Huawei CloudMatrix 384

What it is

CloudMatrix384 is a Huawei Cloud service instance that Huawei Cloud built on top of Atlas 900 A3 SuperPoD systems^[1]. The name needs unpacking. A SuperPoD is not a single server, and it is not a rack of ordinary machines joined by a standard network. It is a group of accelerator chips wired together with a dedicated high-speed interconnect so that software can treat the whole group as one computer. In the Atlas 900 A3 SuperPoD that group is made of up to 384 Ascend 910C chips; Huawei says the 384 interconnected chips work as a single computer and deliver up to 300 PFLOPS^[1]. A petaflop is a thousand trillion floating-point operations per second, so 300 PFLOPS describes the aggregate throughput of the whole system rather than of one chip.

CloudMatrix384 is the cloud layer on that hardware: a customer rents the system as a service instance instead of buying machines and building a cluster.

Where it comes from

Huawei launched the Atlas 900 A3 SuperPoD in March 2025^[1]. By September 2025, Eric Xu said the Atlas 900 A3 SuperPoD remained the largest SuperPoD in the world^[1]. Huawei said it had deployed more than 300 Atlas 900 A3 SuperPoDs, serving over 20 customers in sectors including internet service providers, telecoms and manufacturing^[1]. CloudMatrix384 is that same hardware delivered in cloud form, built on top of those SuperPoDs^[1].

The interconnect is the point

The hard part of a SuperPoD is not only how many chips it holds but how those chips talk to each other. The SuperPoD interconnect uses Huawei's UnifiedBus protocol, whose 2.0 specifications Huawei released at the same event^[2].

Why does this matter so much? Training a large model is a collective job. A chip finishes part of a calculation and then has to exchange results with the others, over and over. The more chips take part, the more often data has to move, and the movement can become the bottleneck rather than raw arithmetic. That is why a system like this is designed around bandwidth, latency and topology as much as around the chips themselves: several hundred accelerators have to behave like one machine, not like several hundred machines that occasionally send each other files.

The roadmap around it

Huawei presented the CloudMatrix384 service alongside a longer plan. The Atlas 950 SuperPoD carries 8,192 Ascend 950DT chips, and Huawei claims a 17-fold training performance improvement over the Atlas 900 A3 SuperPoD^[1]. Above that sits the Atlas 950 SuperCluster, which Huawei says will link more than 520,000 Ascend 950DT chips across over 10,000 cabinets for 524 EFLOPS in FP8, going on sale in the fourth quarter of 2026^[1].

The Atlas 960 SuperPoD is described as holding up to 15,488 Ascend 960 chips in 220 cabinets, 176 for compute and 44 for communications, occupying 2,200 square metres, for the fourth quarter of 2027^[1]. Huawei says it will deliver 30 EFLOPS in FP8 and 60 EFLOPS in FP4, with 4,460 TB of memory and 34 PB/s of interconnect bandwidth^[1]. FP8 and FP4 are low-precision number formats: they store each number in fewer bits, which lets a given amount of hardware do more operations per second, at some cost in precision. That trade-off is common in AI computation.

All of these performance figures are Huawei's own announcements rather than independent measurements^[1].

Where it sits

CloudMatrix384 represents one way of competing in AI infrastructure: rather than arguing only about single-chip specifications, build very large systems, invest in the interconnect that holds them together, and expose the result through a cloud service. For users, the interesting part is the delivery model. A rented instance means a team can reach the aggregate capacity of hundreds of chips without constructing and operating a cluster itself.

The deployment numbers Huawei reports — more than 300 SuperPoDs and over 20 customers across internet service providers, telecoms and manufacturing — indicate that systems of this shape are already in production use rather than only on slides^[1]. The announced follow-on systems extend the same idea in two directions at once: more chips inside a single SuperPoD, and clusters that link many SuperPoDs into installations of hundreds of thousands of chips^[1].