Your server room isn't real estate.
Depreciate it accordingly.

Why this is simpler than a cost segregation study

A cost segregation study exists to find the personal property hiding inside a building — the carpet, the specialty electrical, the decorative lighting — and reclassify it out of 39-year real property into 5, 7, or 15-year life. A rack of GPUs doesn't need to be found. It was never real property to begin with.

The IRS's own asset class list (Rev. Proc. 87-56, Asset Class 00.12, "Information Systems: Computers and Peripheral Equipment") puts computers and related hardware at a 5-year recovery period from the day they're placed in service. Servers, GPUs, networking gear, storage arrays — all of it starts life in the fast lane. No engineering study, no cost allocation, no Whiteco factors. It's a classification question with an answer the code already gives you.

That matters more than it used to, because the compute itself is no longer incidental. A law firm's file server used to be a rounding error next to the office lease. A room of ten to forty GPU servers, purchased so a bank, a clinic, or a startup can run its own models instead of renting someone else's, is often the single largest capital outlay the business makes that year.

A room of ten computers, for very different reasons

"Running your own models" sounds like a startup thing. In practice, the businesses with the strongest reasons to own the hardware outright are often the least startup-like.

None of these are real estate businesses. That's the point. Equipment depreciation doesn't require one. Any business that buys and places tangible personal property in service can use it — the mechanics below apply exactly the same whether the building is owned, leased, or a single rented office suite.

Standard MACRS vs. full expensing

Left alone, 5-year property is written off on a declining-balance schedule over six tax years (the half-year convention stretches a "5-year" life into six partial years). Almost nobody leaves it alone.

The One Big Beautiful Bill Act permanently restored 100% bonus depreciation for qualified property acquired and placed in service after January 19, 2025 — the same provision that makes cost segregation studies land their full value in year one now applies directly to the compute hardware, automatically, with no study required. Bonus depreciation isn't elective by asset; it applies to the whole class unless the business affirmatively elects out.

Where Section 179 fits

Section 179 is the other lever, and it solves a different problem. Where bonus depreciation is close to automatic, Section 179 is an elective, per-item deduction capped at your taxable income for the year — which makes it the more precise tool when a business wants to control exactly how much it deducts rather than take the full amount. Under OBBBA, the Section 179 cap rose to $2.5 million, with the phase-out starting at $4 million of total equipment placed in service.

The room is a different question entirely

The servers are personal property from day one. The room built to hold them is where cost segregation actually re-enters the picture — because that room is an interior buildout, and interior buildouts are exactly what a cost segregation study is built to take apart.

A rack of GPUs draws serious power and throws off serious heat. Housing it properly usually means a dedicated electrical subpanel and circuits sized well above office standard, precision cooling (in-row or CRAC units, not a bigger office HVAC unit), raised or reinforced flooring, fire suppression rated for electrical equipment, and often a card-access security layer. None of that is "the building." All of it is a leasehold improvement or, if the taxpayer owns the shell, a component of the building that a cost segregation study can pull out of 39-year (or 15-year QIP) life into 5, 7, or 15-year categories — dedicated electrical and precision cooling land in 5- or 7-year property; reinforced flooring and standalone fire suppression often land in 15-year.

Why this skews so much higher than a typical study

We've written before about manufacturing buildouts — a bakery outfitting a new production warehouse is a good comparison, because it shows why an AI compute buildout behaves so differently.

In a typical manufacturing cost segregation study, the building shell dominates the budget. A warehouse might reclassify 25–35% of its basis into 5-, 7-, and 15-year components — ovens, refrigeration, specialized electrical, dock equipment — with the rest locked into the 39-year shell. Once bonus depreciation is applied to the reclassified share and equipment is added on top, it's common to see a business recover something in the neighborhood of 40–50% of total project cost in year one. That's a genuinely strong result, and it's the number most cost segregation case studies report.

An AI compute buildout inverts the ratio. The hardware isn't a small slice of a much larger shell — it's usually the majority of total project cost, and it's already short-life property before a single reclassification decision gets made. Run the same logic on a compute-heavy buildout and the year-one share climbs well past the manufacturing case:

That's roughly 92% of a $2 million buildout expensed in year one — not because the tax planning is more aggressive than the bakery's, but because so much more of the spend was 5-year property from the start. This is illustrative, not a promise; the actual split between hardware and infrastructure, and the reclassifiable share of the infrastructure layer, depends entirely on the specific buildout and should be modeled against real invoices, not a template.

Recapture, and the one place the timeline actually helps

Fast depreciation isn't free. Everything expensed under bonus or Section 179 is subject to Section 1245 recapture — sell or dispose of the equipment above its (now near-zero) tax basis, and the gain comes back as ordinary income, not capital gain.

On a building, that creates a real tension: cost segregation shortens the tax life dramatically while the physical components often keep working for decades, so the recapture exposure can sit on the books for years after the deduction is spent. GPU hardware doesn't have that problem, or at least not as badly. The AI hardware market's real-world obsolescence cycle — typically three to five years before a cluster is replaced, resold, or written off entirely — tracks reasonably close to the tax life it was already assigned. The asset that got the fastest write-off is also the one most likely to actually be gone, functionally, by the time recapture would otherwise bite.

Two questions, not one study

If your business owns the compute — a room of GPUs, a bank's on-prem inference cluster, a clinic's local model server — the hardware itself needs no engineering study to depreciate fast; it was 5-year property the moment it was placed in service, and current law lets you expense all of it immediately. The room built to house it is the real cost segregation question, and it's usually where a second, sizable deduction is sitting unclaimed. Model both layers, not just the invoice for the servers.