Provenance Record

Well-structured News piece (634 words, within the 400-1200 range). Clear Overview / What We Know / What We Don't Know / Analysis sections. Each spec is attributed to a named source inline, and the article carefully separates vendor claims from independent verification.

All 4 snapshots read from disk (status 200). source-0 (TrendForce, 2026-05-21): confirms M890 144GB HBM memory, 800GB/s inter-chip bandwidth, 3x the 810E; 810E unveiled Jan 2026 with 96GB HBM2e and 700GB/s; V900 3Q27 (3x M890, 216GB, 1,200GB/s); J900 3Q28; 128-supernode server with ICN Switch 1.0; 560,000 cumulative shipments / 400+ customers / 20+ industries. source-1 (The Next Web, 2026-05-20): confirms M890 positioned against Nvidia H100 generation not Blackwell, and T-Head proprietary GPU chips in scaled mass production. source-2 (WinBuzzer, 2026-05-20): confirms May 20 launch, 144GB on-chip memory, 800GB/s, 3x 810E, FP32-to-FP4 precision, ICN Switch 1.0 rated 25.6 Tbps across clusters of 64 accelerators, 128-card supernode linking 128 chips, 560,000+ shipped / 400 customers / 20 industries, and the undisclosed-figures list (absolute FLOPS, process node, TDP, MLPerf, training-throughput). source-3 (SCMP, 2026-01-29): an article about the PREDECESSOR Zhenwu 810E ('M890' does not appear anywhere in the snapshot). Article attributes ONLY 810E/predecessor context to SCMP: 810E described as a parallel processing unit (PPU) / ASIC for both AI training and inference, a single card handling 700GB/s, on par with Nvidia's H20 tailored for China to comply with US export restrictions. All three SCMP-attributed claims verified verbatim.

Every number, name, and date traces to an allowlisted source. The critical SCMP citation is used strictly for predecessor (810E) context with no M890 spec attributed to it. The 144GB / 128-chip supernode / 3x-prior-generation / May 20 unveil specifics all trace to TrendForce and/or WinBuzzer. Memory type is left conservative: the article body says '144GB of memory' / '144GB of on-chip memory' without asserting an HBM generation, even though TrendForce calls it 'HBM' generically. No HBM generation is claimed for the M890. The bot correctly dropped specs lacking allowlisted support (PCIe 5.0 x16, the Panjiu/AL128 supernode name, <150ns latency) - WinBuzzer references 'AL128 rack networking' and 'hundreds of nanoseconds' latency but the article does not assert the AL128 name or a sub-150ns figure.

APPROVE. All four sources are allowlisted (trendforce.com, thenextweb.com, winbuzzer.com, scmp.com) and were personally read from the committed snapshots. Every cited specific verified. The SCMP source is correctly confined to 810E predecessor context with no M890 spec attributed to it. Memory type left unspecified, no fabricated quotes, originality confirmed. Ready for publication.