Ninety Percent Foreign Silicon: China's SASAC Surveys Broadcom's Grip on State Data Centers

Deep inside the AI stack, below the GPUs everyone argues about, sits a layer of silicon that almost nobody discusses: the switching fabric that stitches thousands of accelerators into a single coherent machine. On Wednesday, September 23, 2026, that layer was suddenly dragged into the geopolitical spotlight. The Financial Times reported that China’s State-owned Assets Supervision and Administration Commission (SASAC) — the body that oversees China’s roughly 100 largest state-owned enterprises — has been surveying how many Broadcom switches operate inside state-backed data centers. The preliminary answer, according to people familiar with the findings, is startling: at some state-owned companies, Broadcom equipment could account for as much as 90 percent of the networking hardware in use.

What the survey found

SASAC’s review, described by sources speaking to the FT and relayed by Reuters and other outlets, asked state-run data center operators to inventory their switching infrastructure. The exercise was not academic. Based on those preliminary findings, SASAC may issue informal guidance directing state-run data centers to scale back their use of Broadcom switches, according to a Reuters summary of the report. The framing matters: this is not (yet) a formal ban, an export-control retaliation, or a public regulatory filing. It is a quieter instrument in Beijing’s toolkit — informal guidance that state enterprises read as a directive, part of what the FT describes as Beijing’s broader “domestic drive.”

The 90 percent figure is the number that should stop readers. Broadcom’s Tomahawk and Trident merchant silicon families, along with its Jericho routing chips, have become the de facto backbone of Ethernet networking worldwide — and China’s state-owned cloud and telecom giants built their AI factories on top of that foundation alongside everyone else. Even as Washington spent years restricting China’s access to leading-edge GPUs, the networking layer beneath those GPUs remained overwhelmingly American. A survey result showing near-total dependency on one U.S. supplier, inside facilities funded by the Chinese state itself, is exactly the kind of dependency Beijing has been systematically hunting down since the U.S. entity listings and export controls began biting.

Why switches, and why now

Three strands of context make this moment legible. First, Beijing has already moved on compute: in late 2025, Reuters reported guidance requiring new state-funded data centers to use domestically produced AI chips, a policy that supercharged Huawei’s Ascend line, Cambricon’s accelerator production, and Alibaba’s in-house Zhenwu silicon. Networking is simply the next layer of the stack to come under the same logic — you cannot credibly claim “autonomous and controllable” (自主可控) infrastructure if the traffic between your domestic GPUs flows through foreign switch ASICs.

Second, AI clusters have made the network a first-order performance component. Modern training runs push terabits per second across thousands of endpoints; the switch fabric is no longer plumbing but part of the computer itself. Broadcom’s Tomahawk 6, with its 102.4 Tbps of per-chip bandwidth, and its co-packaged optics roadmap have kept it generations ahead of most challengers. That technical dominance is precisely why the dependency is so deep — and why unwinding it is so hard.

Third, the survey lands amid escalating tit-for-tat. Just weeks ago, U.S. outlets reported a drafted ban on Chinese data center components in American facilities, citing malware and espionage risks. Beijing’s inventory of Broadcom gear reads, in part, as the mirror image of Washington’s own audit instinct: each side is now mapping the other’s silicon inside its critical infrastructure.

The domestic bench

If state operators are pushed to diversify, the immediate candidates are well known to China’s networking industry. Huawei’s CloudEngine data center switches, built on in-house silicon, are the most mature alternative at the high end. H3C (New H3C, now controlled by Tsinghua Unigroup-aligned capital) fields a broad enterprise and data center portfolio. Centec (盛科通信), a Suzhou-based merchant switch-chip designer listed on the STAR Market, is frequently named as the closest domestic analog to Broadcom’s merchant ASIC business, though its flagship parts still trail Tomahawk-class bandwidth. Tsinghua Unigroup’s and ZTE’s switching divisions, plus whitebox ODMs like Celestica-built platforms re-branded domestically, round out the bench. Domestic optical and PHY vendors have advanced quickly as well — an FT-cited ecosystem that Chinese trade press has chronicled under the “breaking the Broadcom chip lock” narrative for the past two years.

The honest caveat: matching Broadcom at 51.2T and 102.4T generation scale is a multi-year engineering problem. Deep buffer management, mature SDK ecosystems, and the operational tooling that hyperscale network teams live inside are sticky advantages. Informal guidance will likely bite first at the edges — new procurements, expansion pods, less performance-critical segments — rather than triggering a wholesale rip-and-replace of fabric that currently trains frontier Chinese models.

What it means for Broadcom

For investors, the knee-jerk read is revenue risk. Broadcom generated roughly $10.8 billion in AI semiconductor revenue in a recent quarter (up 143 percent year over year), spanning custom accelerators for the largest hyperscalers and the networking franchises at issue here. China has historically represented over $10 billion of Broadcom’s annual revenue across segments. Yet the immediate financial exposure is narrower than the headline suggests: the survey covers state-owned data centers specifically, not the private Chinese cloud giants (Alibaba, ByteDance, Tencent) whose AI buildouts drive the bulk of Broadcom’s China networking sales — though nobody believes those customers would stay exempt if policy hardened. Analysts also note that Broadcom’s growth story now rests overwhelmingly on U.S. hyperscaler custom-silicon deals, a backlog measured in tens of billions, which insulates the top line even in a worst-case Chinese state-sector squeeze.

The more consequential signal is strategic. Every layer of the AI stack is now a policy object. Two years ago the question was “can China buy Nvidia GPUs?” A year ago it became “can China build its own accelerators?” This week it is “who makes the switches, and can Beijing tolerate a 90 percent single-supplier dependency in facilities the state itself funds?” The trajectory of that question is not in doubt — only its speed.

The bottom line

SASAC’s survey is best understood as due diligence for decoupling: measure the dependency, publish nothing, and let informal guidance do the work. For state operators, it portends a gradual re-procurement toward Huawei, H3C, and Centec-class silicon, with performance trade-offs absorbed in exchange for supply security. For Washington, it is one more datapoint that export controls have not just restricted China — they have redirected enormous state capital toward rebuilding the entire semiconductor and networking stack domestically. And for anyone mapping the AI supply chain, it is a reminder that the next chokepoint conversations will be about the unglamorous layers — switches, optics, PHYs — where a handful of companies, Broadcom chief among them, still hold ninety-percent-grade shares.

The switching fabric was the last quiet layer of the AI boom. As of this week, it is quiet no longer.