#AMD预计2030年AI市场达2万亿美元 AI bigwigs call to “hit the brakes,” yet OpenAI stops accepting $200 subscriptions: Is this AI rally ending or being extended?

Over the weekend, two major pieces of news from North America’s AI sector clashed, and I’m sure everyone saw them. Looking at them side by side without further analysis can easily lead to a misjudgment.
On one side, Anthropic CEO Dario Amodei published a lengthy article titled “We Must Pace the Frontier,” advocating that frontier models “slow down”: Give third-party evaluation organizations such as METR badges, workstations, and company laptops, have them stationed permanently with employee-level access to inspect safety compliance, report incidents, and audit training alignment; at the next level, democratic countries should jointly establish safety benchmarks, followed ultimately by global coordination. OpenAI’s Altman and Musk subsequently echoed that “speed limits should be imposed.”
On the other side, starting September 10, OpenAI suspended new subscriptions and upgrades for the $200 ChatGPT Pro tier (Pro 20X). Existing users are unaffected when renewing, while $100 Pro, Plus, Go, and the API continue as normal.
The official explanation was very direct: Demand for the new Astra model is “unprecedented,” and high-usage plans are pushing the system to its limits, so the company is first preserving the experience for existing users.
One is static and the other is active, leaving many people confused: Are they slowing down, or continuing to sprint?

Let’s first try to break down Amodei’s article. The original text explicitly says that “pace” does not mean stopping training, but rather making time for alignment, interpretability, and third-party verification; the first step is Anthropic’s unilateral commitment, the second requires governments to grant antitrust exemptions for industry coordination, and global coordination in the third step is even more difficult. In other words, this is a combination of safety rhetoric, institutional grandstanding, and a commercial moat.
Put bluntly, when leading closed-source companies face pressure from open-source catch-up, price wars, and public-market valuations, they package “capability growth” as a “safety issue.” This allows them to respond to criticism from departing researchers and regulators while potentially laying policy groundwork for future chip export controls, distillation restrictions, and evaluation access.
A full halt to training across the board? Impossible.
Now look at OpenAI stopping new subscriptions. We don’t see this as bearish; on the contrary, it is hard evidence of a computing shortage. Astra includes computer use, deep research, and Codex, with inference workloads far higher than the previous generation. $200 heavy users receive 20 times the Plus quota and consume the most computing resources, so enterprise APIs and B2B contracts are prioritized to preserve revenue, while high-end consumer plans are restricted first. Historical comparison: When GPT-4 first became popular in 2023, new Plus purchases were also suspended, then resumed after capacity was expanded. This episode looks more like temporary rationing caused by “demand exceeding expectations while computing capacity expansion lags,” rather than the industry reaching a peak.

So the question is: Will the United States truly slow down for safety and allow other countries to overtake it?
No. AI already carries significant macroeconomic weight in the United States: Some 2025 estimates show that AI-related investment in information-processing equipment, software, R&D, power, and communications could contribute 0.5—1.5 percentage points to annualized quarter-on-quarter real GDP growth, with its contribution particularly striking in certain quarters; even after excluding imports of computer capital goods, the net contribution remains significant. Another estimate for the first three quarters of 2025: AI-related spending excluding imports contributed an average of about 0.9 percentage points and accounted for about 40% of growth, falling to about 20%—25% after imports were included. Regardless of which methodology is used, AI capital expenditure, computing infrastructure construction, and wealth effects in capital markets have already become tied to growth expectations. Amid great-power competition, verbal speed limits are possible, but capital expenditure will not stop collectively; whoever genuinely stops first will hand over its models, customers, and standards.
For computing-power pullbacks, look at earnings; for application launches, watch retention; don’t chase broad-based concept rallies.
In short, speed limits are rhetoric, GPU shortages are reality, and the rally is far from over. But it is also important to note that the market may shift from “everything rising” back to “seeing who can actually deliver.”$AMD