A Million Chips for ByteDance: China Moves to Clear Nvidia's RTX Pro 5500

The US-China chip war has produced yet another twist, and this one could be worth billions to Nvidia’s bottom line. According to a report by The Information’s Qianer Liu published September 27, 2026, China’s Ministry of Industry and Information Technology (MIIT) has told domestic firms — including ByteDance and Alibaba — that it intends to approve purchases of Nvidia’s new RTX Pro 5500 chip. ByteDance alone reportedly plans to order one million of the chips to power its AI models.

If the orders materialize, they would mark one of the largest single-chip procurement programs in China’s history, and a striking reversal for a market that spent much of the past year actively pushing its own giants away from Nvidia hardware.

What MIIT is actually proposing

The MIIT has asked companies to submit purchase volumes and use cases before it formally green-lights orders, according to the report. Reuters, which picked up the story, said it could not independently verify the claim — a standard caveat for Beijing policy reporting that often runs ahead of official announcements.

The distinction matters: this is not yet a formal approval, but a ministry signaling intent to approve. In China’s regulatory environment, that signal is often the load-bearing step. It gives firms cover to negotiate supply contracts with Nvidia and its distribution partners while the paperwork catches up.

The RTX Pro 5500 launched this month. It is Nvidia’s newest workstation-class Blackwell GPU, and its launch timing now looks less like coincidence and more like diplomacy. A chip that lands in the market right as Beijing softens its stance is a chip designed to be sold to China.

The chip at the center of the deal

The RTX Pro 5500 Blackwell Workstation Edition, introduced September 14, 2026, is a cut-down GB202 part with 21,760 CUDA cores across 170 SMs — roughly 88 percent of the full chip’s 192 SMs. The headline spec for AI buyers is memory: 84 GB of GDDR7 with ECC, connected over a 416-bit interface delivering 1,398 GB/s of bandwidth. Power draw sits at a 600 W ceiling.

That memory capacity is 2.6 times the 32 GB on Nvidia’s consumer flagship RTX 5090, at matching core counts. For inference workloads — running large language models in production — memory capacity is frequently the binding constraint, and 84 GB in a single card moves workloads that previously required multi-GPU rigs into single-card territory. The card also supports Multi-Instance GPU (MIG) partitioning, splitting its resources into two isolated 42 GB instances — a feature aimed squarely at cloud providers slicing hardware between tenants, including the kind of regional GPU-rental businesses ByteDance and Alibaba both operate.

Pricing has not been officially confirmed, but community tracking places the card between $10,000 and $13,000, positioning it well below the data-center Blackwell line while delivering workstation-class compute. For ByteDance, a million units at that price implies a procurement program in the range of $10 billion or more.

A year of whiplash

To understand why this report matters, you need the timeline. In September 2025, China’s internet regulator ordered the country’s biggest tech companies to stop buying Nvidia AI chips, and specifically instructed ByteDance and Alibaba to cancel orders for the RTX Pro 6000D — the China-specific variant Nvidia had built to comply with US export controls.

Then came the thaw. In January 2026, China gave the green light for ByteDance, Alibaba and Tencent to import a first batch of H200 chips — more than 400,000 units in total. In May 2026, the US cleared sales of the H200 to ten Chinese firms, including Alibaba, Tencent, ByteDance and JD.com, each eligible to order up to 75,000 units.

Now, in September 2026, MIIT is signaling approval for the RTX Pro 5500 — a chip that is newer, and for inference workloads arguably more practical, than the H200 batches that preceded it. The pattern is consistent: Beijing’s stance toward Nvidia hardware oscillates with the broader US-China relationship, and each cycle of restriction-and-release has moved larger volumes.

Why ByteDance wants a million of them

ByteDance’s appetite is not mysterious. The company operates Doubao, one of China’s most heavily used consumer AI assistants, alongside a sprawling portfolio of video, advertising and enterprise AI systems. Inference demand — the cost of actually running models for hundreds of millions of users — scales with usage, not with training cycles. Recent financial disclosures showed ByteDance pouring tens of billions of dollars into AI infrastructure spending this year, funded partly by massive syndicated loans.

The RTX Pro 5500 fits that profile precisely. It is an inference-optimized card: enormous memory, MIG partitioning for multi-tenant serving, and a price point that makes fleet economics workable at scale. A million of them would give ByteDance roughly 84 petabytes of aggregate GDDR7 capacity — a figure that would rank it among the largest single-company inference fleets on the planet.

The open questions

Three things could still derail this. First, MIIT’s “intent to approve” is not an approval; policy in Beijing has reversed on shorter notice than this. Second, US export rules could intervene — the RTX Pro 5500’s exact export-control classification relative to the China-specific thresholds matters, and Washington has shown willingness to pull licenses retroactively. Third, Reuters’s inability to independently verify the report is a real caveat: single-source Beijing policy stories have been wrong before.

But the direction of travel is clear. China’s domestic AI chip push — Huawei’s Ascend line, Cambricon, Moore Threads — remains strategically central, yet the practical reality is that Chinese AI labs keep choosing Nvidia silicon whenever the door opens even slightly. Each approval cycle widens that door a little more, and Nvidia’s willingness to keep building market-tuned SKUs like the 5500 shows Jensen Huang’s company intends to walk through it every time.

For the AI industry’s supply chain, a cleared million-chip order would also ripple outward: GDDR7 supply from Micron and Samsung, advanced packaging capacity, and server integration partners would all see demand pulled forward. The chip war isn’t over — it has just entered a phase where the weapons are procurement contracts instead of bans.