US export controls on advanced semiconductors accelerated a global scramble for chip independence, with China completing a 1-gigawatt data center on domestic chips, producing its first DUV lithography machines, and breaking ASML's monopoly on advanced lithography, while the US poured hundreds of billions into TSMC Arizona and South Korean fabs. The Trump administration briefly restarted Nvidia H200 shipments to China before tightening controls again, illustrating how economic pressure and security logic pull in opposite directions even within a single administration.
The global semiconductor industry is undergoing its most consequential restructuring since the invention of the integrated circuit. What began as a US policy decision to restrict advanced chip exports to China has metastasized into a full-spectrum sovereignty race, with every major economy now treating chip supply chains as a matter of national security rather than comparative advantage. The logic is straightforward: whoever controls the compute controls the AI, and whoever controls the AI shapes economic and military outcomes for decades.
What makes this moment distinctive is the speed at which the race has moved from rhetoric to concrete infrastructure. Within a span of weeks in mid-2026, trillion-dollar commitments landed in Arizona, South Korea, and Japan simultaneously, while China demonstrated it no longer needs to wait for Western permission to build frontier compute. The old assumption that export controls could freeze China's capabilities in place has been visibly falsified.
China Closes the Gap It Was Supposed to Stay In
The most strategically significant development in this period is China's demonstrated ability to build and operate frontier AI infrastructure without Western components. On July 20, 2026, Z.AI Completes 1-Gigawatt AI Data Center on Chinese Chips, a facility in Beijing powered entirely by domestically manufactured semiconductors. At one gigawatt, it is the largest known AI compute facility built without Western hardware. Partial operations were already underway at announcement, meaning this is not a paper milestone.
That same day, China's Sugon 8000 100,000-card AI Supercluster Goes Live, described as China's first domestically developed 100,000-card AI supercluster. The Sugon Dengfeng system represents a different kind of proof point: not just raw power capacity, but the ability to orchestrate massive parallel compute at the cluster level, which is where most real training workloads live.
The software layer completed the picture. On April 24, 2026, DeepSeek V4 Released on Huawei Chips, 1.6T Parameters, marking the first major DeepSeek models natively optimized for Huawei Ascend hardware rather than Nvidia. The models, including V4-Pro at 1.6 trillion parameters with a 1-million-token context window, were priced at approximately $0.87 per million output tokens, well below US rivals. The combination of domestic chips, domestic clusters, and domestic models optimized for that stack is precisely the closed loop that export controls were designed to prevent from forming.
Then came the lithography news that rattled markets. On July 27, 2026, China-Backed Firm Breaks ASML Monopoly on Advanced Lithography, with a state-backed Shanghai company beginning production of advanced lithography machines. The following day, reports confirmed China Begins Production of DUV Lithography Machines, triggering a global tech market selloff. Deep ultraviolet lithography is the workhorse technology for manufacturing the chips that fill those data centers. ASML's monopoly on this equipment was one of the last structural chokepoints the West retained. Its erosion does not mean China can immediately manufacture leading-edge chips at scale, but it means the trajectory of the chokepoint is now downward.
The US and Allies Pour Concrete
The Western response has been to accelerate domestic manufacturing investment at a scale that would have seemed implausible five years ago. The anchor is TSMC's Arizona expansion. On July 18, 2026, TSMC Expands Arizona to $265B with CoWoS AI Packaging Capacity, adding dedicated CoWoS advanced packaging across four fabs. A few days later, TSMC Adds $100B to Arizona Investment, Total Hits $265B confirmed the full commitment. CoWoS is the 2.5D packaging technology required by every major AI accelerator, including Nvidia's H100 and H200 series, making this expansion directly relevant to the AI chip supply chain rather than general semiconductor capacity.
AMD's product announcements in late July illustrated what that investment enables. At its Advancing AI 2026 conference, AMD Launches EPYC Venice: First x86 CPU in Volume on TSMC 2nm, alongside the Instinct MI450 accelerator series. Volume production on 2nm is a process node milestone that translates directly into performance per watt advantages for data center operators.
Japan moved simultaneously. On July 16, 2026, Japan's Noetra Launches 140MW National AI Factory with 27,500 Rubin GPUs, a $6.2 billion facility backed by SoftBank, NEC, Sony, and Honda, targeting physical AI for robotics, manufacturing, and healthcare. This is explicitly framed as national infrastructure, not a commercial data center, which signals how Japan's government is thinking about AI compute access.
South Korea Bets Its Industrial Base
South Korea's commitments in the final week of July 2026 were extraordinary in scale. On July 25, Samsung and SK Hynix Sign $950B AI Chip Supply Deals Through 2030, locking in high-bandwidth memory supply for years. The $950 billion figure represents the total combined commitments from both Samsung and SK Hynix across multiple partners: Samsung signed a deal with Broadcom worth over $200 billion, while SK Group committed $750 billion with Nvidia and a broader coalition of US technology firms. HBM is the memory technology that determines how much data an AI accelerator can process per second, and South Korea's two dominant chipmakers have now contractually tied their output to the US AI hardware ecosystem through the end of the decade.
The same day, SK Group and NVIDIA Sign $500B+ AI Factory and Memory Deal, covering a 2-gigawatt AI data center in South Korea using Nvidia's DSX platform and Vera Rubin GPUs, plus a long-term HBM4 co-development agreement with SK Hynix. NAVER added to the picture with NAVER, NVIDIA and Brookfield Triple Korea AI Factory to 200MW, a $10 billion expansion of the GAK Sejong facility, with Nvidia taking a binding $1 billion equity stake in NAVER. Nvidia taking equity in a national AI infrastructure operator is a structural alignment, not just a hardware sale.
Also on July 25, Samsung Wins $200B Broadcom AI Chip Foundry Partnership, further cementing South Korea's position as the memory and advanced packaging backbone of the US-aligned AI hardware stack.
The Trump Administration's Contradictory Signal
Against this backdrop of hardening supply chain blocs, the Trump administration sent a confusing signal. On July 14, 2026, Nvidia H200 AI Chip Shipments to China Confirmed Restarted, with a US trade official confirming that high-end AI chip exports to China had resumed under evolving export control rules. The H200 is Nvidia's most capable chip that had previously been subject to export restrictions.
This reversal is significant and should not be buried. It illustrates the fundamental tension between the economic logic of chip sales, Nvidia's revenue, US fab utilization, and the security logic of denying China advanced compute. The restart came on the same day the Sugon 8000 supercluster went live, a coincidence that underscores how the policy environment and the technical reality are moving in different directions simultaneously. Export controls that are lifted under commercial pressure and then potentially retightened again create exactly the uncertainty that accelerates China's domestic substitution efforts.
Europe Enters Late
The European Commission announced on August 1, 2026, EU Pledges $11.5B for Seven AI Gigafactories, complementing 19 existing AI factories with advanced processors, cloud infrastructure, and high-speed connectivity. At $11.5 billion, the EU commitment is an order of magnitude smaller than the TSMC Arizona expansion alone, which raises real questions about whether Europe is building sovereign capacity or building dependency on a different set of foreign suppliers.
What to Watch
Whether the H200 restart becomes permanent policy or gets reversed again. The July 14 restart is the clearest indicator that economic pressure can override security logic within the Trump administration. Watch for any formal rule change or Congressional response that locks in either direction.
The yield and volume ramp of China's DUV lithography machines. Beginning production is different from producing at competitive yield and cost. The next 12 months of output data from the Shanghai facility will determine whether ASML's monopoly erosion is a slow leak or a rupture.
Whether DeepSeek V4's Huawei-native optimization translates into third-party adoption. If other Chinese labs begin training on Ascend hardware using V4-style architectures, it signals the domestic stack has crossed a credibility threshold.
TSMC Arizona's CoWoS capacity timeline. Advanced packaging is the current bottleneck for AI accelerator supply. When Arizona CoWoS comes online and at what yield will determine how much of the AI hardware supply chain is actually onshored versus still dependent on Taiwan.
The EU's ability to attract anchor tenants for its gigafactories. Announced funding is not deployed capacity. Watch for which hyperscalers or national champions commit to the seven facilities and on what terms, as that will reveal whether European AI sovereignty is a real industrial project or a subsidy program in search of customers.
This piece was originally published on Present of AI, where we cover what AI is actually doing in the world, no hype. Read more or get it in your inbox.
Top comments (0)