An overlooked high-profit sector: single-chip prices increase by 20%, with a 47.5 billion order backlog into next year!

After Chen Xiaoqiong, another new stock legend has emerged![Taoguba]
And it’s directly said that this is the only person in China’s 36-year history who lets retail investors make money lying down!
You’re saying will the final result be the same as Xiaoqiong’s results?
And more importantly, the direction of the new layout is what we wrote about in our March 29 article:
《The next 10x track! New King of Weight-Loss Drugs is born: Menowave JH389! 80% lower cost, 3-year 10x growth in penetration!》
But we suggest you still be careful about traps. You still need to improve your own understanding overall to obtain your own wealth!

And today I don’t want to write about innovative drugs. Today I’ll further think about the research: China-made chips! In 2026, a major explosion may be coming. The core thing is to pay close attention: Huawei Ascend GPU chips!

I. The 2026 boom of domestically produced AI chips (GPUs): five key catalysts

2026 is a historic turning point for domestic AI computing power—from “usable” to “large-scale commercial use.” Five factors resonate together to ignite the explosion:
1、Token demand explodes: China’s AI computing demand leads globally

Token boom: In 2026, China’s average daily Token calls will exceed 140 trillion, with annual growth of over 10x; in mid-February it reached 5.16 trillion Tokens, soaring 127% in three weeks, and inference demand shows “inflation-style” growth.
Application-driven: Business commercialization of large models such as Doubao, Qianwen, DeepSeek, etc. is rolling out across multimodal, Agents, and enterprise-level services; the compute gap on the inference side reaches a million-card level.

2、Technical revolution: Ascend + DeepSeek V4 establishes a new “domestic-first” paradigm

Ecosystem upheaval: DeepSeek V4’s entire-stack underlying code is rebuilt, migrating completely from CUDA to Huawei CANN, granting Ascend several weeks of exclusive tuning optimization; it marks the opening of the Co-design era where “models define chips, and chips feed back into models.”
Performance domination: Ascend 950PR decodes 1920 Tokens/s per card with latency of 50ms; inference performance reaches 2.87x that of H20; OptiQuant achieves INT8 = FP8 precision, with cost down 70%+; Lingqu 2.0 supports ten-thousand-card clusters, with bandwidth ×3 and power consumption -85%.

3、Domestic substitution is an urgent necessity: sanctions forcing + supply-chain security
U.S. blockade: Nvidia H100/H200 are embargoed, and H20 performance is throttled; the risk of high-end compute supply being cut off becomes a normalized issue.
Self-reliance and controllability: DeepSeek V4 discards Nvidia 100% and Internet tech giants incorporate domestic compute into their R&D baseline (not an alternative). Supply-chain security priority surpasses short-term cost.
4、Major breakthrough in domestic substitution for 2025: laying the foundation for the explosion
Share leap: In 2025, shipments of domestic AI accelerators reached 1.65 million cards, with a market share of 41%; Nvidia fell from 95% to 55%.
Leading enterprises: Huawei Ascend shipments of 812k cards (49% domestic, 20.3% overall); Pingtouge 265k cards (6.6%); Kunlun Xin and Cambricon each 116k cards (2.9%).
5、Policy + order cycles drive from both sides: full commercialization
Policy push: the third National Big Fund, “East Data West Computing,” and intelligent computing centers prioritize domestic products; government procurement bans imported chips.
Massive orders: ByteDance purchases Ascend for 40 billion (200k cards); Tencent 60–70 thousand cards, Baidu 30 thousand cards; Huawei Ascend’s 2026 target is 1.2–1.5 million cards.

II. DeepSeek V4 adaptation: “chain reaction” that sparks domestically made GPUs

1、Why it’s a decisive turning point? (three major qualitative changes)
Breaking the lock on the ecosystem: it ends the dead loop of “hardware strong → software weak → developer exodus.” With CANN, from “can be used” to “can be used well,” the adaptation rate is 98%, training speed +60%; migration cost <10%, fully covering the CUDA wall.
Co-design flywheel: V4 customizes MoE parallelism for Ascend. The chip is optimized according to model needs, forming a positive loop of “model → chip → more models.”
Commercial verification: top-tier global large models run natively on domestic chips, proving that it can be optimal even without Nvidia; it completely removes concerns from large enterprises.
2、2025–2028: demand, production, market size (institutional forecast)
Shipments and domestic substitution rate (10,000 cards)

(2) Market size and unit price assumption scenario analysis

Logic upgrade: domestic substitution has moved from the stage of “fighting over market share” into “reconstructing pricing power.” A 20%+ price increase for the 950 series marks that domestic compute has officially bid farewell to “low-price price wars” and entered the era of “performance premium + full-stack value substitution.” The total system market space is about 3–5 times that of chips, and the value amount expands step by step along interconnects, heat dissipation, and server racks.
3、Comparison of China AI-GPU market share TOP5 (2025 actual vs 2026 forecast)

III. The held-back trump card: DeepSeek V4 reshapes the entire domestic semiconductor chain

1、Chip design: from “simulation” to “data-driven”
Stress testing: the V4 trillion-parameter MoE provides an industrial-grade workload for Ascend 950, and optimizes bandwidth, HBM, and quantization units in reverse.
Iteration acceleration: design cycle -50%, R&D efficiency ×2; the next-generation Ascend 960 is directly customized to meet V4 demands.
2、Wafer manufacturing: advanced-process capacity prioritized for compute
SMIC 18nm (1.8nm) mass production: annual capacity 500k wafers, with 70% allocated to AI chips.
HuaHong 7nm breakthroughs: yields 90%+; in 2026, an added 16k wafers per month will be prioritized for Ascend/Cambricon.
Capacity tilt: mobile chip wafer starts -30%; Huawei’s 47.5 billion order locks in foundry capacity.
3、Advanced packaging and HBM: domestic substitution speeds up
2.5D/3D packaging: utilization rate ＞85%; domestic silicon interposer layers and underfill epoxy move from backup to mainstream.
Self-developed HBM: Ascend 950 integrates HiBL 1.0 (112GB, 1.4TB/s), eliminating dependence on overseas supply.
4、Software ecosystem: CANN becomes the domestic standard
Full coverage of adaptation: PyTorch/TensorFlow compatibility ＞90%; migration of major large models in 3 days.
Developer explosion: CANN developers exceed 1 million in 2026; domestic AI frameworks (MindSpore/Paddle) have a market share ＞30%.

IV. Advanced process + capacity tilting: the cornerstone of domestic substitution

No matter how good the technology is, capacity is king. In 2026, three major breakthroughs on the manufacturing side
1、Process breakthrough
18nm (1.8nm) mass production, 7nm mature, 28nm yield 95%+
Core equipment domestic substitution 75%, avoiding ASML bans
2、Capacity explosion
In 2026, AI chip capacity ×3, reaching 15 million chips/year
Three new plants dedicated to AI chips: SMIC South, HuaHong, and JCET/Changdian
3、Allocation revolution
Compute priority: advanced process 70% → AI chips; mobile 30% → 15%
Big enterprise orders lock capacity: ByteDance/Tencent/Baidu prioritize supply
Conclusion: with dual breakthroughs in process and capacity, the pain point of “able to design but unable to manufacture” is fundamentally solved. Domestic GPUs will have no capacity bottlenecks starting in 2026.

V. Ascend servers: 10 institutional “consensus” viewpoints—increment and upside surprises

1. Demand side: from pilots to scaling (three major upside surprises)
   In 2026, the timeline and delivery rhythm for major internet companies’ Ascend server introductions (latest institutional viewpoint)

From Table 1, you can see three marginal “upside surprise” changes:
Import volume targets double: experts clearly point out that in 2026, the import targets for Ascend at big companies such as Alibaba, Tencent, Baidu are “all doubled in import volume,” not just a modest expansion.
Delivery rhythm moves forward and paths are verifiable: ByteDance 910C is fully delivered before April, and goes live in mid-to-late May; 950 moves into large-scale mass production starting mid-April and conducts mid-term evaluations at 10,000-card clusters when reaching 10,000–15k cards. After passing the evaluations, it enters long-term framework orders. The path for large orders landing is clearly defined.
Tencent/Baidu: shifting from “pilots” to “production validation + framework orders”: Tencent requires completing 1–2 rounds of closed-loop testing in production environments before batch imports; Baidu already has orders at the ten-thousand-card scale and long-term framework agreements, reflecting increased confidence in Ascend’s stability and faster procurement decisions.

2、Product strength: threefold turning points in performance/cost/ecosystem
Performance: 950PR H20 2.87x, FP4 memory -75%
Cost: TCO -60%, single-Token cost -70%
Ecosystem: CANN 98% compatibility, ten-thousand-card stability 99.7%
4、Ecosystem maturity: from “can be used” to “can be used well”
Model adaptation: full migration of LLaMA 3/GLM-5/Qwen to Ascend
Cluster deployment: ByteDance 160k-card Ascend cluster; Baidu ten-thousand-card inference cluster go live
Toolchain: end-to-end domestically built from training → inference → deployment → operations and maintenance

VI. Benefiting companies across the entire industrial chain

Summary: an industry inflection point from “usable” to “irreplaceable”
1、A logic closed loop with 100% certainty: Token demand index explosion (pull) + V4 native adaptation + CANN ecosystem (carry) + 7nm/2.5D capacity tilting (guarantee) + internet giants locking orders worth a trillion (validation) = domestic AI chips entering the main uptrend in 2026.
2、Switching investment paradigm: from “trading expectations/wafer-spinning” to “watching deliveries, watching yields, watching TCO.” Order visibility stretches out to 2–3 years. Performance fulfillment replaces concept-driven speculation as the valuation anchor.
3、Value migration path: compute value is no longer limited to “silicon wafer area,” but diffuses into “advanced packaging density, interconnect bandwidth, liquid-cooling efficiency, and software stack maturity” across four dimensions. Manufacturers with system-level integration capability will obtain stronger pricing power.
2026 is the first year of domestically made compute: from “substitution” to “leadership,” from “usable” to “optimal.” The entire domestic compute chain led by Huawei Ascend + domestic chips will dominate China’s AI compute market, and by 2028 achieve full self-sufficiency and global exports.
But you need to understand: who has more value? And think about valuation! For example, Cambricon, with shipments of 110k, has a value of 400 billion. If its shipments are 8x lower than Huawei Ascend, should Huawei Ascend be valued at over 3 trillion? Then should its suppliers also be valued around 10%? Could there be companies around 300 billion? Or even 200 billion? The key is that right now there isn’t any company with over 100 billion. Do you think such value could happen in the future? Who might it be? First: Huawei’s new “son,” meaning a company that is either invested in or controlling！（Who would it be?）
Second: Huawei’s new incoming supplier. Thinking about low valuation!（Who would it be?）
If it’s not that you know: like + share + comment: **A previously overlooked ultra-profitable track: single-chip price up 20%+, 47.5 billion order backlog scheduled until next year, and the leader’s Q4 performance has already exploded by 400%! **