Markets have started trading the idea of slower AI development, but actual corporate spending has not clearly followed.
In its Sept. 15 note, WhiteLine Daily, published by WuBlockchain, said Anthropic CEO Dario Amodei called on frontier AI companies to slow the pace of model capability gains to leave more time for safety testing, third-party evaluations, and governance mechanisms. He did not call for a stop to model training. OpenAI CEO Sam Altman and Elon Musk later also expressed support for slowing frontier AI development.
Chip stocks fell as investors weighed slower AI development and higher yields
By the close of U.S. trading early on Sept. 15, the Philadelphia Semiconductor Index had dropped 5.9%. Nvidia fell 3.4%, Micron lost about 5.3%, and AI chip names including Broadcom and AMD also posted notable declines.
At the same time, the U.S. 10-year Treasury yield moved above 5% during the session. Reuters said the main concern in the market was that if frontier model development slows, growth in training compute demand and AI infrastructure spending could also cool. Even so, JPMorgan and Jefferies said they had not yet seen clear signs that AI infrastructure investment had actually been cut.
WhiteLine Daily said the selloff reflected two pressures at once: concern that cloud providers could delay GPU and storage purchases, hurting revenue growth expectations for companies such as Nvidia and Micron, and the valuation pressure that comes when the 10-year Treasury yield rises past 5%.
The Information says Anthropic is tied to a six-year, $13.7 billion GPU contract
According to The Information, citing people familiar with the matter, Anthropic is the customer behind a previously disclosed six-year, $13.7 billion GPU compute agreement signed by RUM Group.
RUM Group, formerly known as Rumble, entered the AI cloud and data center business after acquiring Northern Data. SEC filings show the agreement was signed on Aug. 23 and disclosed on Aug. 24, though the customer was not identified at the time.
The contract calls for GPU services to be delivered in three tranches from a data center under construction in Maysville, Georgia. The total order value is split evenly across the three tranches. The third tranche will only take effect after the customer approves the delivery timeline. The project already has about 120 MW of grid capacity, with room to expand to 180 MW.
RUM also said when it disclosed the contract that it had not yet secured all the funding needed to complete the facility buildout and purchase GPUs and related equipment. It expects most of that spending to be financed through additional debt or equity fundraising.
The Information estimated that Anthropic’s compute arrangements signed over the past year involve at least 14.8 GW of potential capacity and as much as $517 billion in potential spending over the next decade. The report also made clear that those figures do not mean the capacity has been delivered or that the money has been paid.
WhiteLine Daily’s takeaway was direct: Anthropic is publicly calling for slower frontier AI development while still signing long-term compute deals worth $13.7 billion. Based on actual procurement behavior, the company does not appear to be slowing its compute expansion in step with its public position.
Samsung Electronics and SK Hynix reject roughly $18.7 billion in proposed power prepayments
Korea Electric Power Corporation, or KEPCO, wanted Samsung Electronics and SK Hynix to prepay about five years of future electricity bills to help finance transmission and grid construction for a large semiconductor industrial cluster in South Korea.
The proposal was worth about 25 trillion won, or roughly $18.7 billion, in total. Samsung’s share was about 20 trillion won, while SK Hynix’s portion was about 5 trillion won. Both companies declined to participate.
WhiteLine Daily said the issue was not a lack of need for additional power. Instead, the companies were unwilling to take on that level of upfront funding pressure in one move. Korean media said they were also weighing future chip-market conditions and their own investment plans.
If the prepayment structure does not move forward, KEPCO may need to rely more heavily on bond issuance and other debt financing to complete the related grid buildout.
The publication argued that AI memory demand remains strong, but the refusal to lock in even five years of electricity payments shows that once expansion extends beyond fabs and into grid infrastructure, questions around who funds that capital expenditure and how long it takes to earn a return become a fresh constraint.
BlackRock points to Asia’s AI supply chain and three major bottlenecks
In its latest autumn investment outlook, BlackRock said it still prefers emerging markets over developed markets outside the United States and sees the main opportunities in Asia. Taiwan and South Korea, it said, remain central to the AI semiconductor and memory supply chain.
In a separate report on Asia’s AI supply chain, BlackRock identified memory, power transmission, and cooling as the three main bottlenecks. It estimated that memory demand currently exceeds supply by about 20% to 30%, with shortages likely to last for at least two years. But if demand weakens as new capacity comes online, the market could swing back into oversupply in 2027 and 2028.
BlackRock also said global power demand from AI data centers could rise to about four times current levels over the next decade. As chip power consumption climbs, data centers are also shifting from traditional air cooling toward liquid cooling.
The report noted that these figures are BlackRock research projections, not confirmed supplier orders.
WhiteLine Daily’s main conclusion was that markets are already trading an AI slowdown, while real-world investment has not slowed in tandem. Anthropic’s public stance and its long-term compute procurement are moving in parallel, and storage, power, and cooling remain immediate bottlenecks in the AI infrastructure buildout. The next signal to watch is not rhetoric about slowing down, but whether capital expenditure, long-term compute contracts, and infrastructure orders are actually being reduced.