What Does It Actually Cost to Build AI?

Introduction
Why Is the Bond Market Suddenly a Tech Story?
What Does It Mean That Compute Itself Has Become a Tradeable Commodity?
Why Is Electricity Becoming AI's Most Strategic Asset?
Can AI Solve Its Own Water and Heat Problem?
Why Are AI Chips Getting More Expensive, Not Less?
What Does This Mean for Leaders Outside of Tech?
FAQs

1. Introduction

Forget the chatbot demos for a second. The real AI story this month is not happening on a screen. It is happening in power plants, cooling pipes, and bond markets: the unglamorous physical infrastructure that determines whether "intelligence at scale" is even possible.

Amazon, Alphabet, Microsoft, and Meta are projected to spend roughly $750 billion this year building AI infrastructure, more than 80% higher than last year. That single number reframes the entire conversation. Building AI is starting to look less like writing software and more like building railroads, oil pipelines, or steel mills: capital-intensive, physically constrained, and dependent on resources that can run dry.

This is not a story leaders can afford to file under "technical." It is a story about capital markets, energy contracts, and supply chains, and it is reshaping decisions far beyond the data center.

2. Why Is the Bond Market Suddenly a Tech Story?

There is an assumption that has held for two decades: Big Tech is cash-rich, so interest rates are someone else's problem. That assumption is breaking down in real time.

Tech investors who once ignored Federal Reserve announcements because megacap companies sat on mountains of cash are now watching bond yields the way energy or airline investors do.

The reason is straightforward. Four of the most profitable companies on earth are increasingly funding their AI buildout with debt rather than reserves alone, and when you borrow at this scale, the price of money starts to matter to your bottom line.

It is not only the household names taking this path. Morgan Stanley has been pitching data-center developers on leveraged-buyout-style debt structures, the kind of financing historically associated with private equity takeovers rather than tech campuses. Separately, SpaceX is reportedly preparing a bond deal worth up to $20 billion. Even a rocket company is now turning to Wall Street's debt machinery to fund AI compute.

The companies building AI are increasingly financed like industrial companies, not software companies. That shift changes how every investor, lender, and competitor should be reading their risk.

3. What Does It Mean That Compute Itself Has Become a Tradeable Commodity?

This is the detail that should reframe how anyone thinks about "AI companies." SpaceX's Colossus data center recently signed a compute deal with AI startup Reflection worth up to $6.3 billion, with Reflection paying SpaceX $150 million a month from July 1, 2026 through 2029 for immediate access to Nvidia's newest GB300 chips and related hardware.

SpaceX is not selling Reflection a product. It is renting out raw computing power, the way a utility sells electricity or a landlord leases office space. Access to advanced chips has become so scarce that companies are signing multi-year contracts for it, structured less like a software license and more like a factory lease.

When computing power itself becomes the asset companies fight to secure, the balance of leverage in the AI industry shifts toward whoever controls the physical hardware, not whoever has the most advanced model.

4. Why Is Electricity Becoming AI's Most Strategic Asset?

Chevron recently signed a 20-year power agreement to supply a Microsoft data center in West Texas. Twenty years is an unusually long commitment in any industry, and it signals something important: AI companies are no longer worried only about chips. They are worried about whether the lights will stay on.

Reliable, long-term electricity access has become a strategic asset that technology companies are willing to lock in decades in advance, the kind of deal historically struck between a refinery and an oil supplier, not a software company and a utility.

This matters because it quietly redraws the map of where AI infrastructure gets built. Regions with abundant, stable power generation are becoming as valuable to the AI economy as regions with deep talent pools once were.

5. Can AI Solve Its Own Water and Heat Problem?

Every AI chip generates enormous heat, and cooling that heat has historically required large volumes of water through air-conditioning-style chillers. Nvidia recently unveiled next-generation liquid-cooled systems that can run coolant at 113°F, hot enough that many data centers could reduce or even eliminate the mechanical chillers responsible for much of that water and energy use.

The mechanism is similar to how a car radiator works, just engineered for a machine that generates far more heat and runs continuously at industrial scale.

It is a genuinely promising fix, but it is not a finished one. How fast the technology gets adopted, whether older facilities can be retrofitted at all, and how much water is still consumed indirectly through the electricity needed to power these systems remain open questions. Innovation at the chip level does not automatically resolve the resource pressure created at the grid level.

6. Why Are AI Chips Getting More Expensive, Not Less?

For two decades, the operating assumption in technology was simple: computers get cheaper every year. That assumption has now reversed.

Quality-adjusted import prices for computers and semiconductors are up 14.4% year-over-year, with a 3.6% jump in May alone, before tariff effects were even factored in, according to analysis from SemiAnalysis. The chips that power AI have become so sought-after that scarcity, not efficiency gains, is now the dominant force setting prices.

This is a quiet but significant shift. It means the cost curve every business has relied on for budgeting technology investment over the past twenty years no longer applies cleanly to the hardware behind AI.

7. What Does This Mean for Leaders Outside of Tech?

Step back, and the pattern across all five of these developments is the same: AI's binding constraint is no longer creativity, talent, or even algorithms. It is physical infrastructure — money, electricity, water, and chips.

For business leaders, that is a useful correction to the prevailing narrative. The "AI revolution" is not only unfolding inside software products and chat interfaces. It is actively reshaping energy contracts, debt markets, and global hardware supply chains, and those shifts will determine which companies can actually afford to compete at the frontier, regardless of how good their models are.

Leaders who track AI only through product launches and chatbot benchmarks are reading half the story. The other half is being written in bond prospectuses, power purchase agreements, and cooling system patents.

Understanding that distinction is what separates strategic foresight from hype-following.

8. FAQs

1. Why are tech investors suddenly paying attention to bond markets and interest rates?

Because Big Tech's AI buildout, projected at roughly $750 billion this year across Amazon, Alphabet, Microsoft, and Meta, is increasingly being financed through debt rather than cash reserves alone. When companies borrow at this scale, the cost of borrowing (set by interest rates) starts to materially affect their financial performance, which is why bond market movements now carry the same weight for tech investors that they long have for capital-intensive industries like airlines or energy.

2. What does it mean that "compute" is being sold like a commodity?

It means raw computing power, specifically access to advanced AI chips, has become valuable and scarce enough that companies sign long-term contracts to lease it, similar to a utility contract or a real estate lease. SpaceX's $6.3 billion compute deal with Reflection AI, structured as a monthly payment for hardware access through 2029, is a clear example: the asset being sold is computing capacity itself, not a finished product.

3. Why are AI companies signing decades-long electricity contracts?

Because reliable power has become as critical a constraint as chips. Chevron's 20-year agreement to supply a Microsoft data center signals that AI companies are securing long-term energy access the way industrial companies have historically secured raw material supply, recognizing that data centers cannot run, regardless of how advanced their hardware is, without guaranteed power.

4. Has Nvidia actually solved the water problem in AI data centers?

Not entirely. Nvidia's next-generation liquid-cooled systems, which run coolant at 113°F, can significantly reduce or eliminate the mechanical chillers that consume large amounts of water and electricity. However, adoption speed, retrofit limitations for existing facilities, and the indirect water use tied to the electricity that powers these systems remain unresolved, meaning this is a meaningful improvement rather than a complete fix.

5. Why are semiconductor prices rising after years of getting cheaper?

Because AI chip demand has reversed a multi-decade pricing trend. According to SemiAnalysis, quality-adjusted import prices for computers and semiconductors rose 14.4% year-over-year, with a 3.6% increase in May alone, before tariffs were even factored in. Demand for AI-grade chips is now outpacing the efficiency gains that traditionally pushed computing costs down, meaning businesses budgeting for AI infrastructure can no longer assume hardware costs will fall over time.

Sources: CNBC, the Financial Times, the Information, Chevron Newsroom, Axios, and SemiAnalysis.