Acrostic poem

Many people stare at the order book every day for short-term up-and-down moves, agonize over yesterday’s market fluctuations, and endlessly scrutinize already-released earnings reports—but very few take a moment to think ahead and forecast the real order-book and industry momentum for the second half of the year. Today, combining firsthand industry observations, this piece breaks down the core logic of the 2026 second-half compute-power industry, distinguishing which sectors have long-term incremental growth and which are only short-term price-increase bubbles. [Taoguba]
I. Compute-power infrastructure’s long-term logic will not end; domestic substitution is a rigid main line
During on-site exchanges, many peers asked: when will metaverse applications usher in a breakout? My view has never changed: don’t obsess over the timing of application deployment—compute-power infrastructure will always move ahead of applications. The newly deployed 100,000-card supercluster has already been adapted for more than 100 AI research and industrial applications, enough to support current large-model training and inference demand. Meanwhile, the tendering and planning of domestic “intelligent computing” data centers continues to be rolled out. The infrastructure build cycle is long, far from a short-term speculative market. The tightening of overseas high-end compute-power supply has become a settled fact; overseas chip prices have kept rising. As a result, market demand can only shift comprehensively to domestic compute-power supply chains—this is a change everyone in the industry can directly feel.
From feedback on orders at the front line of industry, for domestic servers, supernodes, and the “xinchuang” (indigenous innovation) sector, this main line’s momentum will only be reinforced. The tightening of demand openings is merely a short-term surface phenomenon. What runs through multiple years of the long-term industrial trend is independent, controllable domestic compute power.
II. The server industry is entering a performance inflection point; storage dividends drive profits to double
Last year I consistently didn’t look favorably on the purely server-assembly sector. The key reason was that the industry business model was thin and profits were meager, with no substantive performance inflection point. But this year, the market has undergone a complete qualitative shift. All-in-one equipment manufacturers have benefited from the upcycle in storage: storage product prices are rising, and OEM/integration manufacturers’ profitability has been lifted dramatically. Net profit margins have achieved significantly enhanced earnings elasticity, and the industry’s net profit margins are seeing a repair-and-upward trend.
Many sectors have already delivered impressive answers with first-half performance, but investors can’t only look at the “results already achieved” in the first half. The core is to track the incremental order growth that continues into the second half. The industry’s traditional peak seasons of “golden September and silver October” are about to arrive. During the back-to-school season, government and enterprise orders, as well as internet-sector procurement, will be released in concentrated fashion. Combined with the traditional promotional peak for consumer electronics in September to October, full-year performance increment will be concentrated in the second half. The industrial inflection point is the key signal for positioning. For sectors without fundamental changes, it is hard to sustain a continuous rally.
III. Distinguish two types of sectors: price-increase bubbles versus true industrial increments—vastly different
At present, market sectors can be clearly divided into two categories, with completely different profit logics—do not mix them up:
The first category: optical communications and PCB sectors. The first-half momentum was extremely hot and performance was well realized. However, the early-stage surge has already fully priced in future expectations of the industry’s prosperity. Market trading prices have already counted in industry boom ahead of time, leaving little room for upside beyond expectations afterward. Even if the sector’s long-term fundamentals are fine, the short-term capital “value for money” has already fallen sharply.
The second category: domestic servers, domestic supernodes, and “xinchuang” hardware sectors. Currently they are still in the undervalued range. Industry orders continue to expand, but the market has not yet fully priced in the incremental earnings growth for the whole year. These sectors offer ample room for expectation gaps.
Here we also need to be objective: the overseas compute-power industry chain hasn’t completely lost value. Optical modules and PCBs still have long-term industrial value; it’s just that the short-term market has priced in too much, and adjustment is needed. Meanwhile, domestic compute-power sectors represent incremental new demand, so there is no problem of stock being “priced in” or “wiped out.” Like water flowing to lower places, capital always switches among highs and lows within the main line. Avoid the overrun sectors and dig into sub-segments with undervalued expectation gaps.
IV. Key constraints of the industry: domestic chip capacity and yield
The biggest divergence in the domestic compute-power sector today isn’t insufficient orders, but rather limited supply due to constrained chip capacity and yields.
Whether it’s domestic DCU, inference chips, or new-generation switch chips, manufacturers may have sufficient order backlogs—but the ramp-up speed of capacity can’t keep up with downstream demand. The market even shows stockpiling behavior, which further intensifies supply-demand tightness.
This Zhengzhou conference focused on the BW1100 new-generation inference switching chip. Its chip parameters are adapted to agent inference scenarios, and compared with the prior-generation product, it significantly reduces compute-power losses. However, the biggest uncertainty at present is whether mass production can be achieved on a large scale. Once mass production is realized and deployed, it will open an entirely new growth space for domestic inference compute power. Also, I’ve consistently maintained an objective stance: compute-power leasing and general cloud computing are not high-quality business models. In China, top internet cloud vendors have long struggled to achieve stable profitability—requirements for asset investment are heavy, depreciation cycles are long, and it’s extremely difficult to model performance. I will actively avoid these kinds of sectors. The largest expectation gap in the industry always comes from uncertainty in technical deployment and capacity release.
V. Objective cycles of market: alternating between crisis and hope
The market constantly loops back and forth between pessimism and optimism. Today the order book looks bleak in every corner; the next day it will likely see a broad-based rally. It’s easy for market participants to forget the regularity of the cycle.
The best way we can respond is: when the market hits extreme panic, stay rational and proactive; when the whole market is celebrating, know when to withdraw quickly. At present, the market is in a phase of steady adjustment and risk prevention. Volatility risk objectively exists, but the industrial logic of domestic compute-power substitution has not changed in any way. At the same time, we should also objectively look at the market’s capital structure: the market’s main funds are not a monolith. There are major internal disagreements—different teams have sharply diverging judgments about industry momentum. Don’t simply use a single mindset to forecast where funds will go. For long-term positioning, you also need patience: factory expansions, capacity ramp-ups, and enterprise growth all require long cycles. Expansion pace for rigorous manufacturing enterprises is slower. Only by holding onto the long-term industrial logic can we ignore the noise of short-term market fluctuations.
The deployment of this Zhengzhou 100,000-card supercluster is only a milestone in the domestic compute-power infrastructure wave, not an endpoint. If you’re observing the industry, you must look ahead: forecast the orders, capacity, and momentum half a year from now, not repeatedly review past market. Distinguish short-term cycle-driven price-increase dividends from long-term industrial increment sectors; avoid popular sub-segments that over-discount expectations, and deeply explore low-position sectors like domestic compute power and xinchuang that have sustained order-book increment and have not yet been fully priced. Market short-term sentiment can influence fluctuations, but the true long-term supply-demand reality of the industry and the major trend of long-term domestic substitution will ultimately be reflected in solid, real earnings growth for companies.