ARTIFICIAL INTELLIGENCER-Nvidia drama revives chip export ban debate

Nov 11 (Reuters) - We’ve heard plenty from CEOs about how AI will change everyone’s job. But how does it change the CEO’s own job?

Ali Ghodsi, CEO of Databricks, told me AI has made him roughly 10 times more productive than a year ago — useful enough that he now sits through far fewer meetings.

The tool behind that claim is Genie, an AI agent Databricks just overhauled this week and now pitches as an always-on assistant for businesses. Ghodsi uses it to stay on top of a $190 billion company with 12,000 employees. He showed me how, and it offers a glimpse of the future of AI-run organizations.

In a ChatGPT-like interface, he asks about spending trends among his biggest customers, which engineers are performing best, or how to think through a reorg, and get an answer within minutes. It also runs in the background unprompted, drafting briefing notes by pulling from things like his calendar.

With the “context layer” underneath it, Genie builds a working understanding of who every employee is and what they do, drawing on everything from how much code someone writes to meeting transcripts and sales records.

That also means Genie’s usefulness has a ceiling set by the organization, not just the AI: how much it can offer depends on how much data a company already keeps on its employees, and how willing it is to feed that into the system.

In this week’s issue, we talk to a former White House AI adviser who had his eye on Megaspeed — the company buying up Nvidia chips that’s now at the center of allegations involving its leadership and Nvidia CEO Jensen Huang. Scroll down to learn more.

OUR LATEST REPORTING ON TECH AND AI

Anthropic IPO could draw record retail buying, Vanda Research says

Chinese university bought Nvidia H200 chips despite US blacklist

Explainer - Could Huang, Nvidia and company's board be held liable over affair scandal?

Anthropic eyes $1.4 trillion IPO valuation

Baird, Morgan Stanley sour on Nvidia as CEO affair controversy deepens

TSMC to invest up to $1 billion in Anthropic, OpenAI

THE CASE FOR A CHIP EXPORT BAN

When Ben Buchanan was writing about Megaspeed for his new book, he didn’t expect the Asian cloud computing firm to become the centerpiece of this week’s Nvidia drama.

In his former role as White House AI adviser during the Biden administration, he drafted policy on preventing China from accessing top-performing US chips. What had drawn his attention then was simpler: Megaspeed, a Singapore-based firm that spun off from a Chinese company in 2023, had become Nvidia’s largest Southeast Asian buyer almost overnight.

Buchanan, who now teaches at Johns Hopkins University, is hoping to channel the attention into something he’s been advocating since 2021: stricter export controls on chips to China.

As his new book “The Bitter Struggle” was released this week, I spoke with him about where he stands on the debate and whether he sees a path forward for policy change. Our conversation has been edited for length and clarity.

Why has chipmaking become such a central national security issue?

Chips are the lifeblood of modern AI systems. AI is governed by what’s called the scaling law: the more computing power you put into training a system, the more powerful it becomes. When we came into the White House in 2021, we believed AI would be critical to national security, and that computing power would be critical to AI. That was the basis of our policy. Six years later, I’d argue that’s been borne out. AI is important for national security, and computing power is important for AI — there’s now bipartisan recognition of that.

In 2021 and 2022, we were making this argument without nearly as much evidence. We had strong conviction, and I thought the evidence pointed in the right direction, but talking about things like AI’s role in cyber operations back then was still somewhat speculative. Now it’s obviously the case that AI matters for cyber operations, and for national security more broadly. That makes the case considerably easier to make today than it was a few years ago.

Nvidia has been making the case that banning exports to China will just push China towards developing its own chips. What do you think?

China has been starting its chip indigenization efforts since 2015. When we came into the White House in 2021, it was already clear Chinese indigenization wasn’t going well. We thought we could make it harder by denying them access to additional chipmaking equipment, which we did, along with denying them chips in the interim. Our read wasn’t that China’s indigenization problem was a lack of focus, or even a lack of money — it’s that it’s incredibly hard to do, and they need Western equipment to do it, which we were in a position to deny them.

I have enormous respect for China’s engineering capability and scientific talent, and China has successfully indigenized other industries. But I think advanced chipmaking is the hardest thing we do as a species — and for the record, something the United States struggles with too. China needs key inputs, including advanced machinery from the Netherlands, the US, and Japan, to indigenize production. Without those inputs, they have a very long way to go to reach frontier chip production.

You don’t have to take my word for it: Secretary (of Commerce Howard) Lutnick testified in 2025 that China would produce only 200,000 chips, each substantially worse than its Nvidia counterpart. The Institute for Progress found roughly a 100-to-1 advantage for the US and its democratic allies in compute production versus China. There’s never been a single Chinese AI system of real power trained on Chinese chips. Chipmaking is hard, China can’t yet do it at the frontier, and export controls on manufacturing equipment can make it harder.

How do you think about the balance between creating the market for American businesses and national security?

Nvidia’s stock has gone up by a factor of 20 since the original chip controls in 2022. That’s because the chip supply coming off TSMC is basically fixed, while demand has grown tremendously — Nvidia’s margins are 70 to 80%. American companies, as (CEO) Jensen Huang himself has said, can’t get enough chips from Nvidia. Our conviction going in was that AI would matter enough to create huge demand, that Nvidia would do very well, and that the policy would still work if we simply said they couldn’t sell chips to China. That’s exactly what’s happened.

Export controls haven’t stopped China accessing chips through other countries. How far down the supply chain do you think enforcement needs to go?

We created what’s called the diffusion rule in 2024, specifically for this purpose, and it was finalized in 2025. It was meant to guard against smuggling and remote data center access through third countries like Singapore and Malaysia. President Trump has said he won’t enforce that rule. I think he should. [Editor’s note: The Trump administration rescinded the rule in May 2025, calling it overly complex and bad for American competitiveness.]

We were very worried about smuggling back then. We knew China would do everything it could to get access to chips, and that remote access to data centers was a significant risk. We wanted to get ahead of it. Under US export control law, Nvidia, as the exporter, has real responsibilities, alongside government enforcement. It’s pretty clear both failed here.

You didn’t restrict open-weight models while you were in government, even though most leading open-source models now come from China.

We understood the debate, the pluses and the minuses. Our view at the time was that the benefits of open-weight models outweighed the costs, so we didn’t constrain them at all.

Does the attention on Nvidia right now, combined with the midterm results, make you more optimistic about policy change?

I am. Republicans on the Hill have said pretty clearly they see real danger in shipping this technology to China, and they want to stop it. So yes, I think there’s a potential path toward stricter export controls.

There is a path, but I’ve given up making predictions about this administration. The clearest venues are the MATCH Act and the Outbound Act in Congress — both are reasonable bills that would force the administration to act on controls for chip manufacturing equipment and chips themselves. If anything happens, that’s probably where it starts.

NEW RESEARCH ON AI'S DUAL ROLE IN CYBERATTACKS

As AI-enabled cyberattacks make headlines, can AI help defenders just as much as attackers? Probably not in the near term, according to new research from the team led by Dawn Song, a computer science professor at UC Berkeley.

The asymmetry is structural: an attacker only needs one exploit to work, while a defender has to protect every system and deploy a patch everywhere it’s needed.

Song’s benchmark focuses on “one-day” attacks: once a patch is released, an attacker compares it to the prior version to find the fixed vulnerability, then builds an exploit for every system that hasn't been updated yet. This was possible before AI, but took real expertise and effort. In Song’s tests, AI generated working exploits straight from patches in about two hours, for under $30.

That speed makes slow patch deployment especially dangerous. In hospitals, for example, where testing and downtime risk slow things to a crawl, the mean deployment time is roughly 500 days. If AI can weaponize a patch within hours, the real question isn’t whether a fix exists but how fast anyone can install it.