2026.09.23 | AC | Semis | Agentic CPU Demand: Hygon, Loongson, Montage

摘要

1. Which part of CPU demand Agentic actually lifts, orchestration not the whole chip；
2. Whether x86's near-term edge and ARM's hyperscaler-only gains are real or overstated；
3. Why P-cores and E-cores benefit from opposite dynamics within the same CPU；
4. How much of the $220B CPU TAM survives historical realization-rate discounting；
5. Whether Hygon's CPU or Montage's memory-interface chips offer more certainty in A-shares.

> Agentic AI is not lifting CPU demand across the board; it is lifting the slice of compute that handles orchestration. That demand fragments along three axes — architecture, core type and vendor — and the incremental dollars that actually materialise accrue to a handful of links with mature ecosystems, fast time-to-delivery and a chokehold on orchestration. In A-shares, Hygon is the CPU maker best aligned with the orchestration thesis, but the highest-conviction exposure is not the CPU itself — it is the memory interface, whose content scales with core count.

The market's new consensus is unambiguous: agentic AI is repricing the CPU. CPU-to-GPU ratios are swinging back, TAM estimates are doubling, and the sell side is busy rotating its picks. The logic sounds airtight: agents do reasoning, decision-making and tool calling — work that GPUs cannot handle and must hand back to the CPU. What this rerating tends to overlook, however, is that agentic AI does not lift all CPUs.

1. Orchestration, Not the Whole CPU, Captures the Upside

To identify which CPUs agentic AI actually lifts, start with what training and inference are really computing. Training is essentially SIMD — single instruction, multiple data — in which massive datasets are fed repeatedly through a largely fixed set of instructions. Such highly parallel workloads are the GPU's home turf, with the CPU playing little more than a supporting role. Agents do something different: reasoning and decision-making, tool calling, and the orchestration and scheduling of memory, applications and the operating system. This branch-heavy, serially dependent work is precisely what CPUs — and mainly CPUs — are built to carry. The swing in ratios therefore rests on hard logic. We believe the CPU:GPU ratio is reverting from the GPU-dominated ~1:8 of the training era towards 1:1. AMD management has said explicitly that for certain agentic workloads the ratio will move to 1:1 or tilt even further towards the CPU, and TrendForce's published research points the same way, showing a shift from 1:4–1:8 to 1:1–1:2. In inference, the ratio works out to roughly one CPU for every two to three GPUs; at the edge and in "physical AI", it falls to 1:1.

Yet a swing in ratios does not mean every CPU benefits — the first flaw in the consensus's indiscriminate view. There is an easily missed bifurcation here. High-concurrency, Alexa-style simple inference barely taxes CPU compute; what saturates the CPU is complex agentic queries — tool calls, web retrieval, multi-document synthesis, and state management across dozens of concurrent agents. This splits CPU demand in two: one half scales with agent complexity, while the other — where the CPU merely acts as a "host" for simple inference — is actually being flattened as inference grows ever more efficient. In other words, "agentic AI reprices the CPU" is only half right: what gets lifted is the specific capability of orchestration, not every core in the data hall. Divergence is the real starting point of this report; the question that follows is which architectures, and which cores, it lands on.

2. x86 Inertia Wins the Near Term; Arm Pays Off Only for a Few Custom-Silicon Hyperscalers

Slicing further along the orchestration axis, the first divergence is at the ISA level — and in the near term, it is x86's ecosystem inertia that absorbs the demand. The reasoning is simple. Over four decades, almost every application, database and precompiled binary has been optimised for x86; under deployment constraints where timelines are tight and failure is not an option, x86 is the lowest-risk default. Binary migration from x86 to Arm is extremely difficult and, in many cases, cannot be done reliably, while hyperscalers prize fleet uniformity and software stability. Together, these give x86 a structural near-term edge. At the company level, AMD is the structural winner within x86 this cycle. Its Zen 6 EPYC "Venice" scales to 256 cores and, on the official 2027 launch cadence, lands roughly a year ahead of Intel's Diamond Rapids. According to Mercury Research, AMD captured a record 46.2% of x86 server revenue in 1Q26, with about two-thirds of the growth coming from unit shipments rather than pricing, and its data center revenue doubled YoY in 2Q26. The share shift was not bought through a price war but earned through genuine product competitiveness — which makes it far higher quality.

Arm's gains are real, but they are concentrated in a few hyperscalers with custom silicon rather than spread across the Arm camp — the second flaw in the consensus's indiscriminate view. In our June report on Arm servers, we unpacked a critical definitional issue: of the widely cited 21% Arm share of server shipments, the bulk consists of CPUs attached passively to Nvidia GB/Grace AI hosts. Stripping those out, we estimate Arm's genuine competitive share in general-purpose cloud rental and enterprise on-premises deployments at only about 4–5% — a conclusion that still holds this cycle. When Arm management cites "roughly 50% at the leading hyperscalers", it is referring to in-house custom silicon (Graviton, Axion, Cobalt), not merchant share won in open competition; the two differ by an order of magnitude. More consequential still is Nvidia's Vera. Every NVL72 rack that ships with this 88-core Arm processor expands Arm's footprint rather than x86's, structurally capping AMD's and Intel's share of AI-attached CPUs. Divergence is therefore at work inside the Arm camp too: the winners are the few hyperscalers able to absorb the migration burden and wall off their agentic clusters — not merchant Arm at large.

3. P-Cores vs. E-Cores: Opposite Payoffs Inside the Same Chip

The second divergence is more granular and sits inside the same chip: performance cores (P-cores) and efficiency cores (E-cores) benefit in opposite directions. Arm management has said publicly that agentic AI could push CPU core demand up as much as fourfold — from about 30mn to 120mn cores per gigawatt — and stresses that this is a "conservative estimate". Multiplying core counts structurally favours smaller, denser E-cores. But there is a trade-off that is easy to gloss over: high-performance orchestration and single-threaded decision-making still depend on P-cores, and E-cores typically trail P-cores by roughly 30–40% in single-thread performance. Different parts of the same CPU thus serve different tasks — decision-making and orchestration demand high single-thread frequency, while MoE-style multi-context parallelism demands core density. Their payoffs pull in opposite directions, which makes "more cores is better" far too blunt a rule.

Product roadmaps already trace this fork clearly. Intel's Clearwater Forest (marketed as Xeon 6+) packs 288 E-cores for high-density inference, while Diamond Rapids returns to the P-core route — up to 256 P-cores, per Hot Chips 2026 — for serial workloads. Arm's own data center CPU (136 Neoverse V3 cores, no SMT) goes all-in on high-frequency cores dedicated to "the serial reasoning that agents need". The split among these three roadmaps maps neatly onto the P-core/E-core payoff divergence. Our earlier report offered evidence from the other side: warp divergence from sparse activation and the PCIe memory wall can push effective GPU utilisation below 10%, handing the advantage to high-core-count CPUs with directly attached memory. Put the two together and the implication is clear: E-cores carry thinner margins than P-cores, so vendors must scale them carefully to keep low-priced E-cores from cannibalising premium P-core profits. The strength of the benefit cannot be generalised; it has to be filtered layer by layer — first by architecture, then by core type.

4. The US$220bn CPU TAM: Right Direction, Inflated Magnitude

With the beneficiary structure mapped, the next question is how big the pie really is — and the oft-quoted US$220bn TAM is right on direction but needs a haircut on magnitude. Estimates of the 2030 server CPU TAM have been revised up relentlessly this year. Narrower definitions, dominated by non-AI workloads, still put it at just US$55–60bn; AMD guided to US$120bn in March, and by its August earnings call management had lifted the figure to roughly US$220bn — implying a CAGR above 50% — while explicitly naming agentic ("sandbox"-style) workloads as the largest and fastest-growing slice of the 2030 TAM. An ~83% upward revision within a single quarter is itself a textbook overshoot signal: once a direction is confirmed, sentiment tends to run ahead of fundamentals.

This is exactly where historical calibration should cool the enthusiasm. In our reports of 7 May and 8 May, we applied the median realisation rate (55–65%) from eight comparable cases over the past two decades and concluded that the 2030 TAM would more likely land at US$72–102bn. Our view then was that the market was in the early stage of an overshoot and should be priced on a 12–36-month timeline rather than on linear optimism. With the headline figure having nearly doubled again since, the same discount still applies: the direction — agentic AI rerating the CPU — is solid, but the realisation rate on magnitude is what determines returns. That judgement drives the investment implication directly: own the links with the highest certainty of delivery rather than spreading exposure evenly across every CPU name on the most aggressive TAM. It also brings us to the final step — in A-shares, which link is the most certain?

5. Among CPU Makers, Hygon Sits Closest to Orchestration

Turning to A-shares, start with the CPU makers themselves: Hygon is the best aligned with orchestration — but one old narrative must be struck off first. China's domestic CPU landscape has converged through market forces rather than policy mandate. Arm (Kunpeng) and x86 (Hygon) dominate the main battlefield, while LoongArch (Loongson) anchors the government-security base. In probability terms, we rank the likely winners Kunpeng > Hygon > Loongson, and a "two-giant balance" should prevent any single player from dominating. Layering the agentic divergence on top requires one distinction: the overall ranking is unchanged, but on the narrower axis of orchestration fit, Hygon's x86 ecosystem inertia maps most closely onto the beneficiary logic. Professional software, databases and legacy binaries all sit on the x86 side, making Hygon the closest domestic analogue to AMD. Kunpeng and Phytium pursue Arm's high-core-count, performance-per-watt route — the right direction, but equally held back by the application ecosystem. That is why, among CPU makers, we place Hygon closest to the thesis.

Hygon's fundamentals support the thesis, but two issues need watching — otherwise "well aligned" is easily misread as "risk-free". In 1H26, Hygon's revenue rose 66.52% YoY to RMB9.099bn and net profit attributable to shareholders climbed 49.69% YoY to RMB1.798bn, driven by large-model iteration, AI-agent deployment and the commercial rollout of Xinchuang (IT localisation) — genuinely impressive growth. The first issue is the old narrative to strike off: the share-swap merger by absorption with Dawning was formally terminated on 9 December 2025. The "chips + systems + compute services" full-stack synergy story is gone, and Hygon must now be judged on its own product lines. The second is quality and timing. Gross margin was about 55% in 1H26, down roughly 5ppt YoY on higher costs for upstream materials and memory-related components, while operating cash flow swung from an inflow a year earlier to an outflow of about RMB428mn. Beneath the top-line growth, both earnings quality and cash flow are under pressure. The real catalyst awaits the ramp of the C86-5G (128 cores, 4-way SMT), expected across 2026–27. "Hygon is well aligned" and "Hygon is risk-free" are therefore two different claims; the stock should be viewed on the premise of no merger story, close monitoring of margins and cash flow, and patience for the volume ramp.

6. Memory Interfaces: The Highest-Conviction A-Share Expression of the Thesis

Beyond the CPU itself lies a higher-conviction landing point — and the one this report most wants to emphasise: memory interfaces, whose content scales with core count. Why is it more certain than the CPU itself? Because CPUs are constrained by both process technology and ecosystem, leaving a long path to realisation with many variables. Memory-interface chips, by contrast, are priced per module and scale per core; their demand is independent of the direction of memory-chip prices, and doubling core counts translates almost directly into higher unit volumes. We analysed this in our 16 September report. The core conclusions were that the increment corresponds to the 2027 CPU platform refresh, and that MRDIMM lifts value per module rather than value per server — so it shows up as revenue leverage, not gross-margin leverage. Rather than repeat those findings, we simply position them within the agentic divergence framework: this is the most certain A-share expression of the thesis. Montage posted 1H26 revenue of RMB3.335bn, up 26.66% YoY, and net profit attributable to shareholders of RMB1.997bn, up 72.33% YoY. In both scale and certainty, no CPU name comes close.

Stringing these threads together yields a map of landing points ranked by certainty — and our answer to the core question. Most certain are memory interfaces whose content scales with core count: Montage, plus Giantec, whose exposure runs the other way and is tied purely to module count. Next come x86 CPUs with the strongest ecosystem inertia, where Hygon sits closest on the orchestration-fit axis, though margins, cash flow and the volume ramp need watching. Third is the Arm route of Kunpeng and Phytium (with listed exposure via China Greatwall): the direction is right, but it is doubly constrained by ecosystem and process technology, and the benefits concentrate in a few closed clusters. Last are Loongson and RISC-V. The former holds the government base but remains loss-making; the latter is virtually absent from the server market and looks more like a long-dated option beyond 2027. The ranking rests not on "whose story is bigger" but on "whose benefits can be realised — and falsified". That is how the agentic CPU rerating gets brought down to earth.

7. Watch Points and Risks

Five watch points. First, and most fundamental: can the CPU:GPU ratio in production agentic deployments genuinely tighten beyond 1:4, i.e. fewer than four GPUs per CPU? If not, the premise of the entire TAM rerating collapses. Second, Mercury Research's quarterly x86 server revenue share — can AMD hold above 45%? A reversal would signal that share gains have stalled. Third, Intel's execution on 18A and Clearwater Forest volume production; a clear upside surprise would rerate Intel and slow AMD's share gains. Fourth, how many DIMMs per channel Diamond Rapids supports — the single most important variable in determining whether memory-interface content per server rises 1.5x or 3.0x, and one that ties directly into the Montage thesis. Fifth, the pace of Hygon's C86-5G ramp and whether operating cash flow turns positive, along with actual progress on ISA convergence in the second phase of Xinchuang. Each of these maps onto a judgement in an earlier section; should any be falsified, the corresponding conclusion must be revised.

Upside risks lie in faster-than-expected agentic complexity and domestic substitution; the downside hinges on a single gate on the demand side. The upside is straightforward: if agentic complexity outpaces expectations, ratios tilt further towards the CPU, or domestic substitution and Hygon's ramp surprise positively, the beneficiary links will realise gains faster. The real concern is the downside. The bullish narrative that compresses the enterprise CPU refresh cycle to 1.5 years rests on an unproven assumption — that enterprises are willing to hand their core data to external AI. Four barriers — regulatory compliance, the SaaS trust gap, internal data governance and geopolitical data sovereignty — form the real gate, as discussed in an earlier report. The pace of the hyperscaler arms race cannot simply be extrapolated to enterprise demand, and this is the biggest demand-side constraint on our thesis. Two further items bear watching. The 50% Affiliates Rule under the US Entity List is set to be automatically reinstated on 10 November 2026; any tightening on Hygon and its affiliates around that date — as well as any upside surprise in Intel's 18A execution — would redraw the ranking above.

The above content does not constitute any investment advice. Invest at your own risk.