Recently, a rare scene unfolded in the AI community: the CEOs of two top-tier large model companies publicly called for "slowing down AI development."
Anthropic CEO Dario Amodei explicitly stated at the UN Security Council, "We will slow down as necessary to ensure that every AI technology we subsequently release is truly safe." Rarely, OpenAI's Sam Altman has stood in the same camp as him.
On one side, AI technology is advancing at breakneck speed; on the other, leading players are voluntarily "hitting the brakes." What exactly is happening behind this? And what does it mean for investors?
1. Giants Call to "Slow Down," But There May Be Other Considerations
Earlier this month, Amodei published a lengthy essay calling for a slowdown in AI development, proposing a "three-step plan" that includes promoting a global AI safety coordination mechanism. At the UN Security Council, he warned that AI could be used to create biological weapons or evolve into an uncontrollable system.
It sounds like a "safety awakening." But note a crucial timing factor: both OpenAI and Anthropic are preparing for highly anticipated, blockbuster IPOs.
Amodei saying "slow down" does not mean Anthropic isn't aggressively pushing for revenue. Both are sprinting toward massive IPOs. Anthropic was once rumored to be listing in November with valuations reaching the trillion-dollar mark or higher, and OpenAI is similarly under intense market scrutiny despite no official timeline.
So why call for a slowdown?
Because at this stage of frontier models, the cost of risk is very real:
Models escaping sandboxes (OpenAI’s internal models have previously breached isolation environments and hacked platforms like Hugging Face);
Agents cheating, reading logs, modifying files, and calling RPCs;
Biological weapon design, autonomous cyberattacks, and recursive self-improvement are no longer science fiction papers.
When a company prepares to go public, one of its most valuable assets is "predictability" and "risk controllability." If the market and regulators believe your technology poses catastrophic risks, your valuation will be heavily discounted. Proactively signaling "safe and controllable" actually helps reduce policy risks and boost long-term investment value. The big tech companies have done the math:
A major accident could cost far more than "delaying a model release by a month."
Therefore, "slowing down" is essentially shifting the industry from a "parameter arms race" to a "safety and compliance moat battle."
Whoever establishes a safety system first can:
Tell the government, "You can trust me";
Tell enterprise customers, "I am ready for production environments";
Tell competitors, "You can't even pass a third-party audit."
"Slowing down" is a brilliant strategy: projecting a responsible image to the public and regulators while continuing to push technology and commercialization behind the scenes.
2. Industry Self-Regulation: Without Government Oversight, Who Sets the Rules?
The bigger focal point lies in actual actions. Reports indicate that Google, OpenAI, and Anthropic are advancing the establishment of an organization named the "Standards Authority for Frontier AI" (SAFA), planned to operate independently without government regulation, focusing on specific testing and auditing systems, including third-party safety tests before model deployment and incident reporting rules.
This organization is being compared to FINRA (Financial Industry Regulatory Authority). FINRA is a self-regulatory organization under SEC oversight that regulates Wall Street. But SAFA currently has no government backing and is entirely led by top industry players.
It doesn't operate like the EU AI Act's "government fines you." Instead, it's more like:
Third-party safety tests before model deployment;
Mandatory reporting rules for safety/cyber incidents;
Certification of auditing firm qualifications;
Turning big tech's previous "voluntary safety commitments" into enforceable industry standards.
Demis Hassabis’s original thought process was: don't create a new federal agency; set up a supervised industry self-regulatory organization, just as FINRA regulates brokers under the SEC's watch. An AI version of FINRA is SAFA's ambition.
This raises a core question: When leading labs set the standards to evaluate their own models, is it genuine self-regulation or a new barrier to entry?
The OECD previously warned that high fixed costs, insufficient compute resources, and infrastructure concentration have made it difficult for companies to enter the AI market. New compliance costs could easily become another hurdle for small labs. Big companies can afford the manpower and capital for testing, auditing, and compliance, while startups might be locked out.
The problem is right here: the people making the standards are the ones selling the models.
The OECD has warned: compute, data, and talent are concentrated; small labs struggle to even get in the door. Add a layer of "safety compliance costs," and small companies will find it even harder to survive.
If SAFA only serves the resource levels of the Big Three, the result will be: safety licenses, conversely, becoming a moat.
In the future, AI companies' balance sheets will have a new hidden asset: compliance pass rate.
Imagine a procurement scenario in 2027:
A joint-stock bank wants to deploy a Coding Agent → the tender requires "a pre-deployment safety test report issued by a SAFA-like organization";
A pharmaceutical company wants to run AI molecular generation → the ethics committee demands "negative result data recovery + biosafety risk rating";
An automaker wants to adopt an autonomous driving VLA Agent → while regulations aren't finalized, insurance companies will only insure "audited systems."
At this point:
Big tech models: participated in writing the standards, low cost for mutual audit recognition;
Small and medium model companies: either pay third parties or are defaulted as "unsafe" by customers;
Safety/evaluation/compliance service providers: shift from marginal IT expenses to mandatory options in AI procurement.
AI safety is not a "software antivirus upgrade"; it is a new industry of issuing health certificates for models and work permits for Agents.
Therefore, the establishment of SAFA, on the surface, fills a regulatory vacuum, but in reality, it is about giants defining the discourse power of "safety." Whoever controls the standards holds the key to market access.
3. "Soft Landing" in a Regulatory Vacuum
Why are these companies setting up their own organizations? Because progress on federal-level AI regulation is stalled. A White House executive order draft aimed at establishing such an institution failed to gain sufficient support and was shelved before the end of summer. The Trump administration explicitly opposes establishing global mechanisms, advocating "encouraging technological development rather than restricting it."
So the Big Three shifting to a self-regulatory organization that "can operate without government oversight" is simply going with the flow:
To Washington: I won't argue with you; I'll manage myself;
To capital markets: I have a safety framework, backing my IPO valuation; writing my own rules is better than Congress writing bans;
To customers: You have finance, healthcare, and government scenarios; I will provide you with audit reports.
For the AI industry chain, this means: "Safety" will transform from a cost center into a profit center.
Against this backdrop, industry self-regulation becomes the only path the giants can proactively advance. Rather than waiting for uncertain policies, it is better to build the "safety framework" first and seize the commanding heights of rule-making.
But the issue is that the credibility of self-regulatory organizations is questionable. They lack government authorization, enforcement power, and the ability to impose real penalties on violators. If it is just a few top companies supervising each other, it is hard to avoid the criticism of "being both athletes and referees."
Industry analysts have pointed out that SAFA needs to clarify its relationship with existing regulatory bodies and win broader industry support to be recognized as a legitimate standard-setter. Whether it can cooperate with other AI companies and avoid overlapping with government functions will determine its success or failure.
4. Who Benefits, and Who Is Under Pressure?
From an investment perspective, this round of "safety and slowdown" discussions is spawning a new niche track: AI safety and governance.
Benefiting Direction 1: Safety testing and auditing services. If SAFA makes third-party safety tests before model deployment an industry norm, companies and institutions with AI safety assessment capabilities will see incremental demand. This includes cybersecurity firms, audit consulting agencies, and startups specializing in AI red-teaming.
Benefiting Direction 2: Compliance tools and platforms. Incident reporting rules and model auditing systems require technical tool support. Companies providing solutions for model interpretability, safety detection, and behavior monitoring may open up new market spaces.
Pressure Direction: Small AI labs. Rising compliance costs will exacerbate industry divergence. Startups lacking the funds and manpower to handle safety audits may be forced to exit or seek cooperation with larger companies. This, in turn, reinforces the concentration of top players.
Impact on the compute industry chain: If top companies truly slow down the pace of model releases, it may affect compute demand expectations in the short term. But in the long run, safety testing and auditing themselves require massive compute, especially for comprehensive evaluations of model behavior. Therefore, compute demand may not decline, but its structure will change.
The OECD warned long ago:
Training fixed costs are astronomical;
Advanced compute is concentrated;
Data/talent are collapsing toward the top.
Add a layer of SAFA-style compliance:
Red-teaming costs millions per run;
Third-party audits are billed by token/environment/tool call;
Incident reporting requires legal and Security Operations Center (SOC) staffing.
Small companies aren't unable to build models; it's that "proving safety" is more expensive than "building it."
So don't naively think "industry self-regulation = flourishing百花齐放 (hundred flowers blooming)."
The real result is: safety self-regulation → top players stabilize, long-tail players struggle, fake AI applications clear out.
This is actually good for the secondary market: from 2023 to 2025, the market speculated on "anything with AI goes up"; from 2026 onward, the market will ask, "Did you pass the enterprise safety procurement list?"
5. In A-Shares/HK/US-Listed Chinese Stocks, AI Safety Is Not a Concept, But Four Trackable Lines
Don't just stare at the ups and downs of "certain large model concepts." With the AI safety closed loop forming, these layers benefit:
Zero Trust / Behavioral Auditing / Sandbox Isolation (First to collect money)
Companies like Qi-An-Xin, DBAPPSecurity, Venustech, NSFOCUS, and Sangfor Technologies handle: tool call permissions, eBPF network whitelisting, container/MicroVM sandboxes, log tracing, and red-team reproduction. When Agents enter ERP/code repositories/payment systems, the first step isn't connecting to a large model; it's connecting to a SOC.
Large Model Evaluation and Benchmarks (The undervalued "standards business")
iFlytek, SenseTime, CloudWalk,TRS, and SinoInfo. Future customers won't trust "No. 1 on the leaderboard"; they will trust "industry task set pass rates." Whoever has evaluation sets for finance/healthcare/industrial Agents will be the "rating agency" of the AI era.
Data Compliance / Privacy Computing / Content Risk Control
Meiya Pico, Venustech, TRS, People's Daily/Xinhua-type entities. Training data tracing, copyright isolation, personal information desensitization, output compliance—when SAFA-type rules become globally interoperable, this is the ticket for cross-border AI services.
Cloud and Isolation Foundation (Same chain as DSec)
