DeepSeek Doubles API Prices During Peak Hours as the Subsidy Era Winds Down
Today, the State Council held a special meeting on artificial intelligence, South Korea is betting 20 trillion RMB on semiconductors, and what you might care about more is: DeepSeek's API has quietly raised its prices.
#1 DeepSeek V4 Pro/Flash API Prices Double During Peak Hours, ⬆️ I Got the Email Last Night
The V4 Pro and V4 Flash models keep their usual prices during off-peak hours, but API prices are adjusted to 2x the normal rate during peak hours.
To put it bluntly, the market DeepSeek captured with its "cabbage prices" is now being harvested. Doubling prices during peak hours means your production environment call costs will rise significantly—most business requests happen to be concentrated during peak hours. If your project heavily depends on the DeepSeek API, it's advisable to recalculate your costs and consider whether you need to schedule traffic between peak and off-peak periods.
DeepSeek API Pricing | Source: CLS
#2 State Council Special Meeting on AI: Accelerate Ultra-Large-Scale Intelligent Computing Cluster Construction, Implement "AI+" Action
On June 29, the State Council executive meeting heard a report on the development of artificial intelligence. Three keywords: accelerate the construction of ultra-large-scale intelligent computing clusters, implement the "AI+" industry-wide action, and improve the tiered safety supervision system. At the same time, increase scientific research funding and talent support for domestic large models and computing chips.
This is a top-level policy signal, a major positive for companies working on domestic large models and computing infrastructure. "Tiered safety supervision" also means more detailed regulatory rules will be implemented next, and companies providing large model API services need to prepare for compliance in advance.
Policy State Council | Source: CLS
#3 South Korea's Samsung + SK Total Investment of 4,755 Trillion Won (Approx. 20.9 Trillion RMB), All-In on Semiconductors
Even more aggressive than the previously reported Samsung solo investment of 648 billion USD—adding SK Group's investment, the two South Korean giants' total future investment in semiconductors and physical AI reaches 4,755 trillion won, approximately 20.9 trillion RMB. South Korean officials also stated they expect DRAM production capacity to double within five years, and the global memory market to grow fourfold within five years.
Meanwhile, Samsung, SK Hynix, and Micron have already been hit with a class-action lawsuit for allegedly manipulating memory prices. On one hand, they are frantically expanding production; on the other, they are being sued for price manipulation—the memory chip business is deeper than you think.
Semiconductors South Korea | Source: Yicai
#4 EU Council Gives Final Approval to AI Simplification Regulation, Rules Officially Take Effect
The EU Council gave the final green light on Sunday, approving new regulations aimed at simplifying and streamlining parts of the artificial intelligence rules. This is the final confirmation of the previous AI Act Omnibus amendment—high-risk system obligations are postponed to the end of 2027, and AI-generated content watermarking obligations are postponed to December 2026.
For developers with business in Europe, the compliance timeline has been pushed back again, reducing short-term pressure. But the regulations are now in place, the direction will not change, and preparing early is better than reacting passively.
EU AI Regulations | Source: EU Council
#5 Ford Rehires 350 Veteran Engineers After Finding Large Model Quality Inspection Replacement Unworkable
A few years ago, Ford tried to replace senior engineers with a large-model-driven quality control system, and it backfired—the training data lacked decades of accumulated engineering judgment experience, and the model actually amplified design flaws. So Ford rehired 350 veteran engineers, known as "greybeards," to optimize those "gone-astray" automated systems.
The result? Ford took first place in the J.D. Power 2026 Initial Quality Study, topping the list for the first time in 16 years. Translation: Large models are not a panacea, and expert experience in some fields truly cannot be replaced.
Ford Large Model Implementation | Source: Sina Finance
One-Liner
Two pieces of news today are interesting to read together: DeepSeek's price hike shows the era of "per-thousandth-of-a-cent" pricing subsidies is winding down, and Ford rehiring old engineers shows large models are not a silver bullet. The industry is moving from the excitement phase of "large models can do everything" into a cooling-off phase of "calculate the costs and boundaries clearly."