Have you noticed that the hottest topic in 2026 has changed?
Last year everyone was still discussing "Can AI write poetry?" This year, no one is asking that question anymore.
This year, everyone is asking something else—Can AI make decisions for me?
This is not a hypothetical. Over the past 30 days, I tracked nine hot topics, and like puzzle pieces, they assembled a complete picture.
Let me tell you about them one by one.
First, the first piece of news.
A couple hoarded memory chips for 5 months and had a paper profit of 32 billion.
You might say: Isn't this just speculating on chips? What does it have to do with AI?
It has a lot to do with it.
Memory chips are AI's "blood supply system." Large model training requires massive HBM high-bandwidth memory, and inference requires large-capacity SSDs and DRAM. From the end of 2025 to now, global AI computing power chips have been in short supply, HBM production capacity has been monopolized by SK Hynix, Samsung, and Micron, and prices have tripled.
People speculating on memory chips are actually betting on one judgment—
AI is fully shifting from the "training era" to the "inference era."
These two eras correspond to completely different business logic.
From 2022 to 2025, the whole world was doing the same thing: benchmark scores.
Whoever had more parameters won. GPT-4 had a trillion parameters, Claude-3 had 2 trillion, and Gemini Ultra was on par. China also launched DeepSeek-V3 and Tongyi Qianwen.
But in 2026, the turning point has arrived.
When large models are already "smart" enough, people no longer care about who has the most parameters—they care aboutwho can make money with AI.
The first step to making money is to make AItransform from an auxiliary tool into an executor.
The second piece of news is even more direct.
Unitree Technology released a manned mecha.
Not a PPT, not a concept video, but a real bipedal robot that you can actually sit on and that can walk. Investors responded: "This is not a toy, it is a productivity tool of the future."
This sentence made me think for a long time.
Have you noticed that in 2026 robots no longer "look like robots"? They are starting to look likeObserve the environment, understand instructions, and act autonomously like a human.
What is the core technology of Unitree mecha? It's not the mechanical structure—that's just the shell. What's truly valuable is the built-inAI Agent。
What can this AI Agent do?
· Understand vague instructions you say in natural language ("Help me move that box to the door")
· Autonomously plan action paths (bypass obstacles, avoid pedestrians)
· Adjust movements in real time (switch grip if the box is too heavy, change gait if the road is uneven)
Sounds simple? But this is exactly what AI couldn't do in the past five years.
This is the essence of an enterprise-level AI Agent: not "follow instructions to do work," but "understand intent, execute autonomously."
"In the industrial age, humans replaced animals; in the information age, the internet replaced intermediaries; in the AI age, intelligence replaces 'execution' itself."
In 2026, 30% of China's Top 500 enterprises have deployed AI Agents.
They don't write copy or draw pictures—those are too basic—they do just three things:
· Decision support:Help sales judge customer intent, help procurement predict price trends
· Process automation:From order to shipment, the system runs it all itself, and humans only review exceptions
· Risk warning:24/7 monitoring of business data, and when anomalies appear, it investigates the cause itself
Three scenarios, each one "thinking on behalf of people."
The third piece of news is DeepSeek.
This Chinese AI company open-sourced their latest model.
What does open source mean? It meansthe threshold drops to zero.
I talked with the CTO of an SME. He said:
"Last year we had to use the GPT-4 API, and every million tokens cost 30 dollars. We also had to worry about data privacy. Now, using DeepSeek's open-source model deployed on our own servers, the cost is 1/20 of before. The key—data never leaves the building."
He is not the only one doing this. More and more Chinese companies have chosen the route of "open-source deployment + fine-tuning."
You see, this is the typical playbook of the Chinese market:
When a technology product goes from "only big companies can afford it" to "anyone can use it," innovation explodes exponentially.
Over the past six months, those who have felt this deeply are developers on low-code platforms.
Low-code + AI, development efficiency increases by 300%. A feature that used to take a team two weeks now takes one person two days.
The WeChat Mini Program ecosystem is doing the same thing—opening AI capabilities to developers. Want to build an AI chatbot? No need to tune the model yourself, just connect to the API. Want to build an intelligent customer service? Just drag and drop a few components.
What chemical reaction will these three trends produce when stacked together?
Cost decreases → tools spread → scenarios explode → data accumulates → models optimize → cost decreases again
This is a positive flywheel. And May 2026 is the starting point of this flywheel.
The good news is over. Now for the bad news.
Not long ago, an accident shocked the tech world.
An AI Agent deployed by a certain company, while executing an automatic cleanup task, misunderstood the scope of the instruction—and deleted the production database. Not the test database, not the backup database, but the production database. Millions in losses, and recovery took three days.
Many people see the news and their first reaction is: "See, AI really isn't reliable after all."
My view is completely different.
This isn't an AI problem, it'sa governance problem.
Looking back at the history of human technology, every revolution has come with new risks:
We didn't give up electricity because of fires, and we didn't ban driving because of car accidents. The same goes for AI Agents.
Every enterprise deploying AI Agents must answer three questions:
1. Permission boundaries—can it do it?
Draw clear "behavioral boundaries" for AI Agents: what can be touched, what cannot.
2. Ethical boundaries—should it do it?
Even if it's technically possible, you must judge whether it's appropriate. For example, can AI directly fire employees? Technically yes, but ethically no.
3. Attribution of responsibility—if it makes a mistake, who's responsible?
Design flaw? Operational error? Training data problem?Without a clear accountability mechanism, don't launch it.
The good news is that systems and rules are catching up.
The EU AI Act officially took effect at the end of 2025, giving rise to a brand-new profession—AI ethics compliance officer. This position starts at a million in annual salary, but there are very few qualified candidates on the market.
The Musk and OpenAI court case is also ongoing. But note one detail: what both sides are arguing about is not "whether AI should be developed," but"who should hold control over AI."
The world's focus has shifted from "whether to use AI" to "how to use AI safely."
Having said all this, let's bring it down to concrete questions:As enterprises, as individuals, what should we do now?
Let's start with enterprises.
First, take stock of your "AI-able" assets. Don't ask "what can AI do for me," but ask "in my business, which links are repetitive, clearly rule-based, and data-rich"—these are most worth transforming with AI Agents.
Second, don't start from scratch. DeepSeek's open-source models are already there. Use them directly. Fine-tune your own version. Put them to use, iterate.
Third, establish an AI governance mechanism as soon as possible. Don't wait until something goes wrong to patch it up. Permission tiers, human review, anomaly circuit breakers—these must be designed before launch.
Now about individuals.
Three types of people are being eliminated:
The first type: those still asking "what use is AI." Living in an information gap. Tools of the same era—others are using them, while they're still watching.
The second type: those still treating AI as a toy. Chatting with ChatGPT, drawing with Midjourney—fun is indeed fun, but have you ever thought about turning it into productivity?
The third type: those who think installing AI means "everything is fine." The most dangerous. AI is not something that "works as soon as you install it." It requires data cleaning, process adaptation, personnel training, and continuous operation.Technology implementation is always 10% technology + 90% organizational change.
That day I saw a comment that was particularly interesting.
Someone said: "Is an AI Agent that dream super employee of mine—one that doesn't need a salary, doesn't take sick leave, and doesn't complain?"
I replied: "Yes, but it also makes mistakes, needs management, and needs boundaries. It's like the smartest new hire on your team—huge potential, but needs a good manager."
That couple in memory chips wasn't betting on chips, but ontrends。
Unitree's mech wasn't showing technology, butdirection。
What DeepSeek open-sourced wasn't code, butopportunity。
What the AI Agent deleted wasn't a database, buta wake-up call。
Trends + direction + opportunity + wake-up call. This is the most complete business puzzle of May 2026.
The remaining question isn't "will AI replace me"—
but: will I become the one who is first to make good use of AI?
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
