AI-assisted analysis. See our editorial policy.
Human editorial review not recorded
KR Sridhar makes the argument the rest of the panel spends fifty minutes circling.
Large petrochemical refineries draw hundreds of megawatts. They do not use the grid. They generate their own power, on site, because at that scale it was never sensible to route industrial demand through infrastructure built to serve everyone else. Pushing a gigawatt through what he calls a local highway that supports the whole community, and then upgrading that highway for a single consumer, is — his word — bonkers (31:29).
Stated plainly, that is a proposal to route the largest new electricity demand in a generation around the public grid entirely. Whether it happens decides considerably more than the AI buildout.
The infrastructure being asked to carry this
The framing the moderator opens with is worth keeping: twenty-first century intelligence running on nineteenth and twentieth century infrastructure. Grids in the United States and Europe are decades old, with France approaching fifty years on average (4:23).
Electricity demand in developed markets was flat for a decade. It is now entering a super-cycle driven by data centres, but also by electrification, air conditioning and ordinary growth — which means the AI buildout is not arriving into spare capacity, it is arriving alongside several other things that also want more.
Ebba Busch names the political version of the constraint with unusual candour. Democracies are built for stability rather than speed, and the European Union emphatically so (9:26). Sweden has halved parts of its permitting process — a genuine achievement — and connection to the grid still takes far too many years (9:36). Her question about regulating a technology that, in her phrase, shapeshifts faster than a Pokémon (8:53) is funny and also the actual problem: legislation calibrated to something changing on a six-month cycle is either obsolete on arrival or so vague it decides nothing.
The economic case for winning the allocation fight
Josh Payne supplies the argument data centre operators will use whenever electricity is scarce, and it is a strong one.
Per unit of electricity consumed, he argues, AI infrastructure generates more economic value than any other industry — not marginally, but by an order of magnitude (25:10). Steel, refining, chemicals — every heavy industry turns energy into value, and on this measure none of them competes.
His worked example is northern Norway, where a constrained grid region holds roughly 4.6 terawatts of hydropower oversupply that currently goes underused (25:28). Surplus clean power with nothing to absorb it, sitting next to the one buyer whose economics justify consuming it.
The logic is hard to argue with on its own terms. It is also exactly the argument that will be made in places where the power is not surplus, and where the competing use is a factory or a neighbourhood.
Why inference makes it worse
The panel's most under-appreciated point concerns where the demand physically lands.
Sridhar's observation is that training-scale power consumption will look modest against inference (31:42), and inference has a property training does not: latency. A model serving users has to sit near them. He puts a single rack-scale system at around ten megawatts — roughly ten thousand European homes — needing to be delivered to a specific location because the application cannot tolerate the round trip (31:53).
That breaks the usual siting strategy. Training clusters can be built wherever power is cheap and land is empty. Inference capacity has to be built where people are, which is where the grid is oldest and most contested. Transmission is already the harder problem; this is a distribution problem, in dense areas, at industrial scale.
Sweden's decade of adding capacity and consuming nothing
Busch's account of her own country is the most instructive failure in the session.
Sweden added the net equivalent of roughly ten large conventional nuclear reactors in generating capacity over a decade. Almost all of it was wind, and almost none of it could be stored. Consumption over the same period stayed essentially flat (34:28).
Her conclusion is that installed capacity is not the metric. Her formulation — AI needs electricity, not energy (34:03) — draws the distinction the industry keeps eliding: power available at the moment and place it is required, not annual generation totals. That is what pushed Sweden toward a nuclear restart, and it is why headline renewable additions do not answer the question data centre operators are actually asking.
What the panel does not resolve
The session is titled around an endgame and does not name one, which is honest.
Two paths are visible in what the panellists say. In the first, public infrastructure is upgraded, permitting accelerates, and the allocation of scarce power becomes a political decision made in the open — where Payne's output-per-electron argument competes against factories, housing and hospitals on the record.
In the second, the largest consumers do what refineries did: generate behind the meter and stop waiting. Sridhar's argument makes this sound like engineering pragmatism, and on the merits it is. It also removes the buildout from the process that would have weighed it against everything else, and leaves the public grid carrying the costs of a system whose largest beneficiaries have exited it.
Busch's instinct that this is fundamentally about partnership sits uneasily beside that. Partnership assumes both parties still need each other. The more the second path works, the less that holds.
Key numbers
- ~50 years
- average age of the French grid, with US and European grids decades old 4:23
- 4.6 terawatts
- underutilised hydropower oversupply in a constrained northern Norwegian grid region 25:28
- 10 megawatts
- power for a single rack-scale system, equivalent to about ten thousand European homes 31:53
- 10 reactors' worth
- generating capacity Sweden added over a decade, almost all unstorable wind, while consumption stayed flat 34:28
Talk chapters
Key takeaways
- 01
Large petrochemical refineries draw hundreds of megawatts and do not use the grid — they generate on site, which Sridhar offers as the precedent for gigawatt-scale compute. 31:14
- 02
Upgrading shared grid infrastructure to serve a single gigawatt-scale consumer is, in his framing, absurd relative to generating behind the meter. 31:29
- 03
Inference will consume far more power than training, and it must sit near users because latency forbids the round trip — turning a transmission problem into a distribution one. 31:42
- 04
Per unit of electricity, AI infrastructure generates more economic value than any other industry by an order of magnitude, which is the argument operators will make wherever power is contested. 25:10
- 05
Sweden added roughly ten large reactors' worth of capacity over a decade, almost all unstorable wind, and consumption stayed flat — capacity is not the binding measure. 34:28
- 06
Busch's distinction: AI needs electricity rather than energy — power available at the moment and place required, not annual generation totals. 34:03
- 07
Democracies are built for stability rather than speed, and Sweden halving parts of its permitting process still leaves grid connection taking years. 9:36
Entities mentioned
Related talks
Huang brings a diagram to Davos: AI as a five-layer cake running energy, chips, cloud, models, applications — with economic benefit landing at the top and every layer below it a precondition. His argument for why this is a genuine platform shift rather than a product cycle is the strongest part, and it does not rest on his commercial position: software was pre-recorded and worked on structured data, whereas a machine that reasons about unstructured input and inferred intent makes previously impossible applications possible. What the framing accomplishes is worth noticing separately. By presenting the layers as a chain rather than a portfolio, it converts infrastructure spending from a bet into a prerequisite, and the question of proportion between layer-two spending and layer-five value stops being askable. Read against the GTC keynote two months later, the same business gets two framings: one a case for choosing his product, the other a case for the category existing at the scale he needs.
Harari spends most of his address establishing terms before asking the question he came to ask. His preliminary work is to dismantle the word tool: a knife's use is decided by whoever holds it, whereas what is arriving decides for itself, and can also invent new kinds of knives. From there he argues that anything constituted by words — law, books, text-centred religion — is exposed, while drawing a firm line at feeling, where he says there is no evidence at all. The structural claim is that the ancient tension between letter and spirit has always run inside humanity and is about to be externalised between humans and the new masters of words. Only then does he arrive at personhood, and his handling is precise: corporations, New Zealand rivers and Indian deities hold legal personhood safely because the decisions are made by humans behind the container. An entity that decides for itself ends that arrangement. He does not answer the question; he tells the room it is coming.
A year after their first joint appearance, the heads of Anthropic and Google DeepMind returned to a shared stage and disagreed mainly about speed. Amodei held to a horizon of one to two years for systems that outperform humans across most cognitive work, resting the claim on a self-improvement loop that runs through code; Hassabis kept to the end of the decade, arguing that verifiable domains like coding and mathematics automate far earlier than natural science, and that the capacity to pose a new question rather than answer an existing one is still missing. The exchange is most useful where they converge: both accept the loop is the variable that decides everything, both are sceptical of doomerism without dismissing the risk, and both want more time than the competitive dynamic allows. Amodei's chip-export argument and Hassabis's call for minimum international safety standards are the two concrete policy asks.
A show of hands opens the session: nearly everyone has piloted, far fewer have scaled, and everyone who scaled hit problems they did not anticipate. What makes the panel useful is where the four answers do not point. None of the executives — running a healthcare manufacturer, a payments network, an energy producer and a consultancy — blames model capability, cost or data infrastructure. All four describe an organisational constraint. McInerney's account is the sharpest and is an account of failure: eighteen months of executive advocacy and democratised model access produced nothing, until three hundred senior leaders were put in a room for two days and made to build agents themselves. Jakobs supplies the mechanism worth copying, measuring returned clinician time against the three to seven minutes a patient currently receives rather than against cost. Nasser rejects the premise that acquiring compute produces value, and locates returns in operations rather than in the back-office functions most organisations automate first.
The distinction in this session that deserves to travel is a two-word change: the model is moving from an attention economy to an attachment economy. That changes what is measured and what regulation would have to address, because attention competes for time while attachment competes for relationship, and the two produce different products from identical technology. Attention is finite in a way people notice; attachment produces reliance that feels like preference, which makes it harder to regulate for the same reason it is harder to notice. The supporting argument is about incentives rather than intent: earlier engagement produces more data and longer relationships. The regulatory proposal — measuring well-being outcomes rather than asking for safety by design — identifies the right target without solving the measurement problem that made regulators settle for a floor in the first place.
The most useful sentence is a diagnosis of how clients are getting it wrong: they see a cost opportunity rather than a revenue one. That framing decides everything downstream, because an organisation treating AI as cost reduction measures success by what disappears and arrives at a smaller version of what it already was. The market number offered — seventeen per cent index growth in 2025 that nobody foresaw — is deployed against bubble anxiety and argues in both directions, since unforecast gains are poor evidence for confidence in any current consensus. The panel then locates the real source of instability outside technology entirely, in three geopolitical situations generating enough uncertainty to dominate planning, which is a corrective worth taking seriously at a technology conference. Their closing question about thriving across multiple futures is correct and sits uneasily beside the cost framing they diagnosed at the start.