On June 4, Sam Altman said something at an OpenAI corporate event that many people overlooked. He stated that AI products go through three stages: chatbots, AI agents, and the third — "Proactive AI."
On the same day, at the other end of the Earth, no one noticed that global enterprise SSD purchase orders were surging. By June 11, TrendForce's data came out: in the first quarter of 2026, the revenue of the world's top five enterprise SSD brands reached $18.46 billion, a quarter-on-quarter surge of 86.1%, hitting a record high. There was only one reason: the deployment of AI agents was devouring storage capacity like crazy.
These two signals have simultaneously landed on the desks of enterprise CIOs. One is "What will AI look like next," and the other is "AI is already eating your infrastructure right now." Sandwiched between them is a third countdown: the EU AI Act takes full effect on August 2, and most enterprises have not yet begun preparations.
Altman's Three Stages: From "I Ask, It Answers" to "It Does Things Without Me Asking"
I translate the three stages Altman mentioned into plain language:
| Stage | What's your name | What are you doing | What is AI doing |
|---|---|---|---|
| Phase 1 | Chatbot | Ask questions, wait for replies | Answer one question at a time |
| Second stage | AI Agent | Assign tasks, review results | Execute multi-step tasks, but require human initiation |
| Phase 3 | Proactive AI | Only review, possibly not even that | Run it in the background yourself, fix any issues you find on your own, and notify you afterwards |
Altman's original words were: "Users don't need to know what the AI is doing in the background; they only need to know the results."
That sounds easy enough. But if you’re a CIO, your first reaction upon hearing this should be: who’s responsible if something goes wrong?
Altman himself obviously knows this is a problem. He admits that OpenAI's current product line is itself quite "fragmented"—users can't figure out when to use ChatGPT, when to use Codex, or when to call the API. And the goal of "proactive AI" is precisely to eliminate this fragmentation—so that users don't need to switch tools, and the AI itself decides in the background which capability to use.
This direction is actually pointing to the same thing as Microsoft IQ (unified intelligence layer) promoted at Microsoft Build and the AI features embedded in Google Workspace: AI retreats from a "front-end tool" to a "back-end operating system".
An underestimated risk: When AI becomes the "back-end operating system," its operations no longer appear on the user interface. Today, you can still see what the AI agent is doing, which system it has accessed, or which data it has modified. In the proactive AI stage, these actions will most likely all be in back-end logs—and in most enterprises, current log systems cannot even fully audit human operations.
Hardware tells the truth: Enterprise SSD surges 86%
If Altman is talking about the future, then TrendForce's data from June 11 is what is happening right now.
$18.46 billion in Q1 2026 total revenue of the top five global enterprise SSD brands, +86.1% QoQ, a historical high
86% is not a small number. Nearly doubling in a single quarter is extremely rare in the history of the storage industry. The driving factors are very specific: CSPs (Cloud Service Providers) are purchasing storage on a large scale for AI agent services, and enterprises are also expanding capacity for AI workloads.
Here is a very practical logic chain: For an AI agent to run → it needs to store a large amount of context, tool call records, and intermediate results → read speed directly determines the agent's response speed → enterprise SSDs go from "sufficient" to "bottleneck".
I asked a friend who works in data centers, and his exact words were: "The storage requirements of AI agents are different from traditional applications. Traditional applications are 'write once, read many times,' while AI agents are 'heavy writing, random reading, latency-sensitive.' When an agent runs a complex task, it might generate several gigabytes of intermediate data. Just imagine, if a company has hundreds of agents running simultaneously..."
In other words, the 86% surge is not an isolated data point. It means that AI agents have transformed from "a concept in vendor PPTs" into "a physical reality that is consuming global enterprise hardware capacity".
What Altman said
AI will evolve from "What do you want me to do" to "I know what to do myself"
Pointing to the future
As shown by TrendForce
AI is already consuming $18.4 billion worth of storage, and supply is starting to tighten
Happening now
The Third Card: EU AI Act Countdown to August 2
In addition to Altman's predictions and TrendForce's data, another matter looms overhead: Most of the rules of the EU AI Act will officially take effect and begin enforcement on August 2, 2026.
This bill requires all enterprises that serve European users or have business dealings with European customers to classify AI systems, conduct risk assessments, disclose transparency, implement human review mechanisms, and retain logs. The maximum penalty for violations can reach 7% of global annual revenue (heavier than GDPR's 4%).
And what time is it now? June 12th. There are less than two months until August 2nd.
The European Commission is expected to release the final version of AI-generated content labeling and code of conduct this month (June). However, most companies have not even completed an inventory of their AI systems.
Looking at these three things together, I believe the core contradiction that enterprise CIOs face is the same:
The physical demands of AI (SSD, computing power, storage) are already growing exponentially, the form of AI (evolving from a tool into a backend operating system) is rapidly changing, and the compliance framework for AI (EU AI Act) is about to take effect—yet most enterprises' IT governance systems remain stuck in the stage where "AI is a pilot project."
What to do: not to chase every model release, but to establish the three pillars
To be honest, Claude Opus 4.8 just launched, GPT-5.6 is reportedly coming next month, and NVIDIA GTC just wrapped up — model updates are coming so fast it's hard to keep up. But for most enterprises, what they really need to focus on right now isn't model parameters, but three things:
First, conduct an AI system inventory, do it now. Which business processes have AI embedded? Whose model is it? Where does the data go? Is there a human review step? How long are logs retained? Don't wait until August 2nd when compliance forces you to do it—by then, time and cost will have doubled.
Second, incorporate AI's "silent operation capability" into infrastructure planning. Whether your enterprise SSDs are sufficient, whether network latency can handle hundreds of intelligent agents running simultaneously, and whether log storage capacity can hold up—these issues won't appear in any AI vendor's sales PPT, but they are the real bottlenecks. TrendForce's 86% data isn't meant for investment banks; it's an infrastructure warning for CIOs.
Third, establish a governance model for "AI silent operations." Altman's "proactive AI" currently sounds like science fiction, but if you look back two years—to June 2024—who would have imagined that by 2026, companies would be seriously deploying hundreds of AI agents? The arrival of proactive AI is not a question of "if," but "how fast." Before it arrives, design the permission boundaries, operational audits, and emergency shutdown mechanisms in advance. The cost of patching after the fact has always been several times that of designing upfront.
