In the spring of 2026, a Beijing-based AI startup quietly decommissioned its last batch of NVIDIA A100 servers. Not due to hardware failure or budget cuts—but because their new Ascend 910B cluster delivered 37% lower cost per unit of compute for local large model training. This isn’t an isolated case. In Shenzhen, Chengdu, Hefei, similar replacements are happening at a pace of tens of thousands of units per week. Meanwhile, Jensen Huang smiles onstage in San Francisco: “I believe the Chinese market will eventually reopen.”
He’s entitled to that belief. But markets don’t wait.
Five years into U.S. export controls on AI chips to China, the strategy has backfired spectacularly. What began as a ban on A100/H100, then morphed into neutered “China-specific” variants like A800/H800, and now extends even to edge inference chips like the L20 requiring case-by-case approval—this escalating blockade was meant to stall China’s AI arms race. Instead, it forged a more resilient domestic compute ecosystem under pressure.
Huawei is the chief executioner. Its Ascend chips are no longer just “good enough” backups. The Ascend 910B now delivers 85% of H100’s real-world INT8 performance—at less than half the price. Crucially, Huawei has vertically integrated its entire software stack: from the CANN heterogeneous computing architecture to the MindSpore framework. This closed-loop system slashes migration barriers. “We used to think CUDA was an unassailable moat,” a CTO who’s deployed both NVIDIA and Ascend clusters told me. “Now we realize moats can be dug by hand.”
This is no accident. It’s a meticulously orchestrated “de-Americanization” campaign. Rather than direct chip subsidies, Beijing channels demand through national AI initiatives, the “East Data West Compute” infrastructure push, and mandatory inclusion in government procurement lists. Huawei leverages this captive demand, bundling Ascend into full-stack domestic AI solutions—even offering free dev kits to universities. The playbook echoes Huawei’s early-2000s assault on Ericsson and Cisco: surround the city from the countryside, then breach with price-performance and hyper-localized support.
And TSMC? The world’s dominant foundry appears neutral but stands at the epicenter. It remains Huawei’s lifeline for sub-7nm nodes (albeit via non-U.S. toolchains) and NVIDIA’s sole supplier for H100/H200. One major client is being undercut by chips TSMC itself helped fabricate; the other faces existential risk if geopolitical tensions sever supply. TSMC’s silence is a masterclass in balancing—but how long can it straddle this divide?
Recall: the U.S. Commerce Department ordered TSMC last year to halt all shipments of chips made with American technology below 7nm to Huawei. Rumors suggest SMIC’s N+2 process can now produce limited volumes of Ascend 910B, but yields and capacity pale next to TSMC’s. Huawei’s growth ceiling remains tethered to an invisible technological umbilical cord.
NVIDIA’s predicament borders on absurdity. Its China-tailored H20 chip offers just one-third the performance of an H100, yet costs only 15% less. Chinese customers vote with their wallets: buy Huawei or wait for homegrown alternatives. The “eventually reopened” market Huang envisions may have already shut its doors psychologically. Even if restrictions lift tomorrow, habits, ecosystems, and trust—once shifted—are rarely reversed.
History rhymes. In the 1980s, Japanese DRAM makers collapsed under U.S. pressure. Today, China isn’t copying—it’s reconstructing. It doesn’t ask you to comply with CUDA; it invites you to win under new rules. That’s what truly unsettles Silicon Valley.
So here’s the question nobody dares voice aloud: as computational sovereignty becomes core to national security, can the semiconductor industry sustain the illusion of “one world, one standard”? Or are we sliding toward a bifurcated future—split by divergent tech stacks, instruction sets, and ecosystem alliances? The answer may already be etched onto TSMC’s wafers.