# "GPU demand" — X 热门讨论 (2026-09-13 08:56 UTC)
## @IREN_Bull (IREN Bull) · 09-12 20:02 · ♥31 ↻6 💬5 Asked Chat GPT how the recent news from Anthropic and Open AI about slowing the pace of their AI development could impact the broader growth trajectory of AI and data centers. Well worth reading (and highly reassuring) for $IREN investors:
1. Yes. There is a very important distinction between slowing the frontier and slowing AI overall, and that distinction matters enormously for data-center demand.
What Anthropic CEO Dario Amodei is talking about today is not stopping AI development. He is arguing that the handful of companies building the world’s most capable models should slow the rate at which they push model capabilities forward so safety systems, independent testing and containment can catch up. Sam Altman has expressed agreement with the basic idea, and OpenAI recently temporarily slowed parts of its own scaling while strengthening safeguards after the Hugging Face incident and evidence that its upcoming Astra model had reached what OpenAI calls a “Critical” cybersecurity capability level.
The key distinction is between frontier training and the much larger AI ecosystem. Frontier AI is the race by companies such as OpenAI, Anthropic, Google and xAI to train increasingly powerful general-purpose foundation models. Those enormous training runs are the area most likely to be slowed or gated by safety evaluations. Enterprise AI is different: banks using AI to analyze documents, software companies embedding copilots, hospitals processing records, companies running customer-service agents, manufacturers optimizing factories, developers writing code with AI, and millions of people querying existing models. Anthropic itself is simultaneously pushing enterprise deployment and just announced new “Enterprise Frontier Safeguards” developed with more than 100 corporate customers. In other words, “slow down the frontier” does not mean “stop businesses from deploying AI.”
And that’s why I would not interpret these developments as fundamentally bearish for data centers. Think of AI compute demand as three increasingly important buckets: training + post-training/reasoning + inference. A frontier slowdown could reduce the rate of acceleration of the first bucket—for example, perhaps the next gigantic training run happens several months later while safety work is completed. But the enormous installed base of models still has to serve customers, perform reasoning, run agents, generate video, write software and process enterprise workloads. Every one of those activities consumes inference compute.
In fact, the industry is gradually shifting toward inference becoming the bigger driver. McKinsey projects that by 2030 inference will represent more than half of AI compute and roughly 30–40% of total data-center demand. Deloitte similarly expects overall AI compute requirements to continue rising rapidly even if growth in training compute slows, because inference, post-training and test-time reasoning consume enormous amounts of computation.
This produces a counterintuitive possibility: better AI safety could actually increase long-term data-center demand. If enterprises become comfortable allowing AI agents to access corporate databases, write and execute software, perform financial analysis, operate equipment and make increasingly autonomous decisions because the models have better containment and monitoring, usage can explode. Anthropic essentially makes this point itself: as agents become more capable, the economic cost of not deploying them becomes increasingly large, provided they can be deployed safely.
So I’d visualize the situation like this:
Frontier model development → potentially somewhat slower
AI capability → still improving rapidly
Enterprise adoption → accelerating
AI inference → accelerating dramatically
Data-center/GPU demand → still growing https://x.com/IREN_Bull/status/2098864736037179463