America’s AI race is accelerating at a blistering pace, and with it, the construction of the most expensive computing infrastructure in history. But behind the headlines about eye-watering data center buildouts lies another, quieter challenge that’s been shaping the economics of U.S. AI growth: tariffs.
In 2025, according to industry estimates, U.S. server manufacturers and hyperscale cloud companies are expected to collectively pay several billion dollars in tariffs on imported components that power AI systems. Tariffs of this nature are unprecedented, as historically, electronics were duty-free.
The question emerging among manufacturers, economists, and supply chain leaders is increasingly straightforward: how much is America paying in tariffs to build its AI future — and at what cost to U.S. competitiveness?
How We Got Here: The Tariff Stack Behind Every AI Server
First, some quick orientation. If you’ve ever seen a picture of a data center, you’ve seen the rows of 6-8 foot tall electronics cabinets – fed with wires and pipes. These cabinets are called “racks” – and each rack contains multiple “tray assemblies”. Trays contain GPU, CPU, or TPU chips, as well as networking equipment and thermal management equipment, which enable the systems to function. For example, a top-of-the-line AI rack such as NVIDIA’s NVL72 has 18 compute trays and nine switch trays.
A modern, top-of-the-line AI server — such as those that run NVIDIA or AMD GPUs or Google TPUs — is a marvel of global supply chains. Tens of thousands of components for each tray flow into assembly lines from places around the world, many with deep ties to Asia. Final assembly is done in factories around the world, with particular concentration in Mexico, Taiwan, China, and other countries in Greater Asia. The United States is also in the process of bringing multiple tray and rack-level assembly factories online; however, capacity has not yet caught up to that of other regions. Most of these top-of-the-line systems are destined for U.S. data centers. The Dell-Oro Group reports that the top four US-based cloud service providers – Amazon, Google, Meta, and Microsoft – account for nearly half of global data center capex in 2025. These racks, with their tens of thousands of individually taxable components, are subject to existing and new tariffs on electronic imports to the United States.
For many Chinese-sourced components and assemblies, U.S. importers continue to face the Section 301 tariffs of either 7.5% or 25% introduced during the first Trump administration. In addition, Section 232 duties apply to steel at 25% and aluminum at 10%, directly affecting server chassis, power supply housings, and thermal frames. The Biden administration did not reverse or scale back these tariffs, and their decision to maintain these duties reflects the extent to which U.S. importers had already adapted to the tariff structure over the previous years.
Many electronics companies relocated their assembly operations to othe countries to help mitigate the duty impact. With the new 2025 tariffs and geopolitical climate, this has resulted in a spectrum of success.
Companies that reshored to Vietnam in response to the 2018 wave of tariffs on Chinese goods were caught by the 2025 tariffs stating that tray and rack assemblies deemed “Made in Vietnam” are subject to a 20% duty, wiping out much of the savings those manufacturers had hoped to gain from relocation.
Taiwan, by contrast, received a significant break. On April 11, 2025, the U.S. granted tariff exemptions for Taiwanese servers and computer parts, substantially lowering the total cost of systems built in factories in Taiwan versus those in China – assuming their components come from Taiwan, and aren’t sourced from other countries.
While the USMCA was supposed to make Mexico a nearshore, duty-free option for electronics assembly, due to the complexity of the global supply chain, a server and rack assembly that has been assembled in Mexico from printed circuit boards made in Taiwan, with connectors, cables, enclosures, and thermal systems made in China, may still be subject to significant duties because the assembly may not qualify for a substantially changed country of origin to be called “Country of Origin Mexico.” To further complicate it, under the USMCA rules, certain assemblies could be marked as “Made in Mexico” under the rules of the agreement, but not qualify for a true country of origin shift to avoid the Section 301 or Section 232 impact
The result is a challenging tariff landscape where most supply chain leaders are forced to make difficult tradeoffs, but essentially cannot avoid some amount of tariff on each multi-million dollar rack imported to the United States.
Running the Numbers: What Tariffs Add to a Modern AI Server
Top-of-the-line AI racks can cost as much as $3-4M. Most of that cost is in the advanced AI chips, which when assembled to circuit boards, can represent 70-80% of the total Bill of Materials cost. Many of these advanced chips are manufactured by TSMC in Taiwan. If purchased on a printed circuit board, which is common, generally, that assembly will have been done in Taiwan. Both would be “duty-free”.
Assuming a $3M rack assembled outside of the U.S. or Taiwan, if 80% is effectively assembled in Taiwan, then $600K of the remaining materials may be subject to duties at rates between 20-25% if they originate outside the U.S., Taiwan, or Mexico. Common origins for cables, connectors, cooling equipment, and chassis are China and countries in greater Asia subject to reciprocal tariffs. If all of the other parts in the rack aren’t duty-free, the duties could be as much as $120K per rack.
Per-Rack Tariff Impact:
- Top-of-the-line AI Rack cost: $3-4M
- Percentage of the Bill of Materials that is Duty Free: 80%
- GPU/CPU/TPU chips mounted on printed circuit boards
- Components subject to tariffs after April 2025 exemptions: 20%
- Tariff exposure calculation: 20-25%
- Total tariff cost per rack: $120,000
Industry-Wide Impact for 2025: In May 2025, UBS analyst Timothy Arcuri reported that NVIDIA stated the hyperscalers were deploying 1000 NVL72 racks per week. Assuming 50K-60K AI racks across the whole industry:
- Low estimate: $6B in tariffs (50,000 racks × $120,000)
- High estimate: $7.2B in tariffs (60,000 rack × $120,000)
The floor for modern “AI campuses” in 2025 is usually around 100MW, with many that are much bigger – such as Stargate in Michigan (1 GW), Meta El Paso Data Center Campus (1 GW), and Homer City Redevelopment Campus (4.5 GW). 100 MW represents about 400-600 AI racks – which means tariffs may be adding $48M-$78M to the data center price tag.
Across the sector, with U.S. companies expected to deploy tens of thousands of AI racks in 2025, the industry-wide tariff burden reaches $6 billion — even under conservative assumptions.
What Manufacturers Are Saying: ‘They Will Charge Whatever the Highest Price Is’
For server OEMs, tariff uncertainty is as damaging as tariffs themselves. Importers frequently have to assume the highest possible rate because final tariff rulings on individual components can arrive months after the hardware ships.
Procurement analyst Scott Bickley describes the ripple effect clearly:
“Server makers are going to charge whatever the highest price is, and it’s the Dells and Supermicros that have to pay that,” Bickley says. “And then they’re going to pass that on to the Microsofts, to the Amazons and the Googles of the world.”
Because many AI components now have lead times of 36 to 52 weeks, manufacturers must commit to orders long before knowing the ultimate tariff liability. The result is cautious procurement, higher prices, and growing costs for big infrastructure build-outs.
The Bottom Line
Even with relatively conservative estimates, AI data centers in 2025 will pay billions of dollars in tariffs — an unavoidable, largely invisible tax on America’s bid in the AI Race. While final figures will depend on each server’s bill of materials and the evolving tariff schedule for imported components, the directional impact is clear. Tariffs are adding meaningful cost, complexity, and uncertainty to the construction of U.S. AI infrastructure.
Whether those costs help catalyze a stronger domestic manufacturing base, or simply slow the country’s AI momentum, is now one of the central strategic questions for policymakers and technology leaders navigating America’s next era of innovation.
