On October 13, Reuters obtained a report from IDC. It contained a set of numbers on accelerator shipments to Chinese data centers in the first half of this year. Two of those numbers crossed, and their order switched for the first time.
One is 991,000 units, the shipment volume of GPUs, down 26 percent from a year earlier. The other is 1.016 million units, the shipment volume of non-GPU accelerators, up nearly 80 percent year on year.
The second overtook the first. According to IDC, this is the first time non-GPU accelerators have exceeded GPUs in shipment volume. Before that, the order of those two numbers never changed. In the first half of this year, the order was scrambled.
The total was 2.008 million units. In other words, a little more than half of the AI accelerators sold in China in the first half were no longer the general-purpose graphics processors we know best, but another architecture.
It should be noted that these figures count units, not money. General chips carry higher prices and specialized chips lower ones, so ranked by sales value the order might not be the same. They are two different rulers.
The Two Numbers Crossed
The significance is not in the ranking but in the structure, a shift in what the market is made of. A GPU is general: it can do any job. A non-GPU accelerator is specialized: it does one kind of job, such as inference, or matrix math for a particular model.
The non-GPU category here includes ASIC, FPGA and NPU architectures. The names differ; what they share is that each is designed for a class of computation, not able to take on anything the way a GPU can, and not as power-hungry. Software is the hardest part.
The advantages of specialized chips are lower power, lower cost and lower latency. The drawback is inflexibility: if the model changes, the chip may not keep up. In the past everyone preferred general chips precisely because models changed so fast that no one dared bet.
Now it has reversed. Model architectures have gradually stabilized, inference demand has surged, and the cost-performance of specialized chips has become visible. This is not a sudden leap in technology; demand changed shape.
The Part That Disappeared
The missing part of GPU shipments came mainly from Nvidia. In the first half it shipped 678,000 units in China, down 44 percent year on year, and its market share fell to 34 percent, still first.
That share is still number one, but the distance to the company behind is only six percentage points. A year ago, the two were not at the same table. Numbers do not talk, but the trend is plain: the gap was worn away in a year.
The reason is not complicated. The most advanced models cannot come in, and the models that can often depend on approvals. When buyers place orders, the first question is not how fast the chip is but whether it will arrive on time, and whether the delivery date holds.
When supply is unstable, purchasing turns. This is not sentiment but the instinct of a procurement department: a production line cannot be built on a promise that the goods "might arrive," and model training cannot wait for a schedule with no date.
Why Specialized Chips
Inference and training are two different jobs. Training has to move vast amounts of data back and forth, valuing generality and bandwidth; inference runs a model that has already learned, valuing unit cost, power draw and response speed. The two are almost opposites.
When a model begins serving hundreds of millions of people, the inference ledger grows thicker than the training ledger. At that point, saving a little power or a few cards, multiplied across a year of days, is real money. That is the economy of scale.
Specialized architectures push at exactly this point. They do not try to do everything, only to do that one thing more cheaply. For buyers, good enough is enough, and cheap is the hard truth. Anyone can do this arithmetic.
IDC's research says these non-GPU accelerators are mainly adopted by government agencies, internet companies and several major telecom operators. The three buyers share one trait: large purchase volumes and extreme sensitivity to cost.
Six Points Apart
Behind Nvidia is Huawei. It shipped 572,000 units in the first half, against just over 300,000 a year earlier, nearly double, with a market share of 28 percent, a step up from where it stood.
The gap between the two is now only six percentage points, according to IDC, and it narrowed fast. The change took just half a year, quicker than many in the industry expected, and harder to ignore.
Further back are several Chinese vendors: T-Head with 7 percent, Cambricon with 5 percent, Kunlunxin with 4 percent and Hygon with 3 percent. None is large, together nearly 20 percent, but all are firmly on the list.
Of the eleven Chinese semiconductor companies IDC tracks, nine increased shipments year on year in the first half. That number says more than any single share: growth is not one point but a whole area.
Not Only Huawei
It is worth noting that Huawei did not take the whole non-GPU segment. On the shares above, Huawei, T-Head, Kunlunxin and Cambricon together account for nearly 90 percent, with the remaining tenth or so split among other vendors; outside those four there is hardly a voice.
Four companies holding nearly 90 percent shows the bar in this market is high. Chips, software stacks, cluster networking: miss any one and you cannot deliver. The barrier was not built in a day. Those who got in have been at it for years.
But a high bar also means a stable lineup. Once a vendor enters an operator's procurement list, switching suppliers is costly, and the next tender usually involves the same few names. New faces struggle to squeeze in.
Whether a chip is easy to use also depends on whether it can work alongside tens of thousands of others. In a cluster, the network, storage and scheduling can each hold things back; if one link lags, the performance of a single card never shows. This is the easiest part to overlook.
That is also why share changes are always slow and then suddenly fast. In the slow phase, vendors are adapting, rewriting software and tuning operators; in the fast phase, adaptation is done, orders arrive together, and the number jumps a notch. That is the usual rhythm in hardware.
What Changed Is the Structure
Return to the two crossing numbers. 1.016 million against 991,000: a gap of only 25,000 units, less than 2 percent of the total, barely visible inside two million, like a statistical error.
But the direction changed. For a decade, AI computing was almost synonymous with GPUs; in the first half of this year, among the accelerators sold in China, the non-GPU kind outnumbered GPUs for the first time. That is a structural swap, and the equals sign had stood for ten years.
IDC's judgment for the full year is that total AI accelerator shipments in China will grow 19 percent, while GPU shipments fall 9 percent. In other words, the crossing is not a one-off; it will continue, and the full-year numbers will write the same direction.
When the structure changes, so does the way companies compete. IDC says the focus is shifting from the peak performance of a single chip to the overall efficiency of large clusters. Whoever can make ten thousand cards work like one card will be worth more.
For an ordinary person, these numbers are far away. But the assistant in your phone that keeps getting better at talking is computed by these machines. The kind has changed, and you probably will not notice, except that the bill rises a little more slowly.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.