On June 2, in Las Vegas, Workday released something called "Agent Passport" at DevCon. On June 9, in Beijing, a white paper on China's AI agent industry quietly went online. On the afternoon of June 15, in Shanghai, the Shanghai Stock Exchange Listing Committee approved Enflame Technology's IPO on the STAR Market. The same week, Bain & Company released an AI investment return rate research report that sent chills down the spines of many CIOs.
Four pieces of news, belonging to four completely different dimensions—governance, market, computing power, and ROI—are densely packed into the first two weeks of June 2026. I read these news items several times, and the more I read, the more I felt this was not a coincidence in timing—they all point to one thing: enterprise AI is falling from the "validation period" into the "governance period", and this turning point is much steeper than most people expected.
To put it more bluntly: the technology has been proven, enough money has been invested, and the chips are almost ready — but the value hasn't arrived yet. The core issue to be solved next has shifted from "can it be done" to "how to manage it, how to account for it, and how to keep it from causing trouble."
I will explain these four signals separately and then look at them together.
1. Workday's Agent Passport: Enterprise AI Takes "Causing Trouble" Seriously for the First Time
On June 2, Workday released Agent Passport at the DevCon developer conference in Las Vegas. Translating this name — "Agent Passport." It's not some flashy marketing concept, but a solid governance tool.
It does three things: tests whether each Agent passes the most severe risk tests, retains records of continuous verification, and gives enterprise managers auditable control over every AI Agent running online. The first launch partner is Cisco AI Defense, with Cisco's security team directly involved in joint debugging.
To be honest, my reaction was muted when I first saw this news. A governance tool launched by an HR and financial software vendor—how big of a deal could it be? But after thinking more carefully about the identity of the publisher—Workday, one of the world's largest SaaS providers for human resources and financial management—I realized the weight of this matter.
Key Judgment: This is the first time an enterprise management software vendor has turned "Agent governance" into a platform-level product. It’s not a security company doing it, nor a cloud vendor—it’s an HR + financial systems vendor. What does this mean? It means the entry point for governance is not "technical security," but business process security—will your payroll calculation Agent send out the wrong salaries? Will your financial approval Agent bypass internal controls? These are the things enterprises truly fear.
Also released at the same time are Agent Builder, Workday API Hub, and Workday AI Gateway — a complete developer toolchain. In simple terms, Workday has not just provided a single lock, but a full system for building doors, installing locks, and inspecting locks.
My judgment is: this marks the official transition of enterprise AI from the "can it be used" stage to the "dare to use it" stage. And the answer to "dare to use it" depends on whether you have governance tools—which is exactly what Workday wants to sell.
Think deeper: Why is Workday doing this at this particular time? Because agents have already started running in enterprises. The market size of 21.2 billion in 2025 is no joke—it means thousands of agents are already operating within corporate HR systems, financial systems, and supply chain systems. Once they run, they need to be managed; if not managed, problems arise. This is not a "future need"—it is an "immediate necessity." Workday is releasing it by mid-2026, perfectly timing the move.
II. 44.9 Billion and 332 Billion: China's AI Agent Market is "Forcibly Soaring"
On June 9, the relevant institution of the China Internet Network Information Center released the "White Paper on China's AI Agent Industry." There is a set of data in it that made me read it over and over again:
In 2025, the domestic enterprise-level AI agent market size is 21.2 billion yuan. In 2026 (forecast), the market size is 44.9 billion yuan, a year-on-year increase of 111.8%, entering the "first year of scaling". In 2029 (forecast), the market size will exceed 332 billion yuan, growing 15 times in four years.
The compound growth rate from 2024 to 2029 is 107%. What surprised me about this figure is not the "fast growth," but rather that the slope of the growth curve suddenly steepens in 2026. From 21.2 billion in 2025 to 44.9 billion in 2026, it more than doubles in one year; but from 2026 to 2029, it will increase by more than seven times over three years. What does this mean?
This means the industry believes 2026 is the turning point — not "gradual acceleration," but "sudden explosion." The white paper directly defines 2026 as the "first year of large-scale adoption," with very clear wording.
| Year | Market size (100 million yuan) | year-over-year growth rate | Stage Characteristics |
|---|---|---|---|
| 2024 | ~102 | — | Exploration period |
| 2025 | 212 | ~108% | Verification period |
| 2026E | 449 | 111.8% | The first year of scaling |
| 2029E | 3,320 | Composite 107% | Full implementation |
But I have to pour cold water on this: the market size doubling doesn’t mean your business can get a piece of the pie. The white paper counts the "market," not the "profits." How much of that 44.9 billion goes to hardware procurement, cloud resource consumption, and consulting fees? How much actually translates into efficiency improvements and cost savings for the businesses themselves?
The answer to this question points directly to Bain's chilling report.
3. The Truth of Bain's Report: 1 Trillion Invested, "Technology Works, but Value Hasn't Arrived"
The core takeaway of Bain's June AI investment return study can be summed up in one sentence: "The technology works, but the value hasn't arrived."
More than $1 trillion in global capital has flowed into AI. One trillion. This is not an abstract number; it's real money — VC funds, corporate IT budgets, government special funds, and cloud providers' capital expenditures. But what about the core metric of ROI?
Core Findings: PoC (Proof of Concept) has matured, but commercial value has yet to be realized. Enterprises are generally stuck at the "last mile"—it’s not that AI is inadequate, but that enterprise organization, processes, and data haven’t kept up.
Let me translate: A model that runs perfectly in the lab falls apart as soon as it hits real business scenarios. It's not because the model lacks capability, but because your processes don't support it, your data isn't clean, your organization doesn't accept it, and your employees don't know how to use it.
To put it bluntly, after spending this 1 trillion, the vast majority of companies are still in the stage of "spending money to verify feasibility," far from the stage of "making money with AI."
This reminds me of a classic technology maturity curve: the speed of technological maturity always outpaces the speed of organizational digestion. Gartner calls this the fall from the "Peak of Inflated Expectations" to the "Trough of Disillusionment." In June 2026, we are likely standing at the starting point of this fall.
What is the most striking point in the Bain report? It points out that the root cause of the missing ROI lies not in technology, but in the organization. This is isomorphic to the logic behind Workday's Agent Passport—the problem is not AI itself, but the environment in which AI operates. A city without traffic lights, even if every car runs very fast, will only end up in a pile of collisions.
4. Suiyuan Passes Review, Four Little Dragons Gather: The "Hard Landing" of Domestic Computing Power
On the afternoon of June 15, the Shanghai Stock Exchange Listing Committee approved the IPO application of Suiyuan Technology on the STAR Market, planning to raise 6 billion yuan.
Enflame Technology was established in March 2018, with its headquarters in Shanghai Lingang, focusing on cloud AI chips. The industry significance of this approval lies in a landmark fact: "The 'Four Little Dragons of Domestic GPUs' have now all entered the capital market."
| Company | Listing Board | Core Direction | Key Features |
|---|---|---|---|
| Moore Threads | Sci-Tech Innovation Board | Full-featured GPU | Dual-track graphics rendering + AI inference |
| Muxi Co., Ltd. | Sci-Tech Innovation Board | High-performance computing GPU | Focus on training scenarios |
| Biren Technology | Hong Kong stocks | General-purpose GPU | Once set a global computing power record |
| Suiyuan Technology | Sci-Tech Innovation Board | Cloud AI Chip | Deeply integrated with Tencent |
Four companies went public intensively between 2024 and 2026, which is no coincidence. The driving force behind this is very clear: NVIDIA export controls.
The United States has been gradually tightening export restrictions on high-end GPUs, making it increasingly difficult for Chinese companies to purchase A100/H100 chips, with prices becoming more and more outrageous. Domestic alternatives have shifted from "optional" to "mandatory." The logic of the capital market is also straightforward—only because of the ban can I buy domestic products, so I must first prop up the domestic ones.
But here's a point worth pondering: Enflame Technology is deeply tied to Tencent, with Tencent contributing over 80% of its revenue. A chip company's lifeline being tied to a single internet giant is both an advantage and a risk. The advantage is having a stable major client and a path for technical collaboration; the risk is that if Tencent's demand fluctuates, Enflame's revenue will be like a roller coaster.
Let's look at the 6 billion fundraising purpose: R&D of fifth and sixth-generation training and inference chips, industrialization of intelligent computing systems, and construction of a software ecosystem platform. Note the last item — "software ecosystem." What domestic GPUs have always lacked is never the chips themselves, but a replacement for the CUDA ecosystem. Nvidia's moat is not in hardware, but in the software development ecosystem accumulated over a decade. The four little dragons have gone public collectively and raised funds. The real battle ahead is the tough fight for the software ecosystem.
If we cannot win this battle, domestic chips will only be stuck in mid-to-low-end inference scenarios, while training scenarios will still have to rely on NVIDIA.
There is another noteworthy layer: 6 billion in fundraising is no small amount, but in the context of global chip investment, it is actually quite meager. Nvidia's annual R&D spending exceeds 8 billion USD. The total fundraising of the Four Little Dragons combined may not even match Nvidia's quarterly R&D budget. The breakthrough of domestic chips cannot be solved by just one or two companies going public; it requires the migration of the entire software ecosystem—from CUDA to domestic frameworks. This process is more difficult and slower than making the chips themselves.
5. Underlying Logic: Why These Four Signals Are Crowded into the Same June
Looking at the four signals together, one can discover a very interesting symmetrical structure:
| Signal | Dimension | Core meaning | Direction |
|---|---|---|---|
| Workday Agent Passport | Governance | Enterprise AI needs a "passport" to be deployed | From laissez-faire to control |
| AI Agent Whitepaper | Market | 44.9 billion, the first year of large-scale implementation | From niche to mainstream |
| Bain ROI Report | Value | 1 trillion investment, ROI not realized | From optimism to anxiety |
| Suizhi IPO / Four Little Dragons gather together | computing power | Domestic chip base completed | From chip shortage to availability |
Governance, market, value, and computing power—qualitative changes occurring simultaneously across these four dimensions are no coincidence; this is the industrial cycle.
My own judgment framework is as follows:
In Q2 2026, enterprise AI moves from the "validation phase" to the "governance phase". The core question of the validation phase is "whether it works"—whether the model is viable, whether the chips are sufficient, and whether the Agent can run. The core questions of the governance phase are "whether we dare to use it, how to manage it, and how to account for it." The nature of these two sets of questions is completely different: the former is a technical issue, while the latter is an organizational issue. And the pace of solving organizational issues is always slower than that of technical issues.
This is why the market is experiencing explosive growth (44.9 billion → 332 billion), computing power is rapidly catching up (the "Four Little Dragons" going public), and governance tools are beginning to be implemented (Agent Passport)—but value has yet to be realized (Bain report). Because the first three are all supply-side progress, while ROI is a demand-side outcome. No matter how fast the supply side runs, if the demand side cannot keep up, even 1 trillion invested will still result in losses.
To put it more bluntly: In June 2026, the industry is undergoing a "split"—the supply side is surging ahead, while the demand side is stumbling. This split will not last, but it will make companies caught in it feel very conflicted: You know you have to adopt AI, but you don't know how to make AI truly generate value; you have domestic chips to choose from, but you're not sure if the ecosystem is mature enough; your peers are already using Agents, but you haven't even set up a governance framework.
And it’s not just companies that are conflicted. Chip manufacturers are also conflicted—they produce chips but don’t know who will use them; AI companies are conflicted too—their models are powerful but clients won’t pay; investors are equally conflicted—they’ve put in money but see no return cycle. The entire industry chain is waiting for one thing: the awakening of demand. And the prerequisite for the awakening of demand is precisely the three things mentioned in the Bain report—organization, processes, and data. If these aren’t fixed, a 44.9 billion market will be nothing but a 44.9 billion bubble.
Three Judgments for CIOs/CTOs/Enterprise Digital Leaders
1 Governance first, not Agent first
Workday's release of Agent Passport is not about selling a feature, but about selling a paradigm—first build a governance framework, then deploy agents. If your enterprise does not yet have a basic framework for agent governance (permission levels, audit trails, risk testing), don't rush to launch more intelligent agents. Launching an agent without a passport is like letting someone without a driver's license drive on the highway—an accident is only a matter of time.
2 Domestic computing power is a "strategic option," not a "cost-effective option"
The listing of the Four Little Dragons means that domestic GPUs have moved from the stage of "whether they exist" to the stage of "whether they are good to use." However, don't use the logic of cost-effectiveness to choose domestic chips—today's domestic GPUs still lag behind NVIDIA in training scenarios, and this gap cannot be immediately closed with money. The correct reason to choose domestic computing power is supply chain security, not cost savings. If you have a rigid demand for training scenarios, the optimal strategy at this stage is a dual-track system of "using NVIDIA for training and domestic chips for inference."
3 The bottleneck of ROI lies in the organization, not the model—first improve the process, then implement AI
Bain's report makes it clear: the bottleneck for ROI is not technology, but organization. My advice is counterintuitive: stop increasing budgets for AI teams and instead invest in business process reengineering and data governance. In an organization with chaotic processes and messy data, deploying more AI only amplifies the chaos with 10,000 times the computing power. Only by streamlining processes and cleaning up data can AI's value move from PoC to production.
In June 2026, four signals appeared simultaneously, all conveying the same message: The good days for enterprise AI have not yet arrived, but the bad days won't last long either—provided you first address the three key issues: governance, computing power strategy, and organizational capability. Whoever passes these three hurdles first will be able to shift from "spending money on validation" to "making money with AI" amid the 15-fold growth from 44.9 billion to 332 billion over four years.
