You didn’t buy a chip. You bought a block of time.
A microchip is the expensive accelerator that does the heavy AI math, whatever the brand on the box. There are two ways to get one, and underneath they turn out to be the same kind of purchase: a fixed block of earning-time that starts counting down the second you commit.
Buy the chip outright and it’s a capital asset, $25,000-40,000 for a top-end unit, that loses value in depreciation over time.1 What ages it isn’t specifically wear. It’s the next chip. The dominant maker has moved to a roughly yearly release cadence, each generation a big step up, so last year’s chip becomes the slower, costlier way to do the same work the moment a faster one ships.2 That decline runs on the calendar, not the meter. Your chip is worth less in a year because a better chip exists in a year, whether it ran flat out or sat cold.
Lease the chip through a committed cloud term and the shape doesn’t change. The cheapest cloud rates come from reserved commitments of one to three years, and the whole point of a reservation is that you pay for the capacity across the entire term whether you use it or not.3 You’ve bought a block of capacity-time, fixed up front, running down on the calendar. The same cadence risk rides along: a multi-year commitment locks you to current-era hardware and pricing while newer, cheaper chips arrive over the life of the term.4
So whether the compute capacity sits on your balance sheet as capex or on your operating statement as reserved opex, you’re holding the same asset: a window of earning-time already counting down, on a clock you don’t really control.
Past idle time is gone but future work recovery is the opportunity
Now let’s dig into how the owned or leased compute capacity actually gets used. A single AI workload, one whole piece of AI work, runs in stages, and commonly only one stage needs the expensive chip. The standard way to execute a workload is to hold that chip for the entire run, so across real production fleets the expensive chips do useful work about 5 percent of the time.56 No matter how you acquired the compute, you’re paying for the full capacity and using a twentieth of it.
The idle time behind you is gone. It’s sunk, and a sunk cost isn’t a decision anyone can act on.
The opportunity window that’s left is a different story, and it’s where every recoverable dollar sits. For a purchased chip, the remaining months of its short economic life are still productive time you can claim, if you put it to work correctly now. For a reserved cloud term, the months left on the commitment are capacity you’ve already paid for and can still use, if you load it now. The value is real and forward-looking but it’s also perishable as time passes.
Why waiting forfeits the part you can still get
Treat the compute window as a fixed budget of productive time. Every quarter the capacity sits idle, the budget shrinks and nothing comes back for it. What’s left to claim next year is smaller than what’s claimable today, smaller still the year after, until the window closes and the remainder is zero. The chip is fully depreciated, or the reserved term ends, and whatever you didn’t capture is just gone. Plus, the next chip’s arrival and the contract’s end date are both fixed too.
That’s what makes capturing the compute window an urgent, forward decision instead of a cost to regret. The money can be recovered by keeping the capacity on well-matched work for the rest of the window that remains, starting now, on hardware or contracts you’ve already paid for. Claiming it takes no new purchase, only using what’s already counting down.
There’s a fair counterpoint. Some operators argue an older chip keeps earning well past that point, because work cascades down to it: as new chips take the frontier jobs, the older ones move to inference and steady tasks and stay useful for years.7 Where that holds, the window is longer than the critics claim. But the cascade is itself a fit argument. An aging chip keeps earning only while its work is matched to what it’s now best at; a chip left idle, or run on work that doesn’t suit it, captures none of the cascade.
Two questions to ask about your committed capacity
For any capacity you acquired by purchase or by a committed cloud term, two questions turn a closing window into something a budget owner can act on.
- How much of this capacity’s window is left, on the depreciation schedule or the contract term, and how much of it is sitting idle? The remaining months are what’s still claimable; the idle share of them is the value at risk of being forfeited.
- What’s our plan to keep this capacity on well-matched work for the rest of its window, without buying more? It’s the only move that captures the remaining value before the window closes, and it works the same whether the capacity is owned or reserved, because it uses what’s already been paid for.
---
A note on the numbers
The purchase price and reserved-term structure are reported as ranges current to Q2 2026: top-end chips at 25,000 to 40,000 dollars, reserved cloud commitments at one to three years with payment owed for the full term regardless of use. The release cadence is reported as the maker’s stated direction toward yearly generations, with performance-gain figures as vendor and analyst estimates. The useful-life question is an active, unsettled accounting dispute, with operators defending five-to-six-year schedules and critics arguing for a two-to-three-year economic life; that dispute sets how short the capex window is, and it’s presented as contested rather than settled. The 5 percent usage figure is from direct measurement across tens of thousands of production systems. The earning-window framing is a way of reading these figures, not a separate measured claim. The argument doesn’t rest on any single number. It rests on the shape of all of them together: capacity acquired as a fixed block of time, a calendar that spends it down regardless of use, and a remaining window that’s recoverable only until it closes.
---
References
- NVIDIA H100 at roughly 25,000 to 30,000 dollars (PCIe) and 35,000 to 40,000 dollars (SXM5), a capital asset carried on the balance sheet and depreciated over time. CloudZero (May 2026). CloudZero
- NVIDIA moving to a roughly yearly architecture cadence (Hopper 2022, Blackwell 2025, Vera Rubin slated for late 2026), with each generation a large step up, so prior-generation chips become the slower, costlier way to do the same work. Interesting Engineering analysis (Nov 2025); Stanley Laman Group (Nov 2025). Interesting Engineering Stanley Laman Group
- Reserved cloud commitments of one to three years offering 30 to 60 percent discounts but requiring payment for the capacity for the whole term whether or not it is used, with early-termination exposure if needs change. Petronella Cybersecurity (Mar 2026); FPT AI Factory (May 2026). Petronella Cybersecurity FPT AI Factory
- Multi-year reserved commitments locking buyers to current-era hardware and pricing while newer, more efficient chips arrive over the term, leaving some long-term reservation holders paying above later on-demand rates. CloudZero cloud GPU pricing comparison (Apr 2026). CloudZero
- 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. Cast AI
- Independent reporting on the same finding: about 5 percent average use across roughly 23,000 clusters. ITBrief. ITBrief
- The value-cascade case for a longer earning window: older chips moving from frontier training to inference and other steady workloads and staying productive for years, with operators citing utilization and secondary-market pricing as evidence. CNBC (Nov 2025); MBI Deep Dives (Oct 2025). CNBC MBI Deep Dives
The capacity you committed to is a window of earning-time the calendar is closing. The hours behind you are gone; the ones ahead are recoverable only if you claim them now.