China's First Nvidia Chip Approval Skips the H200 — And Quietly Sidelines SK Hynix

Beijing has spent nine months saying no to the H200. The carve-out dates back to January 28, 2026, when Chinese regulators approved imports of more than 400,000 H200 chips for ByteDance, Alibaba, and Tencent. By May 14, 2026, the US Commerce Department had widened the channel to roughly ten Chinese buyers — capped at 75,000 chips each — with Lenovo and Foxconn cleared as distributors. Through August 2026, however, Nvidia confirmed Hopper shipments to China covered less than 1% of its $89.0B data-center quarter; NVIDIA guided to no data-center China revenue for Q3 FY27.

The RTX PRO 5500 solves several problems at once. It is a workstation-class product — 21,760 CUDA cores, 1,398 GB/s of memory bandwidth, up to 600W — wrapped around84GB of ECC GDDR7 rather than the HBM3e stack that drives every NVIDIA data-center accelerator. For Beijing, that distinction matters because GDDR7 does not pull from the constrained HBM supply chain dominated by SK Hynix, the memory partner that supplied roughly70% of HBM4 for NVIDIA's Rubin platform in 2026.

The chip's memory architecture is the part most investors miss. The RTX PRO 5500 pairs a cut-down GB202 Blackwell die with 28 GDDR7 chips in a clamshell layout — the same memory type used in NVIDIA's consumer RTX 5090, scaled up to 84GB. It carries no HBM stack. That single fact re-routes the entire China revenue stream around SK Hynix, the company whose HBM business is built around NVIDIA accelerators.

The economic gravity of the NVIDIA supply chain sits in HBM. SK Hynix reported FY2025 revenue of roughly $65B, with DRAM — heavily weighted toward HBM — contributing about $44B of that total. Micron recorded $41.5B of revenue in Q3 FY2026 (a record quarter), fueled by HBM ramp. Both companies were positioned to capture a meaningful slice of any China H200 thaw. The RTX PRO 5500 approval routes that demand into GDDR7 instead — a market where Samsung Electronics and Micron are the principal suppliers, but at price points a fraction of HBM.

The Information's reporting names only ByteDance and Alibaba. Tencent — included in the January 2026 H200 approvals and one of the few firms to actually receive chips in August — does not appear in the new MIIT request. Baidu, another named AI buyer in prior waves, is absent. The selection signals a deliberate triage by Beijing: the two firms with the largest near-term inference capacity gaps, and the two whose capex commitments already dwarf the rest of the sector.

Alibaba's quarterly capex of RMB 67.7B — roughly $10B, up 75% year-on-year — is the headline number. The company has guided to a 20GW data-center buildout by 2032 and unveiled the Zhenwu V900 chip on September 22, 2026, timed for mass production in Q1 2027. That plan requires Nvidia silicon to bridge the gap. ByteDance, privately held but reportedly committing $5.6B to Huawei Technology Ascend orders in 2026, needs Nvidia for the same reason: domestic supply is not yet large enough to fill its training-and-inference footprint.

TSMC is the clearest near-term beneficiary. The RTX PRO 5500 is fabbed on the foundry's 4nm process — the same node family as the rest of the Blackwell line — and Alibaba, ByteDance, and Lenovo all source finished servers through TSMC's packaging partners. TSM reported Q2 FY26 revenue of TWD 1,270B on Aug 14, 2026; any volume addition from a multi-hundred-thousand-unit RTX PRO 5500 run flows directly into that line. NVIDIA's Q2 FY27 disclosure filed Aug 21, 2026 (10-Q for the period ended July 26) carried no China-revenue assumption in its Q3 guide — leaving pure upside if approval converts to orders.

In the days and quarters ahead, the catalysts stack narrowly. Reuters was unable to confirm the MIIT request when it ran the original report; NVIDIA, Alibaba, and ByteDance did not respond to requests for comment. The next datapoint that converts \"weighing approval\" into hard orders will be the first RTX PRO 5500 shipment confirmation — likely visible in NVIDIA's Q3 or Q4 FY27 disclosures and in Alibaba's next quarterly capex commentary. The Reuters September 27 article also noted that NVIDIA's spokesperson still cited ongoing US export-control restrictions, leaving the product in a gray zone where China has said yes but Washington has not formally approved the chip variant for sale into China.

Over a one-to-three-year horizon, the thesis pivots on three structural forces. First, Alibaba's Zhenwu V900 ramps to commercial volume in Q1 2027, narrowing the window in which NVIDIA is the only vendor for high-end Chinese AI compute. Second, Huawei Technology's Ascend 950PR targets 750,000 units in 2026, with ByteDance's reported $5.6B commitment covering roughly 350,000 of those cards — a domestic alternative at roughly $6,900 per card, an order of magnitude below H200 pricing. Third, the NVIDIA Blackwell roadmap into Rubin pulls SK Hynix HBM4 demand into 2027-2028 even if China-bound volumes stay capped; that is the broader tailwind HBM bulls rest on, and it does not depend on this week's news.