At the Baidu World 2026 opening ceremony on November 12, founder Robin Li put a thesis on the record about AI silicon. Compute demand, he said, was once dominated by training, but the balance has shifted to training and inference side by side, and the scale-up of agent applications has triggered an explosion in inference demand. Inference, he argued, will become the main engine of long-term AI chip growth.
On the same day, Kunlunxin, Baidu's chip unit, laid out its product cadence for the next two years.
Two Generations on a Schedule
The official schedule: M100, positioned for extreme cost-effectiveness, is already on the market. The flagship M300, with across-the-board jumps in compute, memory and interconnect, arrives next year. By 2028, multiple new chips based on an entirely new architecture will reach the market, matched to different workloads. Li said the chip is the result of more than a decade of patient investment whose commercial value is gaining recognition.
The Supersystem Matters More Than the Die
More consequential than any single chip is the system product unveiled the same day: Baidu Intelligent Cloud's Tianchi 1,024-card super node, scheduled for official release in 2027.
The official parameters: optical engine technology moves data through optical interconnect, with bandwidth up threefold versus the previous generation. The G1.5 memory acceleration cabinet doubles long-context inference performance. The G3.5 KV-cache cabinet cuts first-token waiting time by 90 percent and lifts throughput sevenfold in cache-hit multi-turn dialogue scenarios.
One super node can train a model with five trillion parameters and scale to ten thousand cards; with the new HPN architecture, clusters can reach 500,000 cards. Tianchi super nodes have already been deployed at thousand-cabinet scale in multiple locations nationwide with a fully domestic stack, and Baidu has delivered more than 30 Kunlunxin clusters to date.
Why Inference
Li's judgment tracks a change in workload structure. Training is a concentrated megaproject; inference lives inside every conversation and every agent call, arriving in fragmented bursts that prize latency, concurrency and per-query cost. The benchmark for chips shifts from peak specifications on paper to stable delivery under real load.
That also explains why a super node premiered before the flagship chip: in the agent era, the contest is measured at the rack level, not on one die's datasheet. A single agent task can involve multiple model calls, tool invocations and result aggregations, which multiplies memory traffic around long contexts and caches. Chips designed only for peak training throughput handle that profile poorly; systems designed around cache hits, memory bandwidth and optical interconnect handle it directly, which is the design logic on display in Tianchi's specifications.
Baidu Group executive vice president and Intelligent Cloud head Shen Dou said the cloud business is reorganizing its products and services into a full-stack AI cloud spanning agent infrastructure and compute infrastructure.
For two years, domestic AI chips competed on tape-outs and specs. In the 2027 window, the competition becomes software ecosystem, system delivery and supply-chain coordination. Whether the M300 ships on schedule will be the most visible first answer.
Disclaimer: This article is based on public information and is for reference only. It does not constitute any investment advice.
