Baidu AI Cloud's 1,024-card system now arrives a year early, and Kunlunxin's Hong Kong listing pitch can point to a chip roadmap that has been pulled forward.
Baidu will put a 1,024-card version of its Tianchi supernode on the market in 2027, a year ahead of the 2028 date it set twelve months ago. Shen Dou, executive vice president of Baidu Group and president of Baidu AI Cloud Group, announced the system at the opening of Baidu World 2026 in Beijing on November 12, IT Home reported. A supernode is a rack-scale machine that links many AI accelerator cards so they behave like one large computer. Kunlunxin, Baidu’s AI chip business, said at the same event that its M100 inference chip has shipped and that its flagship M300 is due in 2027.
Baidu says one Tianchi 1024 supernode can train a model with 5 trillion parameters and can be extended to a cluster of 10,000 cards. Baidu also said a next-generation network design would allow superclusters of up to 500,000 cards, and that several chips built on a new N architecture would reach the market by 2028, according to Lanjing News.
The schedule is the news. Baidu’s 2025 roadmap placed a thousand-card supernode in 2028 and the Kunlunxin N series in 2029; both have now moved up by a year. Baidu has not published prices, a customer list or detailed M300 specifications, and all performance figures so far are its own.
The new supernode uses optical links and a separate memory cabinet
Tianchi 1024 connects its cards with optical engine technology, which Baidu’s press release describes as an all-optical interconnect. Against the previous generation, Baidu claims three times the bandwidth between cards, 2.5 times the training performance and three times the inference speed, IT Home reported.
Baidu pairs the supernode with a G3.5 KV cache cabinet, extra memory that stores the intermediate results a model builds up during a conversation so they do not have to be recomputed. According to IT Home, Baidu says the cabinet doubles performance on long-context inference. Where a multi-turn conversation can reuse cached data, it cuts the wait for the first word of a response by 90% and raises total throughput sevenfold.
On the chip side, the release says the M100 "has shipped as planned" and offers strong cost-effectiveness, and that the M300 will bring a major leap in computing power, memory and interconnection. Baidu gave no figures for either claim.
Baidu’s 2026 dates track the M-series plan and accelerate the rest
At Baidu World 2025, Baidu set out a five-year plan, as reported by IT Home at the time. The M100 was due in early 2026 and the M300 in early 2027. A Tianchi 256 supernode was set for the first half of 2026 and a Tianchi 512, able to train trillion-parameter models within one node, for the second half. The thousand-card supernode followed in 2028, the N series in 2029, and a single million-card Kunlunxin cluster under Baidu’s Baige platform in 2030.
China Daily reported then that the M100 was optimized for large-scale inference and the M300 was meant for training and running very large multimodal models. Measured against that plan, the M100 arrived on time and the M300 remains on track for 2027, though this year’s statements say only "2027" rather than early 2027. The acceleration falls on the systems layer, the 1,024-card supernode, and on the N-architecture chips.
Baidu framed the timetable around inference demand. "We believe inference will be the primary driver of sustained long-term growth in the AI chip market," chief executive Robin Li said in the release. Communication World reported Li’s view that demand for AI computing has moved from being led by training to training and inference growing side by side, a split that favors an inference-first part such as the M100.
Kunlunxin’s listing plans give the roadmap a financial audience
Kunlunxin is preparing to list in Hong Kong, and Baidu this month denied a report that the offering could slip to 2027. A roadmap that delivers on time and pulls products forward is part of the case any prospectus would make to investors, alongside revenue figures Kunlunxin has not yet published.
The concrete markers are on the calendar. The M300 and Tianchi 1024 are both scheduled for 2027, and the first independent benchmarks, rental prices on Baidu AI Cloud and named customers for the 1,024-card system will show whether the bandwidth and speed claims hold up. The next check on the accelerated plan comes in 2028, when Baidu has promised several N-architecture chips.