Hardware Refresh, Depreciation Strategy, Decommissioning & ITAD

Refresh is where the depreciation debate from Chapter 1.8 stops being a line on a pro-forma and becomes a loading dock full of accelerators two generations old. They are a security liability, an environmental obligation and, handled well, a meaningful slice of recovered capital. Every earlier chapter in Part 14 keeps the asset alive. This one decides when to retire it and what happens to it next. One question runs through the chapter, as it does through AI-infrastructure finance: is a retired GPU an asset that cascades to a lower-value workload and earns its way to the book-life, or a depreciating liability you need off your floor before it costs more to keep than it is worth? The answer depends on the workload and the site, and a wrong answer in either direction is expensive.

This chapter covers refresh as execution, not refresh as economics. Chapter 1.8 is the canonical home for the contested 2–3-year-bear-case-vs-4–6-year-estimate argument, the residual-value evidence, and the training-to-inference cascade as a depreciation defense. This chapter does not re-argue it. It owns the operational consequences: the refresh-trigger decision, the decommissioning-and-sanitization workflow, the ITAD/resale channel and the circular economy it feeds, and the migration discipline that keeps the factory earning while you swap its engines. Decommissioning the facility itself (generators, BESS, coolant, the slab) is a separate lifecycle stage in Chapter 14.10. Here we retire IT, not concrete.

There is no single refresh clock. There are four, they fire at different times, and the one you let dominate sets your residual-value outcome. Most operators inherit a default ("refresh on the book schedule") without realizing it is a choice with a downstream cost.

Calendar trigger. Refresh when the depreciation schedule says the asset is fully written down. It is simple and auditable, and it is the implicit default for anyone who let accounting set the cadence. Its failure mode is that silicon does not obey the ledger: a part can be economically dead two years before its book life ends, or perfectly productive two years after.

Efficiency trigger. Refresh when the next generation's perf-per-watt makes the incumbent uneconomic to run against the power it consumes. This matters most in a power-bound world. When megawatts are the scarce input, the question is not "does this GPU still work" but "is this the highest-value use of this megawatt." A Hopper-class rack drawing ~40 kW is a refresh candidate even if it is fault-free, when a Blackwell rack at ~130 kW or a Rubin-class rack beyond that can out-earn it per kW. The opportunity cost is stranded interconnection capacity you cannot get back. → density trajectory in Chapter 14.7.

Failure-rate trigger. Refresh when the fleet slice crosses the wear-out knee of the bathtub curve, and the annualized failure rate plus RMA logistics cost more in lost goodput than redeployment is worth. The spares and RMA machinery in Chapter 14.6 tells you where that knee is. Refresh is the exit when sparing a generation stops paying.

Density-ramp trigger. Refresh to capture the revenue-per-MW step of a new generation. This is the density-ramp logic Chapter 1.1 treats as a scoping decision, now executed mid-life. It works only if the irreversible substrate (floor loading, water, electrical headroom) was reserved for it. A hall scoped for 40 kW air cannot absorb a 130 kW liquid generation, and the refresh becomes a rebuild.

The retired-asset workload cascade is the mechanism that, if it holds, stretches economic life toward book life. Refresh is where the cascade is either executed or exposed as a story. A part retired from the current liquid-cooled tier does not have to leave the building. Frontier pre-training, test-time scaling and agentic inference can all occupy that current tier. Retired parts can move only to workloads whose performance and power limits they still meet, such as smaller post-training or fine-tuning, latency-relaxed enterprise inference, batch or internal work, and earn revenue there. Each step down relaxes a requirement the part can no longer meet (top-end perf-per-watt, the largest scale-up domain) and exploits what it still does well (raw memory bandwidth, adequate FLOPS for a smaller model).

The catch is that the cascade is finite and power-gated. There is only so much lower-tier demand, and in a power-bound site every cascaded part occupies a megawatt that a newer part could use more profitably. So the real refresh decision per part is a three-way choice. Redeploy it down the cascade if you have the lower-tier demand and the megawatt is not the binding constraint. Resell it into the secondary market if the residual price beats the cascade value net of the power it would consume. Retire or shred it if security obligations forbid resale or the part is uneconomic to run. The cascade defense of the long book life is honest only if a real, liquid resale market exists to absorb the parts that do not cascade internally. That is the secondary-market-depth question Chapter 1.8 flags as load-bearing.

For rack-scale systems the unit of disposition is the rack, not the GPU. A GB200 or GB300 NVL72 is one 72-GPU NVLink domain: its compute trays take power from the rack busbar and coolant from the rack manifold, and the GPUs in those trays form a single NVLink fabric only by way of the NVLink switch trays and their copper cable spine. Pulled out and sold alone, a tray loses that domain, and it will not run unless another rack supplies busbar power, manifold coolant and NVLink switching. Plan cascade and resale per rack (or per NVLink domain). The buyer pool is narrower too, because whoever takes the rack must cool it with liquid at the rack, fed either by facility water or by a liquid-to-air sidecar CDU. Drain the loops and blow them dry before freight so nothing leaks or freezes in transit, and crate the rack whole.

Decommissioning an IT asset at scale is a chain-of-custody discipline, not a teardown. The workflow is roughly invariant across operators, and skipping a step is how a retired GPU becomes a data-breach headline or a stranded-capital line. The sequence is: de-provision (drain workloads, remove from the scheduler and fabric, update the CMDB/DCIM asset record); power down and physically extract (de-cable, drain liquid loops on DLC racks, palletize); sanitize all data-bearing media using the approved NIST SP 800-88 Rev. 2 method, then verify execution, validate the result and retain a serial-linked certificate per asset; account and tag every serial through a chain-of-custody manifest; route each part to redeploy, resale or destruction; and document the disposition for audit, environmental compliance and, increasingly, embodied-carbon and Scope 3 reporting. The asset record is the spine of all of it: an untracked serial is an unsanitized serial and an unrecoverable dollar.

IT Asset Disposition is the industrialized channel that turns a retired part into either recovered capital or certified e-waste. The 2026 reality is that accelerator generations now land on a roughly annual rhythm (GB200 → GB300 → Vera Rubin NVL72, shipping since June 2026; Rubin Ultra/Kyber is NVIDIA's H2 2027 target, which SemiAnalysis reports as 2028), against the legacy 5–7-year enterprise cycle. That compressed cadence can expand secondary-market supply when owners actually retire equipment. That supply creates opportunities for smaller inference buyers, and price risk when supply outruns qualified demand. Product launches alone do not establish those retirements. The lever practitioners actually pull is resale velocity: obtain dated, executable bids per SKU and track realized value net of fees, freight, sanitization cost, warranty state and export-control status as each new generation resets the market. If a bid expires while the parts wait at the dock, the delay can become a loss. Holding a part on purpose wins when what it still earns in place, or the value of keeping it for rollback or repair, exceeds the carrying cost and the sale alternative.

Export control can turn a resale into a violation. Advanced accelerators remain subject to the US Export Administration Regulations after their first sale. The chips are ECCN 3A090, and the boards and servers that carry them are 4A090 or a .z item such as 5A992.z. Exporting, reexporting or transferring them to a destination in Country Groups D:1, D:4 or D:5, which include China, needs a license. So does a sale to a buyer headquartered in Macau or in a D:5 country such as China, or to one whose ultimate parent is headquartered there, wherever the hardware sits. A broker’s resale abroad does not lift the liability off you if you knew, or ignored red flags that should have told you, where the parts were going. The controls remain in force as of October 2026, and the Bureau of Industry and Security’s diversion guidance of May 13, 2025 doubles as the due-diligence checklist. Before you accept a bid, get the buyer’s end-user and end-use certification, screen the buyer and the final destination against the Consolidated Screening List, and put a no-re-export clause in the ITAD contract. → how controls differ by jurisdiction in Chapter 11.11.

Vendor selection is a compliance decision with teeth. The certifications that matter are R2v3 and e-Stewards for responsible recycling and chain-of-custody, ISO 14001 for environmental management and, increasingly, i-SIGMA’s NAID AAA certification for data destruction. They help you vet the provider and its audited scope. The per-serial custody record, verified execution, validation decision and disposition receipt prove what happened to a particular asset. A provider certificate alone does not prove the part was sanitized or destroyed rather than quietly resold. Skip certified ITAD and the downside is concrete: a tracked serial surfacing on the secondary market with recoverable tenant data on it, or an e-waste liability. The Basel Convention’s e-waste amendments, in force since 1 January 2025, require prior informed consent before any e-waste crosses a border, hazardous or not. National law decides whether used equipment shipped for repair or reuse counts as waste. So record at the dock whether each pallet is tested reuse stock or waste, and check that call against both the exporting and importing countries’ rules before the ITAD provider ships it. The circular-economy framing (refurbish, reuse, recover materials, certified-destroy only as a last resort) is also where refresh meets the embodied-carbon argument: a part kept in service one more cascade rung is embodied carbon you do not have to re-manufacture. → embodied carbon and circularity in Chapter 15.6.

On a live campus, refresh is a migration. It must protect goodput and the power limits while it happens, rather than fit into a single maintenance window. What the swap strands is revenue: every megawatt taken offline to receive the new generation is a megawatt not earning against a depreciation clock that is still running. The dominant pattern is therefore rolling, hall-by-hall (or row-by-row) migration. Drain and de-provision one fault domain, swap it, re-commission it, and only then move to the next. The cluster never loses more than one domain’s worth of capacity, and the new generation is validated against the live workload before the old one is fully retired. The alternative, a forklift cut-over of a whole hall at once, is faster on paper. It strands the entire hall’s revenue during the transition and concentrates commissioning risk into a single window, which is the high-risk density step-up that re-commissioning discipline exists to de-risk.

The power limit is what makes this delicate. A denser successor generation needs more total megawatts, and during the overlap window you may be running both generations, so the facility power and cooling plant must absorb the transient. The migration plan and the capacity/power plan are therefore the same plan. Where you have stranded power headroom to land the denser part dictates the order in which you swap halls. A refresh executed without that headroom reserved becomes a rebuild of the power chain, not a swap of the silicon. The re-commissioning needed to accept the new density step follows the changed thermal, structural, power and control dependencies, and the continuous re-commissioning discipline in Chapter 14.14 owns it. → capacity and power management in Chapter 14.7; risk-based density-step retesting in Chapter 14.14.

Trace. Cascade value = internal contribution + later net sale − incremental migration/holding cost = about $9k per asset from the inputs above. Sell-now value is about $10k, so choose sale once the custody, export-screening and sanitization gates are met. That is about $1k more per asset, plus immediate release of the old asset’s power and service burden. The original purchase price is sunk and stays out of this comparison.

Flip. Solve contribution + later net sale − incremental cost = sell-now value. Contribution must exceed about $4k over the three-month horizon for cascade to win, with all other assumptions fixed. A support or purge-validation failure overrides the financial ranking and sends the asset back to hold or to an approved destruction path. The comparison is incremental cash. NIST SP 800-88 Rev. 2 governs the media-specific verification and validation, and Chapter 1.8 covers the buy/rent and earning-life economics behind the inputs.