Understanding the Four Barriers from Policy Signals to Real Orders in AI Computing Power

You’ve certainly come across this phrase in various research reports recently:

2026 is the year when domestic AI computing power will be fully realized.

Wujiang Securities said it, Huayuan Securities said it, China Galaxy Securities said it. They spoke with certainty, as if it were an industry consensus.

But I’d like to ask a simple question: what exactly is being exchanged in "cash out"?

If you're investing in this space or working in this industry, this question deserves a thoughtful answer.

The term "Year of Realization" hides a pitfall—it blurs a crucial distinction: policy procurement, pilot orders, large-scale deployment, and software ecosystem maturity are four entirely different milestones, each with its own timeline and distinct value to the industry.

Mixing the four gates together is called "cash out," which can easily lead to systemic misjudgments about real-world progress.

As usual, I’ll try to help you clearly understand these four gates in one article.

Before discussing whether computing power has been realized, we first need to understand: what stages does a computing power product go through from being "developed" to actually creating value?

I’ve distilled it into a chain of progression: policy procurement → real-world deployment → maturation of the software ecosystem → scalable replication

The First Gate: Policy-Driven Procurement—Procurement driven by government funding or policies, where computing power is purchased and machines are deployed, but not necessarily out of genuine business needs, rather to "fulfill deployment targets."

The Second Gate · Real Deployment: The purchased computing power is genuinely used to run business operations, not left idle in the data center gathering dust. This requires companies to have authentic AI needs and be willing to connect them to this computing infrastructure.

The Third Gate: Mature Software Ecosystem—Developers can seamlessly write code, deploy models, and debug optimizations on this computing power without having to perform costly, custom migrations each time.

The Fourth Gate · Scale Replication: This computing power solution can be extended from leading enterprises to mid-sized companies, penetrating from government and enterprise markets into commercial internet markets, creating economies of scale.

These four gates are progressive; without opening the next gate, the progress before it may appear strong on financial statements, but the true value of the industry remains unrealized.

This door has indeed opened wide in 2026.

Galaxy Securities believes that the significant launch of DeepSeek-V4 has shifted market expectations from policy-driven substitution to the realization of genuine demand orders. Dongwu Securities believes that in the first quarter of 2026, the compute leasing industry experienced a "quantitative change" with increased orders and price hikes, marking a "qualitative transformation" in its business model.

The science and technology re-lending program has been expanded to 1.2 trillion yuan, specifically supporting AI and semiconductors, and the National Development and Reform Commission’s equipment upgrade fund of 91.5 billion yuan is also being directed toward computing infrastructure.

Reports indicate that Alibaba, ByteDance, and Tencent have placed bulk orders totaling hundreds of thousands of units for Huawei’s upcoming Ascend 950PR chip, causing its price to rise by approximately 20% due to surging demand.

This number means: this is no longer a "symbolic purchase," but a real, large-scale order.

However, note that the opening of policy-driven procurement does not equate to full realization across the entire industrial chain. The number of AI chips procured and the amount of real business workload those chips actually run are two different things.

This door is the critical turning point in 2026—but it’s only been cracked open, not fully swung wide.

The core evidence of real-world deployment is DeepSeek V4.

On April 6, 2026, DeepSeek V4 officially announced a complete transition away from NVIDIA's CUDA ecosystem, fully migrating to Huawei Ascend chips and the CANN software framework, becoming the world's first trillion-parameter MoE large model trained and deployed entirely on domestic Chinese computing power. In breaking industry norms, DeepSeek did not grant early access to U.S. chip suppliers for V4 testing, instead prioritizing adaptation opportunities for domestic chip manufacturers such as Huawei and Cambricon.

What does this mean? It proves that domestic computing power can fully support the training and inference of world-class large models—not just "barely functional," but actually running in practice. This is the strongest evidence yet that the second door has been opened.

However, fully opening the second door requires more than just adaptation by leading large model vendors—it demands real-world business deployment by a broad range of enterprises. It’s one thing for internet giants to run their own models; it’s another entirely for traditional enterprises to integrate AI into their production processes—the latter moves much more slowly.

DeepSeek V4 disrupts the industry pricing model with its "cent-era" pricing, driving AI applications from pilot projects to widespread adoption. In the second half of 2026, the core focus of China's AI industry will shift: low-cost models will ignite a surge in inference demand, and domestic compute infrastructure will enter its phase of realization.

But there’s a subtle feedback loop here: lower model prices → more businesses are willing to try it → actual usage increases → demand for computing power grows stronger → computing power supply expands → model prices fall further. This positive cycle has just begun and has not yet fully taken off.

Second door judgment: A crack has opened; the head scenario is already underway, while the mid- and long-tail scenarios are still on the way.

This is the easiest of the four gates to overlook, yet it’s the most critical for true realization.

NVIDIA's CUDA is an ecosystem built since 2006 and cultivated over two decades to accumulate millions of developers. Huawei's CANN currently supports over 160 mainstream AI models, while NVIDIA's CUDA ecosystem covers more than 23,000 models. This gap cannot be bridged in just a few months.

But this door is opening quickly.

The strongest signal is DeepSeek V4’s adaptation strategy. DeepSeek stated that, due to limitations in high-end computing power, the service throughput for Pro is currently very limited, and the price of Pro is expected to drop significantly after the bulk release of the Ascend 950 super nodes in the second half of the year.

This sentence hides an important signal: DeepSeek isn’t just “using domestic computing power”—it’s actively waiting for the scale of domestic computing supply to grow, so it can convert that capacity into lower API pricing and drive broader application adoption. This is a symbiotic relationship between a model provider and a computing provider, not a passive adaptation.

Caitong Securities believes that 2026 will be the year when domestic ultra-nodes for inference scale up significantly. Numerous domestic manufacturers have already launched next-generation ultra-node solutions: Huawei's Atlas 950/960 support 8,192/15,488 computing cards, while Suguang, Moxi, Kunlun芯, and Alibaba Panjiu have all made strategic investments in ultra-nodes. With demand and supply converging, the industry is poised for a major scaling-up phase.

The assessment of the third gate: Top-level compatibility has been achieved, but the developer ecosystem further down the stack requires 1–2 years of systematic development to truly mature.

This is the farthest of the four doors and the ultimate form of "cashing out."

Scale replication means that it's not just Huawei, ByteDance, and Tencent using domestic computing power, but thousands of mid-sized enterprises—whose IT systems, AI-powered quality inspection in manufacturing, and hospital-assisted diagnostic systems—all run on domestic computing power, with these customers experiencing no noticeable migration costs.

This step has not yet arrived in 2026.

The core reason: Mid-sized enterprises' IT teams lack the capability to migrate computing power on their own. Leading tech giants have AI infrastructure teams of hundreds of people who can dedicate resources to custom adaptations; a 500-person manufacturing company, by contrast, may have only three to five IT staff—they need a "plug-and-play" solution, not a computing platform requiring a six-month migration project.

This issue isn’t about chip performance or software frameworks—it’s about the level of solution integration. A complete end-to-end service offering, from computing hardware to the application layer, is needed so that mid-sized enterprises can leverage domestic computing power for their AI applications without needing to understand the underlying technology.

The judgment on the fourth gate: Scale replication is unlikely to be seen before 2026; this may not occur until 2027–2028.

Next time you see any report about "computing power redemption," use this checklist to verify:

First Door · Policy Procurement

Verification metrics: Scale of policy funding implementation / Number of major domestic chip transactions

2026 status: Open, and wide open

Risk Warning: Purchase volume ≠ Deployment volume — Do not confuse them.

Second Gate · Real Deployment

Verification Metrics: Q1 Price Increases for Hashrate Leasing / Actual Adaptation by Large Model Vendors / Hashrate Utilization

2026 Status: A crack has opened, head scenarios have been validated, mid- and long-tail use cases are still on the way

Risk Warning: Focusing on the top does not mean understanding the big picture.

The Third Gate: Mature Software Ecosystem

Verification metrics: Number of models covered by CANN / Developer migration cost / Number of adaptation cases for medium-sized enterprises

2026 Status: Top-layer compatibility achieved; mid-to-downstream ecosystem requires 1–2 years.

Risk Warning: This gate determines how deep the moat of computing power is.

The Fourth Gate · Scaling Replication

Verification metric: Number of projects by medium-sized enterprises purchasing domestic computing power / Number of AI application implementations in vertical industries

2026 status: Almost never opened

Risk Warning: This door is the final state of "cash-out"—don’t celebrate too soon.

The term "the year of realization" isn't entirely wrong. From the perspective of the first gate—policy procurement—2026 is indeed a real year of realization. Domestic computing power has transformed from being something that required policy subsidies to attract buyers, to becoming a supplier actively sought after by major tech companies—this qualitative shift is genuine.

But if you interpret "the year of cashing out" as "the full-scale explosion of the computing power industrial chain and comprehensive realization of related companies' performance," then it's dangerous.

The fourth gate has not yet opened, meaning the current industry landscape is still a competition among a few leading players. True economies of scale will only emerge as the third and fourth gates gradually open—that’s when a larger, more sustained market explosion will occur.

After completing my research for this article, I gained two insights for your reference:

First, within the hashpower industrial chain, the timing of realization varies significantly across different segments. Chip design and manufacturing (first segment benefits most directly), hashpower leasing (second segment benefits), software toolchains (third segment benefits), and vertical industry solution providers (fourth segment benefits)—the time windows for realization across these four areas may differ by as much as two full years.

Second, DeepSeek V4’s deep integration with domestic computing power is the most significant industry signal of 2026—without exception. It has transformed the question from “Can domestic computing power be used?” to “When will domestic computing power be able to meet demand?”—a fundamental shift in narrative.

This article is from the WeChat public account "BT Finance" (ID: btcjv1), authored by BT Finance.