Rent and Age: Five Generations

01 What This Measures — and What It Doesn't

This is the income leg only: what a chip earns per hour, by generation, across operator segments. It is not a residual value. Resale-market evidence (American Compute's 2026 residual-value report, built on secondary-market transactions) shows rental rates and resale prices can diverge — H100 rentals fell through a period in which H100 resale values rose on datacenter-shell scarcity. CCIR publishes observed prices; valuation belongs to the reader.

An A100 renting below an H100 is not "the same asset, older." The chip has migrated down the use-case ladder (frontier training → fine-tuning → inference → batch → hobbyist) and often down the operator ladder — in the current panel, 3090 / A40 / L4 / L40 list only in the Marketplace segment, while A100 still spans all three. The curve deliberately measures total earning-power decline, reassignment included, because that is what a lender finances — the same convention by which an aircraft's lease-rate curve includes its migration from passenger to cargo.

A second boundary: every figure here is gross rental revenue — dollars per GPU-hour before power, cooling, space, and operations. The per-watt figures in Section 03 are revenue per unit of power envelope, not margin. A cash-flow or debt-service coverage ratio (DSCR) model nets operating cost against this series; the cost side — power prices, power usage effectiveness (PUE), colocation rates — is published market data of its own and belongs in a separate note, as does the boundary it implies: the point at which a generation's gross rent meets its cash operating cost is its economic end-of-life.

One caveat governs everything below: a cross-section (many generations, one date) is not a cohort curve (one generation through time). The −26%/yr is a fit across differently-aged chips at current prices — it describes how the market prices age today, not the rate at which any chip's price actually fell, and it says nothing about where each generation's pricing started at launch. Reading it as a decline path assumes the next generation ages like the last — exactly the assumption an inference-led demand shift would break. Section 04 lists the observable markers.

02 The Age Curve (2026-07-06, Guaranteed On-Demand, $/GPU-hr)

Neocloud rate, log scale, on age: ≈ −26% per year, R² 0.96 across the five datacenter generations (ages approximate). All points are one date's posted prices (2026-07-06); the line is a fit across generations of different ages, not the price history of any one chip.

The regularity is the finding: five generations spanning five years of silicon, and a single log-linear slope explains 96% of the variance in the Neocloud rate. The largest deviation is H200, which trades below the line — consistent with the bandwidth pricing in the next section.

The two hollow points are the tell. RTX PRO 6000 is three-months-old Blackwell silicon priced at $1.97 — where 5.7-year-old A100 trades — because its memory bandwidth (~1.8 TB/s of GDDR7, no high-bandwidth memory (HBM), no NVLink) is A100-class. A brand-new chip prices at the old chip's level when it carries the old chip's bandwidth: age is not the priced variable. Section 03 makes that explicit.

03 Three Normalizations, Three Markets

Standing version of this section's tables: /chip-economics.

Divide the Neocloud rate by the chip's capability and the curve splits. Per unit of memory bandwidth — the binding resource for large language model (LLM) serving:

Datacenter silicon from eight months to 5.7 years old clears inside $0.87–0.98 per TB/s-hour — ±6% around $0.92. After bandwidth normalization there is essentially no age discount left.

Per unit of dense compute in the BF16 number format — the binding resource for frontier training:

Per FLOP, older chips look expensive — not because anyone charges more (absolute rates are far lower), but because the age discount does not keep pace with the FLOP gap. Each generation improves compute faster than it improves bandwidth, so a market that prices bandwidth leaves older chips above FLOP-parity automatically: a renter who actually wants FLOPs buys the frontier. Compute- normalization does not flatten the cross-section; bandwidth-normalization does.

Per unit of nameplate power — thermal design power (TDP), the facility's constraint, which determines which datacenters can physically host a chip:

The watt-rent sorts the panel into three bands: consumer silicon with no datacenter bundle at ~$1/kW-hr, air-cooled datacenter silicon flat at $4.50–4.66 across five years of age (A100 = H100 = L40S per watt), and liquid-class silicon at $6–7. When the shell is the binding constraint, chips that fit the same shells converge on the same watt-rent — and part of the Blackwell premium is rent on scarce liquid-cooled capacity. The breadth of facilities that can host a chip is also the residual-support mechanism visible in the resale data: old air-cooled generations fit the largest installed shell base.

Reading. The marginal renter across generations is paying for bandwidth-hours, not FLOP-hours — the pricing signature of a serving-dominated market. If frontier training set the cross-section, old chips would trade at a steep per-FLOP discount; instead they trade at a per-FLOP premium and a per-bandwidth par. At the margin, the repricing an "inference shift" is expected to cause appears to have already happened.

04 If Volume Keeps Moving to Inference — the Watchlist

Stated as observables CCIR already publishes, so each can be checked rather than argued:

1. The $/TB/s band is the price. Serving demand deepening → the band tightens and new generations enter at it (B300's 0.98 decaying toward 0.87–0.90 as scarcity fades). Training re-tightening → Blackwell decouples upward per-TB/s.
2. Term structure loads the short end. Inference buyers resist tenors beyond one year; the 1Y cell becomes the market's effective clearing tenor.
3. The guaranteed–interruptible spread narrows on older generations. Today: H100 Neocloud guaranteed $3.22 vs interruptible $2.00 (−38%); A100 $1.80 vs $1.14 (−36%). Structural narrowing on old generations means bursty serving demand is reaching down the ladder.
4. Old-generation floors hold. In power-market terms, depreciated chips are peakers — low carrying cost, dispatched into demand spikes — while new fleets are baseload needing committed offtake. A100 holding ~$0.90/TB/s at 5.7 years is the floor holding; watch it as Blackwell supply normalizes.
5. Utilization becomes the swing variable. Serving is diurnal, training is flat; realized-vs-posted gaps widen for serving-heavy operators even where posted rates hold.
6. Contract quality shifts. Many short-tenor customers replace one five-year offtake: diversification up, committed coverage down — and underwriting moves from offtake credit toward market-rate reference.
7. The liquid watt-premium compresses. As liquid-cooled shells build out, $6–7/kW-hr should converge toward the $4.6 air band — a facility-driven repricing of new silicon independent of age. If instead the air band decays first, old-generation floors break from the facility side.

05 Limitations

Single cross-section, not a cohort. List asks, not executed transactions. The headline stat mixes mean (deep panels) and median (thin) per the published rule; form factor is pooled and the mix of the SXM (socketed) and PCIe (slotted) form factors differs by generation; bandwidth is a crude capability proxy (ignores HBM capacity, interconnect domain, FP4 precision-format paths — the B300 BF16 figure is approximate); the A100 Hyperscaler cell is a slow-moving rate card; thin cells (n<3) are indicative. Sources and series per the Methodology; underlying cells are visible in the Explorer.