Yazılar

US Implements New AI Chip Regulation to Control Global Access

/in /tarafından

The U.S. government has introduced a new regulation to restrict global access to U.S.-designed artificial intelligence (AI) chips and technology. This regulation targets the export of advanced graphics processing units (GPUs), essential for building AI models, and aims to ensure that cutting-edge AI capabilities are developed and deployed securely and in trusted environments.

Which Chips Are Restricted?

The regulation focuses on GPUs, which were initially created to accelerate graphics rendering but have become critical for AI due to their ability to process large amounts of data simultaneously. U.S. companies, particularly Nvidia, dominate the production of these chips. GPUs like Nvidia’s H100 are used extensively in training advanced AI models, such as OpenAI’s ChatGPT.

What Is the U.S. Doing?

To regulate global access, the U.S. is extending restrictions on advanced GPUs, specifically those used in AI training clusters. The new rule sets limits based on compute power, measured by Total Processing Performance (TPP). For most countries, the cap is set at 790 million TPP until 2027, equivalent to roughly 50,000 H100 GPUs. These restrictions are meant to control access to the computing power required for large-scale AI research and applications.

However, certain companies, like Amazon Web Services and Microsoft Azure, that meet the requirements for special authorizations (called “Universal Verified End User” status) are exempt from these caps. Additionally, countries with “national Verified End User” status are allowed more advanced GPUs—about 320,000 over the next two years.

Exceptions to Licensing

There are exceptions for small GPU orders, such as those for universities or research institutions. Orders that do not exceed 1,700 H100 chips only require government notification and do not count toward the caps. This exception is designed to facilitate the global flow of AI technology for low-risk purposes.

GPUs intended for gaming are also excluded from the restrictions, ensuring that the gaming sector remains unaffected by the new rules.

Which Places Can Get Unlimited AI Chips?

Eighteen countries are exempt from the country-specific caps on GPUs. These countries include the U.S., Australia, Canada, Japan, South Korea, the European Union members, and Taiwan. This list reflects nations the U.S. considers aligned in terms of AI development and security.

What Is Being Done with ‘Model Weights’?

In addition to GPUs, the U.S. is regulating “model weights,” which are numerical parameters used in training AI models. These model weights, essential for refining the performance of AI algorithms, are considered sensitive information. The new rule establishes security measures to protect these parameters, ensuring that only trusted entities manage the most advanced AI systems.

Conclusion

The U.S. regulation reflects growing concerns over AI technology’s potential misuse and aims to ensure its responsible development. By controlling the flow of critical AI resources like GPUs and model weights, the U.S. seeks to maintain dominance in the AI field while preventing sensitive technology from reaching adversarial nations.

 

TSMC Fourth-Quarter Profit Expected to Jump 58% Due to AI Chip Demand Surge

/in /tarafından

Taiwan Semiconductor Manufacturing Co. (TSMC), the world’s leading producer of advanced chips for artificial intelligence (AI) applications, is set to report a 58% increase in fourth-quarter profit, driven by strong demand in the AI sector. The company, which counts Apple and Nvidia among its clients, is benefiting from the AI megatrend but faces challenges such as U.S. government technology restrictions on China and potential tariffs under President-elect Donald Trump’s administration.

Analysts estimate that TSMC will post a net profit of T$377.95 billion ($11.41 billion) for the quarter ending December 31, compared to T$238.7 billion in the same period the previous year. This projection follows TSMC’s recent revenue report, which exceeded market expectations. The company will release its revenue outlook in U.S. dollars during its quarterly earnings call on Thursday.

Arete Research analyst Brett Simpson believes TSMC’s growth in 2025 will continue to be driven by AI customers. He is optimistic that TSMC can establish a strong relationship with the incoming U.S. administration, especially given its $65 billion investment in three plants in Arizona. TSMC’s overseas expansion, however, is not expected to diminish its Taiwanese manufacturing base.

Fubon Financial’s Edward Chen noted that the company’s progress in Arizona, including chip yield rates, would be critical for its future performance. He also highlighted the uncertainty regarding how tariffs from the Trump administration may impact demand.

TSMC is expected to provide updates on its current quarter and full-year outlook during its earnings call, including capital expenditure plans. The company has already projected capital expenditure for 2024 to be slightly over $30 billion and indicated that 2025’s capital spending could surpass 2024 levels.

The AI boom has driven up TSMC’s stock, with the company’s shares soaring 81% last year, significantly outperforming the broader market’s 28.5% gain.

 

Nvidia Faces Setbacks as Major Customers Delay Orders of Latest AI Racks Due to Overheating Issues

/in /tarafından

Nvidia is encountering challenges with its new ‘Blackwell’ AI racks, with major customers delaying their orders due to overheating issues, as reported by The Information on Monday. Shares of the Santa Clara-based company dropped more than 4% following the news.

The overheating problems reportedly affect the initial shipments of the racks, which house Nvidia’s chips in data centers. The glitches include issues with how the chips are connecting to each other. This problem has led major customers such as Microsoft, Amazon’s cloud division, Alphabet’s Google, and Meta Platforms to reduce their orders for the new racks.

Delayed Orders and Shift to Older Models

The affected customers, often referred to as hyperscalers, had placed substantial orders for the Blackwell racks, with each company initially committing $10 billion or more. Some are opting to delay their orders until a later version of the racks is available, while others are returning to older AI chip models.

Microsoft, for instance, had planned to deploy at least 50,000 Blackwell chips in a Phoenix facility, but due to the delays, OpenAI, one of its key partners, requested that Microsoft provide older ‘Hopper’ chips instead.

Despite these delays, it remains unclear how much this will impact Nvidia’s overall sales, as the company may still find other buyers for the affected racks. In November, Nvidia’s CEO Jensen Huang had expressed confidence that the company would exceed its target of generating billions of dollars in revenue from Blackwell chips during its fourth fiscal quarter.

Nvidia and Amazon declined to comment, while Microsoft, Google, and Meta did not immediately respond to Reuters’ inquiries.