Huawei Connect 2026 – Still No Contest

Huawei fixes the bits that it can.

• Huawei is holding its annual convention that showcases its upcoming innovations, and while it is clearly making progress on removing the bottlenecks from its AI SuperPoD system, the gap between the Ascend series and Nvidia continues to widen.
• There are bottlenecks all over the data centre, which result in underutilisation of GPUs as they wait for data from other GPUs and memory, which is carried by the networking system that ties the chips together.
• These bottlenecks are substantial which means that there are substantial performance gains to be had from eliminating them.
• Huawei feels this problem more acutely than anyone else because, in order to create a meaningful compute cluster where GPUs are all working together, it has to connect far more GPUs than anyone else does.
• For example, Huawei’s 2027 chip, the Ascend960, when used in the SuperPod can produce 8 ExaFLOPS of FP8 (16 ExaFLOPs of FP4) compute when 4096 chips are connected in a single system, which is in line with what the company announced last year.
• This is where Huawei’s problems begin, as Nvidia Rubin will produce the same amount of compute with 640 Rubin GPUs (6.4x fewer).
• This also means that the Huawei system is much more difficult to build and manage, as 6.4x more chips need to be connected together, managed and communicated with to produce the same amount of compute.
• The net result is that Huawei’s SuperPod system costs more than 3x to build and consumes 9x the power to produce an equivalent amount of compute compared to Nvidia.
• However, the far greater complexity of the Huawei system creates opportunities for improvement by finding and removing bottlenecks, which is precisely what Huawei has been doing for the last 12 months.
• This is why Huawei’s announcements at Connect 2026 are focused on fixing the bottlenecks, because when it comes to improving the compute production per chip, there is very little that Huawei can do other than make the chip bigger.
• This, on its own, is fraught with peril because the larger one makes the chip, the risk of imperfections that make the chip unusable increases exponentially.
• Huawei’s manufacturing process is a 7nm multi-patterning process which has more steps than EUV and already has much lower yields.
• This is why David Wang, Deputy Chairman of the Board, spent his keynote (see here) announcing UnifiedBus, which improves networking speed and efficiency, OceanStor M900 that improves memory performance, and AI-scale Clos architecture that allows up to 1m Ascend chips to be networked together.
• All of these put together will help the Huawei Ascend-powered SuperPod AI cluster to produce more compute, but it is very unlikely to make a meaningful dent in the chasm of performance that exists between it and Nvidia, AMD or any of the other Western players that have access to advanced silicon chip manufacturing.
• Hence, I continue to think that the SuperPod system will have very little appeal outside of China unless it is accompanied by enormous subsidies.
• Furthermore, I would not be surprised if the one system I am aware of outside of China (Malaysia) ends up replacing the older Huawei chips with Nvidia or something else which its system has been designed to be able to support.
• Hence the “over 1,000 Atlas 900 A3 SuperPods” that have been deployed are almost certainly all in China, where companies have very little choice, as they still are unable to buy Nvidia or anyone else.
• The net result is that Huawei Connect 2026 leads to no changes in RFM’s conclusions that were published in February 2026 (see here).
• Hence, an investor in a Huawei data centre would expect to lose 98% of their investment over 10 years at a price of $10bn/GW.
• At $40bn / GW, this improves materially to a 5-year IRR of 16%, but this is way behind Blackwell which offers a 5-year IRR of 65% at the same price.
• No rational person will choose Huawei over Nvidia or anyone else who has access to advanced silicon.
• This means that other countries are more likely to build with Nvidia, meaning that the US strategy to contain China within its borders still has the advantage for now.