Buying and allocating GPUs and hardware for AI data centers

In short

AI data center operators buy GPUs outright, lease them through operating or capital leases, or rent them by the hour from cloud providers. The right model depends on how much you’ll use the GPUs and your tax and capital structure. Federal tax rules cut the cost sharply. Section 179 lets a buyer deduct up to $2,560,000 of qualifying equipment cost in the year it is placed in service. 26 U.S.C. § 179. For larger fleets, 100 percent bonus depreciation has no dollar cap and applies to qualified property acquired and placed in service after January 19, 2025. IRS Form 4562 Instructions, 26 U.S.C. § 168(k). Export controls under the Export Administration Regulations restrict where the most advanced GPUs can be sent. A license is required for certain countries, but a January 2026 rule moved NVIDIA H200 and AMD MI325X chips to China from a denial to a case by case review with strict conditions. BIS Final Rule. Large GPU fleets are often financed through GPU backed debt and special purpose vehicles, and off balance sheet SPV structures alone have moved more than $120 billion in under two years. Law firm analysis. Supply remains tight. NVIDIA controls about 85 percent of the accelerator market and allocates most new production to the largest cloud companies. Next generation Rubin GPUs are scheduled for the second half of 2026. High bandwidth memory and advanced packaging are the main supply bottlenecks. Arc Compute, Fusion Worldwide, Spheron, Tom’s Hardware

What are the main ways to acquire GPUs for an AI data center?

There are four basic models and one newer hybrid. An operator can buy GPUs outright, lease them, rent them by the hour from a cloud provider, or commit to reserved instances. NVIDIA has also begun directly leasing GPUs to large customers.

Introl report

An 8 GPU H100 server costs $320,000 plus $80,000 for networking, storage, and infrastructure, and total deployment costs reach $500,000 per node when including AI data center space, power, and cooling. Introl report. That scale explains why most operators finance, lease, or rent instead of paying cash for large clusters.

NVIDIA has started offering direct GPU leasing to large customers like OpenAI. OpenAI saves 10 to 15 percent on chip costs compared with buying. NVIDIA keeps the residual value and can use the lease contracts as collateral for its own borrowing. News

How much does a GPU cost, and when is buying better than renting?

A single H100 GPU cost between $25,000 and $40,000 in late 2025. The H200 carries a 15 to 20 percent premium. Introl report. In the cloud, the same GPU can be rented for as little as $1.49 an hour from a budget provider or $3.90 an hour from AWS. Introl report

A common rule of thumb is that buying makes sense only when the GPU runs continuously more than 60 to 70 percent of the time. If usage is less than about 12 hours a day, cloud rental is usually cheaper. Introl report

Tax savings from bonus depreciation can tip the scales. A $500,000 node that qualifies for 100 percent bonus depreciation can generate a $500,000 immediate deduction. That effectively lowers the after tax cost by the operator’s tax rate. For a well capitalized AI data center operator with sustained workloads, direct purchase often produces the lowest net cost.

How do operators finance large GPU purchases?

Banks and private credit funds now treat GPU fleets as collateral, similar to aircraft or real estate. Special purpose vehicles, or SPVs, and operating lease structures help sponsors move GPU spending off their balance sheets.

Several recent deals show the scale.

• CoreWeave announced a $7.5 billion debt facility backed by its GPU fleet in May 2024, led by Blackstone. Law firm analysis
• Google backstopped Fluidstack’s lease payments to TeraWulf in a $3.2 billion deal and received warrants for up to 14 percent of TeraWulf. Google has repeated this structure, committing over $5 billion in similar backstops. Law firm analysis
• Meta’s Hyperion AI data center used a $30 billion SPV financing. Law firm analysis
• Oracle raised $18 billion in corporate bonds to fund Stargate AI data centers. Law firm analysis
• In total, technology companies moved more than $120 billion in AI data center spending off their balance sheets in under two years through SPVs and operating leases. Law firm analysis

Four US Senators urged regulators in January 2026 to look at the growing reliance on debt in the AI sector, warning that opaque and complex debt markets could cause destabilizing losses. Law firm analysis

The CHIPS and Science Act allocated $52.7 billion for semiconductor manufacturing and research. CHIPS Act. That law supports building domestic fabrication plants, which could eventually ease the GPU supply crunch, but it does not directly subsidize the purchase of GPUs for AI data centers.

What federal tax rules help reduce the cost of GPU hardware?

Two federal provisions let a buyer deduct the cost of hardware in the year it is placed in service. This lowers the after tax cost of acquiring GPUs and related equipment.

Section 179 expensing

For tax year 2026, a business can immediately deduct up to $2,560,000 of the cost of qualifying equipment placed in service during the year. 26 U.S.C. § 179. The deduction begins to phase out dollar for dollar when total qualifying purchases exceed $4,090,000. It is fully eliminated at $6,650,000. 26 U.S.C. § 179. GPU servers, networking switches, and storage arrays are tangible personal property that qualifies.

Section 179 cannot create a net operating loss. The deduction is limited to the taxpayer’s taxable income from business. But it is elective, so a taxpayer can choose to use it or not.

100 percent bonus depreciation

The One Big Beautiful Bill Act, signed July 4, 2025, restored 100 percent first year bonus depreciation for qualified property acquired after January 19, 2025. 26 U.S.C. § 168(k). Unlike Section 179, bonus depreciation has no dollar cap and can create a net operating loss that offsets other income. GPU servers are five year Modified Accelerated Cost Recovery System (MACRS) property, so they qualify.

A large cluster will quickly exceed the Section 179 cap. For a hypothetical deployment of 1,000 H100 GPUs at roughly $30,000 each, the total cost is $30 million. Section 179 would cover only a fraction. Bonus depreciation can deduct the full $30 million in the first year, saving perhaps $6.3 million in federal tax at a 21 percent corporate rate.

Both provisions can apply together. Section 179 is taken first, and bonus depreciation applies to any remaining basis. IRS Form 4562 Instructions A cost segregation study can separate out shorter lived components like cabling and power distribution, accelerating deductions further. Cost segregation guide

What export controls apply to GPU purchases and leases?

The Bureau of Industry and Security (BIS) controls exports of advanced GPUs under the Export Administration Regulations (EAR). BIS advanced computing export controls. So an exporter needs a license to ship these chips to certain countries.

GPUs are assigned Export Control Classification Numbers (ECCNs). The most advanced data-center GPUs, including the NVIDIA H100, H200, B200, and AMD MI300X, fall under ECCN 3A090.a. BIS Commerce Control List. A license is required to export these chips to Tier 3 countries including China, Russia, Iran, and North Korea. Before January 2026, the policy was generally a presumption of denial. Lower-tier chips fall under ECCN 3A090.b, and less restricted ones under 3A991.

In January 2026, BIS issued a final rule that moved exports of NVIDIA H200 and AMD MI325X to China and Macau from a presumption of denial to a case by case licensing review. A license is still required, but the application is reviewed on its merits instead of being denied automatically. BIS Final Rule. To qualify, several conditions must be met.

• The chip’s Total Processing Performance must stay below 21,000.
• DRAM bandwidth must stay below 6,500 GB per second.
• The exporter must show that US domestic supply is adequate.
• China shipments cannot exceed 50 percent of the exporter’s US shipments by volume.
• Third party US testing is required.
• Know your customer procedures must be in place.
• A separate Presidential Proclamation on January 14, 2026 imposed a 25 percent tariff on imports of certain advanced computing chips. CNBC, White House

Even with US approval, China blocked H200 imports at customs in early 2026. Chinese authorities warned domestic companies not to purchase the chips unless necessary. Reuters. ByteDance nonetheless prepared H200 orders worth about $14 billion for 2026 after the rule change. Tom’s Hardware

An earlier arrangement, announced in August 2025, allowed sales of NVIDIA H20 and AMD MI308 chips to China in exchange for 15 percent of revenue paid to the US government. Data Center Knowledge. Chips above that performance threshold, including the H100, H200, and Blackwell series, remained barred.

The Remote Access Security Act passed the House on January 12, 2026 and would extend export controls to cloud GPU rentals. Tom’s Hardware. If it becomes law, a Chinese company that rents GPU capacity from a US cloud provider could trigger export licensing requirements. As of May 2026, the bill had not passed the Senate.

AI data center operators must comply with the EAR permanently. Even a non US operator can face liability if restricted GPUs or sanctioned end users are present in its facility. Civil penalties can reach $374,474 per violation or twice the transaction value. Criminal penalties can reach $1 million and 20 years in prison. BIS regulations, Law firm analysis. The BIS Affiliates Rule, which would have extended restrictions to foreign entities majority owned by Entity Listed persons, was suspended for one year in November 2025. Operators still must respond to know your customer and end use screening requests from sellers and maintain due diligence safeguards. Data Center Knowledge

How do supply bottlenecks affect buying decisions?

NVIDIA supplies about 85 percent of the AI accelerator market. Persistence Market Research. Its data center revenue reached $115.2 billion in fiscal year 2025 and $193.7 billion in fiscal 2026. NVIDIA, NVIDIA. Hyperscalers lock up most of that output.

Cloud giants secure roughly 65 percent of GPU production through multi year purchase agreements. Introl report. In the first year after a new architecture launch, NVIDIA allocates 60 to 70 percent of production to the largest cloud providers. Enterprise and government buyers compete for the remaining capacity. Introl report

In 2026, the five largest cloud providers plan to spend between $660 billion and $690 billion on capital expenditures, with the vast majority going to AI infrastructure. The Futurum Group. That spending drives the intense competition for GPU supply.

Supply improved for the Hopper generation in 2025, which led to sharp price cuts. AWS reduced H100/H200 pricing by 44 percent in June 2025. Introl report. Blackwell GPUs, however, were sold out through 2025 with a 12 month waitlist. About 3.6 million units were allocated in advance before Meta’s own large allocation was included. News

The next generation, Vera Rubin, entered full production in the first quarter of 2026. The Vera Rubin NVL72 rack contains 72 GPUs and 36 CPUs. It delivers 3.6 exaFLOPS of FP4 inference performance and draws 120 to 130 kilowatts per rack, requiring direct liquid cooling. Partner availability is expected in the second half of 2026. StorageReview, Grokipedia NVIDIA estimates committed demand of at least $1 trillion through 2027. Arc Compute

Two components constrain supply. One is TSMC’s Chip on Wafer on Substrate (CoWoS) advanced packaging. CoWoS capacity is not expected to meaningfully expand until 2027. The other is high bandwidth memory. HBM4, used in Rubin GPUs, is about twice the die size of equivalent DDR5 DRAM. Yields are still ramping and will likely limit Rubin availability for the first 12 to 18 months. Arc Compute

For any buyer, the practical lesson is to plan orders well ahead. A buyer who needs a few hundred GPUs this year should expect to compete with hyperscalers and may need to place multi year commitments or work through a system integrator with allocation.

How do operators allocate GPUs across workloads and users?

Inside an AI data center, GPUs are usually treated as whole devices. Kubernetes, the most common orchestration tool, assigns GPUs to containers through device plugins. Native Kubernetes does not split a GPU between containers without NVIDIA’s Multi Instance GPU technology (MIG) or application level multiplexing. vCluster guide

Workload patterns are shifting. In 2025, inference accelerators accounted for 54.23 percent of AI data center GPU revenue, and they are forecast to grow at a 15.37 percent compound annual rate, faster than training GPUs. Mordor Intelligence That means more operators will need to allocate GPUs for mixed inference and training clusters, often on the same hardware.

For cloud users, reserved instances can cut cost sharply. A one year commitment for an AWS p4d.24xlarge instance (8 A100 GPUs) costs $13.60 an hour, down from $32.77 on demand. A three year commitment drops to $8.14 per hour. Introl report

Technology refresh cycles are short. When a new GPU generation ships, the previous generation can lose about 40 percent of its value. Generations have been arriving every 18 months or so. Introl report That rapid depreciation makes leasing and cloud models attractive for workloads that might otherwise buy hardware and hold it too long.

Key takeaways

• Match the acquisition model to your utilization. Direct purchase with tax deductions usually wins for sustained, high utilization workloads. For sporadic use, cloud rental or reserved instances cost less.
• Maximize first year tax deductions. The full 100 percent bonus depreciation is available for GPU hardware placed in service now. Section 179 can supplement, but its cap makes it a partial solution for large clusters.
• Structure financing to keep debt off the balance sheet. GPU backed loans, SPVs, and operating lease structures are widely used. The market is large but regulators are watching the build up of opaque AI infrastructure debt.
• Screen every GPU transaction for export controls. Even a cloud rental can trigger EAR obligations if Chinese end users are involved. The rules are changing quickly, and civil penalties are severe.
• Plan around supply constraints. Most new production goes to the largest cloud companies first. Smaller buyers should negotiate allocations early and understand the HBM and packaging bottlenecks that will limit next generation Rubin chips.
• Allocate GPUs with Kubernetes and MIG for efficient sharing. Reserved instances offer predictable pricing for inference workloads. Factor in the fast obsolescence cycle when choosing between lease and purchase.

Frequently asked questions

Q:Can I deduct the full cost of GPU servers in the first year?

A:Yes. If the GPUs are new and acquired after January 19, 2025, 100 percent bonus depreciation allows an immediate deduction of the full cost. There is no dollar cap. 26 U.S.C. § 168(k)

Q:What is the difference between Section 179 and bonus depreciation?

A:Section 179 is elective and capped at $2,560,000 for 2026. It phases out as total purchases rise and cannot create a net operating loss. Bonus depreciation is automatic, unlimited, and can produce a loss. Both apply to GPU hardware. 26 U.S.C. § 179, 26 U.S.C. § 168(k)

Q:Do export controls apply to renting GPUs from a US cloud provider?

A:If the Remote Access Security Act passes the Senate, they will. Even now, if a cloud customer is a sanctioned end user or located in a restricted country, the cloud operator must comply with EAR licensing. BIS Know Your Customer guidance recommends that operators investigate red flags when a customer’s circumstances suggest potential diversion of restricted hardware. 15 CFR Part 732, Supplement No. 3, BIS Know Your Customer guidance

Q:How does NVIDIA allocate new GPUs?

A:NVIDIA gives 60 to 70 percent of early production to hyperscale cloud providers through multi year purchase agreements. Enterprise and government customers compete for what remains. Supply of HBM memory further constrains availability. Introl report, Arc Compute, Introl report, Fusion Worldwide analysis, Supply chain analysis

Q:What happens to GPU values when a new generation launches?

A:An H100 system can lose roughly 40 percent of its value when the next chip arrives. Generations refresh about every 18 months. That steep depreciation favors leasing or cloud rental over long term ownership for many users. Introl report

Q:Is GPU backed debt safe or risky?

A:Lenders now accept GPU fleets as collateral, but the asset class is untested through a full downturn. Senators have asked regulators to examine the risk of complex, opaque AI infrastructure debt. Borrowers and lenders should monitor loan to value ratios and understand that resale values can drop fast when new chips ship. Law firm analysis

Q:Which is cheaper, buying or renting GPUs?

A:It depends on utilization. At twenty four seven usage above 60 or 70 percent, buying and taking bonus depreciation usually delivers the lowest after tax cost. At part time usage, cloud on demand or reserved instances are cheaper. A break even analysis should include power, cooling, and staffing. Introl report

Q:What is the status of the Chinese market for advanced GPUs?

A:The US has eased some restrictions for H200 and MI325X chips on a case by case basis with conditions. China, however, blocked H200 imports at customs and warned companies not to buy. The situation is fluid and any deal involving a Chinese end user requires careful legal review. Network World, BIS Final Rule, Reuters

Subscribe to The Compute Law Brief

The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.

Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.