Why Does the Term 'Year of AI Computing Power Realization' Have Pitfalls? —Understanding the Four Hurdles from Policy Signals to Actual Orders in One Article

marsbitОпубликовано 2026-05-08Обновлено 2026-05-08

Введение

This article critiques the phrase "The First Year of AI Computing Power Cashing In," arguing it oversimplifies a complex, multi-stage process. It proposes a "Four Gates" framework to assess the true commercialization of domestic AI computing power (like Huawei's Ascend chips): 1. **Policy Procurement:** Widely open in 2026. Significant government funding and large bulk orders from tech giants like Alibaba and Tencent exist. However, purchasing hardware is not the same as deploying it for real use. 2. **Real Deployment:** A crack has opened. The key evidence is DeepSeek V4, a top-tier AI model fully migrating from NVIDIA's CUDA to domestic computing platforms. This proves the capability for real, high-level tasks, but widespread adoption beyond leading tech firms is still nascent. 3. **Mature Software Ecosystem:** A narrow crack has opened. While frameworks like Huawei's CANN are progressing, they lag far behind NVIDIA's vast, established CUDA ecosystem in terms of supported models and developer ease-of-use. Building this middle-to-downstream developer environment is estimated to need 1-2 more years. 4. **Scalable Replication:** Essentially closed. This final gate, where thousands of mid-sized enterprises across various industries can easily adopt the technology without major migration costs, is not expected before 2027-2028. The core risk is conflating these stages. While 2026 marks a real turning point in policy-driven procurement and proving technical viability (Gates...

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

Связанные с этим вопросы

QAccording to the article, what are the four stages (or gates) that AI computing power needs to pass through from development to truly creating value?

AThe four stages are: 1) Policy-Driven Procurement, 2) Real Business Deployment, 3) Mature Software Ecosystem, and 4) Large-Scale Replication.

QWhat is the key evidence mentioned in the article that the second gate (Real Business Deployment) has begun to open?

AThe key evidence is DeepSeek V4's official announcement to completely abandon the Nvidia CUDA ecosystem and 100% migrate to Huawei's Ascend chips and CANN framework, proving that domestic computing power can support the full training and inference of a world-class large-scale model.

QWhy does the article say the third gate (Mature Software Ecosystem) is the narrowest and most critical for true 'realization'?

ABecause Nvidia's CUDA ecosystem has been built over 20 years, covering over 23,000 models, while Huawei's CANN currently supports around 160. Bridging this gap in developers and model coverage takes time, and a mature ecosystem is crucial for reducing migration costs and achieving widespread adoption.

QWhat is the main reason given for the fourth gate (Large-Scale Replication) not being open yet in 2026?

AMid-sized enterprises lack the in-house IT capabilities (unlike tech giants with large dedicated teams) to handle the customized migration work required to adopt domestic computing power. They need fully packaged 'plug-and-play' solutions, which are not yet widely available.

QWhat does the article identify as the most important industrial signal of 2026 regarding domestic AI computing power?

ADeepSeek V4's deep binding with domestic computing power. It transforms the industry narrative from 'Can domestic computing power be used?' to 'When can domestic computing power supply meet the demand?', representing a fundamental shift.

Похожее

AI Relay Stations Spark Heated Debate on Zhihu: Behind Cheap Tokens, What Are Users Really Worried About?

A discussion on Zhihu about "AI relay stations" shifted the niche developer topic of "cheap tokens" into broader user awareness. Users moved beyond simply questioning the legitimacy of these services to focus on practical concerns: Where do cheap tokens truly come from? Is the model being accessed the real one? Can relay stations see prompts, code, and API keys? For occasional users, are the risks worth it? The core debate centered less on price and more on trust. A primary worry is model authenticity—the risk of "model swapping," where users paying for a premium model might be routed to a cheaper one, creating an information asymmetry. Others argued that cost comparisons matter; while cheaper than official pay-as-you-go APIs, relay stations may not be the lowest-cost option versus subscriptions, domestic models, or free tiers, making user needs assessment crucial. Speculation about token sources ranged from legitimate bulk discounts to gray-area methods like account sharing or exploiting regional pricing. This opacity makes risk assessment difficult for users. Data security emerged as a critical concern, especially for enterprise use. When processing sensitive information like code, contracts, or client data, the inability to verify a relay station's data handling, retention, or access policies poses significant compliance and confidentiality risks. The evolving consensus suggests relay stations can be used cautiously for low-sensitivity, disposable tasks (e.g., summarizing public info, simple translation). However, they should not be the default for sensitive, professional, or production workflows involving proprietary data, Agents, or automated systems. Recommendations include avoiding large prepayments, not relying on a single service, using test prompts to monitor quality, anonymizing data where possible, and keeping official channels as backups. Ultimately, the discussion framed tokens not just as a billing unit but as a measure of real cost encompassing price, model integrity, data security, and service stability. The popularity of relay stations highlights user demand for affordable access, but the debate underscores a key trade-off: the savings from cheap tokens may come at the price of trust, transparency, and control over one's data and AI experience.

marsbit32 мин. назад

AI Relay Stations Spark Heated Debate on Zhihu: Behind Cheap Tokens, What Are Users Really Worried About?

marsbit32 мин. назад

In-Depth Research Report on TradFi: The Convergence Wave of Crypto and Traditional Finance

In 2026, the crypto industry is undergoing a profound infrastructure-level transformation—TradFi assets are migrating on-chain at an unprecedented pace. According to CoinGecko's Q1 2026 report, the total value locked (TVL) of tokenized real-world assets (RWA) has surpassed $31 billion, a nearly 4x increase from $7.8 billion at the beginning of 2025, with the sector’s aggregate market capitalization reaching $19.3 billion. Among these, the market cap of tokenized stocks surged from $2 million to $486 million, with Q1 spot trading volume reaching $15.1 billion—a single quarter already surpassing the entire second half of 2025. RWA perpetual contract Q1 trading volume reached a staggering $524.8 billion, far exceeding the $313 billion for all of 2025. Meanwhile, BlackRock's BUIDL fund has reached $2.3 billion in scale and has filed for two new tokenized funds, signaling that the world's largest asset manager's tokenization strategy is evolving from pilot to product suite expansion. HTX, as a core participant in the crypto exchange sector, officially launched TradFi perpetual futures products including NVDA, AAPL, MSFT, META, and SPY in 2026, enabling crypto users to gain 24/7 trading access to core U.S. equities. Boston Consulting Group predicts that global tokenized asset scale could reach $16 trillion by 2030, while McKinsey offers a conservative estimate of approximately $2 trillion. The on-chain migration of TradFi assets is no longer a "future narrative" but a structural transformation unfolding in real time, as crypto exchanges evolve from single crypto asset trading platforms toward "multi-asset-class trading infrastructure."

HTX Learn35 мин. назад

In-Depth Research Report on TradFi: The Convergence Wave of Crypto and Traditional Finance

HTX Learn35 мин. назад

Blocked Its Own Treasure, WeChat AI Steps Up

Tencent's stock surged over 10% on June 2nd amid reports that WeChat, with 1.43 billion monthly users, is finalizing tests for a native AI Agent. The reported feature, accessible by swiping right from the main interface, allows users to issue commands in natural language. The AI then decomposes tasks and automatically calls upon relevant Mini Programs within WeChat to complete actions like ordering food, booking tickets, or making payments, creating a closed-loop service execution system. This strategic shift follows the internal conflict and subsequent "blocking" of Tencent's standalone AI app, Yuanbao, by WeChat for violating sharing rules during a 2026 Spring Festival promotion. The incident highlighted a lack of internal consensus and exposed the weakness of competing in the standalone AI assistant arena against rivals like ByteDance's Doubao (345M MAU) and Alibaba's Qianwen. The new WeChat AI Agent aims to leverage WeChat's unique assets—its massive user base, standardized Mini Program APIs, WeChat Pay, and identity system—to move from simple content generation to actual task execution. Analysts note this changes the competitive landscape from model benchmarks to which AI can connect to more real-world services. However, success depends on key variables: the capability of Tencent's underlying Hunyuan model, managing massive inference costs, and redesigning incentives for Mini Program developers whose traffic might be bypassed. The move is seen as an attempt to keep user service intent within WeChat's ecosystem as AI begins to redefine how users access services.

marsbit1 ч. назад

Blocked Its Own Treasure, WeChat AI Steps Up

marsbit1 ч. назад

ByteDance Adopts Arm CPUs, Jensen Huang: So Sad I Didn't Buy Arm

**Summary:** At Computex 2026, Arm CEO Rene Haas announced that ByteDance and Oracle have adopted Arm's self-designed Arm AGI data center CPU. The company expects significant revenue growth from this product, projecting $20 billion in demand for the 2027/2028 fiscal years. Haas noted that restricting AI-capable CPUs from the US to China is nearly impossible due to their widespread applications. Arm's stock has surged dramatically this year, notably rising 16% after NVIDIA's Arm-based Vera CPU and RTX Spark announcements. A highlight was the informal, humorous on-stage conversation between Haas and NVIDIA CEO Jensen Huang. Huang joked about NVIDIA's failed attempt to acquire Arm and playfully lamented selling his Arm shares. Both executives showed a clear sense of camaraderie and shared regret over the missed merger. Key technical topics were discussed: 1. **AI PC Design:** Huang explained NVIDIA's RTX Spark superchip (with a 20-core Arm CPU) is designed for future AI agents that will autonomously run and use tools on PCs, blending local and cloud processing. 2. **Agent vs. OS:** Huang emphasized the operating system remains crucial, as AI agents rely on its APIs and tools to function. 3. **Growth Constraints:** He identified the shift to "useful AI" that generates profitable tokens as a primary driver for immense, almost limitless, computational demand. Haas outlined Arm's strategy across PC and data centers. For PCs, Arm collaborates with partners like NVIDIA and MediaTek, offering its compute subsystem (CSS) for custom SoCs. In data centers, its Arm AGI CPU (built on TSMC's 3nm process) has gained major partners including OpenAI, Meta, and now ByteDance and Oracle. Arm presented a multi-year roadmap for its in-house CPU line. The article concludes that while GPUs dominated the AI training race, the explosion of AI agents is shifting significant focus to CPUs for inference, state management, and tool orchestration. The industry is trending towards vertical integration, with companies like cloud providers designing chips and chip/IP firms offering full solutions, all competing to deliver more efficient computing per watt.

marsbit1 ч. назад

ByteDance Adopts Arm CPUs, Jensen Huang: So Sad I Didn't Buy Arm

marsbit1 ч. назад

Торговля

Спот
Фьючерсы

Популярные статьи

Неделя обучения по популярным токенам (2): 2026 может стать годом приложений реального времени, сектор AI продолжает оставаться в тренде

2025 год — год институциональных инвесторов, в будущем он будет доминировать в приложениях реального времени.

1.8k просмотров всегоОпубликовано 2025.12.16Обновлено 2025.12.16

Неделя обучения по популярным токенам (2): 2026 может стать годом приложений реального времени, сектор AI продолжает оставаться в тренде

Обсуждения

Добро пожаловать в Сообщество HTX. Здесь вы сможете быть в курсе последних новостей о развитии платформы и получить доступ к профессиональной аналитической информации о рынке. Мнения пользователей о цене на AI (AI) представлены ниже.

活动图片