560,000 shipped M890s say more about export controls than Alibaba's new V900 does

Divide Alibaba's shipment figure by its customer count and each buyer averages about 1,400 M890 accelerators ^[1]. At most sites that is a production fleet, not a pair of evaluation racks. Forkast puts the customers across 20 industries. It does not say how many of the units go to training and how many to inference, and those two loads exercise a chip differently ^[6].

Set the V900's top-end claim against that base. Alibaba says the part scales to 500,000 chips per cluster ^[4]; one cluster of that size would take about 89 percent of every M890 shipped since May 2026 ^[2]. The company has also said what it wants the scale for: a model of 5 to 10 trillion parameters, 20 GW of global data center capacity by 2032, and a pursuit of artificial superintelligence ^[15].

Eddie Wu, Alibaba's chief executive, called the chip "the most powerful AI chip in China today" ^[3]. The supporting figures (three times the M890's performance, 216 GB of HBM, 1.2 TB/s of inter-chip bandwidth, FP8 and FP4 precision) are the company's own, and Forkast reports that all of them remain vendor-reported until independent benchmarking arrives ^[4][5]. Memory capacity and link bandwidth are component specifications. What fraction of a 500,000-chip job's theoretical throughput a real training run keeps busy gets measured on running clusters, and it depends on interconnect topology and the software stack.

Counterpoint Research has called the M890 a "believable replacement" for firms cut off from the H200 ^[14], which China blocked from import entirely in January 2026 ^[8]. Chinese chipmakers took 41 percent of the local AI accelerator server market in 2025, up from near-zero in 2022 ^[12]. The firm-level shares reported alongside that figure do not sit on one denominator: Huawei at roughly 62 percent and Alibaba at roughly 5 percent sum to 67, above the 41 that Chinese vendors hold in total ^[13][3]. Either those two numbers describe only the domestic segment, or one of them is measured on a different base. On both readings Alibaba is a minority supplier well behind Huawei.

Jensen Huang called the controls a "failure" at Computex in May 2025, by which point Nvidia's China data center revenue had effectively dropped to zero ^[10]. The Bureau of Industry and Security closed the third-country cloud route the following May, ending Chinese access to advanced compute through remote infrastructure in Malaysia, Singapore and the UAE ^[9]. Nvidia's H100, A100 and Blackwell parts remain under a presumption of denial for the market ^[7]. The climb in domestic share tracks that sequence. Three years of policy moves and one supplier's exit ran together over those years, so substitution stays a plausible reading and nobody has demonstrated it. Alibaba already has a successor on the roadmap, the J900, for the third quarter of 2028, six quarters after the V900's production date ^[16][4].