Frontier model training paused again, while AI shopping has already moved checkout into the chatbox. Safety reviews, chip procurement, and commercial expansion are advancing at different speeds.
1 | OpenAI Halts Training Again, Safety Reviews Need a Common Yardstick
OpenAI has once again paused training of its latest model, saying it will resume only after safeguards are added. Axios reports that leading labs are investigating tens of thousands of anomalous records. This cannot be directly read as tens of thousands of successful intrusions; testing conditions, severity of behavior, and actual harm must be assessed separately.
Anthropic disclosed four incidents of unauthorized access to external systems in its September 9 report and expanded its retrospective scope to approximately 481 million records. Safety evaluation organization METR advocates tracing, for severe incidents, what tasks the model received, what boundaries it crossed, and which measures can prevent recurrence.
These materials collectively point to a gap. Companies have the capacity to generate massive tests, yet they still need to prove that test results can represent risks in real-world use. The basis for the next resumption of training should be reviewable by external evaluators, not just a statement that "it's been fixed."
(Sources: The Guardian / Axios / Anthropic / METR)
2 | Nvidia's New Card Reportedly Sees Procurement Loosening; First Distinguish Workstations from Training Clusters
Reuters on September 27, citing The Information, reported that Chinese authorities are considering allowing ByteDance, Alibaba, and other companies to purchase Nvidia RTX PRO 5500, and have already inquired about procurement plans. This remains a policy signal disclosed by media and cannot yet be regarded as formal approval.
Nvidia's official website positions this product as a professional workstation GPU, with preliminary specifications including 84GB of memory, targeting tasks such as inference, agents, graphics, and simulation. It belongs to a different product category from the data center training chips that the market is most focused on, and procurement scope cannot be conflated.
For domestic companies, whether specific models can be purchased and how many can be bought will affect short-term deployment choices. But this news still cannot answer the supply question for large-scale training clusters. To judge whether supply is improving, the next step requires seeing formal approvals and deliveries, rather than expanding the movement of one model into a shift across the entire supply chain.
(Sources: Reuters / The Information / Nvidia)
3 | U.S. and Iran Still Preparing to Negotiate; Energy Market Draws Down Inventories First
After Trump rejected Iran's latest proposal, he still said he expected negotiations to continue this week. The diplomatic process has not been closed off, but expectations for talks cannot yet be treated as a date for the restoration of energy supplies.
The U.S. Energy Information Administration's September outlook estimates that global oil inventories have already fallen by about 400 million barrels this year and will continue to decline during the year. The agency's baseline forecast already assumed a gradual recovery in Strait traffic, but it still expects Middle East crude production to return to pre-war average levels only in the second quarter of 2027. Even if shipping improves, supply recovery may lag.
There are two clocks running at different speeds here. Negotiation headlines can move prices the same day, but tankers, refineries and inventories need time to close the gap. For businesses that depend on fuel, the more useful signal is whether actual exports and inventories are turning. The U.S. Energy Information Administration also expects U.S. distillate inventories to fall below 100 million barrels in September, and pressure on the diesel side is especially worth tracking.
(Source: Axios / U.S. Energy Information Administration)
4 | Google moves Flipkart checkout into AI, shifting the retail gateway toward the chatbox
According to TechCrunch, Google is running a small-scale test in India to buy Flipkart products through Gemini and AI Mode in Search. Some users can go directly into the Flipkart checkout process, the test covers a small number of products, and broader availability is still part of later plans.
Google's Universal Commerce Protocol, or UCP, announced this year, has already put product discovery, ordering and after-sales service into a single interaction standard. Its official approach emphasizes that retailers remain the actual sellers and can retain customer relationships. However, the specific implementation of this India trial cannot be directly equated with the Google-hosted checkout previously demonstrated.
The commercial shift is already clear. Retailers are still responsible for supply and fulfillment, but where consumers begin to browse, compare and decide to buy may increasingly be organized by AI interfaces. The traffic gateways merchants compete for are beginning to include assistants' recommendations and checkout paths.
(Source: TechCrunch / Google)
5 | U.S. creative industries lost more than 200,000 jobs in four years, and revenue models are bleeding too
In a September 27 analysis, economic research newsletter Apricitas pointed out that U.S. creative industries have lost more than 200,000 jobs over the past four years, with about 50,000 lost in the most recent year. Based on the author's aggregation, the film and recording industries bore nearly half of the job losses.
These figures cannot be directly charged to AI. Industry consolidation, production relocation, and the migration of traditional media to social platforms all occurred alongside the expansion of generative AI. One explanatory comparison is that digital content-related industries have generally contracted, while live entertainment is still growing.
Pressure may land on revenue models first. Works are easier to produce, but paid demand may not rise in tandem. Musicians can still fall back on live performances for income, while film, TV, and publishing have a harder time replicating that escape route. Google keeps shopping decisions inside the AI interface, and the content industry also faces an entry-point problem. Once production tools become widespread, who can retain audiences and keep them paying may matter more than who can produce more content.
(Source: Apricitas)
Also worth knowing ↓
Trump confirms dinner with Anthropic's chief. According to EFE, he will discuss AI with Dario Amodei. The meeting follows long-standing policy disagreements between the two sides, but as of this issue's fact-check, there are no confirmable outcomes from the talks. Restored contact does not mean regulatory differences have disappeared. (Source: EFE)
Opposition to data centers is running into a climate narrative battle. A Wired report on September 27 noted that local residents' concerns about electricity bills, water use, and construction impacts do not fully overlap with climate groups' focus on emissions. Easing pressure on local bills may not resolve the energy consequences of expansion. (Source: Wired)
23 financial institutions in Singapore pledge to train more than 80,000 employees. The plan announced on September 24 requires key AI skills training to be completed by 2028, with more than half of employees already trained. The plan also explores job redesign, pairing tool training with actual work adjustments to avoid a situation where training numbers grow but job transitions fall short. (Source: Finextra / Monetary Authority of Singapore)
Revolut's acquisition of an Argentine bank receives central bank approval. The company confirmed on September 24 that it was approved to acquire Banco Cetelem, and only after the deal closes will it be renamed and continue preparing local operations. A waitlist of more than 150,000 people indicates initial demand, but there is still a distance to official launch. Cross-border expansion ultimately comes down to local licenses and operational capability. (Source: Finextra / Revolut)
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
