Four Little Dragons assembled, 44.9 billion market lands — in 2026, enterprise AI reaches the intersection of computing power and governance

Yesterday, June 15, the 37th meeting of the Shanghai Stock Exchange Listing Review Committee approved the IPO application of Suiyuan Technology on the STAR Market, with plans to raise 6 billion yuan. On the same day, the term "Four Little Dragons of Domestic GPUs" can officially come to an end—Moore Threads, Muxi Co., Ltd., Biren Technology, and Suiyuan Technology, all four domestic AI chip companies have entered the capital market.

But what gave me pause was not the number 6 billion, but another figure: Tencent contributed over 80% of Enflame Technology's revenue. A listed company, with one client accounting for 80% of its income. This number will reappear repeatedly later, because I believe it is the most authentic entry point for understanding the current domestic AI chip ecosystem.

There are two other things worth looking at together this week. The "2026 China Enterprise AI Agent Industry White Paper" released last week gave a staggering growth rate: 21.2 billion in 2025, 44.9 billion in 2026, a growth rate of over 111%, and 332 billion by 2029, a 15-fold increase in four years. The white paper directly defines 2026 as the "first year of large-scale enterprise AI agents." Two weeks earlier, Workday released Agent Passport at DevCon 2026 on June 2 — a platform specifically for testing, verifying, and continuously monitoring enterprise AI agents, with Cisco AI Defense as its first partner.

Three lines come together, and what I want to say is: for enterprise AI to move from pilot to scale, a three-layer foundation must be in place simultaneously—computing power that is independently controllable, an application market large enough, and a governance framework that keeps pace. In June 2026, these three layers were densely implemented within the same week; this is no coincidence, but a concentrated eruption after industrial maturity reached a critical tipping point.

But there is one number that I can never let go of — over 80% of Enflame's revenue comes from Tencent. Behind this number lies the customer concentration dilemma of the entire domestic AI chip industry, as well as an unavoidable risk benchmark when enterprises procure domestic computing power.

6 Billion IPO Joins the Four Little Dragons: Domestic AI Chips Enter the 'Ecosystem War'

First, clarify what "the Four Little Dragons joining forces in the capital market" means.

Since the end of 2023, NVIDIA's export controls on China have been continuously escalating, with H100, A100, and H800 successively added to the restriction list. For domestic large model manufacturers, this is not just a "news-level impact" but a "supply cut-off" at the practical level—money cannot buy the goods, or the prices for what can be bought are outrageous, not to mention the new restriction orders that could land at any time. Against this backdrop, the financing capabilities and R&D pace of domestic AI chip manufacturers directly determine whether the training of domestic large models can continue to iterate.

The chip industry has a brutal rule: from the tape-out of one generation of chips to the finalization of the next, tens of billions or even hundreds of billions of continuous investment are required in between. Without capital market funding, relying solely on purchase orders from major clients makes it difficult to sustain this pace. Therefore, the core significance of the Four Little Dragons going public is not "how much money was raised," but rather obtaining a ticket to continuous research and development.

Suiyuan Technology was established in March 2018 with a registered capital of 380 million yuan, focusing on cloud AI chips, AI accelerator cards, and intelligent computing systems. This 6 billion yuan financing round is considered large on the STAR Market, and the funds will be used as follows:

Among the four directions, I read them several times, and the third one is the most worth pondering. "Software ecosystem platform construction" — in plain terms, it's not enough to just have chips; you also need to make developers willing to write code on your platform. Nvidia's moat has never been just hardware, but the developer inertia and depth of operator libraries accumulated over more than a decade of the CUDA ecosystem. If domestic chip manufacturers only focus on hardware substitution without making up for the shortcomings in the software ecosystem, they will at best be a "cheaper substitute," with no pricing power, let alone long-term competitiveness.

So carving out a portion of the 6 billion specifically for building a software ecosystem is the right direction. But whether it can succeed depends on two things: the activity level of the developer community and compatibility with mainstream frameworks (such as PyTorch, DeepSpeed, etc.). This is not a problem that can be solved by throwing money at it; it takes time.

Let’s look at the overall structure of the “Four Little Dragons”:

The four companies each have different positioning, but they share one commonality: focusing on training chips. The reason is simple — training cards have high unit prices and thick profit margins, and the buyers are mainly leading large model companies and cloud providers, making sales highly efficient. Although inference chips are deployed in large quantities, their profit margins are thin and competition is more intense.

Another noteworthy phenomenon: leading domestic cloud vendors and major internet companies are both purchasing domestic chips and developing their own AI chips. Huawei has Ascend, Baidu has Kunlun, Alibaba has Yitian, and ByteDance is also increasing its self-research investment. This means that the long-term competitive barriers for the "Four Little Dragons" ultimately come down to the software ecosystem — hardware can be caught up with, but developers' habits are hard to change.

Returning to Enflame's 80% issue. The prospectus shows that Tencent is Enflame Technology's largest customer, contributing over 80% of its revenue. This figure is already considered high on the STAR Market. After listing, each quarterly financial report will magnify this number—once Tencent's procurement pace slows down, or Tencent achieves a breakthrough in its self-developed chips, the pressure on Enflame's stock price will quickly transmit.

My judgment is: This is not a fatal flaw, but it is indeed a core issue that Enflame must directly address after going public. Deep binding with major clients is not uncommon in the chip industry—Intel relied on IBM in its early years, and AMD relied on Compaq, following similar paths. The key lies in, after securing a major client, where will the second and third clients come from? When can the software ecosystem contribute genuine revenue diversification? This answer is more important than any quarterly report figure.

The Truth Behind 44.9 Billion: Which Industries Are Taking Real Action and Which Are Still Just Paying Lip Service

The "2026 China Enterprise AI Agent Industry White Paper" released on June 9 contains a set of data that has been repeatedly cited:

Most people are sharing the 332 billion figure for 2029 — it's indeed large and eye-catching. But frankly, the confidence level in long-term forecasts is low, with too many variables. Instead, I think the 44.9 billion for 2026 is more worth serious attention.

Why? Because from 21.2 billion to 44.9 billion, the growth rate exceeds 111%, which is an estimate of the market size "at that time," not a long-term forecast for five years later. If the underlying data is reliable, then by 2026, enterprise-level AI agents will not be in the pilot phase but will have already begun large-scale deployment—only real deployment can support this scale.

The white paper defines 2026 as the "first year of large-scale enterprise-level AI agent deployment," a judgment I largely agree with, but with one qualifier: industry differentiation is severe.

Finance, manufacturing, and retail—these three industries are the fastest adopters of enterprise AI. Finance goes without saying, with compliance pressures clearly present, making risk control, auditing, and report automation essential needs. In manufacturing, quality inspection, predictive maintenance, and production scheduling optimization are naturally suited for AI agent intervention. Retail is expanding step by step, from user behavior analysis to personalized recommendations and then to supply chain forecasting.

My own assessment is that among this 44.9 billion, the three industries of finance, manufacturing, and retail account for the majority. Progress in government, education, and healthcare is clearly lagging behind—not because there is no demand, but due to long procurement processes, unclear data compliance boundaries, and high trial-and-error costs. Large-scale adoption in these industries is estimated to wait until 2027-2028.

There is a signal that can indirectly corroborate this: On June 15, the first trading day after the white paper was released, stocks such as Nanxing Co., Ltd. saw a significant rally. The market's capital voting direction pointed to the intersection of "Manufacturing + AI Agents." This kind of market reaction usually indicates that real demand has been identified, rather than being purely sentiment-driven.

For enterprise managers, the figure of 44.9 billion carries a very direct implication: your competitors may already be on the move. Not planning, not running POCs, but actually deploying and generating business data. The gap is widening, not narrowing.

Workday Agent Passport: Governance First is the Strongest Signal for Scaling

On June 2, Workday released Agent Passport at DevCon 2026 in Las Vegas. First, let's talk about what the product itself is:

Agent Passport = Enterprise-level AI Agent "Testing + Validation + Continuous Monitoring" Platform

Capability 1: Verify that each agent passes the most severe risk test

Capability 2: Retain verification records for audit and compliance inspections

Capability 3: Continuously monitor operational status and detect abnormal behavior

First launch partner: Cisco AI Defense (independent third-party security testing)

When I first saw this release, I found it quite interesting, but not to the point of being "shocking." The need to "test AI systems" has been discussed in the industry for quite some time. However, after thinking it over carefully several times, my assessment changed.

What truly matters about Agent Passport is not how revolutionary the product itself is—testing, verification, and similar actions are not new. What truly matters is: this is the first time an enterprise management software vendor has turned "agent governance" into an independent platform-level product line. It is not a small toggle within some feature module, nor a piece of advice in developer documentation, but a product with its own name, partners, and a continuous operation plan.

Who are Workday's customers? A large number of HR and financial system users from Fortune 500 companies. The anxieties of these customers are very specific:

First, what decision did this AI Agent actually make? Was there any record left? If a compliance issue arises, to what extent can responsibility be traced?

Second, internal audit departments and external regulatory bodies require the submission of AI usage reports and risk assessments—what should be used to fill out these forms?

Third, has the agent experienced hallucinations? Has it modified rules that should not have been changed? Can it provide a complete timeline of operations?

The Agent Passport approach is to issue a "passport" to each AI agent — documenting which risk tests it has passed, what known risk points exist, and which third party has verified it. In simple terms, this is not about restricting AI capabilities, but about providing a "chain of evidence" for corporate internal audits and external compliance checks.

Let me translate: Agent Passport is essentially helping enterprises build AI governance archives. If this product can succeed, Workday will occupy a position more valuable than HR and financial SaaS — a platform player for enterprise AI compliance governance. This position is currently unoccupied.

For Chinese companies, this product may not be directly usable for now—Workday's coverage in China is limited, and cross-border data compliance is a practical issue. But the trend it reveals is real: when intelligent agents begin to participate in real business decisions, even if only as "recommendations" rather than "executions," compliance risks already exist.

Companies that proactively establish AI governance frameworks now will encounter far fewer pitfalls when regulations tighten. Those still adopting a "move fast and fix things later" approach will inevitably have to catch up—and the cost of doing so will only keep rising. This isn't just about compliance risks, but also business continuity risks: when an AI system without governance records malfunctions, you won't even have a starting point for troubleshooting.

Another noteworthy detail: Cisco entered as a launch partner to conduct third-party security testing. This means that security testing for AI agents has already begun to have professional service providers. Information security audits started with specialized companies and gradually became a standard requirement for listed companies—AI governance audits are likely to follow the same path.

Five Suggestions for CIOs and Enterprise Managers

Looking at the three things above together, I have five specific suggestions. These are not vague strategic directions, but actionable points that can be directly discussed at the next quarterly review meeting.

1. Computing power procurement list needs updating
   The financing and listing of the "Four Little Dragons" means they now have the ammunition for sustained R&D. From 2026 to 2027, the performance of domestically produced training and inference cards will see a significant improvement. The gap left by NVIDIA's export controls is being rapidly filled by domestic chips. IT procurement managers should now include the product roadmaps of manufacturers like Huawei Ascend, Enflame, and Moore Threads in their evaluation lists. This doesn't mean an immediate switch, but rather having a backup plan — with a backup, your bargaining power when negotiating prices with NVIDIA will be entirely different.
2. The keyword for 2026 is not "watching AI," but "watching ROI."
   A market size of 44.9 billion means a large number of enterprises are already deploying intelligent agents. At this point, blindly chasing trends is wrong, and completely sitting on the sidelines is also wrong. My suggestion is to focus on 1-2 scenarios with clear ROI indicators—such as customer service agents reducing labor costs, or supply chain forecasting optimizing inventory turnover—and first achieve a quantifiable result. One proven ROI is more convincing than ten POC projects.
3. AI governance cannot rely solely on the IT department—bring in legal and compliance now
   Workday's Agent Passport illustrates one thing: when intelligent agents begin to participate in real business decisions, compliance risks arise. Don't wait until regulators ask, "Do your AI systems have audit records?" to catch up. Sit down with the legal and compliance teams now to discuss: Where are the boundaries of AI usage? What is the decision-recording mechanism? How is responsibility allocated when problems arise? Discussing these issues now incurs the lowest cost.
4. When selecting an AI supplier, customer concentration is a health signal
   The lesson from Enflame's 80% concentration reminds us: when choosing a supplier, we should not only look at product specifications but also at the customer structure. A supplier that relies excessively on a single major client raises questions about pricing negotiation space and long-term service stability. If your AI infrastructure supplier is primarily supported by just one or two major customers, you need to reassess your bargaining power and risk exposure.
5. The window of opportunity is opening — building capabilities now is more worthwhile than waiting for SaaS to mature before switching
   China is simultaneously filling gaps across the three layers: computing power, models, and applications. This time window is critical for enterprise tech teams: participating now in co-building private or hybrid-deployed agent capabilities with AI suppliers means accumulating data, tuning scenarios, and operational experience. If you wait until standardized SaaS solutions mature before making the switch, your switching costs and time costs will only be higher. The first-mover advantage is especially evident in this field — whoever gets it running first will be the one accumulating the data flywheel.

Returning to the number at the beginning: 80%. It is not just a risk warning for Suiyuan Technology, but a microcosm of the entire domestic AI industry chain—capabilities are rapidly catching up, but the ecosystem and customer structure are still playing catch-up. In the week of June 2026, the simultaneous implementation of the three layers—computing power foundation, market scale, and governance framework—is a clear signal: enterprise AI infrastructure is shifting from "whether it exists" to "whether it is good."

For CIOs, this means you no longer need to convince anyone "whether AI is worth doing," but instead must answer a more difficult question: "How can we move fast without crashing?"

The answer to this question will not come from any white paper or any vendor. It can only come from your own team, starting with the first real scenario, trying one by one.