AI Intelligencer: Art of the chip deal

Reuters

14 hours ago

Aug 14 (Reuters) - (This was originally published in the Artificial Intelligencer newsletter, which is issued every Wednesday. Sign up here to learn about the latest breakthroughs in AI and tech.)
A week after OpenAI's long-anticipated GPT-5 launch, one thing is clear: newer isn't always better.
Despite benchmark-crushing scores and clear gains in coding and reasoning, many users say they miss chatting with GPT-4o, which OpenAI removed when the new model came out. The backlash was loud enough that the company restored 4o for Plus users within days.
The takeaway: even for OpenAI, the 'best model' isn't simply the biggest or the newest. It's about fitness. Different tasks—math, coding, writing, small talk—sometimes pull models in conflicting directions, and building one that excels at all of them is getting harder.
Strategically, GPT-5 looks like a shot at Anthropic's API business, especially in code generation. Its lower pricing undercuts Claude Opus 4.1 and could kick off a price war in a money-losing market. The good news is demand is booming—Sam Altman says OpenAI plans to double its compute power to train and serve the models in the next five months.
So why do some users prefer the older 4o? In this week's issue, we explore how 'self-affirming' bots might shape user loyalty—and mental health. We also dig into Washington's hottest AI debate: an 'export tax' proposal that could reshape how American AI companies sell advanced tech abroad. Scroll on.
(Email me here, opens new tab or follow me on LinkedIn, opens new tab to share any feedback and what you want to read about next in AI.)
In President Donald Trump's dealmaking world, few things seem off-limits—national security included.
He has cut a deal with Nvidia and AMD that would let them resume some exports of restricted AI chips to China in exchange for the U.S. government taking a 15% cut of those sales.
The White House has called the idea 'creative' and suggested it could extend to other firms, effectively creating a pay-to-export model for sensitive tech. It has also drawn bipartisan concern on Capitol Hill.
Historically, export controls have been non-negotiable national-security decisions: if certain technology is restricted, companies can't pay to get around the rules. The proposal raises a broader question as well: is it a form of state capitalism when U.S. companies must share revenue with Washington to access a foreign market?
Another chipmaker in negotiations with Trump is Intel CEO Lip-Bu Tan, and what's on the line is his job. After Reuters reported in April that Tan made hundreds of investments in Chinese companies over decades, Trump last week demanded Tan's immediate resignation, calling him 'highly conflicted.'
Tan, tasked with turning around the American icon, flew to Washington to meet Trump on Monday. For now, Tan appears to be staying, but expect more commitments from Intel to satisfy the White House.
Beijing is watching Trump's tactics—and responding. With trade talks extended again, everything from tariffs, rare earths, to AI chips and the fate of TikTok is back on the table. Behind the scenes, Chinese authorities have summoned domestic giants including Tencent, ByteDance, and Baidu over their Nvidia purchases, asking them to explain their needs and warning about information risks, according to sources who talked to my colleagues.
U.S. enforcement is tightening its grip on the flow of advanced chips to China, too. Reuters exclusively revealed that U.S. authorities have secretly placed location-tracking devices on targeted shipments of servers containing Nvidia and AMD chips to detect illegal diversion to destinations such as China. The aim is clear: curb the flow of the best AI hardware even as parts of the market reopen.
Bottom line: money might buy market access, but it doesn't buy trust. Even if a revenue-sharing scheme moves forward, the underlying security concerns—and the broader U.S.–China tech standoff—aren't going anywhere.
This chart from a report, opens new tab by Bessemer Venture Partners shows how fast AI startups can now climb to $100 million in annual revenue. The 'Supernova' cohort, like coding tool Cursor, rockets there in roughly a year and a half. 'Shooting Stars' take about four years. Both are hitting revenue milestones much faster than classic cloud companies, which needed closer to seven years.
AI products spread faster because they're easier to try and share, and they improve quickly once people start using them. But speed cuts both ways. The flashiest growth can be fragile. This helps explain the VC fanfare around AI to capture the next 'Supernova,' but the picking game is getting harder, as revenue growth is more likely to be driven by incentives, novelty or low switching costs.
As AI becomes part of our daily life, a mental health concern dubbed 'AI psychosis' is getting more attention. It's not new—delusions have always tracked the zeitgeist, from Cold War CIA paranoia to today's AI-fueled beliefs—but the rapid rise of chatbots like ChatGPT may be amplifying the risk, especially for people under stress from sleep loss or substance use.
One theory points to engagement-optimized models that act more sycophantic, reinforcing users' false beliefs instead of challenging them. San Francisco–based psychiatrist Keith Sakata said on X that he has seen a dozen hospitalizations linked to AI-related psychosis, a concern echoed by recent research on LLMs in mental health settings.
The appetite for self-affirming bots was on display when OpenAI deprecated GPT-4o after the rollout of its GPT-5 last week—and backlash pushed the company to reinstate it for Plus subscribers. As chatbots get more personal, Sakata argues tech companies face a hard trade-off: validate delusions for retention, or risk losing users by challenging them.