Compute Is The New Oil: Export Controls And AI Hardware

Compute Is The New Oil: Export Controls And AI Hardware

Data was supposed to be the new oil. Then generative AI arrived and taught the market a more expensive lesson: data is useful, but it does very little without huge amounts of compute, fast memory, advanced packaging, power, and network links. By early 2026, Nvidia had reached $215.9 billion in annual revenue. At the same time, the company said its near-term outlook assumed no China data center compute revenue, which shows how valuable AI compute has become and how political access now shapes the market.

Since 2022, U.S. export controls have grown from a focused set of chip limits into a wider system touching tools, memory, packaging, end users, and even the location of data centers. Export controls on AI chips now influence roadmaps, sales plans, and factory investment. The old semiconductor business was hard enough; now every product launch also needs a geography lesson.

Why Compute Became Strategic

Compute now means a full industrial stack. TSMC said it saw robust AI-related demand through 2024 and lifted revenue 30% in U.S. dollar terms. SK hynix then posted record 2025 results driven by AI memory and HBM, a high-speed memory used in AI accelerators. When one part of the stack gets tight, demand quickly rushes to the next bottleneck.

Governments noticed the same thing. BIS (the U.S. Bureau of Industry and Security) built its rules around advanced computing chips, supercomputer-related end uses, semiconductor manufacturing tools, and later high-bandwidth memory because these parts matter for both commercial AI and military capability. The semiconductor supply chain is now part of national security policy.

A Short Timeline Of The New AI Trade Rules

• October 2022: BIS launched the first big wave of controls aimed at China’s access to advanced computing chips and semiconductor manufacturing equipment.
• October 2023 And April 2024: The U.S. tightened and clarified the rules, including broader licensing limits and added notification duties.
• December 2024: BIS added controls on HBM, more software and tool categories, and 140 entities.
• January 2025 To May 2025: The AI Diffusion rule tried to create a global framework for advanced compute, then was rescinded before compliance began.
• January 2026: BIS reopened a narrow path for some chips, saying H200-, MI325X-, and similar exports to China could be reviewed case by case under strict conditions.

How U.S. Export Controls Changed The Game

The January 2025 AI Diffusion rule made the new logic very clear. It would have required authorizations for advanced chips in many more countries, created special paths for trusted data center operators, and set country-level allocations for controlled compute. The rule was rescinded in May 2025 before its compliance date.

In January 2026, BIS shifted again and said exports of Nvidia H200, AMD MI325X, and similar chips to China could be reviewed case by case if firms met security conditions, including U.S. testing and customer safeguards. Supply chains hate surprises more than they hate bad news. When the rulebook moves this fast, companies stop waiting for certainty and start designing around politics. 

Why The Market Moved Faster Than The Rules

The new rules are also highly technical. Nvidia told investors that China now controls total processing performance, performance density, interconnect bandwidth, and memory bandwidth. So engineers do not only ask whether a chip is fast. 

BIS’s January 2025 update asked foundries and packaging companies to verify certain designs and transistor counts, while the Netherlands expanded national license rules for more DUV, or deep ultraviolet, lithography systems in 2024 and more measuring and inspection tools from April 2025. A modern chip sale can now depend on who designed the chip, who packaged it, and where the shipment might end up two steps later.

Nvidia, AMD, And The Art Of The Compliant Chip

Nvidia shows how costly this can be. In its 2026 annual filing, the company said April 2025 restrictions on H20-class products led to a $4.5 billion charge tied to inventory and purchase obligations. It also said licensed H20 sales to China brought in only about $60 million. A few months later, Nvidia told investors it was still assuming no China data center compute revenue in its near-term outlook.

Nvidia’s geographic revenue based on customer headquarters showed China, including Hong Kong, falling to $19.7 billion in fiscal 2026 from $25.0 billion in fiscal 2025, even as total revenue rose to $215.9 billion. Nvidia also warned that, by the end of fiscal 2026, it was effectively shut out of China’s data center compute market. In its own words, that gap helped rivals build larger customer and developer ecosystems.

AMD faced a smaller version of the same problem, which was still a very expensive problem. The company said new April 2025 license rules hit its MI308 products, caused about $800 million in inventory-related charges, and later allowed some shipments that helped it reverse about $360 million of those charges. AMD also warned that export controls push Chinese customers toward local alternatives and create uncertainty around any replacement rules. The Nvidia China strategy may get more headlines, but the pressure is spread across the whole U.S. accelerator business.

How Chipmakers Are Adapting

• Use Special Product Lines And License-Based Sales: Restricted markets are increasingly served through narrower product paths and tightly controlled approvals.
• Build Compliance Into The Product Itself: Packaging, testing, customer screening, and end-user checks now sit much earlier in the design and shipment process.
• Shift Attention To Friendlier Markets: Companies are putting more weight on trusted cloud operators, approved customers, and capacity growth outside the most restricted channels.

ASML, TSMC, HBM, And The New Bottlenecks

ASML said Dutch changes in September 2024 moved more DUV immersion systems under national license requirements, and the Dutch government widened the measure again from April 2025 to cover more measuring and inspection technologies. Even with tighter rules, ASML said it expected China business to be around 20% of its 2025 net sales.

Japan is part of this map too. In 2023, Japan said it would add 23 items of semiconductor manufacturing equipment to export control, arguing that advanced chips can be turned to military use and that policy had to match a tougher security environment. The semiconductor supply chain is global by design. The strongest controls are the ones allies can enforce together.

Then there is TSMC. In its 2024 annual report, TSMC said AI-related demand was robust, revenue rose 30% in U.S. dollar terms, and customers wanted geographic flexibility plus government support. The company highlighted expansion in Arizona, Japan, Germany, and Taiwan. Its January 2026 earnings materials also said AI accelerators accounted for a high-teens share of 2025 revenue.

Memory became strategic as well. BIS added new controls on high-bandwidth memory in December 2024, and SK hynix later said HBM revenue more than doubled in 2025 as the company delivered record annual results driven by AI memory. In the AI era, the winning stack is no longer GPU plus software. It is GPU plus HBM plus advanced packaging plus a flag on the right side of the geopolitical map.

Cloud Access And Sovereign AI

The January 2025 AI Diffusion rule proposed special authorizations for trusted operators, divided approved data center users into broad and country-specific categories, and tied licenses to the scale of compute that could be built in each place. Even though that rule was later rescinded, it showed where policy is heading. Governments care about who owns the cluster, where it sits, and who can rent it.

This is one reason “sovereign AI” became a serious business phrase rather than a conference slogan. TSMC openly said customers value geographic flexibility and government support, while BIS later issued guidance on protecting supply chains and on the risks around Chinese advanced computing chips. The new AI compute infrastructure is being built with a mix of engineering logic and diplomatic logic. That is efficient only if you enjoy spreadsheets full of export codes.

What This Means For China

For China, the effect is mixed and very clear at the same time. Nvidia said export controls helped competitors build stronger ecosystems, AMD said the rules encourage Chinese customers to seek local substitutes, and BIS in May 2025 specifically warned industry about risks tied to PRC advanced computing chips, including Huawei Ascend parts. Controls can slow access to the best foreign chips. They also give local champions a stronger reason to improve fast.

Three Trends To Watch

• More Controls On Compute Access: Policy is likely to keep moving from simple chip shipment limits toward control of cloud clusters and data center scale.
• More Value In Bottlenecks: HBM, advanced packaging, and trusted foundry capacity will keep gaining importance because these are the tight points in the AI stack.
• More Regional Product Roadmaps: Chip plans will increasingly be built with export thresholds and country rules in mind from day one.

The New Oil Now Comes With Paperwork

Oil shaped the twentieth century because it powered transport, industry, and armies. Compute is starting to play a similar role in the AI age because it powers models, data centers, research labs, and military-related technology. The companies that win will do more than build faster chips. They will manage licenses, memory supply, packaging capacity, friendly jurisdictions, and political trust.

That is the real meaning of “compute is the new oil.” The value is not only in the raw resource, but in the routes, the refining points, the storage, and the permission to move it. In the AI hardware market, export controls are reshaping all of those layers at once. The new oil still moves, but now it travels with a customs file.