Alibaba's T-Head Unveils Zhenwu M890 AI Accelerator With 144GB Memory and a 128-Chip Supernode, Claiming 3x Its Prior Generation
Alibaba's chip unit T-Head detailed the Zhenwu M890, a 144GB AI accelerator it says triples the prior 810E, alongside a 128-chip supernode server.
Overview
Alibaba’s in-house chip unit, T-Head, has unveiled the Zhenwu M890, a GPU-class AI accelerator the company says delivers up to three times the performance of its prior generation. According to TrendForce, the M890 integrates 144GB of memory and offers inter-chip interconnect bandwidth of up to 800GB/s, with performance roughly three times that of the previous-generation Zhenwu 810E. The launch, made at Alibaba’s cloud summit on May 20, also included a 128-chip server that the company positions as a single-system answer to large-scale AI training and inference, as reported by WinBuzzer.
The disclosure adds detail to China’s continuing effort to field domestic alternatives to Nvidia hardware as U.S. export controls tighten. Per The Next Web, the M890 is positioned against Nvidia’s H100 generation rather than the newer Blackwell parts, and T-Head’s proprietary GPU chips have achieved scaled mass production.
What We Know
The Zhenwu M890 is built to handle both training and inference. According to WinBuzzer, the chip carries 144GB of on-chip memory and 800GB/s of inter-chip bandwidth, and Alibaba lists precision formats spanning FP32 down to FP4 to support both workloads. The same report notes Alibaba states the M890 delivers three times the performance of the older Zhenwu 810E.
The predecessor sets the baseline for that comparison. TrendForce reports the Zhenwu 810E, unveiled in January 2026, featured 96GB of HBM2e memory and inter-chip bandwidth of up to 700GB/s. When the 810E was detailed earlier this year, the South China Morning Post described it as a parallel processing unit (PPU) — an application-specific integrated circuit designed for both AI training and inference — and reported that a single 810E card handled 700 gigabytes per second of bandwidth, putting it on par with Nvidia’s H20, a GPU that Nvidia tailored for China to comply with U.S. export restrictions.
Beyond the chip itself, Alibaba framed the M890 as the building block of a larger system. According to WinBuzzer, the company unveiled a 128-card AI supernode server alongside an ICN Switch 1.0 fabric rated at 25.6 terabits per second across 64-accelerator clusters. TrendForce describes the same system as a 128-supernode server equipped with ICN Switch 1.0 that lets 128 AI chips operate as a single computer, aimed at massive concurrent agent inference and large-model training.
T-Head also pointed to existing traction for the Zhenwu line. Both TrendForce and WinBuzzer report cumulative shipments of more than 560,000 Zhenwu chips to over 400 customers across roughly 20 industries.
The announcement came with a multi-year roadmap. TrendForce reports a successor V900 slated for the third quarter of 2027, expected to deliver three times the performance of the M890 with 216GB of memory and inter-chip bandwidth boosted to 1,200GB/s, followed by a J900 in the third quarter of 2028 described as a major leap in T-Head’s self-developed parallel computing architecture.
What We Don’t Know
The vendor’s performance framing rests on its own figures rather than independent testing. WinBuzzer notes that Alibaba has not publicly disclosed absolute FLOPS, the chip’s process node, thermal design power, MLPerf results, or real training-throughput benchmarks, limiting independent verification of the three-times claim. The relationship between the company’s stated specifications and head-to-head performance against current Nvidia accelerators therefore remains unproven outside Alibaba’s own materials.
Analysis
The M890 lands amid a hardening regulatory backdrop for Chinese access to Western AI silicon. Earlier this month, The Machine Herald reported that the Commerce Department issued weekend guidance closing an AI-chip loophole for overseas subsidiaries of Chinese firms, underscoring why domestic accelerators and home-grown interconnect fabrics like ICN Switch 1.0 matter to Alibaba’s cloud roadmap. By pairing a single-die training-and-inference part with a 128-chip supernode and a published successor cadence, T-Head is signaling a system-level strategy rather than a single-chip release — though, as the sources note, the absence of disclosed benchmarks leaves the competitive comparison open.