Baidu claims more than 30% cheaper AI inference, but the M300 chip behind the prices lands in 2027

A price list on stage in Beijing

At Baidu▼−2.46% — Baidu, down 2.46 percent today World 2026 on November 12, founder and chief executive Robin Li said inference, not training, will be "the primary driver of sustained long-term growth in the AI chip market" ¹². His prices followed. Two new application models, ERNIE Light Deep Thinking and ERNIE Wizshot Micro-Expression, cut the cost of delivering inference by more than 30% and video production cost by 50%, Baidu said ¹².

Customers benefit first: finished video from the Baidu Yijing agent now costs RMB 0.2, about three US cents, a second ²³. The Token Exchange marketplace for model calls↗, one month old, already averages more than 700 billion tokens a day ²³.

The cheap silicon is dated 2027

The roadmap under those prices is a 2027 product ¹². Baidu says the Tianchi supernode, 1,024 of its own Kunlunxin accelerator cards chained over an all-optical interconnect, launches in 2027 alongside the flagship M300 chip, with an "N architecture" generation to follow in 2028; the M100 inference chip has already shipped as planned ¹². The claimed gains, three times the chip-to-chip bandwidth, 2.5 times the training performance and three times the inference performance, arrive without a named comparison machine, a chip price or a capex figure ¹².

Until then, Baidu's data centers keep running a mix of Kunlunxin and Nvidia▼−5.64% — Nvidia, down 5.64 percent today hardware ⁸. Kunlunxin held 4% of China's data-center AI chip market in the first half of 2026↗ ⁷. Nvidia's shipments into that market fell 44% in the first half, to 678,000 units; Huawei, the nearest rival at 28% share, shipped 572,000, up from more than 300,000 ⁷.

Monday prices the wait

Third-quarter results land on November 16↗ ⁵. FactSet▲+0.60% — FactSet, up 0.60 percent today consensus puts revenue at RMB 30.8 billion, slightly below the 31.2 billion yuan a year earlier, and net profit at RMB 1.65 billion ⁵. Citi▼−0.17% — Citi, down 0.17 percent today expects AI cloud infrastructure revenue up 99%; Deutsche Bank has cut its 2026 and 2027 profit forecasts by 15% and 14% on the shift toward lower-margin cloud ⁵.

The second quarter showed both edges of the trade↗: total revenue down 4%, GPU cloud revenue up 283%, AI cloud profit and margin rising ⁶. Bank of America▼−0.20% — Bank of America, down 0.20 percent today reads China's price war as a move from token prices to cost per completed task, with the durable economics in chips and cloud platforms rather than model vendors ⁹. That is the bull case for Baidu's stack, and it stands only if Kunlunxin is genuinely cheaper than the Nvidia silicon it replaces; no third party has tested that yet.

The cut may buy share, not margin

The bear risk runs through rivals: Chinese model vendors have been cutting token prices aggressively since DeepSeek kicked off the race ⁹, so a cheaper floor helps them as much as Baidu, and the 30% cut may buy share rather than margin. Watch the AI cloud margin line on Monday, the November 25 release of the design files for ERNIE 5.5 Flash↗, Baidu's new model ²⁴, and the M300 date. Until the M300 lands, Baidu is selling 2027 economics on 2026 silicon ².