Domestic supernodes now beat Nvidia on total compute, but 5x the power buys no ecosystem
Huawei's CloudMatrix 384 trades roughly 560 kW of power for 1.7x the total compute of NVL72, but what decides the fate of domestic supernodes is not specs — it's production capacity and software ecosystem.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Supernodes became the default exhibit at the show
More than 10 supernode products were shown at this year's WAIC. Huawei, Alibaba and Baidu, chipmakers including Biren, Moore Threads, MetaX and Enflame, and even server vendors like ZTE and Sugon all put supernodes in the most prominent spot on their booths. Brokerages are calling 2026 the first year of the domestic supernode. Behind this homogenized competition is strong FOMO: without a supernode, you're embarrassed to show up. Supernodes have evolved from 8-card machines into rack-scale forms of dozens to over a hundred cards, and the core purpose is to solve the communication wall in model training and inference.
— YaxianSingle chips can't catch up, so make it up at the system level
Domestic AI chips lag clearly behind overseas ones at the single-chip level, and that gap is hard to close quickly, so performance has to come from the system level. Kimi's K3 model has 2.8T total parameters and 896 experts, selecting 16 experts per token, and the official recommendation is a supernode of 64 or more accelerators to solve capacity, bandwidth and latency problems. At the same time, interconnect technology can already make supernodes a product, and vendors are willing to sell a whole rack rather than single cards, because the value is completely different. Together these factors make supernodes the mandatory path for domestic compute to catch up.
— Zhang HaijunThe Scale-up ecosystem is fragmented
Almost every supernode scale-up protocol in the world is proprietary: Nvidia has NVLink, Huawei has UB, and Alibaba and domestic chip companies each have their own. Overseas, UALink was initiated by AMD, Marvell, Broadcom and others, but AMD is the main driver, Marvell makes switch chips yet lacks the motivation, and Broadcom pulled out to do its own protocol, leaving the open protocol loose. Huawei's UB can unify scale-up and scale-out, while Nvidia's NVLink only handles in-rack interconnect, with IB between racks. Fragmented protocols mean ecosystem lock-in, and domestic substitution is not just a hardware problem.
— Zhang HaijunBenchmark comparison: NVL72 versus Huawei 384
NVIDIA NVL72 consists of 18 compute trays and 9 NVSwitch trays, 72 B200s in total, with whole-machine power above 120 kW, and early GB200 shipments were mostly air-cooled, not all liquid-cooled. Huawei's CloudMatrix 384 is China's first supernode, 16 racks with 384 910Cs in total. Single-chip BF16 compute is only 30% of B200's, but total system compute is 1.7x that of NVL72. The price is whole-machine power of about 560 kW, close to 5x Nvidia's. This is buying total compute by piling on hardware, and it also explains why China has to pursue system-level innovation.
— Zhang HaijunSystem vendors refuse to just assemble
Traditional server OEMs like ZTE and H3C also rolled out supernode designs this time. ZTE used its networking expertise to develop its own switch chip, with a protocol called Olink that it says is compatible with several domestic chips, though no details were released. Lightelligence is betting on optical interconnect, positioning itself as a supplier of optical communication and optical switching components. Right now the only supernodes actually in mass production and use in China are Huawei's and Sugon's: Huawei's customers are spread across carriers and cloud vendors, while Sugon is tied to Hygon and focused on xinchuang. Other vendors' claims of "zero adaptation" are still a long way from large-scale commercial use.
— YaxianCan domestic supernodes beat Nvidia
Zhang Haijun argues that domestic supernodes have already surpassed Nvidia on system specs: CM384's total compute is 1.7x NVL72's. But his definition of "surpass" is: if Nvidia opened up the market and cloud vendors, under equal conditions, would rather choose a domestic supernode — that's when you've won. Right now that's hard, because the CUDA ecosystem is too deep, domestic engineers are used to CUDA, DeepSeek can even find bugs in Nvidia chips, and switching to domestic hardware and software means relearning; Nvidia is also more stable. But Nvidia's next-generation product delay gives domestic players an opening, and the software ecosystem has improved noticeably over the past year, so the gap is narrowing. From 0 to 1 China is weak; from 1 to 100 China is stronger.
— Zhang HaijunProduction capacity is the real bottleneck
Many chipmakers does not mean large capacity; the real choke point is AI chip production capacity. The price of an 8-card B300 machine rose from 3.7-3.8 million yuan in last year's Q3 to 12-13 million yuan this year, showing that overseas goods can't get in and are scarce. Big internet companies are buying more domestic chips this year than last, and one important reason is that domestic capacity is relatively reliable. Because each company's software ecosystem is incompatible, big players can't adapt to all chips at once, so if a second-tier vendor hasn't landed a big order from a major player by next year, it will be very hard afterward. The industry landscape may reshuffle like autonomous driving, leaving only five or six players. Huawei and Cambricon are stable, and big-company in-house efforts like T-Head and Enflame also have a shot.
— Zhang HaijunSupernodes pull up power, liquid cooling and switch chips
Supernode power has surged, and Nvidia has been pushing 800V high-voltage DC power since last year; liquid cooling has gone from optional to standard, with GB200 later switching to liquid cooling and GB300 fully liquid-cooled, and Google data shows liquid cooling demand value in 2027 will be 4x the previous level. On the network side, China previously relied mainly on Broadcom switch chips, and now domestic switch chips can rival Broadcom. Broadcom still dominates scale-out, while scale-up is split among proprietary protocols. These supporting technology upgrades will likely release commercial opportunities earlier than supernodes themselves.
— Zhang HaijunIn their own words · checked verbatim
From 8-card machines to supernodes, one of the main changes is really to solve the communication wall problem.
从巴卡机到超节点，它的一个主要变化，其实是为了解决通信墙的问题。
Zhang Haijun6:35
Single-chip compute may be only 1/30 of theirs. But once you put them together into a supernode, the total compute is actually about twice theirs.
单片算力是可能只有它的多少30分之1。但是合在一起会做成超点点以后，其实总算力是它大概两倍。
Zhang Haijun23:21
It's possible for us to surpass Nvidia. Because our characteristic is that from 0 to 1 we may be weaker than the US side. But from 1 to 100, I think we will definitely be stronger than the US side.
我们超NV是有可能的。因为我们的特点是说，我们可能从0到1可能会比美国那边要弱一些。但是我们从1到100，我觉得肯定是要强于美国那边的。
Zhang Haijun32:54
Right now production capacity will be a bottleneck, meaning whoever has capacity may have a bigger advantage.
现在产能会是一个瓶颈，就是说谁油产能，可能谁的优势就会更大一些。
Zhang Haijun37:02
If you still haven't landed a big order from one of these major internet companies, then I think it will be quite hard afterward.
你还没有从这几家大的互联网厂商拿到一个大的订单的话，那我觉得后面可能就比较难了。
Zhang Haijun41:06
Figures
| Number of supernode products (WAIC 2025) | More than 10 | 4:26 |
| K3 total parameters | 2.8T | 11:51 |
| K3 number of experts | 896 | 11:51 |
| NVL72 rack power | Over 120 kW | 22:10 |
| 910CA BF16 compute vs B200 | About 30% | 23:21 |
| CM384 total compute vs NVL72 | 1.7x | 23:21 |
| B300 8-card machine price (last year Q3) | 3.7-3.8 million yuan | 38:02 |
| Google 2027 liquid cooling demand value | 4x the previous level | 44:23 |
Glossary
- Scale-up
- Combining multiple chips into one giant GPU to raise single-task compute.
- Scale-out
- Expanding cluster scale by adding more racks or nodes.
- UALink
- An open accelerator interconnect protocol initiated by AMD, Broadcom and others, whose participants are currently loosely organized.
- CloudMatrix
- Huawei's multi-rack interconnect supernode line, such as the 384-chip design.
How to listen
Cloud procurement teams watching domestic AI compute substitution, chip startups, and investors in the compute sector who need to judge technology routes and the inflection point for orders.
The opening trade-show griping (0:00-4:30) can be skipped; start at 5:28 with the rundown of supernode vendors.