President Trump announced multibillion-dollar AI deals between U.S. companies and Gulf countries this week even as many in Washington continue to fear that China could gain backdoor access to advanced AI chips and worry that critical AI infrastructure could end up based outside the U.S.

The big picture: The president loves big deals with big numbers. Tech interests close to Trump want to see U.S. AI firms win global business. But China hawks in both parties distrust the Gulf states, which have close trade ties with China.

• The dealmaking flurry comes after the White House last week scrapped Biden administration rules limiting AI chip exports. A new policy is in the works but not yet in place.

Driving the news: The House Select Committee on the Chinese Communist Party on Thursday introduced new legislation "to stop advanced U.S. AI chips from falling into the hands of adversaries like the Chinese Communist Party (CCP)."

• The bipartisan Chip Security Act calls for the Commerce Department to require location verification for advanced AI chips and enforce mandatory reporting by chipmakers about any diversion of their chips to restricted companies or countries.

Earlier this week, as Trump and tech titans announced deal after deal, the Bureau of Industry and Security at the Commerce Department issued new guidance saying the department has to approve the deals.

• A Commerce Department official told Axios the guidance aims to prevent U.S. adversaries from using workarounds to access U.S.-made chips.

Catch up quick: Trump's trip to the Gulf has included a flood of AI-related deals.

• Nvidia said it will ship 18,000 of its advanced AI chips to Saudi Arabia for a 500 megawatt data center being built by Humain, a new AI infrastructure startup backed by the country's sovereign wealth fund.
• AMD — an Nvidia rival — also announced a $10 billion partnership with Humain, while Qualcomm also signed a memorandum of understanding with the startup.
• Amazon and Humain are investing more than $5 billion in developing an "AI zone" in Saudi Arabia with Amazon Web Services infrastructure and servers, training programs and AI agents that will be used by the kingdom's government.

The intrigue: The Trump administration is reportedly considering a deal that would allow more than 1 million advanced AI chips to be exported to the UAE.

• That could open the door for other companies to expand their capacity in the Middle East: Bloomberg reported this week that OpenAI is weighing opening a data center in the UAE.

Yes, but: The deal announcements don't come with many details. It's unclear if there will be restrictions on who can access the chips on site and what they can be used for.

• White House AI czar David Sacks this week in Saudi Arabia advocated for a "partner ecosystem" around AI and called for fewer restrictions on chips.

What they're saying: "President Trump and his administration remain committed to making the United States a global leader in AI, while also ensuring our most advanced technology does not fall into the hands of our adversaries," National Security Council spokesperson James Hewitt told Axios in an email.

• "These goals are not mutually exclusive, and we look forward to working with industry leaders to accomplish both these missions."

Flashback: An "AI diffusion" rule issued in the final days of the Biden administration placed restrictions on the export of chips to most countries, including Saudi Arabia and the UAE.

• Nvidia, Microsoft and others pushed back hard on the rule, arguing it would stifle innovation and economic growth.

State of play: The U.S. and China lead the world in AI infrastructure and deploying AI models, but "the Gulf states have massive ambitions to be the third leg of the stool," says Jimmy Goodrich, a senior adviser to RAND for technology analysis.

• "They have the ambition and the funds, and they see AI as central to their economic survival because as the world reduces its carbon footprint, their geopolitical leverage decreases."
• The Gulf deals "suggest a shift away from multilateral coordination toward more transactional, bilateral diffusion diplomacy," Joseph Hoefer, government relations and AI principal at Monument Advocacy, told Axios. "With the diffusion rule being rolled back, it signals a move toward partnerships that prioritize short-term strategic and economic gains."

The near-term concern for some in Washington is that China could access advanced U.S.-made AI chips through Saudi Arabia, the UAE and other third-party countries.

• The UAE and China remain close economic partners in other sectors and maintain military ties.

Scammers are using artificial intelligence to impersonate senior U.S. officials, the FBI warned Thursday.

Why it matters: Many of the targets have been current or former government officials and their contacts, the FBI said.

• "The malicious actors have sent text messages and AI-generated voice messages — techniques known as smishing and vishing, respectively — that claim to come from a senior U.S. official in an effort to establish rapport before gaining access to personal accounts," the FBI alert said.

Context: With seconds of audio, artificial intelligence can mimic a voice that is virtually indistinguishable from the original to the human ear.

• Scammers have weaponized voice cloning tech, and many products lack significant safeguards to prevent fraud or misuse.

Ina's thought bubble: It's another sign that voice cloning has become trivially easy, and that the era of deep fakes is here, not in the future.

• Security systems shouldn't rely on voice as a means of authentication — and all of us, government officials or grandmothers, should stop assuming that someone on the other end of a phone call is who they say they are.

Meta is pushing back plans for a public release of its largest Llama 4 AI model, known as "Behemoth," because of concerns that it may not be enough of an advance on previous models, per a new Wall Street Journal report.

Why it matters: It's another indicator that the AI industry's scaling strategy — "just make everything bigger" — could be hitting a wall.

Driving the news: The Journal says that Behemoth is now expected to be released in the fall or even later. It was originally scheduled to coincide with Meta's LlamaCon event last month, then later postponed till June.

• It's also possible the company could speed up a more limited Behemoth release.

The big picture: Meta is spending tens of billions of dollars on its AI projects, and till recently it had made fast progress catching up with rivals like Google and OpenAI.

Between the lines: Meta's disappointments mirror a broader worry inside the AI industry that progress dependent on scaling up models may be plateauing.

• OpenAI has faced well-documented hurdles in shipping an all-in-one next-generation large-scale model to follow GPT 4.0. Instead, the ChatGPT maker has rolled out a series of separate models, some specializing in reasoning and others in coding and technical work.
• Google and Anthropic have also reportedly encountered setbacks with their most recent efforts to train the largest models.

• Cohere, which specializes in enterprise AI, is falling 85% short of revenue goals. (The Information)
• Elon Musk's xAI said its Grok chatbot's brief fixation on the notion of "white genocide" in South Africa was the result of an "unauthorized modification" by what the chatbot itself called "a rogue employee." (X)

Giphy

A lawyer representing Anthropic in a case brought by the music industry had to apologize after Claude made up an article the company then cited in a legal brief.