TL;DR: Uber's president said over the weekend that the company is questioning the ROI on its AI spending after burning through its entire token budget in the first four months of 2026. It’s an early sign of a potential reckoning around the aggressive AI adoption we’ve seen in the corporate world. The shift in attitude shows that AI use for the sake of it isn’t necessarily a magic wand for boosting overall company performance.

What happened: The company’s president, Andrew Macdonald, admitted on a podcast that it’s unclear whether its devs’ extensive use of Claude Code is actually leading to useful features for its consumers, saying in an interview that the “link is not there yet.”

It’s a notable change for the company, which has heavily encouraged AI use among its employees. Uber rolled out Claude Code for internal use last December, and now 95% of its engineers use AI tools monthly, with roughly 70% of committed code originating from those tools.

Tokenmaxxing falls flat: Uber is one of several tech companies that set up internal leaderboards pushing employees to compete on using the most AI—the logic apparently being that sending more AI prompts means becoming more efficient or producing more work. CEO Dara Khosrowshahi even said on an earnings call earlier this month that the company would slow hiring to put more money into AI investments (though Macdonald’s recent comments indicate that such a trade-off might be “harder to justify” if the AI payoff isn’t obvious).

Agentic AI tasks like those the Uber engineers are doing are more token-hungry, which in turn is driving up how much (and how quickly) companies spend on AI. The revelation in April that the company had already used up its AI budget for the year was, per Macdonald, a “head-exploding moment” that’s led to a recalibration on whether its all-in AI strategy is justified.

So what’s the deal?: Uber’s retreat from tokenmaxxing might look like an early warning that the corporate world is cooling on AI adoption. But here’s one point of evidence that demand overall remains hot: GPU rental prices have more than doubled since January, according to data compiled by AI compute investment firm AMP PBC. In fact, thanks to high demand, compute costs are now eating up a bigger chunk of major AI labs’ overall expenses. Both OpenAI and Anthropic have introduced higher-priced subscription tiers and separate API pricing for agentic AI this year as a result—which could, in part, be driving up costs for companies such as Uber.

Bottom line: Firms like Uber aren’t abandoning AI as a key workplace tool—but they are starting to take a more measured tack on using it, recognizing that an engineer who keeps their AI agent humming around the clock isn’t necessarily building something anyone wants to use. —WK

• Japan’s new jet engine could fly you from Los Angeles to Tokyo in two hours. In a step toward making hypersonic travel a reality, Japan’s Aerospace Exploration Agency (the country’s counterpart to NASA) and a coalition of Japanese universities recently completed a successful ground combustion trial of an engine designed for a Mach-5 flight—a speed five times faster than the speed of sound. At that speed and an altitude of 82,000 feet (the target to make flights shorter), the heat generated exceeds 1,000 degrees Celsius. That’s where the breakthrough comes in: JAXA’s engineered thermal protection system held up under heat, an obstacle the US Air Force has struggled to overcome. If all goes according to plan, commercial hypersonic passenger services—like the Tokyo to LA flight in two hours instead of 10—could be possible by the 2040s.
• A one-time gene edit might be able to replace a lifetime of cholesterol meds. A single infusion of an experimental gene-editing treatment could dramatically lower cholesterol levels, perhaps permanently, researchers in a new preliminary study found. In the small 35-patient trial, one dose cut cholesterol levels by up to 62%. It did this by changing a single DNA letter in the gene responsible for regulating blood cholesterol levels, disabling it and forcing the liver to continuously pull more cholesterol from the bloodstream. A subgroup of patients appeared to sustain the lower levels after 18 months. Researchers hope the findings could lead to a quicker way to prevent cardiovascular disease, an illness that kills about 800,000 Americans a year. —LC

Forget using AI to curate playlists tailored to your tastes—now you can just generate an entire discography of tunes that sound exactly the way you want. Some people are doing just that and exclusively listening to their own AI music. One user on the subreddit for AI music tool Suno described their creations as “album after album of bangers,” according to the Verge, while another said listening to AI music they’d generated was now an "infectious addiction.”

The Verge reached out to several of these Suno users to ask why they put their own songs on repeat, but no one wanted to talk. Churning out an AI discography based on a quick prompt is probably easier than trawling through questionable recs from your friend who loves sending you hyperspecific playlists. But music has always been a back-and-forth exchange between humans—with “wait, I love that song, too” being a time-honored way of bonding. If an AI song drops in the forest and no one but you hears it, does it actually make a sound? —WK

Chaos Brewing Meter: /5