The Org Chart Is the Strategy

In March, Anthropic hired Ariel Horowitz. Her previous job was Deputy Director of Grid Modernisation at the United States Department of Energy. Not a research scientist. A person whose last brief was the American electricity grid.

In March, Anthropic hired Ariel Horowitz. Her previous job was Deputy Director of Grid Modernisation at the United States Department of Energy.

Not a research scientist. Not a safety lead. Not a distinguished engineer with a citation count and a following. A person whose last brief was the American electricity grid.

She was the company's first dedicated energy hire. In April, Sana Ouji followed as Energy Lead, after more than six years at Google working across data centre clean energy strategy, and before that at Lightsource BP and NRG Energy. In February, Tim Hughes had arrived from STACK Infrastructure, where he had been Chief Development Officer with oversight of more than 200MW of critical capacity across hyperscaler build-to-suit projects. At Anthropic he runs leasing, land and energy. He describes his own career with some economy, as a decade and a half spent working out "where to put large buildings full of computers", and how to get enough electricity into them.

Anthropic calls these three its inaugural energy team.

Inaugural is the word doing the work. It tells you the function did not exist until now. A company valued in the hundreds of billions, whose entire product is a model, has just discovered that it needs a grid strategy, and has gone out and bought one.

That is not a personnel story. That is a disclosure.

Hiring is the only honest document

Companies dissemble in press releases. They tell you what they wish were true, in language calibrated for the audience they most fear. Strategy decks are worse, because they describe the world the author would prefer to live in.

Hiring is different. Hiring is where an organisation stops describing its strategy and starts funding it. A headcount requisition is a claim about the future, backed by cash, signed off by someone who will be asked in eighteen months why they made it. It is the point at which intention becomes cost.

So when you want to know what an organisation actually believes, read the org chart, not the keynote.

And the org chart at the frontier of artificial intelligence has been quietly filling up with people who have never trained a model in their lives.

Reading the roster

The energy team is the newest layer, but it sits on top of something that has been assembling for over a year.

Winnie Leung is Head of Data Centre Infrastructure, arriving after more than two decades at Google, including a spell as a director of engineering for data centre operations. Brett Rogers led data centre construction and design at Google for close to six years. Liwen Mao is Data Centre Design Lead. Adam Johnson is Data Centre Electrical Lead. Zach Miller became Data Centre Operations Manager after seventeen years at Google. Peter Sarossy, a twenty year Google veteran, joined in January to work on data centre security engineering.

Design. Electrical. Operations. Security. Construction. Land. Power. Grid policy.

Now set that list beside the coverage. The coverage was about Andrej Karpathy. It was about the Nobel laureate. It was about the researchers, because researchers are legible, and because a research hire fits the story the industry prefers to tell about itself, in which the constraint is genius and the race is intellectual.

Meanwhile, the company hired an electrical lead.

The scale explains the urgency. Anthropic has been in discussion over at least ten gigawatts of data centre capacity. It has committed to a fifty billion dollar United States data centre programme with Fluidstack, with Google providing financial backing. In October 2025 it signed for over a gigawatt of Google cloud capacity, including up to a million tensor processing units. In April it agreed terms with Broadcom and Google for TPUs representing three and a half gigawatts.

Those are not software numbers. They are utility numbers.

And the stakes are understood internally as existential. Dario Amodei has said that being a year out on growth projections could be enough to send the company bankrupt. A rival's internal memo reportedly accused Anthropic of a "strategic misstep to not acquire enough compute". You do not stand up a named energy function in response to a nice-to-have. You do it because the thing you cannot get is the thing that ends you.

The constraint has moved

Here is the claim, stated plainly.

At the frontier of artificial intelligence, the binding constraint is no longer the model. It is land, power and interconnection. It is grid queues, planning consent, transformer lead times, and the willingness of a local authority to permit a substation. It is a production problem wearing a research costume.

The people building the frontier already know this. Their hiring says so, and hiring does not lie, because hiring costs money.

The rest of us have not caught up. We are still holding conferences about model capability, still benchmarking, still arguing about which lab leads by what margin on which evaluation. That conversation is not wrong. It is simply no longer where the constraint lives.

Once you see that, two other conversations look very different.

The first conversation: compute sovereignty

Ireland has spent a decade telling itself a story about being a technology nation. Talent, tax, multinationals, an English-speaking bridge into Europe. It is a good story, and much of it is true.

Here is the number that ends it.

In 2025, data centres consumed 23 per cent of Ireland's metered electricity. Not 23 per cent of industrial electricity. Twenty-three per cent of the metered total. In 2015, the figure was 5 per cent. Data centres now draw more power than every urban household in the state combined, which sits at 18 per cent, and more than twice the rural household share.

Consumption rose 10 per cent last year alone, from 6,973 gigawatt hours to 7,663. All other users, households and businesses together, grew by 2 per cent.

And it did that while a moratorium was in force. The Commission for the Regulation of Utilities had an effective block on new data centre grid connections in the Dublin region, in place since 2021, lifted only last December. The moratorium slowed new connections. It did nothing to the underlying demand curve, which, in the CSO's phrasing, has grown every single year without exception.

The replacement policy is stricter. New large energy users must bring their own flexible generation, and must meet at least 80 per cent of annual demand from renewable sources. That is a serious piece of policy, and it is better than a moratorium. It is also, unmistakably, an admission. The state is now rationing the input.

So consider what compute sovereignty actually means in a country in that position.

It does not mean a research grant. It does not mean a skills programme, a centre of excellence, or a memorandum of understanding with a hyperscaler. Those are all fine things, and Ireland has produced a great many of them.

It means this. Can we generate, transmit and connect the electricity that a sovereign AI capability requires, on land we control, on a timeline that matters?

If the answer is no, everything else is decoration. You can host every AI company in Europe and still be a tenant. Tenancy is a respectable commercial position. It is not sovereignty. And the distinction becomes very sharp indeed on the day the landlord's interests and yours diverge.

I have written before about Ireland's position in the European compute build-out, and about the strange spectacle of a country that markets itself as a digital hub while sitting outside the infrastructure programmes that will define the next decade. I no longer think that is a technology policy failure. I think it is an energy policy failure that has been misfiled.

The frontier labs have worked this out. That is why one of them hired a grid modernisation official out of the Department of Energy, and why they are all negotiating in gigawatts rather than gigaflops. They have stopped competing on models. They are competing on electrons.

A national AI strategy without an energy chapter is not a strategy. It is a brochure.

The second conversation: the deployment gap

Now shrink the frame.

I have spent years watching large regulated organisations, mostly in pharmaceuticals and life sciences, fail to get value from artificial intelligence. The failures are remarkably consistent, and almost none of them are technical.

The pattern runs like this. A capability is procured. A pilot is run, and the pilot works, because pilots always work. Then the thing has to move into an environment with validated systems, audit trails, change control, data lineage, regulatory exposure, and a named individual whose signature carries legal weight. And there it stops. It stops not because the model is weak, but because nobody built the apparatus that a production system requires.

Governance was bolted on afterwards, when it should have been encoded upstream. Ownership was distributed, which is a polite way of saying it was absent. Nobody could answer the only question that matters in a regulated environment, which is: who is accountable when this is wrong?

The organisation bought AI as a product. It then discovered, expensively, that it had acquired something which needed to be run.

This is the same failure as the sovereignty failure. Precisely the same. Both mistake the artefact for the system. Both assume the hard part is acquisition, when the hard part has always been operation. Both treat infrastructure, governance and delivery as details that will resolve themselves once the interesting decision has been taken.

The state does it with electricity and land. The enterprise does it with pipelines, controls and accountability. It is one error, expressed at two altitudes.

The symmetry

What makes the Anthropic hiring significant is not that it is clever. It is that it is early, and that it is honest, in the only way an organisation can be honest, which is with its payroll.

The labs are, at this moment, the only actors in the ecosystem behaving as though AI is a production system rather than a product. They have energy teams. They have electrical leads. They have operations managers, construction directors, and people who understand interconnection queues.

They have built, inside a research company, something that looks unnervingly like a utility with a compliance function bolted to the side.

They did not do this out of foresight or virtue. They did it because the constraint bit them, and it bit them first, because they are furthest along the curve.

It will bite everyone else in turn. It is biting Ireland now, at 23 per cent and climbing. It is biting every pharmaceutical company that cannot get a model out of the sandbox and into a validated environment.

The organisations that come through this will be the ones that stop asking which model to buy and start asking what it takes to run one. That question has an unglamorous answer. Power. Land. Pipelines. Controls. A named human being who is accountable when it fails.

None of that is exciting.

All of it is load-bearing.

The labs have already stopped treating AI as a technology. Read their org charts. They treat it as infrastructure.

Ours still treat it as a pilot.