Accelerator life cycle and when to sell

An accelerator does not lose value the way a truck does, by wear. It loses value in steps, and the steps are set by NVIDIA's release calendar. Owners who understand this sell at the right time; owners who think of the hardware as an asset that depreciates gently over a fixed schedule find that the schedule was written by someone else. This note sets out the mechanism with the figures that are available — one of them from a published residual-value report, two from market commentary rather than a single primary source — and then looks at the decision from each side of the trade.

Market commentary during 2026 put the effect of the B200 launch on H100 values at a cut of 25–35 %. That is a step, not a slope: it is attached to the launch of the successor. From the moment the next generation is announced and confirmed to ship, every buyer of the old generation knows what the alternative is, and the price of the old generation is set against that alternative rather than against its own cost.

Fleet rotations then add a second, slower movement. When operators rotate fleets, the previous generation comes onto the secondary market in quantity, and dealer observations put the fall at the order of 10–20 % per quarter while that stock is being absorbed. The two effects compound: a step on launch, then a slope while the fleet turns over.

NVIDIA Rubin, the VR200 platform, goes into deployment in the second half of 2026. That is the next step for B200. Published residual-value work for 2026 puts the B200, on a conservative view, at 59–67 % of original MSRP. That band should be read as the market's pricing before the Rubin effect has shown in secondary prices, since there is as yet no meaningful used B200 supply on which it could show. Whether the eventual step matches what the H100 experienced is not known, and this note does not estimate it.

Table scrolls sideways

The sale decision for an accelerator fleet has a value of its own, and that value is highest immediately before the next generation is announced and falls sharply afterwards. An owner who has decided in principle to rotate should therefore set the timing against the release calendar, not against the fleet's age or the depreciation schedule in the accounts. Waiting until the successor is available, so that the replacement can be installed before the old fleet is sold, means selling into the step and into the slope together.

Within whatever band the market sets, the seller controls the following:

• Documentation. Purchase invoices, serial-number inventories, service records and a clean statement of origin move a lot from the pool of buyers who can accept unknown provenance to the pool who can pass it through procurement. That is a different, larger and better-funded set of buyers.
• Condition. Complete systems with all original modules, no missing brackets, cables or rails, and no evidence of disturbed labels sell as systems; incomplete ones sell as parts.
• Packaging. Original packaging is evidence of handling and allows insured transport at a lower risk premium. It is routinely thrown away and is worth keeping.

None of these change the step. They determine where in the band a specific lot lands.

A buyer of the previous generation is buying hardware whose price has already absorbed a step. That is the attraction. The software stack is mature, the failure modes are known, and the parts market exists. The buyer should nonetheless price in the next step, which is on the calendar, and should hold the hardware for a workload that justifies it rather than as a store of value.

Two generations back, at n-2, the price has absorbed both steps and the slope of the intervening rotation. What the buyer gets for that is hardware that is well understood and cheap to keep running; what the buyer gives up is performance per watt and per rack unit against the current generation. For inference and for training that fits the hardware, n-2 can be the right purchase. For a facility with a fixed power envelope it usually is not.

One technical point cuts against the assumption that newer is always better. The B300 reduces FP64 performance to roughly one thirty-second of the B200's. For HPC and scientific workloads that depend on double precision, the B200 is therefore the more suitable part, and the newer generation is not a replacement for it. An owner of B200 hardware selling into that market is selling a capability the successor does not carry, and a buyer with such workloads should not assume that the n-1 discount reflects any reduction in fitness for the purpose.

How we handle this at GLO Corp: lots are listed with their origin stated and with serial numbers and purchase documentation available to shortlisted buyers under NDA, so that the documentation premium described above is something the buyer can verify rather than take on trust.

1. Introl, NVIDIA Rubin production and deployment — https://introl.com/blog/nvidia-rubin-full-production-ces-2026-ai-infrastructure
2. amcompute.com, GPU depreciation and residual value report 2026 — https://www.amcompute.com/blog/gpu-depreciation-residual-value-report-2026
3. H100 repricing after the B200 launch (25–35 %) and the 10–20 % per-quarter figure during fleet rotation: market commentary and dealer observations collected in 2026, not tied to a single published source; quoted as indicative