominated by Moscow and Beijing — and the scramble to close that gap is now one of the most strategically important infrastructure programmes in the Western world, conducted almost entirely off the data-centre industry’s radar.

VI. Two models, diverging: the American market and the European state

The most revealing geopolitical contrast of the whole conference was structural, and it surfaced on the nuclear-deployment panel as an almost throwaway comparison.

In the United States, SMR deployment is bilateral, private-sector-driven and hyperscaler-led — technology buyers and reactor developers cutting gigawatt-scale deals, balance sheets and capital markets doing the heavy lifting, the federal government providing pilot programmes, executive orders and a fuel-security push but leaving the commercial risk largely to private actors.

In Europe — the State is in the structure and de-risks new nuclear through debt financing and other support at sovereign rates, power-offtake guarantees, and guaranteed return-on-equity protections for equity investors. These are mechanisms designed explicitly to assemble investable consortiums where the private market alone would balk. The United Kingdom’s regulated-asset-base model, funded through consumer bills, is a cousin of the same instinct: socialise the development risk that capital will not bear, because the strategic stakes justify it.

This is the deepest divide in the room, and it maps directly onto the old transatlantic argument about industrial policy. The American wager is that hyperscaler demand and capital-market depth will pull advanced nuclear into existence the way they pulled shale and cloud into existence — fast, private, and a little chaotic. The European wager is that only the state balance sheet can absorb first-of-a-kind risk at the speed the AI race demands, and that refusing to deploy it is how Europe loses. Both wagers are coherent. Both might even be right, and the AI build-out is about to test them at scale.

A sober European coda was supplied from the operator side. SMRs, the argument ran, will not rescue Europe’s near-term power gap; realistic operational dates are closer to 2033 than the 2028–2030 the optimists quote, with Generation III light-water reactors arriving before the Generation IV designs. If true, the “bridge” — grid upgrades (a ~56 GW European programme is underway), PPAs, storage, flexible demand and transitional gas — is not a three-year stopgap but closer to a decade. And Europe’s specific weakness is not engineering but consolidation: too many competing reactor designs, too little standardisation, too slow a path to bankability. Which is, once again, precisely the ground on which China is strong.

VII. What nobody quite wanted to say

The industry has committed to one hundred per cent clean energy by 2030 — and is, in the same breath, building gas turbines as the only bridge fast enough to matter, hoping that “hydrogen-ready” and “CCS-optional” labels will reconcile the pledge with the practice. The IEA’s modelling is unsentimental about this: gas and coal together are expected to meet over forty per cent of the additional electricity demand from data centres until 2030, with nuclear arriving only afterward.

The industry is betting on reactors with sixty-year lives to power data centres whose useful configuration may not survive the next GPU generation — a life-cycle mismatch one speaker captured by warning, soberly, against a future littered with stranded reactors, and by noting that decommissioning work is already visible in places like Switzerland. The quantum-computing wildcard sits behind that: a single architectural shift could rewrite the demand assumptions underpinning billions in committed generation.

And the industry’s most cited social risk was not technical at all. It was the community — the “CAVE people,” citizens against virtually everything, in one phrase from the floor — and the project that swung from an 8–1 preliminary vote in favour to 6–3 against in six weeks once organised opposition arrived. The honest operators have stopped fighting this. They now say plainly: find out whether a community wants you before you put a billion in the ground, and walk away from the ones that don’t. Social licence has become a hard input, as real as a grid connection.

The race to build artificial intelligence is, at the physical layer, a race to solve four problems the technology cannot solve for itself: electrons fast enough, capital structured cleanly enough, fuel sourced sovereignly enough, and communities willing enough. None of these is a chip problem. All of them are problems of energy, finance and State capacity.

The comforting story is that whoever has the most GPUs wins. Whoever solves risk-allocation engineering and sovereign fuel — whoever can finance a first-of-a-kind reactor and feed it without having to ask permission from Moscow — owns the physical layer on which the entire AI economy will run. The bottleneck has migrated, decisively, from silicon to statecraft. The contest everyone thinks is about intelligence is, underneath, a contest about electrons. And on the evidence of Cannes, the West has not yet decided whether it intends to win it with private capital, with the State, or — as the truth may turn out to be — with neither fast enough.