Baidu unveils a 1,024-chip AI 'supernode' it says can train 5-trillion-parameter models, arriving next year

Baidu used the opening of its Baidu World 2026 conference on Thursday to unveil what it calls the Tianchi 1024-card supernode, a system that ties 1,024 AI accelerator cards together so they behave like one very large computer. Baidu executive vice president Shen Dou said it will go on sale in 2027, and the company claims a single unit can handle training of a 5-trillion-parameter AI model.

Why "supernode" matters: training a big AI model means splitting the math across hundreds or thousands of chips that constantly swap partial results. If the links between them are slow, expensive chips sit idle waiting for data. A supernode wires a large group of chips together with a very fast internal network, so the whole group works more like a single machine than a rack of separate servers.

According to on-site reporting by IT Home, the new system uses what Baidu calls optical-engine technology, and TMTPost describes it as an all-optical interconnect, meaning the chips talk to each other more with light than with copper wiring. Baidu's numbers: chip-to-chip bandwidth is three times the previous generation, training performance is up 2.5x, and inference performance (running a finished model to answer requests) is up 3x. It also comes with a "G3.5 KV Cache" memory cabinet. A KV cache is roughly the model's scratch paper from earlier turns of a conversation; keeping and reusing it means the model doesn't have to redo that work every time. Baidu says this doubles long-context inference performance, and in multi-turn chats where the cache can be reused it cuts the wait before the first word of a reply appears (time to first token) by 90% while raising overall throughput 7x. Baidu also says the design can scale out to clusters of 10,000 cards.

The timeline moved up. At Baidu World 2025, Baidu introduced Tianchi 256- and 512-card supernodes built on its in-house Kunlun P800 chips and said a 1,000-card-class supernode would arrive in 2028. IT Home notes that has now been pulled forward by a year. Baidu did not say which chip powers the new 1,024-card system. At the same event, Baidu's chip unit Kunlunxin said its value-focused M100 chip is already on the market, the flagship M300 is due next year, and several chips on a new "N architecture" will follow in 2028. Baidu co-founder and CEO Robin Li said the spread of AI agents has caused demand for inference computing to explode, and that he believes inference will be the main long-term driver of AI chip growth.

Keep the caveats in view. Every performance figure here comes from Baidu and is measured against Baidu's own previous product. There are no independent benchmarks yet, and pricing and launch customers were not disclosed. "Can support training a 5-trillion-parameter model" describes the system's capacity, not a model that has actually been trained on it.

The bigger picture: as AI models grow, the race is increasingly about how well companies can make thousands of chips cooperate, not only how fast each chip is. Tech companies in both China and the US are now shipping or planning supernodes with hundreds to thousands of accelerators, and Baidu's move to bring its 1,000-card-class product forward to 2027 is one step in that race.

What it means for you: you will never have one of these at home, but systems like this sit behind the chatbots, AI search and coding assistants people use every day. Faster inference and shorter waits in multi-turn conversations would mean quicker answers, and lower costs for providers could eventually make AI features cheaper and more widely available. How much of that Baidu delivers will only be clear once the product ships next year and independent users can test it.

Original sources published: IT Home (Tianchi supernode), Nov 12, 2026, 12:09 Beijing time / Nov 11, 2026, 11:09 PM EST; TMTPost, Nov 12, 2026, 12:18 Beijing time / Nov 11, 2026, 11:18 PM EST; IT Home (Kunlunxin roadmap), Nov 12, 2026, 11:46 Beijing time / Nov 11, 2026, 10:46 PM EST