Tech Tonic | Can China’s 14nm chips really rival Nvidia and AMD’s AI silicon?

If you’re keeping score, that’s like claiming your 2021 Nissan Patrol can outdrag a 2024 Porsche 911 because you installed a really clever turbocharger. Well, more surprising things have happened. Here’s the thing — Wei knows what they are claiming. And that possibility should make the AI companies rethink some fundamental assumptions about the chip race. The claim hinges on architectural innovation rather than brute-force miniaturisation. Instead of pursuing ever-smaller transistors, the path that’s led to 3-nanometer chips and required billions in EUV, or extreme ultraviolet lithography investments towards semiconductor manufacturing, Chinese researchers are betting on 3D hybrid bonding and near-memory computing.

The approach they’re taking is, stacking logic chips directly onto DRAM using copper-to-copper connections at microscopic scales, creating what amounts to a fundamentally different beast than traditional GPU architecture. In an AI chip, a DRAM is fundamental since it provides the high-speed temporary memory necessary to process vast amounts of data.

Here are the headline specs of the Chinese chip approach — 120 TFLOPS of performance at 2 TFLOPS per watt efficiency (TFLOPS or teraflops, measures AI compute power), which would theoretically place this design in Nvidia’s Hopper or even Blackwell territory for certain workloads. The architecture supposedly achieves this by eliminating something called a ‘memory wall’, which is a fundamental bottleneck where GPUs spend enormous energy and time shuttling data between processing cores and memory banks. By fusing logic and memory into the same package with thousands of high-bandwidth vertical connections per square millimetre, the design promises to significantly reduce latency as well as power consumption. Performance per watt — the mission for every chip maker.

The physics may well be checking out but, and this is a substantial but, there’s still a canyon-sized gap between theoretical architectural advantages and shipping silicon that performs as advertised. First, thermal management in 3D stacks can prove notoriously difficult because when logic and memory get sandwiched, it is essentially creating a heat-generating source with limited airflow. Second, manufacturing complexity may be a challenge because hybrid bonding requires near-perfect wafer alignment and surface preparation.

China’s leading foundry, SMIC, has proven capabilities at 14nm production, but scaling hybrid bonding for logic-memory integration is an entirely different ballgame. Yield rates could be low initially, and that’ll drive costs high. That’s to be expected, but what’ll matter at this stage is the appetite for navigating this tough stage.

Critically, there’s the software ecosystem problem. Wei himself identified this as the existential threat, calling Nvidia’s CUDA platform a ‘deep pit’ of dependency that has the entire AI industry reliant to at least some extent. He’s right about the problem, but underestimates the solution’s difficulty. Even if China builds a chip that’s genuinely superior on paper, adoption requires seamless integration with comprehensive developer tools, extensive documentation, and experienced programmers. Nvidia didn’t become dominant just because their hardware was fast—they became dominant because they made it easy to use.

Chinese companies like Cambricon and Huawei are working on this with frameworks like NeuWare and migration tools for CUDA-trained models. But creating a parallel software ecosystem that can genuinely compete is a decade-long project, not a quick fix.

So, what’s really happening here? This announcement is part technical achievement, part strategic positioning, and part aspirational roadmap. US export controls have effectively locked China out of advanced lithography equipment and cutting-edge GPU technology. They can’t purchase Nvidia’s latest hardware, for instance, and use those AI chips as they are meant to be used. In that context, architectural innovation isn’t just clever, but perhaps the only viable path forward. There are calls within the US, by AI companies mostly, that they need to be shielded by the government for future loans to drive investments, else they risk falling behind China. Patriotism can be a viable business call too.

Whether these specific chips can truly rival Nvidia or AMD’s silicon remains to be seen, but the broader pursuit makes sense. If you can’t compete on transistor size, compete on system architecture. If you can’t match CUDA feature-for-feature, build something different enough that direct comparison becomes meaningless. The geopolitical implications remain significant. Wei explicitly framed this effort as achieving technological sovereignty and breaking free from US control over the AI hardware stack. That’s not just nationalist rhetoric but more a strategic necessity for a country that views AI leadership as essential to economic and military competitiveness.

For the rest of us watching this play out, the takeaway shouldn’t be whether China’s 14nm chips can truly match 4nm Nvidia GPUs on every benchmark. The real question is whether the semiconductor industry’s obsession with smaller nodes has blinded us to alternative paths to performance. If Chinese engineers can genuinely achieve competitive AI performance using mature process nodes and advanced packaging, it doesn’t just change China’s position—it potentially democratises high-performance AI by making it accessible without billion-dollar fab investments.

Technical details may be released soon. Until then, healthy skepticism is warranted, but so is attention. The history of technology is littered with moments when everyone assumed the future would be linear, only to be blindsided by someone who changed the playing field entirely. Netflix putting an end to CDs, DVDs and Blockbuster is an example. Toyota’s focus on efficient and compact cars caught traction in the US, a complete opposite to the Detroit boardrooms that thought massive vehicles were the ticket. We don’t know where this Chinese AI chip pursuit will go. But this may well be a moment to reconsider what’s actually possible when you can’t rely on Moore’s Law to solve your problems. Either way, the chip wars just got more interesting.

——————————–

Vishal Mathur is the Technology Editor at HT. Tech Tonic is a weekly column that looks at the impact of personal technology on the way we live, and vice versa. The views expressed are personal.