Baidu's Kunlun Core Unveils 4th-Gen M100 Inference Chip at WAIC, Targets Nvidia H20百度昆仑芯在WAIC发布第四代M100推理芯片，剑指英伟达H20

Baidu's chip subsidiary Kunlun Core Technology used the World AI Conference in Shanghai this week to pull back the curtain on the M100, its fourth-generation AI accelerator. The chip is aimed squarely at large-model inference workloads and, according to BigGo Finance's on-site coverage, is built entirely on a domestic Chinese supply chain — a claim worth reading carefully, because "domestic supply chain" in Chinese chip discourse can mean anything from wafer fab to packaging to firmware, and the precise scope is rarely disclosed in first-wave launch materials.

The WAIC debut is the M100's first public appearance outside Baidu's internal testing environment. That matters for one specific reason: Kunlun Core has been a captive supplier to Baidu's own data centers since its third-generation chips, and the M100's appearance at a public trade floor signals an intention — stated or implied — to sell outward. Whether purchase orders or even letters of intent exist is not yet documented in any source available at time of writing. Watch the press release, not the booth.

Kunlun Core has positioned the M100 as a cost-competitive alternative to Nvidia's H20 in the inference tier. The H20 is itself a deliberately neutered part — Nvidia's export-control-compliant card for the China market, capped in interconnect bandwidth to satisfy U.S. Commerce Department restrictions. Targeting the H20 therefore means targeting a chip that was already designed to lose. That framing is a smart piece of competitive positioning, but it should not be mistaken for evidence of parity with Nvidia's unrestricted data-center silicon.

TechTimes reported that Zhangjiang Science City's section of WAIC hosted more than 100 companies and logged China-debut products across ten competing domestic GPU architectures. Ten architectures on one conference floor is a sign of a market still fragmenting, not consolidating. For enterprise buyers trying to standardize inference infrastructure, that fragmentation is a genuine procurement problem — and it is one of the structural reasons Nvidia, even a crippled Nvidia, retains stickiness in Chinese data centers.

The capital backdrop amplifies the stakes. China has reportedly earmarked approximately 2 trillion yuan ($295 billion) in data-center and AI compute investment over the next five years. Domestic chip vendors, Kunlun Core among them, are in an accelerating race to lock in design wins before that buildout peaks and before any further tightening of U.S. export controls forces a harder architectural choice on Chinese hyperscalers. The M100's inference focus is a deliberate lane selection: inference racks scale differently from training clusters, they turn over faster, and they represent the near-term volume market as Chinese firms deploy large models at the application layer.

BigGo Finance's reporting quotes directly from WAIC floor coverage: "昆仑芯M100采用全国产化供应链，专为大模型推理场景设计，性能对标英伟达H20。" — "The Kunlun Core M100 uses a fully domesticated supply chain, is designed specifically for large-model inference scenarios, and benchmarks its performance against the Nvidia H20." Note the verb: 对标 (benchmarks against) is a marketing posture, not a published benchmark. No independent third-party inference throughput numbers, memory bandwidth figures, or power-efficiency data have been released as of this writing.

What to watch next: Kunlun Core's historical pattern has been to announce at a public venue and release detailed technical specifications — TFLOPs, HBM capacity, NVLink-equivalent interconnect — in a follow-on whitepaper or developer documentation within four to eight weeks. If those numbers do not appear by late August 2025, the M100 is likely still in pre-production. Also watch Baidu's Q2 2025 earnings call for any mention of M100 internal deployment timelines; Baidu's cloud and AI infrastructure capex guidance will tell you more about this chip's real trajectory than any WAIC presentation slide.

The harder question the M100 does not yet answer: can Kunlun Core manufacture at volume? Third-generation Kunlun chips faced yield and supply constraints that limited their reach beyond Baidu's own clusters. The M100's all-domestic supply chain claim, if accurate, means it is navigating the same constrained SMIC and domestic advanced-packaging ecosystem that every Chinese chip designer is competing for simultaneously. Supply chain nationalism is a policy directive; it is not a production guarantee.