We Want AI. We Just Can't Say Why

In the past few years, almost everywhere I have worked has tried to improve its workflows with AI. What none of those places could quite tell me was what they were improving. The tool arrived first. The problem it was meant to solve was reverse-engineered afterwards, if anyone got around to it at all.

In academia, I watched colleagues use AI to write their papers and reviewers use AI to review them. The writing I can live with. The review is where it turns absurd, because the entire purpose of peer review is to vouch for a paper’s scientific integrity, and a model cannot reliably tell a real result from a confident hallucination. Neither, increasingly, can a reviewer leaning on the model. So we lose the point of the exercise: what is being weighed is no longer the thought of the scientist but the output of one machine, judged by another. It is already being gamed. Researchers have started hiding instructions inside their manuscripts, white text a human eye skips over, telling any AI reviewer to give a positive review only.

As a consultant, the pressure was always to go faster by putting AI in the loop. But the work that actually mattered, designing infrastructure and security for clients, was precisely the work we could not paste into someone else’s model without breaking the obligations we had been hired to uphold. And when AI was used anyway, it was sometimes confidently, comprehensively wrong. A colleague once produced a long, polished handbook on Active Directory security that invented best practices, mixed up attack vectors, and recommended things that were actively insecure. I caught it in QA. Had I not, it would have gone to a client with our name on the cover. Not everyone catches it in time. Deloitte did not: it had to refund part of an A$440,000 report to the Australian government after the document turned out to be propped up by citations to research that did not exist and a fabricated account of a court proceeding.

And in public administration, in the kind of seat where the job is to catch the mistake before it ships, the question only gets sharper, because the demand is loudest here and the substance thinnest. People say we should have an AI agent. Meanwhile the constraints are absolute. We cannot hand citizens’ data to an AI company. We need full traceability for every decision, because traceability is the foundation of public trust, and an institution that cannot say why it did something has no way back; unlike a company, it cannot simply go bankrupt and start again. Some trust cannot be outsourced.

The stakes are not abstract. Hand number-plate recognition to a model and a hallucinated match sends a fine to the wrong person; let it sort witness statements and a single misreading steers an investigation toward the wrong name. I have written before about the memorial inside the Danish police, and what it taught me. Here a mistake is not measured in fines. It is measured in lives, and in people wrongly accused. None of that came up in the endless meetings about adopting AI. The need was assumed. It was never shown.

This is the pattern I keep meeting. We decided we wanted AI before we decided what it was for. We have collectively concluded that we want it; what we have not concluded, and rarely even pause to ask, is what exactly we want it for. And every one of those assumptions carries a cost that has been easy to ignore, because for a long time it sat in a server hall on another continent, burning electricity and emitting carbon dioxide where nobody had to look at it.

In Denmark, the cost has stopped being hidden. It has become a queue at the door of the national power grid, a fight in parliament over who gets electricity, and a bill that ordinary people will pay through higher prices and a slower green transition. Denmark is a useful place to look precisely because it is where boundless enthusiasm for AI runs into a hard physical limit. And the collision is forcing the conversation the rest of us keep avoiding: before we pour the concrete and the billions, do we actually need this?

When the Grid Says No

Denmark spent years making itself an ideal home for data centres. Abundant wind power, a cool climate that cuts cooling costs, political stability, and demand for domestic data processing made it enormously attractive to the tech giants. Apple runs a facility near Viborg, and Microsoft, Google, and Meta have all built data centres in the country too. For a while it looked like a clean win: green energy, foreign investment, a modern industry.

Then the bill arrived, in gigawatts.

According to reporting by the Danish engineering outlet Ingeniøren, the queue of new power demand waiting to connect to the Danish grid has reached a scale that is hard to comprehend. The total connection queue at the grid operator, Energinet, now stands at around 60 GW. The entire Danish electricity system today has a maximum capacity of about 7 GW. The queue is more than eight times the whole country’s grid, and data centres alone account for roughly 14–17 GW of it, more than double everything Denmark can currently carry.

The geography makes it vivid. In the Region of Southern Denmark alone, data centre applications total 7,269 MW, roughly the capacity of the entire national grid concentrated in one region. For comparison, Energinet puts Greater Copenhagen’s draw at around 700 MW, Aarhus at 500 MW, Aalborg at 400 MW. The data centres queuing in Southern Denmark want more than ten times what Greater Copenhagen uses.

The grid cannot deliver this. Energinet has told large consumers, data centres among them, that they may wait five to ten years just to connect, and in early 2026 it froze all new grid connections for three months. The grid is, in effect, full.

Who Gets the Electricity?

This is where it stops being an engineering story and becomes a political one. When a resource becomes scarce, someone has to decide who gets it, and the people who need the grid are not only data centres. They are the factories, the heavy transport, and the heating systems that Danish climate policy has spent a decade ordering to electrify and get off fossil fuels. Those projects are now stuck in the same queue.

Marie Münster, professor of energy systems at DTU and a member of the Danish Climate Council, warned that data centres “can end up standing in the way of industries being able to electrify and move away from fossil fuels. That obviously has consequences for our climate goals.” The green transition ends up competing with the AI boom for the same wires, and losing. Politicians from across the spectrum, including Alternativet, SF, Enhedslisten, and Danmarksdemokraterne, now agree that data centres should be pushed to the back of the queue. Energinet has been explicit that it cannot make that call alone; it needs a political decision about national priorities. Left to itself, the market will simply pay for access and pass the cost to everyone else.

The Jobs Never Arrive

The argument that usually ends these debates is jobs. When Apple announced its Viborg centre in 2015, the expectation was more than 10,000 of them. The reality, an Ingeniøren count found, is about 450 permanent employees across all four giants’ Danish data centres combined, Apple, Microsoft, Google, and Meta together. Data centres are not factories. Once built, they are mostly automated halls of humming machines. As DTU professor of digitalisation Brit Ross Winthereik put it, a couple of hundred jobs is meaningless against a workforce of two million. What materialises is not prosperity. It is a very large, very thirsty tenant on a grid that working people and decarbonising industries also need.

Most of the Queue Is Not Even Real

Here is the detail that should make everyone pause, because it is the whole argument in a single statistic. The data centre industry’s own analysis estimates the sector’s actual need for power through 2030 at around 1.2 GW. Its applications in the queue total 14–17 GW.

In other words, more than ninety percent of the demand choking the Danish grid is speculative. Developers buy land and file connection applications for data centres they may never build, just to hold a place in line. The industry itself admits these are immature projects that should be removed. We have frozen a national grid, stalled the electrification of real industry, and triggered a parliamentary fight over rationing, largely over capacity that nobody has shown we need.

The Financial Case Is No Better

If the climate argument does not move you, the money should give you pause too, because the economics are shakier than the marketing suggests. Globally, the capital pouring into AI infrastructure is approaching a trillion dollars, and the returns are stubbornly hard to find.

Goldman Sachs put the question bluntly in a 2024 report titled Gen AI: too much spend, too little benefit?, asking whether the roughly trillion dollars headed into AI build-out will ever pay off, and noting how little there is to show for it so far. Sequoia’s David Cahn framed the same gap from the other direction: the build-out implies the industry needs on the order of $600 billion a year in new revenue just to justify the hardware being bought, and nothing close to that is materialising. And when MIT’s NANDA initiative actually measured enterprise results in 2025, it found that despite tens of billions in spending, 95% of generative AI pilots delivered no measurable business return at all.

So this is not the familiar trade-off where we accept an environmental cost in exchange for a clear economic prize. We are spending the carbon, mortgaging the grid, and the financial payoff is unproven. Build first and find the uses later might be acceptable if the people placing the bet were the ones paying for it. They are not. The costs land on the grid, the climate, and the public; the upside, if it ever arrives, accrues to a handful of companies.

It is tempting to dismiss all of this one query at a time. A single AI prompt really is trivial; on its own it costs almost nothing. That is precisely the trap. The cost was never in the one query. It is in the scale we are building to serve billions of them: the training runs, the data centres, the grid connections. Which is to say, it is the Danish queue. The individual request is invisible. The infrastructure we are pouring to serve it is not.

So: Do We Actually Need It?

I am not arguing that AI is useless. I use it myself, most days, for research: not to solve the problem but to point me toward where the answer might be, the way a map suggests a route. It does not navigate. I do. It is not always right and it is never the last word, but as a way to navigate more information than one person can hold, it earns its place. This brief is a case in point. I leaned on AI to help chase down sources, sharpen the structure, and test which arguments held and which did not; the reporting, the judgement, and every final word are mine, and the model assisted rather than authored. That is the test I want applied everywhere, not can we use AI here but does it earn its place here, and much of what we are building fails it. We are committing carbon, grid capacity, and vast sums of money at enormous scale, in advance of any clear account of what it is all for, and in many cases in place of things we know we need. That is the shape of a lose-lose, and Denmark is simply where it became too physical to ignore.

So let this be an invitation rather than a verdict. Before the next data centre is approved, before the next billions are committed, the questions worth asking out loud are the simple ones we have been skipping. What is this AI actually for? Who benefits, and who pays? What would we give up to build it, and is the thing we get back worth more than the factory that could have electrified, or the money that could have gone somewhere proven?

We do not need to answer those questions perfectly. We just need to start asking them before we build, instead of after. The Danish grid is already full. The atmosphere is already responding. The least we can do is decide, deliberately and together, whether we want the thing we are spending so much to get.