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 moved down the use-case ladder, from frontier training to fine-tuning, serving, batch work and hobby use, and often to smaller operators. The curve deliberately measures total earning-power decline, reassignment included, because that is what a lender finances. It is 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 is published market data of its own: power prices, power usage effectiveness (PUE), colocation rates. It 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 −21%/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. 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 (2026-09-28, Guaranteed On-Demand, $/GPU-hr)

Neocloud rate, log scale, on age: ≈ −21% per year, R² 0.93 across the five datacenter generations (ages approximate). All points are one date's posted prices (2026-09-28); 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 93% of the variance in the Neocloud rate. The largest deviation is H200, which trades below the line.

The two hollow points are the tell. RTX PRO 6000 is Blackwell silicon about 1.4 years old, and it rents at $1.96, below the 5.8-year-old A100 at $2.30. Its memory bandwidth is about 1.6 TB/s of GDDR7, with no high-bandwidth memory (HBM) and no NVLink. That is less than the A100 carries. The newer chip rents lower because it can do less. Capability is the priced variable. Section 03 shows it.

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 0.9 to 5.8 years old rents inside $0.86–1.13 per TB/s-hour, a 27% spread around $0.99. The oldest chip, A100, sits at the top of the range. After bandwidth normalization there is no discount for age.

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

Per FLOP, older chips look expensive even though their absolute rates are far lower. Each generation improves compute faster than bandwidth, so the A100 sits at $7.37 per PFLOP-hour against $3.04 to $4.18 for newer chips. The spread per FLOP is 100%, against 27% per unit of bandwidth.

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 $0.60–1.12 per kW-hr, air-cooled datacenter silicon at $4.49–5.75, and liquid-class Blackwell at $6.85–7.03. Chips that fit the same halls land close to each other per watt. 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. Rent tracks capability. Once capability is taken out, posted rents show no extra discount for age. Bandwidth and 16-bit compute both fit the cross-section well, and these data cannot separate them. L40S shows why: per unit of bandwidth it rents far above the band, while per FLOP it sits inside the Hopper range.

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. On 2026-09-28 it runs $0.86–1.13, with B300 at $1.00. If serving demand deepens, the band should tighten and new generations enter inside it. If training demand tightens, Blackwell should move above it.
2. Term structure loads the short end. Inference buyers resist tenors beyond one year. The 1Y cell becomes the tenor where committed pricing concentrates.
3. The guaranteed–interruptible spread on older generations. On 2026-09-28: H100 Neocloud guaranteed $3.56, interruptible $2.16 (-39%); A100 $2.30 and $1.20 (-48%). A lasting narrowing on old generations would mean 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 at $1.13/TB/s at 5.8 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. Underwriting moves from offtake credit toward market-rate reference.
7. The liquid watt-premium. On 2026-09-28, liquid-class Blackwell rents at $6.85–7.03 per kW-hr against $4.49–5.75 for air-cooled chips. As liquid-cooled halls build out, the gap should narrow. If the air band falls first, old-generation floors break from the facility side.

05 Limitations

Single cross-section, not a cohort. Posted asks, not executed transactions. The headline stat is the operator-equal median. The fit uses SXM (socketed) cells and the witnesses use PCIe (slotted) cells. Bandwidth and 16-bit compute are crude capability proxies: they ignore memory capacity, interconnect domain and faster number formats such as FP4. Specs are NVIDIA datasheet values. Thin cells (n<3) are indicative. Sources and series per the Methodology. Underlying cells are visible in the Explorer.