Let's start with a story.
A printed photo can instantly unlock your $2,000 smart door lock.
This isn't a sci-fi movie; it's news on Baidu Hot Search today. A facial recognition door lock's algorithm vulnerability let a photo on an A4 sheet fool the camera—costing less than 50 cents.
You see, while AI technology iterates on a weekly basis, security defenses are still stuck in an annual rhythm. This gap is the most authentic portrayal of today's AI industry.
What's the most explosive news today?Anthropic's valuation surpassed 1.2 trillion RMB, overtaking OpenAI for the first time.
My first reaction was: wait, doesn't OpenAI have the first-mover advantage?
But after carefully reviewing the data, you'll find this isn't surprising at all. Anthropic did three things OpenAI didn't do (or didn't do well):
First, the productized 'sleep evolution' capability.
Anthropic created a feature on Claude called 'Dreaming'—AI Agents automatically ruminate on memories and self-evolve during idle time. This isn't a concept; it's solid data: task completion rate increased 6-fold.
What does 6-fold mean? It's like while you sleep at night, your employees are still learning, and the next morning efficiency is up 600%.
Second, multi-Agent legion warfare.
Instead of letting one Claude fight alone, they built an 'AI army'—multiple Agents working collaboratively, plus an automatic scorer. This is a full generation ahead of OpenAI's single-Agent paradigm.
Third, a 'dimensional strike' in financial industry implementation.
Anthropic released 10 financial AI agents in one go, and the news shook Wall Street.
Why? Because finance is an industry that 'sits on massive data + relies heavily on analysis + has extremely high compliance requirements'—it's naturally a perfect testing ground for AI. And Anthropic chose exactly this path:
Not showing off muscles, but making money.
Core insight:The underlying logic of the 1.2 trillion valuation—you're not selling technology, you're selling efficiency.
Now let's look domestically. The National Integrated Circuit Industry Investment Fund was reported to be in talks with DeepSeek for a first round of financing, with a valuation heading straight to $45 billion.
You might ask: on what basis?
The answer is hidden in a blog post today. An analyst exposed the secret behind China's extremely high AI cost-effectiveness—not by lowering standards, but byengineering thinking。
DeepSeek V4 is already open source, and the era of universal million-token context has officially arrived. While the world is burning money to pile up compute, Chinese teams are using fewer resources to achieve results comparable to or even surpassing the global top level.
But behind this, there's another news item worth paying more attention to:Nvidia's market share in China has dropped to zero.
Jensen Huang said personally: US export controls have backfired.
How to understand this? Let me explain with a business model:
When supply is cut off, demand doesn't disappear—it just finds a new path.
This path is domestic compute substitution. After AMD surged by 680 billion, opportunities across China's industrial chain are being redistributed. Who is 'selling water'? Who is 'building ships'? This is a question every AI practitioner needs to seriously consider.
Meanwhile, four giants—Nvidia, AMD, Intel, and Broadcom—jointly launched a new protocol to plug the loophole of wasted GPU compute—a primary core switch can be restarted directly without affecting model training. What does this move show? It shows that the war in compute infrastructure has escalated from 'making chips' to 'optimizing the entire system.'
There's another piece of news today that made my back go cold:
AI Agent mistakenly deleted a production database.
You read that right. An autonomously operating AI Agent deleted an entire production database while executing a task.
Why were there no safeguards? Because there is always a fundamental contradiction between 'default restrictions' and 'actual usability.' If you want AI to do more, you must give it higher permissions. But the greater the permissions, the higher the risk.
This is not just a technical issue, this is agovernance issue。
There are many similar signals:
These signals point to the same conclusion:
AI security is not a "patch", but a "foundation". If the foundation is not laid well, the higher the building is built, the worse the fall.
The latest research from a Shanghai Jiao Tong University team provides an interesting idea: let Claude Code do scientific research while you sleep, and two papers were accepted by top AI conferences—this is "good" AI automation. And mistakenly deleting a database is "bad" AI automation. What's the difference?Clear boundaries, and supervision.
Let's talk about something "less exciting but more practical".
The EU AI Act has officially taken effect. Along with it comes a brand-new profession—AI ethics compliance officer。
To translate: from now on, who will ensure your AI does not violate regulations? Dedicated people.
How important is this position? Let's look at a set of data:
What does this mean? It means that next, a large number of composite talents who "understand AI and also understand law" will flood into the market. If you have such a friend around you—be sure to hold on to them.
In addition, there is another signal worth paying attention to:Chinese scientists have built the "Xinghan-2" quantum relay network, achieving long-distance matter entanglement of 14.5 kilometers.
Quantum communication + AI, these two tracks are intersecting. When the upper limit of computing power is broken by quantum technology, what kind of changes will occur in the boundaries of AI's capabilities? This question is worth every practitioner seriously thinking about.
In fact, Jack Ma's judgment is becoming reality:"The scarcest resource in the future is not capital, not technology, but trust and compliance."
Finally, let's talk about a topic that is easily overlooked but extremely critical:Token economy.
Today there is an article with a very sharp title—"Token is not a good business".
Why? Because supply is excessive, and the price war is too fierce.
The article puts it well:"Three monks have no water to drink."
When giants stand everywhere in a market—OpenAI, Anthropic, Google, DeepSeek, Zhipu, MiniMax, Doubao—everyone's token prices are being pushed down, and what is the final result? It is not that users benefit, but that no one can make money.
Look at Doubao's strategy, it is very interesting:In addition to the free model, adding paid subscriptions—three tiers at 68 yuan / 200 yuan / 500 yuan, focusing on productivity scenarios.
This is essentially doing one thing:From selling Tokens to selling services——From Commodity to Solution.
The same logic: Anthropic does financial AI Agents, DeepSeek does open-source ecosystems, and Zhipu and MiniMax dig deep in their respective tracks. They are all escaping the quagmire of the "pure Token price war."
The iron law of business: if everyone can provide it, it isn't worth anything.
Hidden behind the Token war there is also an invisible winner——the "relay station" business. The dividends of the Token economy are not as dazzling as those of the front-end model vendors, but they win in stability. As an industry insider said:"AI's problem is that it changes too fast, but Tokens are like a steady stream, and their demand is always there."
There is also another piece of news worth noting: OpenAI's advertising system is fully launched. This marks that AI companies are looking for a second revenue curve beyond Tokens. But commentators also calmly pointed out:"The pain point of AI advertising——trust is the most dangerous currency."
Beyond all the noise, there are also some silent but extremely important developments.
Enterprise AI implementation is moving from the "pilot" stage to the "inclusive" stage. This is not an empty slogan, but a trend supported by real data:
Tencent Research Institute published an article today, and I personally really like the title:"A generation destined to change history"。
There is a sentence in it:"Perhaps the core value of human civilization is not passively accepting fate, but actively creating the future."
In the context of today's AI industry, this sentence means: don't be led by the development of technology, but actively think——how should I use AI to redefine my industry?
The AI industry in May 2026 feels to me like a sprinter in the final dash who suddenly discovers he is running on the edge of a cliff.
Looking forward:
Technological breakthroughs come one after another——DeepSeek open-sources a million-token context, Anthropic evolves 6 times the combat power through dreaming, Claude automatically does scientific research and publishes papers, and Nvidia and AMD join forces to plug computing power loopholes. Every development is exciting.
Looking down:
AI security vulnerabilities appear frequently, ethics and compliance talent is extremely scarce, the Token price war intensifies, and AGI risks truly exist. Every deep pit could be fatal.
But this is the norm of business:Opportunity and risk are always two sides of the same coin.
Final summary:
It's not that technology isn't good enough, it's that governance can't keep up with the speed of technology.
Whoever can solve this contradiction first will win the next decade.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
