I am an optimist about artificial intelligence. Most of what follows may seem otherwise because of the political weather turning against it, but I am still optimistic about AI, and all the opportunity that comes with it. Including the future of data centres.

Right now, the AI arena is becoming toxic, with ‘AI doomers’ making inroads and influencing development. A year ago, these same people were writing long essays nobody in a marginal seat had read. Now regulation is on the agenda, and politicians shifted their attention to the issue, with a distinct voter-filter. As a mathematician I notice the shape of it more than the content. The anxiety about a technology has found itself a physical object to point at: the data centre.

This matters more than it looks. An abstract fear is hard to legislate and harder to campaign on, but a data centre is concrete. It sits in a real place, draws a city’s worth of power, drinks water from the local supply, and hums. You can drive past it. You can photograph it. You can put it in a press release. A backlash needs something it can see, and right now it can see the data centre.

The optimist’s question is not “how do we win the argument?” The argument is mostly lost the moment it becomes a fight about a building. The optimist’s question is: what if the focal point of the backlash is about to disappear?

Two podcasts, one footpath

I do my thinking on walks, with a podcast in my ears, as readers of The Walking Investor will know by now. On my recent walk, two episodes landed back-to-back, and merged into a single thought by the time I reached home.

The first was Invest Like the Best with the founders of Etched. The second was 20VC with KR Sridhar of Bloom Energy. One is a chip company, one is a power company, and neither is in the business of fixing AI’s image problem. But together, they address two of the big issues the backlash is pointing at: the power the data centre draws, and the heat it throws off. Collectively that is far more interesting than either company on its own.

The Etched story is the kind that makes you feel old and hopeful at once. Two Harvard dropouts who, back in 2023, went around with a thirty-page memo arguing that AI would eventually need chips built for one job rather than general-purpose GPUs, and could barely get anyone to take the meeting. It turns out the memo was right. The bet was to hardwire the transformer (the architecture underneath ChatGPT, Claude, Gemini) directly into the silicon and throw away the flexibility that lets a normal chip do everything else. While other chips struggle to accommodate every other AI architecture, their chip, Sohu, can do precisely one thing. And the claim is that it does that one thing at roughly ten times the throughput of a top Nvidia chip for the matrix maths a transformer runs on. The downside is as blunt as the upside: if the industry abandons transformers, Etched is, in the founders’ own word, toast.

It is an all-or-nothing bet, and the market has decided it likes the odds. The company has raised $800 million and is valued at $5 billion, pre-meaningful-revenue, with a cap table that runs from Peter Thiel to half the people who invented the field. I am not here to tell you whether that is a good price. I am here because of what it does to the building.

The efficiency lever

Inference is now the biggest cost centre in AI. It is the biggest single draw on a data centre’s power budget, and the number Etched is attacking. Simply put, it’s the work a model does every time you use it, as opposed to the training that happens once. And it is exactly the kind of supply-side shock I wrote about in The Multiplex Moment; the breakthrough that reprices the infrastructure layer underneath a boom. In the late 1990s it was DWDM making fibre abundant overnight. The candidate this time is an order-of-magnitude gain in inference efficiency.

If a rack of purpose-built chips does the work of many racks of general-purpose ones, the next data centre will be a smaller, cooler, and less thirsty building than the one the AI doomers are pointing at. Etched may win as a standalone. More likely, given how these things go, it gets bought by a hyperscaler or a chip company, and there are fast followers within eighteen months. Either way the direction is the same.

The power lever

Now for part two. Where Etched shrinks the demand, Bloom changes where the supply comes from.

Bloom Energy builds solid-oxide fuel cells that make electricity through an electrochemical reaction rather than by burning something to spin a turbine. The cells, called the “Bloom Box”, are quiet, compact, and produce power on-site. Which is why data-centre operators facing overloaded grids with years before upgrades, have instead bolted Bloom microgrids onto their facilities. Sridhar’s framing on the podcast was that electricity, not the models, will decide who wins the AI race. The proof of the point is speed: Bloom powered Oracle’s AI data centres in 55 days, better than any timeline a grid connection can offer.

The market noticed. Bloom’s shares rose more than 1,400% in a year, from an overlooked fuel-cell company to a data-centre power play worth around $90 billion US dollars. It’s a market capitalisation that, translated into ours, would put it comfortably inside the top five listed companies in Australia. Yet even with that growth, the same machine that is a near-certainty in one country remains a hard sell in another. Three of the four complaints associated with data centres can be softened considerably when the power is generated quietly, on-site, next to the load. It takes out the strain on the public grid, the noise of diesel backup, and the visible industrial bulk. But whether that’s enough to win depends entirely on what the country in question is optimising for.

In the United States, Bloom is close to a no-brainer. Energy security, reliability, competitive power prices, and geopolitical advantage sit above climate in the running order of national priorities, so a machine that delivers firm, fast, affordable on-site power wins even though it still burns gas. You can watch the calculus play out in a single project: Bloom’s New Mexico campus with Oracle is slated to run on up to 2.45 gigawatts of fuel cells in place of the planned gas turbines and diesel generators. It’s cleaner and faster than the alternative it displaces, and nobody in the approval chain is troubled that the alternative wasn’t zero.

In Australia, the running order is reversed and subsequently, so is the answer. In the current Australian political environment, a technology that runs on burning more gas is not going to be waved through, no matter how clean, quiet, and quick it may be. Which is why the flexibility in the fuel is the whole point. The cells run on natural gas, biogas, or hydrogen more or less interchangeably; the same box that runs on gas today runs on hydrogen tomorrow. Gas gets it built, while hydrogen gets it permitted. In a country with our gas reserves, a machine that can start on one fuel and migrate to the other is better suited to the actual conditions than one locked to either.

The tension is not unique to fuel cells. There are similarities with the micro-nuclear thesis I set out in The Reactor in the Shipping Container, where we ban the cleanest baseload option going while sitting on the world’s largest uranium reserves. We are not short of energy resources in this country. We are short of permission to use them. Different machine, same insight: the future of industrial power may be modular, local, and sited next to the thing that needs it, but here it also must be the politically survivable fuel, and that narrows the field.

Pull the two levers together and something changes in the shape of the thing. If the compute inside shrinks and the power no longer comes from a substation three suburbs away, the data centre is no longer a single enormous shed on the edge of town. It can be smaller, and there can be more of them, tucked wherever the demand is. The mental picture of “a data centre” as the windowless hyperscale box behind a security fence may turn out to be a snapshot of one awkward moment in the technology, before reaching its permanent form.

Which is where the data centre stops being only a data-centre story.

The catalyst hiding in the argument

I have argued before, in “Don’t be idealistic”, that Australia’s mineral-processing crisis will not be reversed by bailouts. It needs a circuit-breaker, and the most plausible one is cheap, reliable, on-site power. But power was only ever one input to viability, and I should be careful not to overclaim it. The thing that most often stops a processing plant here isn’t the carbon in its electricity. It’s the toxic by-product of the process itself: the tailings, the effluent, the heavy-metal streams that have to go somewhere. Cheap, clean power does nothing for that. A hydrogen-fed smelter produces the same waste as a gas-fed one, and the same environmental-approval regime is waiting for both. So, the honest claim is a modest one. On-site modular power removes one obstacle, that being the cost and the emissions of the electricity, from a stack of obstacles, and leaves the hardest one, the by-product, exactly where it was. It doesn’t make the equation easy. It makes one line of it a little more possible.

The data centre backlash and the smelter closures rhyme, but they aren’t the same song. The data centre’s problem really is mostly power — draw, source, siting — and modular power largely answers it. The smelter’s problem is only partly power; the rest is what comes out the other end, and that needs a different fix. What they share is the shape: a viability equation that better infrastructure can shift, even when it can’t solve.

The deal is local

Here is the part an angel should sit with. You could read all the above and conclude the action is in California. That backing this thesis means wiring money to a $5 billion chip company you will never get an allocation in, or a $90 billion listed one you can buy but can hardly call angel investing. That is the wrong lesson, and it is the lesson I keep trying to talk people out of.

The fourth complaint against the data centre is heat, and the water used to manage it. And one of the more interesting answers to that problem is a company we backed. Enaxiom turns data-centre cooling on its head, cutting the energy the cooling consumes, and producing ultra-pure water as a by-product rather than drinking it. Energy, power, heat, water: those are the four things the backlash points at, and the third and fourth of them are being approached head-on by a portfolio company in our own ecosystem. We did not have to go to Sand Hill Road to find it.

Enaxiom will not be the last. There are tens, probably hundreds, of companies currently chiselling away at one corner or another of the data-centre problem. Each of them addressing all those issues of cooling, power, water reuse, waste heat, chip efficiency, load management. Most of them you have never heard of, and some of them will cross our desk, because a problem this large and this politically charged pulls founders toward it from every direction. When a category is being repriced, the interesting companies do not all appear in one postcode. A meaningful number appear in ours.

That is the angel lens on all of this. The global majors are repricing the chip and the power source. You are unlikely to get into Etched, and buying Bloom after a fourteen-fold year is a different sport altogether. The same wave that produces them produces local companies attacking the same problem from the same logic. Those are the ones you can back early, with small cheques (and many of them) into the founders who are building for the world created by the repricing rather than the one it leaves behind.

The optimist’s bet

I’m going to flag this as a bet rather than a forecast, because the line between the two is where people get into trouble. I think that in a few years the data centre will stop generating so much political noise, and not because the politics calms down but because the political focal point changes underneath it. The efficiency lever shrinks what must be built; the power lever cuts the cord to the grid. Between them they point at a data centre that is smaller yet still able to do all the things: more distributed, self-powered, cooler, drawing less because the chips inside it are doing more per watt, producing water instead of consuming it. Not one monstrous shed the whole town organises against, but a scatter of small, dull buildings you walk past every day without blinking. The way you already walk past the substation and the exchange without once thinking about the grid or the phone network they quietly run.

Yes, there will still be a handful of vast training centres but those largely already exist. They are built, sited, and through their approvals. The backlash isn’t going to un-build them. The fight is over the next wave of construction, the hundreds of facilities planned to serve inference as the models reach everyone, and that is exactly the wave that shrinks and scatters. The thing the politics is mobilising against is the thing the engineering is about to make small.

AI doomers will then move on to the next visible object. They always do. The optimist’s job, and the angel’s, is to be already invested in the thing that makes the current fear obsolete. While everyone else is watching California, we will notice that a good deal is being built within driving distance.

Richard Moore — MooCoo Ventures

Richard Moore is co-founder of MooCoo Ventures, an angel syndicate that co-invests alongside Brisbane Angels, one of Australia’s most active angel groups. He has made over ninety personal angel investments since 2013.

moocoo.vc