Rent and MSRP: Five Generations of Posted Prices

01 The MSRP Figures

NVIDIA publishes no MSRP for any datacenter GPU. The figures used here are the closest documented equivalent — system prices (chiefly NVIDIA's DGX systems) divided by GPU count, the best-documented grain available across all five generations — with the confidence grade and the basis stated per row:

Grade = source confidence: A — vendor-documented (a published system price or filing); B — contemporaneous credible press or a dated reseller capture; C — secondary compilation or undated source.

System-grain division embeds CPUs, memory, storage, networking, and chassis — roughly 25–30% of the quotient is non-GPU content (the H100 teardown puts the same silicon at $24,400 at bare-board grain). Alternate grains move the figures in §02 proportionally: bare-board figures shorten the recovery months by about a quarter; all-in installed capex (the IREN-disclosed $46–54K per Blackwell GPU) lengthens them 15–40%. The system grain is used because it is the only one documented for all five generations; the sensitivity is the band, not a footnote.

02 Months to Recovery

The elections first. The net figures — the chart above and the first column below — use three inputs. First, a realization factor of one-half: filings analyzed on /realized show operators realizing well below their posted rate cards once utilization, discounting, and mix are in the number. Second, an imputed facility rent per chip — $0.75 per thermal-design-power (TDP) kW-hr, the mid-point of the cash band published on /chip-economics, times each chip's nameplate power draw. Operator cost bases differ by facility strategy — a lessee's building is cash rent while an owner's is capitalized and invisible to a cash measure — so every chip is charged its facility at arm's length whether its operator owns or rents: the owner-lessee normalization equity analysts applied to airlines for decades, run in reverse because the outlay here is the chip, not the fleet. Third, overhead at 20% of realized revenue — the cash operating cost that remains once facility payments are stripped from the filings that disclose it (CoreWeave's cash opex ex-lease runs 19–20% of revenue; IREN's direct non-power costs plus unallocated corporate land in the same range; Nebius's scale-out spend sits above it). Costs exclude depreciation and financing; months where the net rate falls below zero are floored at zero; months between general availability and archive entry assume the first observed rate held from GA. The question a payback calculation asks of an initial outlay: how long did each generation's net stream take to add up to its own MSRP?

Net of everything, V100 crossed at 31 months and stands at 1.85× its MSRP at the window's edge; A100 crossed at 37 months and stands at 1.53×; H100, at 42 months old, stands at 0.83× and has not yet crossed — paces at or inside the 3-to-5-year amortization schedules lenders in this market actually write, with only the two oldest generations through. H200 and B200 are 24 and 16 months into their windows. These are elections, stated so the arithmetic can be reproduced or re-elected; CCIR publishes the prices, not the elections.

Sensitivity. Across the published facility band ($0.64–0.87 per TDP-kW-hr), V100 crosses at 29–32 months, A100 at 34–40, and H100 ends its window at 0.75–0.90×. A realization of 0.40 moves V100 to 43 months and A100 to 62. Late in life the facility floor binds — a chip's watts do not fall with its rent — and the oldest V100 months run near cash breakeven, visible as the flattening at the right edge of its panel.

The ceiling by construction

The gross figure strips every cost and friction: 100% utilization, 24 hours a day, at the posted panel-median list rate, with no cost line — no discounting, no idle hours, no power, no facility. It is gross revenue, not margin, and not a forward figure; each window contains the demand episodes documented in the cohort note. It is the ceiling the elections above discount — and the figure with no cost elections, a ratio of two posted prices (the pre-entry backfill assumption applies here too):

100% utilization · 24×7 · posted list · before all costs. For nine years, across a 3.5× range of anchors, the idealized answer has stayed inside 8–14 months.

The same panels, gross — cumulative posted-rate revenue divided by MSRP, the other generations ghosted for context:

03 The Ratio Through Time

Entry ratios cluster near one across five generations and nine years — posted pricing has repeatedly placed a new chip's annual rental stream at approximately its launch price. The ratio then decays with age at the cohort rates measured in the companion note, and rises during demand episodes: H100's ratio at the window's edge is roughly its level at launch, and H200's — which entered low — sits above one. H200's entry point reflects a single early repricer's aggressive list; its MSRP figure is also the weakest-documented of the five (§01). The window ends at the archive's edge, April 2026; the live CCIR panel is the continuation.

04 Both Legs, 2023: One Generation, Two Posted Prices

For one generation the record captures the purchase leg and the rental leg moving together, mid-life. Through 2023, A100 — then roughly three years old — saw both of its posted prices firm. On the rental leg, the repricing providers moved lists upward through 2023 into early 2024, and Lambda raised its A100 list outright on January 1, 2024 (+17–19%, shown as side-by-side 2023/2024 columns on the archived page). On the purchase leg, contemporaneous reporting put the US "going price" near $10,000 while gray-market units in China traded near $20,000 — "double the usual price," a premium carrying an export-control confound the record cannot separate out. Both legs are archived, both are posted or reported prices, and both moved upward on an asset three years past launch. The episode ended the way the cohort note documents: rental lists stepped back down 30–56% from their peaks across the following year.

05 MSRP and Rent, per Unit of Bandwidth

The chip-economics tables show current rents are flat per unit of memory bandwidth. The MSRPs are not:

Acquisition cost per unit of bandwidth fell roughly 2.5× from A100 to the Hopper-refresh and Blackwell generations; posted rent per unit of bandwidth did not follow it down. Stated both ways: the rent band is not a constant markup over per-bandwidth acquisition cost, and the per-bandwidth economics of renting out newer silicon have improved relative to older silicon at current posted prices. This note reports the two series; the mechanism — demand-side pricing versus cost pass-through — is not adjudicated here.

06 The Income Leg and the Resale Leg

Resale-side research (American Compute's 2026 residual-value report, built on secondary-market transactions) reports H100 resale at 27–53% of replacement in 2026 and documents the rental and resale legs diverging. Nothing in this note is a residual value: the ratio here is rent against MSRP — two primary-market prices — and the cumulative figures in §02 are gross revenue, not value. The two legs answer different questions, and moved independently in the one episode where both are well-recorded (§04).

07 Limitations

Everything in the cohort note's limitations applies: posted asks, panel composition, carry-forward repair, variant pooling, archive gaps. Additionally: the MSRP figures are graded B or C for three of five generations and embed non-GPU content at system grain (§01); the recovery arithmetic assumes continuous rental at the posted median, which no operator realizes; and every window includes demand episodes — a recovery period measured through a firming market is a property of that window, not of the next one.