The US trade deficit hit a record $1.2 trillion in 2025. The single biggest driver: $450 billion in semiconductor and computing hardware imports, up 60 percent in 12 months. Every one of those chips, servers, and GPUs crossed a border, cleared customs, and settled through a payment rail.

The AI boom is not a technology story. It is a payments story.

Hyperscalers are issuing debt at record pace to fund procurement cycles that span continents. Tariff regimes are shifting month to month, adding compliance cost and FX volatility at every node. A single chip crosses approximately 70 international borders before reaching end use, according to Accenture. Each crossing is a wire transfer, a customs declaration, a currency conversion.

Start with the headline number. US imports of computers and semiconductors exceeded $450 billion in 2025, a 60 percent increase from the year prior. The United States simply cannot produce enough chips domestically to meet the demand that AI data centres require.

The US-Taiwan goods trade deficit doubled, rising 99 percent to $146.8 billion in 2025. US goods imports from Taiwan totalled $201.4 billion, up 73.3 percent from 2024. Taiwan overtook China as a monthly source of US imports for the first time in decades, driven almost entirely by advanced semiconductors and computing equipment.

Behind these imports sits an unprecedented capital expenditure cycle. The five largest US cloud and AI infrastructure providers, Microsoft, Alphabet, Amazon, Meta, and Oracle, are projected to spend a combined $660 billion to $690 billion on capex in 2026, according to Futurum Group's aggregation of individual company guidance. Roughly 75 percent of that spend is directly tied to AI infrastructure: servers, GPUs, data centre equipment. That is approximately $450 billion in AI-specific procurement, nearly doubling 2025 levels.

These companies are not funding this from cash alone. Hyperscalers raised $108 billion in debt during 2025, and analysts at Bank of America project new debt issuance reaching $175 billion in 2026. UBS estimates suggest the number could reach $230 billion to $240 billion. Operating cash flows are being consumed almost entirely by capex. The AI infrastructure build is being financed on credit, and the credit is being spent across borders.

As we covered in our analysis of the $650 billion squeeze on hyperscaler balance sheets, the financial pressure of the AI build-out is real and accelerating.

Here is why payments people should care. Every dollar of that $450 billion in imports represents a real wire transfer, a real FX conversion, and in many cases a real letter of credit or trade finance instrument. When Accenture estimates that a single chip crosses approximately 70 international borders before reaching end use, those are not abstract logistics steps. They are payment events. Customs duties, freight invoices, component purchases, assembly fees, testing charges, packaging costs. Each border crossing generates a financial transaction that settles through the global payments infrastructure.

The cross-border B2B payments market was valued at $31.7 trillion in 2024, according to FXC Intelligence, and is projected to grow 51 percent to $47.8 trillion by 2032. B2B transactions account for roughly 60 percent of all cross-border payment volume. The AI hardware boom is one of the largest single contributors to that growth.

Stripe processed $1.4 trillion in total payment volume in 2024, up 38 percent year over year. Adyen processed €1.4 trillion in 2025. Both companies reported strong growth in US-origin cross-border flows, with the US-Taiwan and US-Korea corridors among the fastest-growing B2B payment lanes. As we examined in our profile of Stripe and Adyen as the two fintech giants processing $250 billion with zero IPOs, these processors are handling volumes that rival mid-sized economies.

Correspondent banking still handles the majority of international B2B transactions. That means the AI hardware boom, the single largest cross-border procurement cycle in commercial history, is running through legacy rails. SWIFT messages. Nostro and vostro accounts. Multi-day settlement windows. The infrastructure was designed for a world where cross-border B2B volumes grew steadily, not one where a single demand driver adds hundreds of billions in annual flow within 24 months.

Consider the mechanics. A hyperscaler placing a $2 billion order for GPUs from a Taiwanese manufacturer does not wire $2 billion in a single transaction. The order fragments into dozens of component purchases, logistics payments, insurance premiums, customs duties, and assembly fees spread across multiple jurisdictions and currencies. Each fragment settles independently, often through different banking relationships. The aggregate payment infrastructure load from a single large hardware order is far greater than the headline purchase price suggests.

The strain is visible at the corridor level. The USD-TWD (US dollar to New Taiwan dollar) corridor has seen volume surge in line with the 73 percent increase in Taiwanese exports to the United States. The USD-KRW corridor is growing as Samsung and SK Hynix supply memory chips for AI servers. These are narrow corridors handling enormous and concentrated flows, exactly the conditions that stress correspondent banking infrastructure and create settlement bottlenecks.

As we explored in our analysis of the physical chokepoints threatening the AI economy, the hardware supply chain has geographic vulnerabilities that compound the financial ones. When the physical and financial infrastructure share the same narrow corridors, disruption risk multiplies.

If the payment volumes are the story, tariffs are the complication that makes every chapter more expensive.

On January 15, 2026, President Trump imposed a 25 percent Section 232 tariff on a narrow category of advanced semiconductors, including chips like NVIDIA's H200 and AMD's MI325X. The tariff specifically targets advanced computing chips not intended for end use in US data centres, with exemptions for domestic R&D and public sector applications.

One month later, the Supreme Court struck down the broader IEEPA tariffs in a 6-3 ruling on February 20, finding that the International Emergency Economic Powers Act does not grant the President authority to impose tariffs of indefinite scope. The Penn Wharton Budget Model estimates the ruling could generate up to $175 billion in potential refunds, though the mechanics of disbursement remain unresolved. The Court of International Trade has ordered refunds with interest accruing at an estimated $650 million per month, but US Customs and Border Protection has told the court it is "not able to comply" with the order due to the unprecedented volume.

The administration responded within days. A 10 percent global tariff under Section 122 of the Trade Act of 1974 took effect on February 24, applicable to products from all countries for 150 days. Section 122 had never been used before. Unlike the IEEPA tariffs, it applies uniformly and contains no country-specific exceptions.

Separately, a bilateral US-Taiwan deal set a 15 percent tariff rate on Taiwanese goods, reduced from 20 percent. Taiwanese chipmakers investing in US production can import up to 2.5 times their planned US capacity duty-free during construction, dropping to 1.5 times capacity once production is operational.

For payments infrastructure, every tariff rate change translates directly into compliance cost, customs reclassification, and payment re-routing. When a chip crosses 70 borders and each crossing carries a different tariff exposure depending on the week's policy, the compounding friction is substantial. Customs brokers must reclassify goods. Trade finance documents must be reissued. Letters of credit must be amended. FX hedges must be recalculated. Every one of those actions is a transaction with a cost.

The compounding effect matters. A 25 percent Section 232 duty on a chip that crosses into the US is straightforward enough. But that same chip may have components that crossed borders in Japan, South Korea, the Netherlands, and Germany before final assembly in Taiwan. Each crossing may trigger a separate tariff assessment under the Section 122 regime. The total tariff burden on a single finished AI accelerator is not 25 percent. It is 25 percent plus the accumulated duties from every upstream crossing, plus the compliance cost of documenting each one.

The concentration of AI hardware procurement in a handful of Asian corridors is creating FX dynamics that the currency markets have not seen before.

Taiwan's exports to the United States grew 73 percent in 2025. The United States became Taiwan's largest export market for the first time in 26 years, accounting for 30.9 percent of total Taiwanese exports. Taiwan holds $604.5 billion in foreign exchange reserves and is the 10th-largest foreign holder of US Treasuries with $310.6 billion in holdings. The scale of the USD-TWD flow means that currency movements in this single corridor can have outsized effects on semiconductor procurement costs.

The same concentration is visible in the Korean won and Japanese yen corridors. Samsung, SK Hynix, and Japanese equipment manufacturers collectively supply critical components for AI infrastructure. The AI hardware cycle is funnelling enormous FX risk into three narrow Asian corridors: TWD, KRW, and JPY. For any enterprise running cross-border procurement at this scale, hedging is not optional. It is a line item.

The quiet beneficiaries are the companies that sit between the buyer and the border. Trade finance providers issuing letters of credit for $100 million GPU shipments. FX hedging platforms managing TWD and KRW exposure for hyperscaler procurement teams. B2B payment fintechs processing the thousands of component invoices that precede every finished server.

Stablecoin infrastructure is entering the picture. Stripe's $1.1 billion acquisition of Bridge, completed in February 2025, was explicitly aimed at cross-border settlement. Bridge enables businesses to move, store, and accept stablecoins, settling transactions near-instantly at lower cost than traditional correspondent banking. As we explored in our coverage of stablecoin settlement entering the mainstream, the regulatory environment is shifting to accommodate these rails. In 2024, stablecoins moved $15.6 trillion in value, putting stablecoin transaction volume on par with Visa's.

Real-time payment rails are expanding domestically as well. FedNow processed over 8.4 million settled payments in 2025, up 460 percent from 2024, with a total value of $853.4 billion. The average payment size of $101,435 suggests FedNow is finding traction in B2B settlement, not just consumer transfers. As the domestic leg of cross-border hardware procurement settles faster, the contrast with multi-day international settlement becomes starker.

Domestic fabrication capacity is years away from meaningfully reducing import dependence. TSMC's first Arizona fab is producing 4nm chips as of 2025, but the second fab will not begin production until 2027, and advanced 2nm capacity is targeted for 2029. TSMC plans up to 12 fabs in Arizona with a $165 billion total investment, but even that scale will not dent the import deficit for years. The $450 billion annual flow is structural for the foreseeable future.

Tariff policy uncertainty creates ongoing FX volatility and compliance cost. Three distinct tariff regimes have applied to semiconductor imports in the first 90 days of 2026 alone. Each change ripples through procurement contracts, customs classifications, trade finance instruments, and payment routing decisions. For payments infrastructure providers, this is recurring revenue. For importers, it is recurring friction.

The hyperscaler debt cycle is the accelerant. With projected debt issuance potentially reaching $1.5 trillion over the coming years according to the Bank for International Settlements, the capital flowing into cross-border hardware procurement will continue to grow. That capital must settle. It must convert currencies. It must clear customs. The payment rails carrying these flows are as critical to the AI build-out as the fabs producing the chips.

Further out, agentic commerce protocols will eventually automate portions of the procurement workflow. As we covered in our analysis of the agentic commerce stack and Stripe's Machine Payments Protocol, the infrastructure for AI agents to initiate and settle B2B transactions is being built now. Automated procurement, automated customs classification, automated FX hedging. That is a 2027 to 2028 story. But the payment rails being stressed today are the same rails those agents will need tomorrow.

The irony is clean. The AI industry needs faster, cheaper, more transparent cross-border payment infrastructure to sustain its own supply chain. The payment industry needs the revenue from processing those cross-border flows. And both are constrained by the same legacy infrastructure that neither has yet replaced at scale. The Stargate project alone targets $500 billion in AI infrastructure investment by 2029, with over $400 billion in commitments within three years. Every dollar of that investment will cross a border, convert a currency, and settle through a payment rail.

If every AI chip touches a payment rail 70 times before it powers a model, who is really building the infrastructure for the AI age: the chipmakers, or the payment networks?