If there were concerns building about an AI infrastructure bubble forming, last week’s Liberation Day tariffs — and the subsequent pause — surely liberated our thinking from that.
As we’ve previously detailed, the ingredients of a bubble were all around: speculators clogging load interconnection queues at utilities, hyperscalers committing hundreds of billions in capex to data centers, and an increasing role of debt financing in the buildout of digital and energy infrastructure. At the time, we thanked the natural recalcitrance of utilities, grid operators, and regulators as a necessary break on runaway investment. The process of securing power was enough of a barrier to unchecked growth to make a true asset bubble unlikely.
But now we have tariffs. As we’re already seeing, they may not be set at any specific rate for long, but they are clearly here and their influence is already profound. For those looking to rapidly scale AI data centers, the challenge a month ago was gaining access to affordable, long-term power contracts. Going forward, the challenge will clearly be the affordability and availability of nearly everything. Consider the modern data center’s inputs:
- GPUs and semiconductors: Semiconductors are currently exempt from tariffs, though the exemption only applies to direct semiconductor imports. GPU board assemblies, like the ones Nvidia imports from Taiwan, currently don’t appear exempt. Additionally, chip manufacturing equipment is subject to tariffs, which can increase costs for any domestic manufacturers. Nvidia and TSMC are making moves to manufacture their chips and solutions in the U.S., but they are years from production.
- Transformers and electrical equipment: Transformers are critical for data center power distribution and may be sourced from abroad, making them susceptible to tariff impacts. The 25% tariffs on steel and aluminum directly impact the cost of transformers, switchgear, and other components for data centers. Transformers rely heavily on steel cores, and with over 80% of large power transformers (LPTs) imported into the U.S., these tariffs exacerbate existing shortages and increase procurement costs. Additional tariffs of nearly 125% have been imposed on China as of Wednesday afternoon, and even during the newly-announced 90-day pause, other key importers like Taiwan and South Korea, which supply critical electrical components and data center equipment, face a 10% tariff.
- Power generation and uninterruptible power supply (UPS) backup systems. Accessing power for AI data centers typically requires new generation be added either to the grid or onsite, at unprecedented scale. Tariffs will now impact the price and availability of solar, batteries, and gas power turbines. In addition, backup power systems, often a mix of fast-response battery systems and longer-duration backup generators, are all very exposed to tariffs and supply chain disruptions.
- Construction materials, such as steel and aluminum, are vital for building data centers and are subject to tariffs, increasing construction costs. CBRE recently estimated that tariffs will raise construction costs for commercial projects by 3% to 5%, and impact nearly all electrical, structural and cooling systems.
- AI server racks and networking equipment: Networking equipment is largely imported from China, Vietnam, Taiwan, and the European Union, (which is set to impose its own tariffs on U.S. imports of up to 25% in response.) Those imports are essential elements of any cloud infrastructure, which will increase costs for hyperscalers, regardless of whether the data center houses AI systems.
The Trump Administration has repeatedly spoken to the importance of developing AI as a geopolitical priority, particularly in competition with China, and promoted Stargate as an example of critical investments to keep pace. Yet while it held a press conference on accelerating the federal government’s commitment to AI, it simultaneously announced sweeping tariffs that will make it increasingly expensive and difficult to build AI infrastructure, with no coherent explanation.
It’s worse than that. The AI market is well on its way to one that prioritizes unit economics over exponential growth metrics. To its credit, most infrastructure investment in AI has come from the balance sheets of major tech companies, so it’s their risk to take. But these are all publicly traded companies (other than OpenAI) and will need to exhibit a path to profitability.
Costs matter. Microsoft’s recent moves to rationalize its data center footprint is less evidence of it pulling back from AI than a cost-conscious approach to capex.
Constraints matter, too. As the DeepSeek saga taught the world, when China is faced with constraints (availability of AI chips) it doesn’t retreat, it innovates around them and devises a solution that may in fact be preferable to those in the West. An emphasis on clever software engineering over brute force computing led to potentially much better energy efficiency, which addressed one of the key hurdles in the AI race. The DOE, expanding on a Biden-era initiative to accelerate the building of AI data centers on federal land, can impact only part of the path to AI growth — land, permitting, and other federal resources — and the rest is left to the market, which is getting an entirely different signal from the administration on tariffs.
