Why Data Centers Are Becoming the New Oil Refineries

For decades, data centers were the most boring buildings in technology. They were essentially warehouses full of blinking servers, managed by facilities teams that most executives never met. They were a cost center — something you outsourced, commoditized, and forgot about. Then artificial intelligence happened, and data centers became the most strategically important real estate on the planet.

The analogy is not perfect, but it is powerful: oil refineries transform crude oil — a raw, largely useless substance — into gasoline, diesel, jet fuel, and petrochemicals that power the modern economy. Data centers transform electricity — a raw energy input — into intelligence: trained models, inference results, and AI-powered services that are rapidly becoming the substrate of every industry. The parallels run deeper than most people realize, and understanding them is essential for anyone trying to navigate the AI economy.

On TBPN, John Coogan and Jordi Hays have been tracking this transformation in real time — from the first hyperscaler GPU orders to the nuclear power deals that would have seemed absurd three years ago. This post lays out the full analogy, explains why it matters, and identifies what it means for founders, investors, and the tech industry at large.

The Core Analogy: From Commodity to Strategic Asset

Consider what oil refineries actually do. They take a raw material — crude oil — and through an energy-intensive, capital-heavy industrial process, produce outputs that are far more valuable than the input. The refining process requires specialized equipment, deep expertise, enormous capital investment, and continuous operation to be economical. You cannot build a refinery quickly. You cannot build one anywhere. And the companies that control refining capacity have outsized influence on the global economy.

Now substitute the terms. Data centers take a raw input — electricity — and through a compute-intensive, capital-heavy process, produce outputs — trained AI models and inference results — that are far more valuable than the electricity consumed. The process requires specialized equipment (GPUs, networking, cooling), deep expertise, enormous capital, and continuous operation. You cannot build an AI-scale data center quickly. You cannot build one anywhere. And the companies that control compute capacity are gaining outsized influence.

Where the Analogy Holds Tight

• Capital intensity: A modern oil refinery costs $5 billion to $15 billion to build. A hyperscale AI data center campus is now in the same range. Both require years of planning and construction.
• Location dependency: Refineries locate near crude oil sources, waterways for shipping, and pipelines. Data centers locate near cheap power, fiber routes, and water for cooling. In both cases, the wrong location is a multi-billion-dollar mistake.
• Operational complexity: Both are 24/7 operations with zero tolerance for unplanned downtime. Both require specialized operators and continuous maintenance.
• Strategic and national security importance: Governments treat refining capacity as strategic infrastructure. The same is now happening with AI compute — export controls on GPUs, discussions of "compute sovereignty," and government incentives for domestic data center construction.
• Dominated by a few large players: Five companies control most global refining capacity. Similarly, a small number of hyperscalers (AWS, Azure, GCP) and GPU cloud providers (CoreWeave, Lambda, Oracle) control most AI compute capacity.

The Three-Year Transformation

It is worth pausing to appreciate how fast this shift happened. In 2022, data centers were widely considered a mature, boring sector of the tech economy. Real estate investment trusts (REITs) like Equinix and Digital Realty traded at modest multiples. Colocation was a commodity business with thin margins. The primary customers were enterprises hosting web applications and databases.

Then, in rapid succession:

1. ChatGPT launched (November 2022), demonstrating that AI had crossed a capability threshold that demanded massive compute.
2. Every major tech company announced multi-billion-dollar AI infrastructure investments within months.
3. GPU shortages made compute access — not just model quality — a competitive differentiator.
4. Power constraints emerged as the binding constraint. You could order GPUs (eventually), but you could not get the electricity to run them.
5. Data center valuations exploded. Equinix, Digital Realty, and especially newer entrants like CoreWeave saw their valuations multiply.

In less than three years, data centers went from boring real estate to the most sought-after infrastructure asset class in the world. As TBPN has covered extensively, this was not a bubble — it was a structural repricing of an asset class based on a genuine shift in demand.

Location Decisions: The New Geopolitics of Compute

Where you build a data center is now a decision with geopolitical implications. The factors driving location selection mirror those of industrial energy infrastructure:

Power Cost

The target is 2 to 4 cents per kilowatt-hour for wholesale electricity. At AI scale (hundreds of megawatts), even a 1-cent difference translates to tens of millions of dollars annually. This pushes development toward regions with cheap power: the US Southeast (nuclear and gas), Scandinavia (hydro), Quebec (hydro), and parts of the Middle East (gas). Locations with expensive power — like California or the UK — are disadvantaged unless they offer other compensating benefits.

Water Availability

Cooling systems — especially evaporative cooling towers — consume enormous quantities of water. A 100 MW data center using evaporative cooling can consume 300,000 to 500,000 gallons per day. In water-stressed regions like the American Southwest, this creates political opposition and regulatory risk. Some operators are moving to dry cooling or closed-loop liquid cooling to reduce water dependency, accepting higher energy costs in exchange for reduced water risk.

Fiber Connectivity

Data centers must be connected to the internet backbone and to each other via high-capacity fiber. Locations that lack existing fiber infrastructure require expensive new builds — $50,000 to $100,000 per mile for buried fiber. This favors established data center markets (Northern Virginia, Dallas, Chicago, Phoenix, Amsterdam, Frankfurt) where dense fiber networks already exist.

Local Regulation and Permitting

Building a large data center requires zoning approval, environmental review, building permits, and utility interconnection agreements. Some jurisdictions are actively courting data center development with streamlined permitting. Others — particularly in Northern Virginia, which hosts the densest concentration of data centers in the world — are pushing back against further development due to noise, water, and power grid concerns.

Tax Incentives

States and countries are competing aggressively to attract data center investment through tax incentives. Sales tax exemptions on equipment, property tax abatements, and electricity tax reductions can save hundreds of millions of dollars over a project's lifetime. Texas, Virginia, Ohio, and Indiana have been particularly aggressive in the US. Internationally, the Nordic countries, Singapore, and several Middle Eastern states offer significant incentives.

The 15-Year Power Purchase Agreement

One of the most striking developments in the AI infrastructure buildout is the emergence of long-term power purchase agreements (PPAs). AI companies are signing contracts to buy electricity for 10 to 15 years — commitments that would have been unthinkable for a tech company a few years ago.

Why? Because AI workloads are fundamentally different from traditional tech workloads:

• Continuous operation: Training runs last weeks or months, running GPUs at full power 24/7. There is no "off-peak" for AI training.
• Predictable demand: Unlike web traffic, which is spiky, AI training demand is flat and predictable — perfect for baseload power contracts.
• Scale: A single large training run can consume 50 to 100 MW for months. At that scale, wholesale power procurement through PPAs is dramatically cheaper than retail electricity.
• Carbon commitments: Many tech companies have net-zero pledges. PPAs for renewable or nuclear energy help meet these commitments while securing cheap power.

Microsoft's deal to restart the Three Mile Island nuclear plant, Amazon's investment in nuclear through Energy Northwest, and Google's geothermal PPA with Fervo Energy are all examples of this trend. These are not PR exercises — they are hardheaded economic decisions driven by the insatiable power appetite of AI.

The Hyperscaler Land Grab

The largest cloud providers are engaged in a land grab for data center capacity that has no precedent in the technology industry. The numbers are staggering:

• Microsoft announced over $80 billion in data center capital expenditure for fiscal year 2025, with plans to accelerate further.
• Amazon (AWS) committed $100+ billion in infrastructure spending over a multi-year period.
• Google is investing heavily in both owned facilities and third-party colocation capacity.
• Meta is building massive AI training campuses, including a 2 GW campus in Louisiana.
• Oracle is aggressively expanding its cloud infrastructure, particularly for AI workloads.

This land grab extends beyond building — it includes acquiring development sites, securing power contracts, and in some cases buying entire power generation assets. The total infrastructure investment by major tech companies is projected to exceed $500 billion between 2024 and 2028.

The Investment Thesis: Picks and Shovels, Upgraded

For investors, the oil refinery analogy points to a specific thesis: the companies that build, supply, and operate AI infrastructure are the durable beneficiaries of the AI boom, regardless of which model or application ultimately wins.

Data Center REITs

Equinix, Digital Realty, and QTS Realty (now part of Blackstone) own and operate large portfolios of data center facilities. As AI demand has surged, these companies have seen vacancy rates drop to near zero and rental rates increase substantially. New development is being pre-leased — often entirely — before construction is complete.

Power and Cooling Infrastructure

Companies like Vertiv (power distribution, cooling), Eaton (power management), and Schneider Electric (integrated data center infrastructure) are direct beneficiaries of data center buildout. Cooling companies — particularly those offering liquid cooling solutions — face demand that vastly exceeds current supply.

Construction and Engineering

Data center construction is a specialized discipline. Companies like Holder Construction, DPR Construction, and Rosendin Electric have deep expertise in building high-power-density facilities. The construction pipeline is measured in years, providing long-duration revenue visibility.

Networking Equipment

Arista Networks, Broadcom, and Ciena supply the switches, optics, and routers that connect data center infrastructure. As AI clusters grow, networking becomes a larger share of total infrastructure spend.

What This Means for Founders

If you are building a startup in 2026, the data-center-as-refinery framing has practical implications:

1. Infrastructure access is a moat. If your product depends on large-scale AI compute, your ability to access that compute reliably and affordably is as important as your technology. Build relationships with infrastructure providers early.
2. Selling into data centers is a massive opportunity. Every layer of the stack — power, cooling, networking, storage, security, monitoring, construction — needs better products. Data center operators are spending aggressively and moving fast.
3. Geography matters again. The cost and availability of power, water, and connectivity in your chosen location directly affect your unit economics. This is not a "cloud-native" consideration — it is an industrial one.
4. Think in decades, not quarters. The companies winning in AI infrastructure are making 10- to 15-year commitments. If your business depends on this infrastructure, you need to think on similar time horizons.

Wear the mindset of a builder — grab a TBPN jacket or hat and join the conversation about what it really takes to build the AI future.

The Refinery Analogy's Limits

No analogy is perfect. A few important ways data centers differ from oil refineries:

• Software flexibility: A refinery is physically configured to produce specific products. A data center can be repurposed from training to inference to traditional workloads with software changes. This flexibility is a major advantage.
• Pace of change: Refining technology evolves slowly. Data center technology — particularly GPUs and accelerators — changes on 2-year cycles, requiring constant hardware refresh.
• Scale of the market: The global refining market generates $3 to $4 trillion in annual revenue. The data center market, while growing explosively, is still an order of magnitude smaller. The question is whether AI will close that gap.

Frequently Asked Questions

Are data centers really as strategically important as oil refineries?

Increasingly, yes. Governments around the world are treating AI compute capacity as critical infrastructure. The US, EU, China, and several Middle Eastern nations have all implemented policies to promote domestic data center construction. Export controls on GPUs are effectively export controls on the raw materials for these "intelligence refineries." The strategic importance will only grow as AI becomes more deeply embedded in economic activity, defense, and governance.

What happens to data centers if there is an AI downturn?

Unlike highly specialized refineries, data centers have significant reuse flexibility. If AI workloads slow, facilities can serve cloud computing, enterprise IT, streaming, gaming, and other workloads. However, the most AI-optimized facilities — those built for 100+ kW per rack with liquid cooling — would require partial reconfiguration for lower-density workloads. The infrastructure would not become worthless, but returns would compress.

Which data center markets are the most important right now?

Northern Virginia remains the world's largest data center market, hosting over 3 GW of capacity. Dallas-Fort Worth is growing fastest due to cheap power and land. Phoenix and Las Vegas offer low-cost power but face water constraints. Internationally, Amsterdam, Frankfurt, London, Singapore, and Tokyo are established hubs. Emerging markets include Johor Bahru (Malaysia), Batam (Indonesia), and several locations in the Middle East, driven by sovereign AI investment.

How can individual investors gain exposure to the data center buildout?

Publicly traded options include data center REITs (Equinix, Digital Realty), infrastructure companies (Vertiv, Eaton, Schneider Electric), networking (Arista, Broadcom), and power companies serving data center markets. Several data center-focused ETFs have also launched. For venture-stage exposure, companies building liquid cooling, modular power systems, and data center management software are actively raising capital.