Tax Implications of GPU Purchases: Section 179, Bonus Depreciation, and R&D Credits

Section 179 of the Internal Revenue Code allows businesses to deduct the full purchase price of qualifying equipment, including GPU hardware, in the year it is placed in service rather than depreciating it over multiple years. For tax year 2026, the Section 179 deduction limit is $1.22 million (indexed for inflation), with the deduction phasing out dollar-for-dollar once total equipment purchases exceed $3.05 million. This means an AI company purchasing $2 million in GPU hardware can deduct approximately $1.17 million in the current year, with the remaining $830,000 eligible for bonus depreciation or MACRS depreciation.

The Section 179 benefit is most valuable for profitable AI companies in high tax brackets. A C-corporation paying the 21 percent federal rate plus state income tax (effective rate typically 25-28 percent) saves $293,000-$327,000 in taxes for every $1.17 million of Section 179 GPU deduction, assuming sufficient taxable income to offset. For pass-through entities (LLCs, S-corps), the deduction flows through to individual owners' tax returns, where the 20 percent Qualified Business Income deduction under Section 199A may apply, further reducing effective tax on the GPU savings.

Timing is critical. Section 179 requires the GPU hardware to be placed in service (operational and available for use) by December 31 of the tax year. Hardware received but not racked and powered by year-end does not qualify. AI companies planning large GPU purchases should schedule delivery and installation 4-6 weeks before year-end to allow for physical deployment, network configuration, and acceptance testing. A November 2026 GPU purchase delivered in December but not powered until January 2027 misses the Section 179 window for that tax year.

Bonus depreciation under IRC Section 168(k) allows businesses to deduct a percentage of the cost of qualifying property in the first year of service, on top of regular MACRS depreciation. The bonus percentage has been phasing down from 100 percent (2022) to 80 percent (2023), 60 percent (2024), 40 percent (2025), and 20 percent (2026), after which it drops to 0 percent for property placed in service after December 31, 2026, unless Congress extends it. For 2026, the 20 percent bonus applies to the adjusted basis after any Section 179 deduction, creating a combined first-year deduction of up to approximately 50-60 percent of GPU hardware cost for qualifying companies.

Bonus depreciation applies only to GPU hardware with a recovery period of 20 years or less (GPU servers are 5-year MACRS property), and the property must be new (original use starts with the taxpayer). Used GPU purchases do not qualify for bonus depreciation, which is a significant consideration for AI companies buying refurbished H100s or secondary-market hardware. For a used GPU purchase, only Section 179 (if eligible) and regular MACRS depreciation are available, reducing the Year 1 tax deduction by approximately 20-30 percentage points versus new hardware.

The bonus depreciation phase-down creates a powerful incentive for AI companies to accelerate planned GPU purchases. A company planning to acquire $5 million in new GPU hardware could deduct approximately 65-75 percent of the cost in Year 1 if placed in service in 2026 ($1.22M Section 179 + 20% bonus on remaining $3.78M = $756K + $378K MACRS Year 1 = approximately $2.35M total Year 1), versus approximately 35-40 percent if delayed to 2027 when bonus drops to 0 percent ($1.22M Section 179 + MACRS only = approximately $1.98M Year 1). The difference of approximately $370,000 in Year 1 deduction at a 25 percent effective tax rate represents $92,500 in incremental tax savings by acting in 2026 versus 2027.

The research and development tax credit under IRC Section 41 can offset GPU compute costs for qualifying research activities. The credit equals 6-14 percent of qualified research expenses (QREs) that exceed a base amount, with the 14 percent rate applying to the alternative simplified method. GPU compute costs can qualify as QREs when they are incurred in a process of experimentation to develop new or improved business components, and when they are technological in nature and substantially all activities constitute elements of a process of experimentation related to a new or improved function, performance, or reliability.

Qualifying GPU activities include: training and fine-tuning models to develop new AI capabilities, running ablation studies and hyperparameter optimization to advance model performance, conducting distributed training research across GPU cluster topologies, and developing novel inference optimization techniques (speculative decoding, KV cache compression, quantization methods). Non-qualifying activities include: running standard inference serving for existing products, performing routine fine-tuning with established techniques and datasets, software development that does not involve technological uncertainty, and maintaining production GPU infrastructure.

The documentation burden for GPU-related R&D credits is substantial. Companies must contemporaneously document: the technological uncertainty being addressed, the alternative approaches considered, the GPU experiments conducted (GPU-hours per experiment, model architectures tested, dataset variations), and the results and conclusions. The IRS has increased scrutiny on AI-related R&D credit claims, with audit rates for R&D credits reaching 12-15 percent in 2025-2026 compared to the general corporate audit rate of 0.5-1.0 percent. The most common audit issue is distinguishing qualifying experimentation from routine product development. A well-documented claim can support $150,000-$500,000 in annual R&D tax credits for a mid-market AI company spending $2-5 million on GPU compute.

The optimal entity structure for GPU-intensive AI companies depends on whether the GPUs are held for internal use, for resale, or for providing GPU-as-a-service to customers. Internal-use GPU companies (training their own AI models) benefit from a standard operating entity that claims Section 179, bonus depreciation, and R&D credits against operating income. GPU-as-a-service companies should consider forming a separate equipment-holding entity (either a corporation or a disregarded LLC) that owns the GPUs and leases them to the operating entity. This structure preserves the tax attributes of the GPUs (depreciation, interest deductions) in the holding entity and may allow them to be allocated to investors or partners with better tax positions.

State tax treatment of GPU purchases varies dramatically. California, New York, and Massachusetts conform to federal Section 179 rules but with lower annual limits ($200,000-$500,000 in some states). States like Nevada, Texas, and Florida have no state income tax, making them advantageous for GPU ownership. Property tax is another consideration: GPU hardware in a California data center is subject to personal property tax at approximately 1.1-1.4 percent of assessed value annually. A $10 million GPU deployment in Santa Clara County would incur approximately $110,000-$140,000 in annual property taxes. Comparable deployment in Oregon (no state-level sales tax but lower property tax at approximately 0.7-0.9 percent) would save $40,000-$70,000 annually, and in Nevada (no property tax on tangible personal property for businesses), the savings reach $110,000-$140,000 annually.

Sales tax on GPU purchases can add 5-10 percent to the effective hardware cost. Companies purchasing GPUs for direct use in manufacturing or R&D may qualify for sales tax exemptions in 30+ states. The manufacturing exemption typically requires that the GPUs be used directly in the production of tangible personal property (which can include software under certain state interpretations). The R&D exemption is narrower but available in states like California and Texas. AI companies should work with a multistate tax advisor to determine exemption eligibility and ensure proper exemption certificate documentation before purchase, as retrospective exemption claims are often limited to 12-24 months.

The tax treatment of leased versus purchased GPU hardware produces materially different outcomes that should factor into the lease-vs-buy decision beyond the direct financing analysis. Purchased GPUs generate depreciation deductions (Section 179, bonus, MACRS) and interest deductions if financed, which can offset taxable income. Leased GPUs generate rent deductions that are typically less front-loaded than depreciation but easier to claim (no complex depreciation calculations, no bonus depreciation phase-down concerns). For a profitable AI company, the net present value of tax benefits is typically 15-25 percent higher for purchased hardware than leased hardware over a 3-year period, assuming the company has sufficient taxable income to absorb the deductions.

The comparison changes for companies without current taxable income (i.e., most pre-Series B AI startups). Depreciation and interest deductions provide no current tax benefit to loss-making companies, and net operating losses (NOLs) may have limited carryforward value depending on jurisdiction and ownership changes under IRC Section 382. For these companies, the tax analysis favors leasing, which avoids creating NOLs that may expire unused or be limited upon a future acquisition or IPO. A loss-making startup should not make GPU purchase decisions based on tax benefits that it cannot currently realize, unless it has a clear path to profitability within the NOL carryforward period (20 years for federal purposes, shorter in some states).

The lease classification (operating vs. finance under tax law, which does not always align with ASC 842 classification) determines the tax treatment. Under tax law, a lease is classified as a true lease (rent deduction) or a conditional sale (depreciation + interest) based on the economic substance of the arrangement. The IRS uses a facts-and-circumstances test focused on whether the lessee bears the economic risk of ownership. Leases where the lessee is obligated to purchase the GPUs at lease end, where the lease term covers substantially all of the useful life, or where the lessee has an option to purchase at a bargain price, will typically be recharacterized as conditional sales for tax purposes, eliminating the rent deduction and substituting depreciation and interest. Proper lease structuring for tax requires careful attention to these criteria.