The Race for Compute

What do land, power grids, and water systems have to do with AI anyway?

I’ve been wanting to write this piece for a few weeks now.

Compute keeps showing up at the edge of every conversation I’m in. Fundraising conversations. Sovereign wealth conversations. A conversation last week with a friend watching his county approve a data center rezoning that almost nobody in the room seemed equipped to underwrite. The story is everywhere, yet few people are fully connecting the dots. I wanted to wait until I had a deep enough grasp of it to actually link it to the broader thesis I’m working through on capital flows, incentives, and where the leverage is moving over the next decade. This is that piece.

If you’ve driven across central Virginia, the outskirts of Phoenix, parts of rural Mississippi, or pretty much anywhere along the corridor from Columbus, Ohio to Dallas in the last two years, you’ve probably passed one. A long, low, windowless building the size of a Walmart. Razor wire. Unmarked doors. A handful of cars in the parking lot. Sometimes a corporate logo on the side, more often nothing.

That building is a data center. There are hundreds of them being built right now in places that didn’t used to be on anyone’s map. Almost every piece of AI you’ve used in the last twelve months — every ChatGPT prompt, every AI-generated image, every “summarize this email” button in your inbox, ran inside one of them.

Most of the coverage you’ve read about AI assumes you already know what’s inside, who’s competing, and why it matters. That’s part of why the story isn’t landing for most people. So let me reframe the whole thing in one sentence.

There are two races happening here, not one. The first is the one everyone sees: NVIDIA, OpenAI, Microsoft, Google, Meta, Amazon. The second is much quieter: counties, utilities, water authorities, landowners, transmission operators, sovereign wealth funds. The first table gets the headlines. The second table may decide who actually wins. That’s what this piece is about.

Compute is the raw processing power that AI runs on. Every time an AI model “thinks,” somewhere in the world a specialized chip is performing trillions of mathematical operations per second to produce an answer. The chips are called GPUs, graphics processing units. The dominant supplier is a company called NVIDIA, which is now one of the most valuable companies in human history specifically because everyone needs its chips and nobody else can make them as well. Compute isn’t only chips, though. It’s the chips, plus the electricity to run them, plus the cooling to keep them from melting, plus the fiber-optic cables that connect them to the rest of the internet, plus the physical buildings to hold all of it. When you read about a “race for compute,” what’s actually being raced is the assembly of all those pieces at industrial scale.

The data centers going up everywhere are physical containers for compute. Each one holds tens of thousands of GPUs running twenty-four hours a day, drawing enormous amounts of electricity (a single hyperscale campus can pull 500 megawatts — enough to power half a million homes), generating enormous amounts of heat, and needing constant water-based cooling to stay operational. These are the new factories. Instead of producing cars or steel, they produce intelligence. Answers, predictions, classifications, generated text and images that gets sold to billions of users and rented by every other industry now rebuilding itself on top of AI.

The reason these factories keep showing up in rural counties is mostly economic. Industrial power rates outside cities are a fraction of what residential customers pay, and rural counties sit closer to the major transmission lines and substations that can deliver hundreds of megawatts without a multi-year grid upgrade. Land is a hundred times cheaper. And the local politics are simpler. A county commission that meets twice a month with three full-time staff negotiates differently against a multi-trillion-dollar company than a major city would. The hyperscaler arrives offering jobs and tax revenue. The county, often economically anxious, often without the expertise to model the trade, says yes.

This is why permission is becoming the new GPU. The chip matters, but the chip is useless without somewhere to plug it in. The cycle will not be won by whoever buys the most silicon. It will be won by whoever can assemble power, water, land, and political consent before everyone else realizes those were the scarce assets all along.

Who is actually racing? The big four American hyperscalers: Microsoft, Google, Amazon, and Meta, plus Oracle, plus a smaller club of AI-native companies (OpenAI, Anthropic, xAI) that rent compute from the hyperscalers at scale. Across the Pacific, China is running a parallel race with Alibaba, Tencent, Baidu, and ByteDance, and is racing to build a domestic chip industry because it can no longer depend on U.S. exports. Underneath all of this, the Gulf sovereigns, Saudi Arabia’s PIF, the UAE’s G42 and Mubadala, Qatar’s QIA, are deploying tens of billions to position the Middle East as a major host of global compute, with cheap power, desalinated water, and political certainty that no U.S. county can match.

Why are they competing so violently? Because compute is becoming a winner-take-most market the way cloud computing did fifteen years ago. The companies that lock up the most chips, the most power, and the most physical capacity get to train the largest models. The largest models get to power the AI products that everyone else has to buy or rent. The losers don’t get “second place.” They get pushed down the supply chain where margins are thin and the winners set the rules. The fear in these boardrooms is not that they spend too much. The fear is that they spend too little and another company captures the rails of the next economy first.

The number that gives the race its shape: the big four American hyperscalers will spend roughly $725 billion on capex in 2026, up about 77% from last year’s record. That is more than the U.S. defense procurement budget.

Of that $725 billion, roughly 60% is going to power, cooling, land, and supporting infrastructure, not chips. The bottleneck on AI in 2026 is not silicon. It is electricity. NVIDIA can make more chips. Nobody can make more megawatts on short notice. The U.S. grid is being asked to absorb a doubling of total data center power demand in three years, from roughly 80 gigawatts in 2025 to a projected 150 gigawatts by 2028. That’s a generation of new utility infrastructure compressed into 36 months. It’s why hyperscalers are buying up gas peaker plants, restarting old coal plants, signing decade-long deals with nuclear operators, and optioning small modular reactor sites that won’t come online for years.

This is where the second table starts to come into view. The chips are the visible asset. The power, water, land, and permitting are the binding constraints. Whoever controls the binding constraints controls the rest of the cycle.

How it shows up in your life depends on where you live. If you’re in a region with heavy data center buildout, your residential electricity rates are likely subsidizing the industrial rate the hyperscaler locked in. Some cluster regions have seen power rates climb sharply over the last five years. Your water utility is competing with cooling demand. Your local property tax base has a giant hole in it where the data center should be paying. Virginia alone is now giving away roughly $1.6 billion a year in data center tax abatements. If you live nowhere near one, your federal taxes are still funding the chip supply chain through the CHIPS Act and the grid buildout. The local trade tends to look something like this: many construction workers come in for two years and then leave, the campus stays and provides a small number of permanent technical and security jobs, and the tax base gets locked away for fifteen to thirty years. The math is rarely as good for the host community as the press release suggests.

The AI you actually use is, for now, being sold to you at a heavy loss. Every prompt you type into ChatGPT, Claude, or Gemini is being subsidized by venture capital, by hyperscaler partnerships, by a willingness to lose money to capture users, the same playbook Uber ran with $4 rides a decade ago. Estimates vary, but the direction is clear: every consumer-facing AI interaction has a real physical cost in electricity, cooling, chips, and infrastructure, and that cost is currently outrunning what users pay. The cost of running these models has fallen dramatically as chips and software have gotten more efficient. Whether that efficiency outpaces demand will determine whether your AI bill goes up, stays flat, or drops.

What gets disrupted by all of this is everything that’s been treated as sleepy and regulated for thirty years. Electric utilities. Water authorities. Transmission operators. Transformer manufacturers (there is a global shortage). Nuclear plant operators. Real estate near substations. Municipal bond markets, because the public infrastructure being built to support these campuses is being financed against tax bases the campuses have negotiated out of. What gets advanced, possibly, is the first significant expansion of American grid infrastructure in a generation. Though whether that benefits the public or strands them with the bill is the question now being fought out in state legislatures.

Who’s actually winning? Today, NVIDIA and the hyperscalers. Underneath them, the Gulf sovereigns are positioning to be long-term winners on the host side. The UAE-US AI Campus unveiled last May spans 10 square miles and 5 gigawatts of capacity — the largest AI infrastructure project outside the United States. Saudi Arabia’s Humain is building a parallel structure backed by roughly $100 billion in committed sovereign capital. Middle East SWFs collectively manage close to $6 trillion, and they treat compute the way they once treated oil as the next strategic resource to host, refine, and export.

The quieter winners are upstream of the data center itself. Power generation. Water rights. Transmission corridor land. Transformer supply. Permitting expertise. Behind-the-meter generation. Nuclear restarts. The hyperscalers know this, which is why 60% of their capex is moving that direction.

Why does it actually matter? Because compute is becoming the strategic resource of the next economy, but the leverage is not only in owning it. The leverage is in owning what it cannot exist without. Power. Water. Land. Permission. Whoever owns those owns the terms on which the AI economy operates. Nations are figuring this out. Sovereigns are figuring this out. Most retail investors and most local governments have not yet.

The first race, the one everyone is watching, is real. It’s just not the whole race. The second race is happening in county supervisor meetings, utility rate hearings, water authority filings, and sovereign wealth offices in Riyadh and Abu Dhabi.

Most people are reading about the companies at the first table.

The next economy may be decided by the people sitting quietly at the second.

Further Reading: How to Track the Second Table

A lot of what’s in this essay doesn’t show up in mainstream coverage until it’s already priced in. If you want to track this in real time, here is the short list of sources I lean on. Some are paid, most are not. Together they give you the picture mainstream business news might sometimes miss.

SemiAnalysis (Dylan Patel) — the single best public resource for understanding hyperscaler capex breakdowns, chip economics, and where the bottlenecks actually sit. Worth the paid tier.

Data Center Dynamics — industry trade publication. Granular, unfiltered, and tracks the things general business press misses (interconnect queues, transformer shortages, individual campus permits).

Stratechery (Ben Thompson) — best strategic frame on the hyperscalers as platforms. Paid, dense, but consistently sharp.

Heatmap News — the cleanest reporting on the intersection of AI, the grid, and U.S. energy policy. Free, lightly editorialized, very current.

FWDStart — three editions a week on MENA tech and venture. The Gulf sovereign positioning around compute and AI is far better covered here than in U.S. publications. Jamie and team are on the ground in the region and writing about real updates in real-time.

Hyperscaler 10-Ks and earnings call transcripts — most underrated source on this list. Microsoft, Amazon, Google, and Meta are now legally required to tell you what they’re spending and where. Reading the actual filings (especially the capex commentary and depreciation footnotes) tells you more than any analyst report.

State utility commission filings — the boring, public, free version of the second table. PJM, ERCOT, MISO. The data center fights show up here first, months before they hit the news.

If you read three of these consistently for six months, you will know more about what’s actually happening with AI infrastructure than 95% of the people who post about AI on LinkedIn. That asymmetry is the entire point.