Over the past two years, the computing power market has been almost entirely dominated by GPUs. When training peaks could not be pushed higher, all other efforts seemed secondary, and the landscape was paced by a few accelerator cards. However, as the industry chain moves past its most boisterous phase, the trend is shifting from "making models larger" to "making models usable." Attention is quietly moving away from the peak computing power of a single chip to whether the entire system can handle massive concurrent daily calls.
This shift has pushed the long-neglected CPU back to the center of the table. It has transitioned from a supporting front-end role for GPUs to the true scheduler and executor of task decomposition, data movement, and multi-round invocations. The change in its status in this round determines not only the financial reports of several companies this year but also a reshuffling regarding "how computing power should be priced" and "where the domestic chain should stand."
01. The Shift to Inference: The Price Anchor for CPUs is Re-established
To clarify the return of CPUs in this cycle, one must first understand that its revenue structure is completely different from that of GPUs. GPUs are priced based on the peak performance of a single card, selling the ceiling; CPUs are priced based on throughput and latency per unit of computing power, selling "affordability." In the past, when training dominated computing power construction, CPUs were merely a part of networking, with low unit prices and procurement tied to complete machines. Bargaining power was always held by OEMs and cloud providers. But when inference demand rises, especially as AI agents start to truly "do the work," the situation reverses.
The operational path of an agent imposes a completely different demand on CPUs. Every time it processes a request, it must first plan tasks, break down a prompt into several steps, invoke tools, read and write caches, and enable collaboration among multiple agents. Almost all these steps rely heavily on general-purpose computing, where GPUs cannot help and should not be used. This means that the CPU occupancy time and implicit value per single request have multiplied compared to the training era. Signals from the industry chain also confirm this: after advanced capacity is squeezed out by accelerator cards, global shortages of CPUs and price hikes have begun to emerge. A product once considered a marginal accessory is seeing its price surge upward for the first time. The segment holding scheduling power suddenly finds itself at the upstream of profit distribution.
A previously rare observation emerges here: the direction of supply-demand mismatch has changed. In the past, GPUs were extremely hard to get; now, the tight supply of CPUs has become a delivery bottleneck, and the shipment of entire AI servers must yield to their arrival time. To get machines on schedule, customers are willing to concede on previously fixed prices, and the profit structure of OEMs loosens accordingly. When "whether it can be delivered on time" replaces "whether more expensive cards can be bought" as the main industry contradiction, pricing power quietly shifts from component prices to delivery and scheduling capabilities. Whoever holds this chain captures the thickest margin in this cycle.
The overseas market's reaction to this change is more direct. Several leading chip companies are making their new-generation server processors alternate between two generations, using the current generation to target benchmark customers while using the next generation to secure architectural positions in next-generation data centers. Others are simply designing server chips specifically for general-purpose workloads tailored to the characteristics of AI agents, treating the penetration of the ARM architecture as a new growth ledger. This momentum, in turn, adds fuel to the supply side. After advanced process capacity is squeezed out by training accelerator cards, the wafer share allocated to CPUs tightens. Shortages and price hikes occur simultaneously in the same period, and the center of gravity for prices is tangibly raised.
The market's pricing for this is also unambiguous: multiple institutions have raised the long-term market size of server CPUs from the tens of billions of dollars to hundreds of billions of dollars, pushing growth expectations into a high range of around 40%. The data sounds spectacular, but it ultimately points to one thing—CPUs have transformed from a "supporting cost item" back into a "sought-after scarce commodity."
02. The Valuation of the Domestic Chain Splits into Two Tracks
The fact that CPUs are valuable again, coupled with the demand side's tilt toward domestic computing power, presents a more complex picture in the Chinese market. The logic is not complicated: since CPUs have once again become the value center of computing power, and both demand and ecosystem are making room for domestic hardware, the domestic chip responsible for scheduling and execution in servers steps on both paths of computing power expansion and domestic substitution. It is complex because these two paths fall on different companies, resulting in vastly different valuation flavors, much like an industry chain torn into two worlds.
Let's first look at Hygon Information, the one with the most solid implementation. It holds two product lines: high-end general-purpose processors and high-performance co-processors. CPUs secure the basic market of server general-purpose computing power, while co-processors (GPGPU architecture) are fully compatible with mainstream large models. These two legs support each other, making it one of the most commercialized samples in the current domestic computing power chain. Hygon has laid out its cards even broader than its peers—on the one hand, it continuously iterates server CPUs to handle the general-purpose computing workloads brought by inference and agents; on the other hand, it extends its reach beyond data centers, launching lightweight edge CPUs for physical world scenarios such as robotics, machine vision, and intelligent manufacturing. As computing power sinks from the cloud to the edge, this is the second growth curve it has found for itself, with boundaries much broader than "selling chips" alone.
Looking in the opposite direction, there is Loongson Technology. It does not rely on any external licensing, building everything from the instruction set system to the supporting general-purpose accelerators from scratch. This "dual instruction set system" (holding two independent instruction sets for both CPUs and GPUs simultaneously) is almost unique in China. The cost is also evident: the ecosystem must be nurtured from nothing, the software stack must be adapted bit by bit, revenue growth is slow, and financial reports have long been under the pressure of the investment period, with the financial figures for the first half of the year still hovering in the loss range. What it fights for now yields little return on financial reports; it bets on the scarcity in a corner of industrial control and industry digitalization in the future after pushing "independent and controllable" to the extreme. The market's tolerance for this track is completely opposite to the logic for Hygon—one looks at when scale will be realized, and the other bets on whether the ecosystem can grow.
Moving further downstream, the situation of OEMs and contract manufacturers becomes clearer when laid out with numbers. Sugon reported a revenue of CNY 7.466 billion in the first half of 2026, a year-on-year increase of 27.62%, and a net profit attributable to shareholders of CNY 971 million, up 33.31%. The gross margin is still climbing at 27.23%. On the one hand, it presses high-end machines on the Hygon track; on the other hand, it relies on self-developed networks like scaleFabric to fill the gap in high-speed interconnection for data centers. The scale of Inspur Information is on another level—revenue of CNY 84.379 billion, up 5.22%, while net profit attributable to shareholders surged to CNY 2.952 billion, a massive increase of 269.59%, fully displaying the profit elasticity of order-based pricing under the shortage of AI servers. However, placing a net profit growth of two and a half tenths alongside single-digit revenue growth clearly shows a business driven by "orders and bargaining," where any fluctuation in shipments will be amplified. One layer further out is channel distribution. Digital China reported a revenue of CNY 86.919 billion, up 21.42%, and a net profit attributable to shareholders of CNY 496 million, up 16.34%. The accounts run very fast, but the profits are thin. Talkweb Information saw its revenue of CNY 1.312 billion basically flat, with a net profit attributable to shareholders of CNY 69 million, down 12% year-on-year. All in the realm of "domestic computing power," Sugon takes the gross margin of designing complete machines, Inspur bets on the elasticity of shipment pricing, and Digital China and Topway earn the turnover of channel distribution—four sets of books, four ways of making money.
Looking deeper into the books, the divergence of the two tracks is essentially a misalignment of the valuation coordinate system. On one side, orders are visible and growth is steep; the market is willing to pay a premium for "certainty," anchoring on the profits of the next two to three years. On the other side, losses are still ongoing and cash flow is negative, so it can only be underpinned by "scarcity," anchoring on a long-term test paper on a ten-year scale. Adding complete machines and distribution into the mix only makes the spectrum more chaotic—they do not rely on the scarcity of chips to make a living, but on the elasticity derived from scale, delivery, and bargaining capabilities. Each holds a different passbook of price differences: those who rely on technological scarcity wait for the compound interest on the day the ecosystem matures; those who rely on scale turnover bet on the explosion in the few quarters at the peak of the boom.
This game of domestic computing power started with making up for shortcomings, and the moves often carried a national narrative. But as it reaches the middle game, the results become clear—what truly determines the seating order in this world has never been about whose story is more compelling, but about who can first convert price differences into cash on the table and draw the gross margin into an upward curve. Stories always have an ending day, but verifying financial reports is a marathon quarter after quarter. Who will sit on the dragon throne in ten years is uncertain today; the only thing that can be confirmed is that the answers in the next semi-annual report are rewriting the beginning of the next story.
- End -