You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
In a new interview on 60 Minutes, President Donald Trump declared that China and other foreign nations would be barred from accessing Nvidia’s most advanced AI chips, such as those in the company’s flagship Blackwell series. Trump stated, “The most advanced, we will not let anybody have them other than the United States,” underscoring his intent to tighten technology export controls for national security reasons. The remark signals a continuation—and possible escalation—of Washington’s semiconductor restrictions, first imposed to curb China’s access to leading-edge artificial intelligence hardware.
Trump’s comments highlight how AI chips have become a strategic asset, comparable to energy or defense technologies. Nvidia’s chips power everything from generative AI systems to autonomous weapons research, and limiting access could preserve America’s technological lead. However, the policy could also strain global supply chains and complicate Nvidia’s international business, as China remains one of its largest markets. The president hinted that less advanced models might still be sold abroad, but the “most advanced” technology would remain exclusively for U.S. use. The statement underscores growing geopolitical tensions over AI dominance and marks another turning point in the global competition for semiconductor leadership.
It's interesting watching this unfold live this year.
First, the administration had serious doubts about the CHIPS act. Then it pivoted to "well, the idea is sound but the implementation is terrible", and now it's "US Chips Uber Alles".
(I personally agree semiconductor security is a strategic concern for any nation.. It would be really fun to be a fly in the wall for semi- supply chain threat assessments that are going on in the Pentagon and other places..)
I just watched the whole thing. Wow. I had always said I would not vote for a president older than 65 and I stand by it. If you are a Captain for United Airlines you are retired on your 65th birthday. POTUS and all politicians for that matter should be the same, absolutely.
NORAH O'DONNELL: The one thing that China wants but it doesn't have is the world's most advanced semiconductors.
PRESIDENT DONALD TRUMP: That's right.
PRESIDENT DONALD TRUMP: Chips in particular.
PRESIDENT DONALD TRUMP: Within-- two years from now, we'll have 40% or 50% of the chip market. What's happening here, the biggest companies are leaving Taiwan. They're coming into the United States because of tariffs. If we didn't have tariffs, they wouldn't be doing it because--
I guess Abu Dhabi is ok, because the most advanced AI chip is really the Cerebras WSE-3. G42, an Abu Dhabi company, is responsible for over 80% of Cerebras' revenue.
That's a complex question. The wafer-scale strategy only works with very small functional units (specialized AI cores, in the case of the WSE-3) which can be disabled if the fabrication process has a defect which caused a core to be bad. If you need large cores, like CPUs or XPUs do, the wafer scale strategy may be an inefficient use of die area, because the die area disabled from a defect is large, and the probability of a core being disabled is higher.
Two big advantages of wafer scale for Cerebras are 44GB of on-die SRAM, which has very low latency, and the on-die mesh interconnect between the cores and the SRAMs is incredibly fast and high bandwidth. Oddly (and annoyingly to me), Cerebras does not publish the bandwidth for each mesh link, or the standard measure for network bandwidth (called bisection bandwidth), they only publish "aggregate bandwidth", which is probably the sum total bandwidth of all of the functional links in the mesh, and that is 27 petabytes per second (27,000,000 gigabytes per second). Pardon my petty annoyance; SRAM speed and capacity, and on-die interconnect bandwidth are both huge advantages, but the advantage depends on the application. For AI they are huge advantages.
I guess Abu Dhabi is ok, because the most advanced AI chip is really the Cerebras WSE-3. G42, an Abu Dhabi company, is responsible for over 80% of Cerebras' revenue.
Actually, didn't they just approve a Microsoft/UAE deal? I read the transaction involves GB300 GPUs, isn't that the best Nvidia stuff right now?
Given where things sit, I am now beginning to think the 15% China surcharge 15% is not as silly as it initially sounded.
Why not open up the whole portfolio and have an escalating surcharge;
they are going to get Blackwells, and Rubins and Feynmans (if Nvidia's reign lasts that long!), they just have to pay through the nose for them. So, why not collect the bootlegging fees officially!
www.cbsnews.com
