China's AI Queue: Six Million Cards, Three Million Slots

ByteDance, Alibaba and Tencent have placed orders for more than three million of Nvidia's RTX PRO 5500 workstation cards, a combined sticker value above 35 billion US dollars, according to a Reuters report published on 20 October. Total orders from Chinese customers exceed six million units. Nvidia has told customers and partners that it cannot guarantee it can fill them, and that some buyers may receive 5 to 10 percent of the quantities requested; the company has indicated it can produce no more than three million RTX PRO 5500 cards globally in 2027.

Read that as a supply chain document rather than a geopolitics story and three things follow. First, the demand is for inference, not training, which changes what kind of board gets built. Second, the constraint is not the GPU alone, it is everything around it. Third, a queue of six million units against a three million unit capacity is the definition of an allocation market, and allocation markets leak into component pricing well beyond the product in question.

Why inference hardware looks different

The RTX PRO 5500 sits below the B300 and H200 in capability. B300 is not available to Chinese buyers at all under current US export controls, and H200 exports are reviewed case by case under licensing conditions. The 5500 is a Blackwell-based workstation card, comparable in performance to the RTX 6000D launched last year, and the buyers in this order are reported to be using it primarily to run trained models rather than to train new ones.

Inference boards are less about peak matrix throughput and more about sustained memory bandwidth, power delivery stability and thermal headroom at 24/7 duty. That has a direct component consequence. Inference servers live on GDDR or HBM capacity per card, on multiphase VRM stages with high current inductors, on high-layer-count HDI boards with tight impedance control, and on thermal solutions that have to work in a dense rack for five years. The bill of materials is unglamorous and it is where the volume goes.

The parts that get pulled

Workstation-class cards compete for the same memory output as everything else. With DRAM contract prices up 20 to 24 percent over three months depending on generation, and with high-bandwidth memory absorbing an increasing share of wafer starts, a six-million-unit order book for cards that each carry a substantial memory complement is not neutral for the rest of the market. It also competes for the power stages: a card of this class draws several hundred watts through a multiphase converter, and each phase needs a DrMOS stage, an inductor and bulk capacitance.

The board side is equally affected. High-layer HDI with fine line widths is already the tightest part of the PCB market, and the AI server segment has been pulling capacity since the first half of the year. Buyers of ordinary industrial boards do not compete for the same stack-up, but they do compete for the same laminate supply and the same plating capacity, which is why CCL lead times move when accelerator demand moves.

What a component buyer should do with this

Three actions, in order of urgency. Confirm the memory content and the power stage on any 2027 design now, because those are the two lines where an allocation decision made upstream shows up as a six-month delay downstream. Ask suppliers for a written statement of allocation status on the specific part numbers you buy, not a category-level answer. And treat any quote with a long validity on power or memory content as provisional, because the supplier who grants it is absorbing a risk that the market has already priced.

There is also a second-order effect worth watching. When six million cards are ordered and three million can be built, the difference does not disappear; it moves into secondary markets, refurbishment and grey channels, and it raises the price of every adjacent product that shares a component. That is how a data-centre allocation becomes a line item on a small manufacturer's BOM six months later.

Where we sit in this

We do not build accelerator cards. We build boards and supply the components around them: the power stages, drivers, memory support parts, connectors and passives that a controller or an industrial compute board actually needs. Our component catalogue lists what we source and hold, our PCBA service covers assembly from a released design, and our earlier note on AI server demand and PCB lead times covers the board-side picture.

If your design shares silicon or memory with the AI build-out and you want to know what that means for your next two quarters, send the BOM through the inquiry form. We will tell you which lines are exposed and which are not, while both answers are still useful.