A company that needs more AI compute has three ways to obtain it: build new capacity, buy more from a cloud provider, or recover the expensive capacity it already owns yet isn't using. Put them side by side on the only two things a buyer actually weighs, price and how long until it's running, and it isn't a close call. Build and buy options are priced at scarcity and measured in years until go-live. Recover is already paid for, and it's available now. The uncomfortable part is that for build and buy options, each have their cost partly because recovery keeps getting skipped. A company shopping for new capacity often pays a scarcity price for a seldom used version of what's already sitting in its own racks.
Can you imagine going to a currency exchange and converting a dollar and getting 5 cents in return? This just happened to me in Croatia (OK, not that bad but still financial game-playing). With compute, you're currently getting 5-15 cents of value for every dollar you've invested in your compute infrastructure.
Overall, compute is driven by microchips. The question is how to get more of them without more spend.
A chip here means the expensive accelerator that does the heavy AI math, whatever the brand on the box. When a company runs out of room to do its AI work, the reflex is to go buy more of them. As noted, there are three ways to do that, and success should be measured in the capacity costs and how long before it's actually running. Let's take them one at a time.
Build: the most control, the worst numbers
Building your own capacity means standing up a data center, or expanding one, to house the chips. You get maximum control. You also get the worst answer to both questions a smart buyer is asking.
On cost and spend, the money goes out the door upfront, and it's a lot. A standard facility runs about $11-12 million per megawatt of capacity as a one-time construction cost. An AI-optimized build, with the dense power and cooling those chips demand, runs $20 million or more per megawatt. That is before a single chip is bought.¹² That money buys the building and its power and cooling, not a year of running it.
Then there's the wait, which is the bigger problem, as it actually isn't about construction today. It's about power. Connecting new capacity to the grid takes four to seven years in the major markets. Even the fastest sites, the ones that can line up their own power, are looking at 24 to 48 months, and plenty of locations can't get a firm power date at any price.³⁴ Build is the option for a buyer with years to spare and capital to burn, but ultimately gated by a constraint no budget can buy its way past.
Buy: faster, but you're renting on someone else's clock
Buying means renting capacity from a cloud. It's what most companies do, and it's much faster than building. The catch is that the price is set by the same shortage that makes building slow, and at the end of it you own nothing. You have rented access for a term.
Look at today's rates and the shortage trend is quickly highlighted. The one-year rate for a top-end chip climbed roughly 40 percent from late 2025 into early 2026, on-demand capacity sold out across providers, and the identical chip rents for several times more on the big clouds than on specialist providers.⁵⁶ Want the cheaper number? You commit to a one-to-three-year term and pay for the capacity whether you use it or not.⁷ You can be running in days if capacity exists. But a multi-year lock ties you to today's price and today's hardware while faster chips are often introduced over the life of the contract.⁷ Buying beats building in terms of speed but still costs a fortune, as you're bidding against everyone else for the same scarce supply.
Recover: already bought, already running
The third option is the one most buyers never price, because it doesn't look like a purchase. It's recovering the capacity the company already has and isn't using.
A single AI workload, one whole piece of AI work, runs in stages, and only one of those stages actually needs the expensive chip. Nonetheless, the standard way of running it pins that chip for the entire job. Measured across real production fleets, the expensive chips do useful work about 5 percent of the time.⁸⁹ It's worth reading that number again: the company is holding roughly twenty times the capacity it actually uses, on hardware it already bought or already committed to rent. The capacity costs no new money, because the money is already gone, so the marginal cost of the work it unlocks rounds to roughly zero. And it's available now, because the hardware is powered on and sitting there. Route each stage to the chip that fits it, hold the expensive chip only for the stage that needs it, and the idle capacity starts doing work, on the same fleet, right now.
The three choices, side by side
Lined up, the decision isn't subtle. Build costs $11-20-million dollars per megawatt and takes two to seven years. Buy costs are driven up by scarcity on a multi-year commitment, riding on capacity that's currently sold out. Recover costs no new purchase and is available now, drawing on hardware already paid for and running at a fraction of what it can do. On both axes that matter, price and lead time, recover wins big, and it's the only one of the three that doesn't ask you to win a bidding war or wait on a utility.
Ultimately, build and buy are expensive largely because demand is outrunning supply, and demand is outrunning supply partly because so much bought capacity sits idle. A company paying for new capacity while running its existing fleet at a twentieth of its potential is paying the scarcity premium twice: once in the idle hardware it already owns, and again in the new hardware it buys to cover for not using the full capacity of the existing fleet. Recovering existing capacity is the option that breaks the loop.
Three questions to ask before you add capacity
Before you sign off on a build or a buy, three questions tell you whether the cheaper option was ever really ruled out.
- What's the price and lead time of the capacity we're considering? Measure it in dollars per unit and months to running. Millions per megawatt and years out, or scarcity rates on a multi-year lock: that's the bar recover is being measured against.
- What's our current utilization, and how much would we get back by closing the gap? If the fleet runs near a twentieth of its potential, the recoverable capacity is large, already paid for, and it stands as the build-or-buy alternative with very favorable math.
- Are we buying new capacity to do work our own fleet could already do if it weren't idle? If the honest answer is yes, you're paying the scarcity premium twice, and recovering first gets you the same output cheaper and sooner.
A note on the numbers
Every figure here is reported with its source and current to Q2 2026. The build costs (11 to 12 million dollars per megawatt standard, 20 million or more for AI-optimized) are construction-industry benchmarks reported as ranges, before the cost of the chips themselves. The power-wait figures (four to seven years to grid connection, 24 to 48 months for the fastest self-powered sites) are separate clocks reported as ranges, not summed. The rental figures (the roughly 40 percent contract-rate move, the sold-out on-demand market, the multi-provider spread) are supply-driven evidence, not a forecast, with the reserved-term structure noted alongside. The 5 percent usage figure comes from direct measurement across tens of thousands of production systems; the "twenty times" framing restates that same figure as a usage ratio, not a speed claim. The argument doesn't rest on any one number. It rests on the shape of all of them together: two options priced at scarcity and measured in years, and a third already paid for and available now.
References
¹: Standard data center construction at roughly 11 to 12 million dollars per megawatt in 2026 as a one-time build cost, on a global average near 11.3 million, before IT equipment. JLL 2026 Global Data Center Outlook via Aptly Tech (Mar 2026); Avid Solutions (Jan 2026).
²: AI-optimized builds running to 20 million dollars or more per megawatt due to higher power density and liquid cooling, and a 100 MW AI facility climbing beyond 1 billion dollars. TrueLook (Mar 2026); ConstructElements (2026).
³: Grid interconnection waits of four to seven years in the major data center markets, with the queue set by the utility rather than the developer. Sightline Climate data cited by Bloomberg (May 2026), via Tech-Insider.
⁴: The market splitting between sites that can energize new load in 24 to 48 months and sites that cannot get a firm grid date at any price; power deliverability, not construction, deciding which projects are real. Archdesk AI data center construction analysis (Apr 2026).
⁵: H100 one-year rental contract pricing up almost 40 percent from late 2025 into early 2026 with on-demand capacity sold out across providers. SemiAnalysis H100 rental index (Apr 2026).
⁶: The same chip renting for two to three dollars an hour on specialist providers and several times that on the big clouds, with published rates reaching about 6.88 and 12.29 dollars an hour for identical hardware. Spheron GPU cloud pricing comparison (May 2026).
⁷: The cheapest cloud rates requiring one-to-three-year reserved commitments with payment owed for the term whether or not the capacity is used, locking buyers to current hardware and pricing as newer chips arrive. Petronella Cybersecurity (Mar 2026); FPT AI Factory (May 2026).
⁸: Average GPU utilization of about 5 percent measured directly across tens of thousands of production clusters, indicating roughly 20 times over-allocation. Cast AI, 2026 State of Kubernetes Optimization Report.
⁹: Independent reporting on the same finding: about 5 percent average use across roughly 23,000 clusters, with billions in committed compute sitting idle. ITBrief; Data Center Knowledge (May 2026).
Build and buy are priced at scarcity and measured in years. Recover is already paid for and available now, and skipping it is what makes the other two cost what they do.