You must have seen this phrase in various research reports lately:
"2026 is the year of full realization for domestic AI computing power."
Dongwu Securities said it, Huayuan Securities said it, Galaxy Securities said it. They said it resolutely, as if it were an industry consensus.
But I want to ask a simple question: What exactly is being "realized"?
If you are investing in this sector or working in this industry, this question is worth answering seriously once.
Because there is a pitfall hidden in the term "year of realization"—it blurs a crucial distinction: policy procurement, trial orders, scaled deployment, and software ecosystem maturity are four completely different gates. Their timelines are different, and their value to the industrial chain is also completely different.
Mixing these four gates together and calling it "realization" can easily lead you to systematically misjudge the actual progress.
As usual, I will try to use one article to help you see these four gates clearly.
First, establish a framework for understanding: What does "true realization" mean?
Before discussing whether computing power is being realized, we first need to understand: What stages must a computing power product go through from being "developed" to "truly creating value"?
I summarize it into a transmission chain: policy procurement → real deployment → software ecosystem maturity → scaled replication.
First Gate·Policy Procurement: Procurement driven by government funding or policies. Computing power is bought, machines are shipped, but it may not be for real business needs, but to "complete deployment tasks."
Second Gate·Real Deployment: The purchased computing power is actually used to run business applications, not left idle in the server room. This requires enterprises to have real AI needs and be willing to connect them to this computing power.
Third Gate·Software Ecosystem Maturity: Developers can smoothly write code, deploy models, debug, and optimize on this computing power, rather than requiring "customized migration" each time with high adaptation costs.
Fourth Gate·Scaled Replication: The solution based on this computing power can be promoted from top-tier large enterprises to medium-sized enterprises, penetrating from the government/enterprise market to the internet/commercial market, forming economies of scale.
These four gates are progressive. If the later gates are not opened, progress in the earlier ones may look good on financial statements, but the real value of the industry is far from being realized.
First Gate·Policy Procurement: Already Open, and Opened Wide
This gate is indeed open in 2026, and it's opened quite wide.
Galaxy Securities believes that the heavyweight release of DeepSeek-V4 is gradually shifting market expectations from policy-driven substitution to the realization of real demand orders. Dongwu Securities believes that in Q1 2026, the computing power leasing industry ushered in a "quantitative change" of increased orders and price hikes, and a "qualitative change" in business model upgrades.
The Sci-Tech Innovation Relending Facility has expanded to 1.2 trillion yuan, targeting AI and semiconductors, and the 91.5 billion yuan NDRC equipment renewal fund is also tilted towards computing power infrastructure.
According to reports, Alibaba, ByteDance, and Tencent have placed bulk orders totaling hundreds of thousands of units for Huawei's upcoming Ascend 950PR chips. Due to surging demand, the price of Ascend 950PR chips has increased by about 20%.
This number means: This is no longer "symbolic procurement," but real large-scale orders.
But beware: The opening of policy procurement does not equal the comprehensive realization of the industrial chain. How many computing power cards are purchased and how much real business these cards run are two different things.
Second Gate·Real Deployment: A Crack Has Opened, But Full Opening is Still Some Distance Away
This gate is the key breakthrough point in 2026—but it is "a crack has opened," not "the door is wide open."
The core evidence for real deployment is DeepSeek V4.
On April 6, 2026, DeepSeek V4 was officially announced to have completely abandoned the NVIDIA CUDA ecosystem, migrating 100% to Huawei Ascend chips and the CANN software framework, becoming the world's first trillion-parameter MoE large model trained and deployed on purely domestic computing power. DeepSeek broke industry惯例 this time by not providing early testing access for V4 to U.S. chip suppliers, only giving priority adaptation windows to domestic chip manufacturers like Huawei and Cambricon.
What is the significance of this? It proves that domestic computing power can support the complete training and inference of world-class large models—not "just barely usable," but actually running. This is the most powerful proof that the second gate is opening.
However, the full opening of the second gate requires not just adaptation by leading large model companies, but the real business deployment by a wide range of enterprises. Internet giants running their own models is one thing; traditional enterprises landing AI into their own production processes is another—the latter is much slower than the former.
DeepSeek V4 broke the industry pricing system with its "cent-era" pricing, promoting AI applications from pilots to普及. In the second half of 2026, the core theme of China's AI industry will shift: low-cost models will stimulate an explosion in inference demand, and domestic computing power adaptation will enter the realization period.
But there is a subtle cycle here: model prices decrease → more companies are willing to trial → real call volume increases → computing power demand becomes stronger → computing power supply increases → model prices further decrease. This positive cycle has just begun and is not yet fully underway.
Judgment on the second gate: A crack has opened. Leading scenarios are already running, but mid- and long-tail scenarios are still on the way.
Third Gate·Software Ecosystem Maturity: A Crack Has Opened, But This Crack is the Narrowest
This is the most easily overlooked of the four gates, but also the most critical one for true "realization."
NVIDIA's CUDA is an ecosystem that started construction in 2006 and took twenty years to accumulate millions of developers. Huawei's CANN currently supports over 160 mainstream AI models, while the NVIDIA CUDA ecosystem covers over 23,000 models. This gap cannot be bridged in a few months.
But this gate is opening quickly.
The most powerful signal is DeepSeek V4's adaptation strategy. DeepSeek stated that, limited by high-end computing power, the service throughput of Pro is currently quite limited. It is expected that after the batch上市 of Ascend 950 super nodes in the second half of the year, the price of Pro will be significantly reduced.
Hidden in this statement is an important signal: DeepSeek is not just "using domestic computing power"; it is actively waiting for the scaled supply of domestic computing power to ramp up, then converting this computing power capability into lower API pricing to promote broader application普及. This is a symbiotic relationship deeply绑定 between a model provider and a computing power provider, not passive adaptation.
Caitong Securities believes that 2026 is also the first year of scaled volume for domestic super nodes on the inference side. Currently, many domestic manufacturers have released new-generation super node solutions. Huawei Atlas 950/960搭载 8192/15488 computing power cards. Sugon, Moore Threads, Kunlunxin, Alibaba Panjiu, etc., all have super node layouts. Supply and demand sides are meeting each other halfway, and the industrial chain is about to enter a volume-expansion phase.
Judgment on the third gate: Top-level adaptation has been achieved, but the mid- and downstream developer ecosystem needs 1-2 years of systematic construction to truly mature.
Fourth Gate·Scaled Replication: Not Yet Opened
This is the gate currently farthest away among the four, and also the final form of "realization."
What does scaled replication mean? It means it's not just Huawei, ByteDance, Tencent using domestic computing power, but the IT systems of thousands of medium-sized enterprises, quality inspection AI in industrial manufacturing, auxiliary diagnosis systems in hospitals—all running on domestic computing power, and these customers do not feel significant migration costs.
This step has not arrived in 2026.
The core reason: The IT teams of medium-sized enterprises do not have the capability to independently complete computing power migration. Top-tier large companies have AI infrastructure teams of hundreds of people who can invest manpower in customized adaptation; a manufacturing company with 500 people might have an IT team of only three to five people. They need "plug-and-play" solutions, not computing platforms that "require six months of migration engineering."
This issue is not about chip performance, not about the software framework, but about the封装 level of the solution—it requires a complete service capability from computing hardware to the application layer, allowing medium-sized enterprises to use domestic computing power to run their own AI without needing to understand the底层.
Judgment on the fourth gate: Scaled replication is not visible in 2026; this might be something that happens in 2027-2028.
"Four Gates of Computing Power Realization" Verification Checklist
Next time you see any report about "computing power realization," you can use this verification checklist for reference:
First Gate·Policy Procurement
Verification Metrics: Scale of policy fund落地 / Number of domestic chip大单成交
2026 Status: Opened, and opened wide
Risk Warning: Procurement volume ≠ Deployment volume. Don't confuse them.
Second Gate·Real Deployment
Verification Metrics: Q1 computing power leasing increased orders & price hikes / Real adaptation status of large model vendors / Computing power utilization rate
2026 Status: A crack has opened. Leading scenarios are running, mid- and long-tail scenarios still on the way.
Risk Warning: Looking at the leaders does not equal looking at the whole picture.
Third Gate·Software Ecosystem Maturity
Verification Metrics: Number of models covered by CANN / Developer migration cost / Number of adaptation cases for medium-sized enterprises
2026 Status: Top-level adaptation achieved. Mid- and downstream ecosystem needs 1-2 years.
Risk Warning: This gate determines how deep the computing power's "moat" is.
Fourth Gate·Scaled Replication
Verification Metrics: Number of projects where medium-sized enterprises purchase domestic computing power / Case studies of vertical industry AI application落地
2026 Status: Basically not opened.
Risk Warning: This gate is the final state of "realization." Don't celebrate early.
A Final Fair Word
Saying the phrase "year of realization" is completely wrong would be incorrect. From the perspective of the first gate (policy procurement), 2026 is indeed a real realization. Domestic computing power has changed from "requiring policy subsidies for anyone to buy" to "a supplier actively competed for by large companies"—this qualitative change is real.
But if you understand "year of realization" as "the computing power industrial chain comprehensively explodes, and the performance of related companies is fully realized," then that's dangerous.
The fourth gate not being open means the current industrial landscape is still a game among a few leading players. True economies of scale need to wait for the third and fourth gates to open one after another—that will be the point of a larger, more sustained market爆发.
After completing the research for this article, I have two takeaways for your reference:
First, within the computing power industrial chain, the "realization progress" corresponding to different segments varies极大. Chip design and manufacturing (most directly benefiting from the first gate), computing power leasing (benefiting from the second gate), software toolchains (benefiting from the third gate), vertical industry solution providers (benefiting from the fourth gate)—the realization time windows for these four directions could differ by a full two years.
Second, the deep binding between DeepSeek V4 and domestic computing power is the most important industrial signal of 2026, bar none. It transforms the question from "Can domestic computing power be used?" to "When can domestic computing power be supplied?"—this is an essential shift in the narrative.
This article is from the WeChat public account "BT财经" (ID: btcjv1), author: BT财经
