A new TechInsights report published this week poses what may be the defining question of China’s AI decade: can the country deliver datacenter-grade artificial intelligence performance without access to Nvidia’s leading chips? The report, titled “The $100 Billion Question: Can China Deliver Datacenter AI Performance Without NVIDIA?” is behind a paywall, but its central thesis is relevant to an ongoing debate among semiconductor analysts, AI developers, and investors on whether China can deliver full datacenter AI performance without Nvidia.
The short answer emerging from that debate is: not yet at parity, but closer than most Western observers assumed two years ago.
What It Is Actually About
The framing matters. Critics of China’s semiconductor strategy often benchmark it against Nvidia’s latest Blackwell architecture, a comparison that, by design, China cannot win. Nvidia’s datacenter GPUs are fabricated on TSMC’s 3nm and 4nm processes, offering transistor densities and power efficiencies that SMIC, China’s leading foundry, cannot currently match. U.S. export controls have restricted China’s access to the advanced lithography equipment needed to close that gap quickly.
But a growing number of analysts argue this is the wrong benchmark. The relevant question is not whether a single Chinese chip can match a single Nvidia chip, but whether China can assemble enough compute, at sufficient cost efficiency, to train and run the AI models its industry needs. On that measure, the picture looks different.
Semiconductor industry observers, including analysts who have reviewed the TechInsights methodology, note that China’s strategy centers on two pillars: scale and packaging. SMIC is aggressively expanding its wafer fabrication capacity, aiming to roughly double output on its N+2 node, a 7nm-class process, over the near term. Meanwhile, Huawei’s semiconductor division has invested heavily in advanced 3D chip stacking and chiplet architectures, allowing multiple smaller dies to be combined into a single high-bandwidth package that delivers more aggregate compute than any individual die could achieve.
Huawei’s Ascend 950 as a Case Study
Huawei’s Ascend 950 AI processor, reportedly manufactured on SMIC’s N+2 node, has become the focal point of this debate. Independent developer testing, including published benchmarks from the AI developer community, suggests the Ascend 950 can run large language models like DeepSeek V4 at commercially viable throughput, even if per-chip FLOPS fall short of Nvidia’s H100 or B200.
The key insight from developers who have worked with both platforms is that DeepSeek V4’s architecture was explicitly designed with hardware efficiency in mind. As EastFrontier reported when DeepSeek V4 launched, the model’s mixture-of-experts design activates only a fraction of its parameters per token, dramatically reducing the raw compute required per inference call. That architectural choice makes it more forgiving of lower per-chip performance, and more compatible with the kind of scaled, packaged hardware China is building.
This is not a coincidence. Chinese AI labs and Chinese chipmakers are co-evolving in ways that their Western counterparts, who design models for Nvidia hardware, are not. The hardware shapes the software, and the software shapes the hardware.
The Limits of the Strategy
Honest assessments of China’s approach acknowledge real constraints. Yield rates on SMIC’s N+2 node remain below those of TSMC’s comparable processes, meaning more wafers are needed to produce the same number of functional chips, raising costs and limiting how quickly capacity can be deployed. Power consumption per FLOP is also higher on older process nodes, which matters in large-scale datacenter environments where electricity costs are a significant operating expense.
There is also the question of the frontier. Even if China closes the gap on today’s Nvidia hardware, Nvidia is not standing still. The Blackwell Ultra and Rubin architectures in Nvidia’s pipeline will extend the performance lead further, and the export control regime means China will not have legal access to them.
What China’s strategy buys is time and independence, the ability to run its domestic AI industry without being subject to foreign supply chain decisions. Whether that is sufficient depends on how quickly Chinese AI labs can continue to extract more from less, and how long the current export control framework holds.
Implications for China’s Broader AI Ecosystem
The datacenter hardware question is not purely technical. It has direct implications for the economics of China’s AI industry. As EastFrontier has reported, companies like SenseTime are pivoting to cost-efficient AI models and overseas deployment precisely because domestic compute costs, while falling, remain higher than in markets with unrestricted Nvidia access. The pricing war now underway among Chinese AI model providers, with DeepSeek V4 at $1.74 per million tokens forcing competitors to cut prices aggressively, is partly a function of this hardware reality: efficiency is not just a product feature, it is a survival strategy.
The silicon wafer self-sufficiency drive, in which China is targeting 70% domestic use of advanced wafers by year-end, is the upstream complement to this story. Chips are only as independent as the materials that go into them, and Beijing is moving on both fronts simultaneously.
A Debate Worth Watching
The TechInsights report has crystallized a conversation that was already happening across the semiconductor and AI investment communities. The consensus, to the extent one exists, is that China has made more progress than export control advocates hoped, and less progress than Beijing’s most optimistic projections suggested. The gap is real, but it is narrowing, and the pace of narrowing depends on variables, including SMIC’s yield improvements, Huawei’s packaging innovations, and the continued evolution of software-hardware co-design in Chinese AI labs, that are genuinely difficult to predict.
For investors and policymakers tracking China’s AI trajectory, the datacenter hardware question is not a binary, China can or cannot do this, but a gradient that is moving in one direction.