|
|
|
|
 |
French president Emmanuel Macron speaks via video link at the Raise AI Summit in Paris on July 8, 2026.
Nathan Laine—Bloomberg/Getty Images |
|
|
|
| Welcome back to In the Loop, TIME’s twice-weekly newsletter about the world of AI. If you want to help us grow, why not forward this newsletter to your friends? |
|
|
What to Know: AI is booming, but no company is safe
|
|
|
|
|
| On the sidelines of a massive AI conference in Paris this week, an executive gave me a glimpse of the industry’s future. |
| He told me he spends much of his day in the office chatting with an AI agent through a tiny microphone pinned to his T-shirt—one that triages his emails, preps him for one-on-ones, and helps him analyze new opportunities. |
| “You get hooked on these things very quickly,” says Jacob Wallenberg, who leads international expansion for the fintech company Ramp. “It’s completely changed the way that I work in the past eight months or so.” |
| For me, the most telling detail wasn’t the AI coworker—versions of which are becoming commonplace across the industry. Instead, it was the taxi-meter-style ticker that Ramp has built into it, which displays in real time the dollar cost of every request an employee makes. “Do I need the frontier model?” Wallenberg now asks himself before he makes each request, “or do I need something a bit more cost-efficient here?” |
| That taxi meter is a pretty good metaphor for the state of AI in July 2026. |
| Almost every conversation I had at the RAISE Summit this week, which was held underneath the famous Louvre Museum in central Paris, seemed to show that the AI industry’s center of gravity has shifted from the training of AI to the running of it. And everywhere I looked, it seemed there was a new narrative about why even the biggest companies in the world are ripe for disruption. |
| The rise of open-source AI— Earlier this year, there was a brief hysteria around “token-maxxing,” or the idea that companies should be spending as much as they could afford on running frontier AI models made by the likes of Anthropic and OpenAI. Michele Catasta, president of the vibe-coding platform Replit, told me this era has come to a close. AI firms, including his own, are now devising clever ways to use cheaper open-source models whenever they can, he says, while using frontier models for more complex tasks, like marshaling swarms of agents. “We are experiencing a U-turn” against token-maxxing, Catasta says. “We’re outcome-maxxing as a company—we try to be as optimal as possible in terms of token usage.” |
| Meanwhile, in the first half of this year, the capabilities of open-source models have reached a level where they can be relied upon for many rote tasks, says Vipul Ved Prakash, CEO of Together AI, a cloud that serves these models to customers. “What our customers are seeing is that open-source models can be 10 to 60 times more cost-effective [than frontier models],” he says. |
| Business is still booming for OpenAI and Anthropic, to be sure, but these coalescing trends may be putting downward pressure on their revenues at precisely the time when each is preparing to go public at near-trillion-dollar valuations. “The most popular workloads that are deployed in the largest number of organizations will get commoditized by open models,” Prakash tells me. “I do not think that closed labs will be able to charge a premium for these workloads.” |
| The chip insurgents— OpenAI and Anthropic aren’t the only companies in the crosshairs. |
| I spoke to several chipmakers in Paris who are taking aim at Nvidia, betting that the world’s largest company—which makes the vast majority of the AI chips on the market today—isn’t well-placed to run models quickly. |
| “Nvidia and all the GPU makers are in an architectural cul-de-sac for speed; they can’t get out,” says Andrew Feldman, CEO of Cerebras. His company’s chips don’t rely on the same scarce type of memory as Nvidia’s, which is currently ballooning in price. Feldman claims Cerebras’ chips can perform AI inference up to 20 times faster than those made by Nvidia—which he suggested had been all but forced into spending $20 billion to license the technology of a Cerebras rival, Groq. “Of course inference is a bigger market than training,” he says. “Training is how we make AI and inference is how we use AI. You train a model and then tens of millions, maybe billions of people use it.” |
| The neoclouds— You couldn’t escape the rise of the neoclouds at this conference, either. Iren was a key sponsor, with its logo plastered everywhere; others, like Coreweave, Nscale, Crusoe, and Nebius, were out in force. |
| As their name suggests, neoclouds are taking aim at the huge cloud giants: Amazon, Microsoft, and Google. (Virtually all the neoclouds benefit from Nvidia’s patronage, as the chip giant seeks to diversify its customer base.) |
| Just like the other upstarts I met, the neoclouds argue that the massive businesses who currently dominate their industry were not built specifically for AI—and are thus ripe for disruption by an AI-first competitor. The neoclouds are optimizing for one thing: “operating and running AI models as efficiently and as fast as we can,” in the words of Ben Richardson, Coreweave’s vice president of strategy. |
| But last week, it seemed that it was the neoclouds themselves who were being disrupted, when the news that Meta was considering moving into the AI cloud business sent Coreweave stock tumbling 14%. Richardson, when asked, put a brave face on it. “I think it’s a vindication of what we’re doing,” he said. “The fact that we are building something and large organizations are actually looking at us and saying: that’s a great idea, maybe we should do some of that.” |
| What it all means— I left Paris with two impressions. The first was no surprise: the AI industry is experiencing an unprecedented boomtime, with companies big and small reaping the rewards. My second impression was more novel, at least for me: that although the biggest companies appear more dominant than ever before, competition is flourishing, too. |
| For every trillion-dollar company, there are dozens of startups that believe they’ve identified its key vulnerability. All of these competitive forces are partly the reason that a given level of AI intelligence is becoming radically cheaper to access over time. They might also explain why even the world’s top billionaires go to sleep with one eye open. As the former CEO of Intel Andy Grove famously put it, in business, “only the paranoid survive.” |
| If you have a minute, please take our quick survey to help us better understand who you are and which AI topics interest you most. |
|
|
|
|
|
|
|
|
|
|
| OpenAI launched ChatGPT Work, a competitor desktop app to Claude Cowork, on Friday, in an effort to further put its tools into the hands of desk workers. The app, which OpenAI said was powered by its new GPT 5.6 models, “gathers context, plans the approach, and takes action across your tools, files, and desktop apps to create polished spreadsheets, docs, and slides,” according to OpenAI’s website. |
|
|
| What if It’s Not the Phones? by Kaitlyn Tiffany in The Atlantic |
| A counterintuitive piece in the Atlantic about a new book by Peter Gray, which argues that well-meaning crackdowns on children’s usage of the internet may be doing more harm than good. “Children’s need for unstructured play and exploration (guided by some safety rules and common sense) applies not just to vacant lots and city parks and backyards in the suburbs, [Gray] says, but to other settings too. It now extends to the wild spaces of the internet,” Kaitlyn Tiffany writes. “Kids should be free to play without their parents’ supervision, Gray insists, even when they go online.” |
|
|
Opava .City™️ 
Diskuze