Alexey Gubarev on the Race to Build AI Infrastructure

Big Tech will spend more than $700 billion on AI infrastructure this year. Yet only about a third of the US data center capacity due in 2026 is actually being built, according to Bloomberg.

The numbers are extraordinary by any standard. Amazon, Alphabet, Microsoft and Meta are on course to spend roughly $730 billion on capital expenditure in 2026, about three-quarters more than in 2025, and most of it goes into data centers, chips and power. Europe is answering with a mix of private capital and public ambition. The EU opened a call in July for up to seven AI gigafactories, after an earlier round in which bidders pointed to more than €230 billion of potential investment across 60 sites in 16 member states. That is a genuine opportunity. Whoever builds this capacity will host the next generation of software, medicine and industrial automation. But it is also a risk, because capital is now moving faster than the physical world can absorb it.

I have spent twenty years on the side of this industry that signs for transformers and waits for grid connections, first with a hosting business I started in 2005 and then with Servers.com, which we built into a global bare-metal platform before selling it in 2023. In that time I do not remember a single site that was late because of the servers. It was always a utility, a permit or a piece of electrical equipment. That is the lens I bring to the numbers below.

The physical world is sending clear signals. A large power transformer now takes around 128 weeks to arrive, compared with 30 to 60 weeks before the pandemic. In Europe's five core data center hubs, only 6.4% of capacity is free. The danger here is not an AI bubble in the usual sense. It is stranded money: billions committed to buildings that wait years for equipment and grid connections, and electricity bills that turn households against the industry. The opportunity belongs to whoever solves the unglamorous side of the problem. That means the builders, the grid operators and the countries that can shorten the distance between a signed check and an energized rack.

Four kinds of builders

The first builders are the hyperscalers. For years they funded their build-out comfortably from cash flow, and that era is ending. Alphabet reported negative free cash flow in the second quarter, its first such quarter since its 2004 listing, and the bond market is now part of the AI infrastructure story. When the richest companies on earth start borrowing to build, it shows how far demand has run ahead of available capacity.

The second group is the neoclouds: GPU-as-a-service providers that rent specialized capacity to model developers and enterprises. Their AI capacity signings in Europe nearly tripled last year. They move fast, but they depend on someone else's buildings and someone else's grid connection.

The third is the state, and Europe's gigafactories are the clearest example. Brussels will fund up to 17% of the computing infrastructure, member states must at least match that share, and private consortia supply the rest. The structure is sensible, with public money reducing risk and private money leading. But it runs on a public-procurement clock in an industry that measures itself in quarters.

The fourth builder is the least visible: real estate and infrastructure capital. These are the developers and funds that buy land, secure power and build shells long before a tenant signs. In practice, they decide where AI will physically live. I know this side best, and its logic is unforgiving: a shell with power is worth several times a shell without it, and the gap between the two is measured in years, not dollars.

America's bottleneck: equipment and the electricity bill

In the United States, money is abundant. The constraint is increasingly physical, and increasingly political.

Roughly four out of five large power transformers used in the US are imported, and switchgear and batteries are in similarly short supply. Of the roughly 12 gigawatts of US data center capacity expected online in 2026, only about a third was under active construction at last count.

The second American constraint is who pays. PJM is the grid operator for 13 states and Washington, D.C., including Northern Virginia's data center corridor. In PJM, capacity prices have jumped from under $29 per megawatt-day to the regulatory cap for three consecutive auctions. The latest auction still fell about 6.8 gigawatts short of what the grid needed. The market monitor attributes about 38% of that auction's $16.4 billion cost to data centers.

Once household bills start carrying the cost of AI, infrastructure stops being an engineering question and becomes an electoral one. Expect more state-level conditions on large new loads, not fewer.

Europe's bottleneck: time, space and scale

Europe's problem is different. Capital is available. What is scarce is connected, permitted land near the places people want to build.

In Frankfurt, vacancy has fallen to just over 3%. The industry is responding by moving outward. Hyperscale campuses planned for 2026 to 2028 sit on average about 175 kilometers from the major hub cities, compared with roughly 46 kilometers for campuses delivered in the previous four years. In the second quarter, less than two-fifths of Europe's new capacity landed in the five traditional hubs of Frankfurt, London, Amsterdam, Paris and Dublin.

That dispersal is healthy, not a defeat. It is how Paris pulled ahead this year, on the back of France's nuclear fleet and grid-ready sites. It is also why Ireland's new rules are pushing developers away from the Dublin area toward sites that can bring their own generation.

But dispersal also exposes Europe's structural weakness, which is fragmentation. The continent has twenty-seven permitting regimes, dozens of grid operators and national energy policies that rarely line up. Together they make it hard to assemble gigawatt-scale projects on the timelines the market demands. The US has a cost problem. Europe has a clock problem.

What separates the builders who deliver

Having watched this cycle from the inside, across several countries and grids, I would watch three things.

The critical path has moved from the chip to the switchgear. A few years ago the scarce item was the GPU. Today it is the electrical equipment that feeds it. The teams that reserve transformer slots and grid capacity before they need them will deploy in 2027. Everyone else will be looking at 2029.

Inference redraws the map. Training clusters favor a handful of very large sites with enormous power. Inference, meaning the everyday running of AI models for users, is expected to overtake training by late 2026. It is more sensitive to latency and more widely distributed. That spreads demand toward secondary markets, including southern Europe and the Eastern Mediterranean. Smaller locations will not host the largest training runs, but they can win a meaningful share of the inference layer if they can offer power and permits quickly.

Social license is now infrastructure. The PJM numbers warn every market. Operators that pay their own way on grid upgrades, bring new generation with them, and can show local communities a tangible benefit will get connected faster. Operators that cannot will meet moratoriums.

The contest is usually framed as America against Europe. I think the more useful frame is capital against time. The US has the capital and the speed, and it is now running into equipment and political limits. Europe has the policy intent and a widening map of new sites, but it needs to shorten its clock. I have seen what happens on both sides when a check is signed and the rack is not energized for three years: the money is not lost, but the market it was meant for has moved on. Whoever closes that gap will define the next phase of AI.

Alexey Gubarev (also spelled Aleksey Gubarev) is a Limassol-based technology entrepreneur and investor. He has worked in the hosting and infrastructure business since 2005 and co-founded Servers.com in 2013; the company was sold to CloudOne Digital in July 2023. He is one of the founders of TechIsland, the Cyprus technology association, and serves on its Board of Directors. He is also building a classic car museum, scheduled to open in 2027. More at gubarev.com.