📈 Data to start your week

Start with the number that actually matters for anyone trying to figure out if AI is a business or a bonfire: quarterly AI revenue has, for the first time, edged past quarterly depreciation on all that GPU and data-center capex. That sounds like progress, and it is. But it's a narrow win. Companies are still nowhere close to covering the mountain of depreciation they've already racked up, let alone generating the margin cushion a normal business needs to survive a bad quarter. Think of it as finally paying this month's rent while still owing three years of back rent — better, not good.

While the model builders sweat over that math, the hardware side is quietly minting winners. SK Hynix just passed Samsung Electronics in market value for the first time ever, and the reason isn't mysterious: memory chips are the unglamorous backbone of every AI cluster being built right now, and Hynix has positioned itself as a primary supplier of the high-bandwidth memory that Nvidia's GPUs need to function. Samsung, historically the bigger, more diversified giant, is getting outrun in the one category investors currently care about most. It's a reminder that in gold rushes, the shovel sellers often outperform the miners.

Over in China, the AI labor market is telling a different but related story about speed versus seasoning. Chinese AI labs are staffing up with engineers who average just 1.6 years of experience, compared to 5.5 years at comparable US firms. That's not an accident — it reflects a strategy built around younger, cheaper, hungrier talent moving fast on well-defined problems, rather than assembling teams of veterans to solve harder, more open-ended ones. Whether that's a genuine edge or a bet that catches up with them later is the real question, but it explains a lot about the pace coming out of Chinese labs lately.

And then there's a data point that has nothing to do with AI directly but says something about how technology reshapes labor anyway: women who were not employed before starting GLP-1 weight-loss drugs are roughly 27 percentage points more likely to be working eighteen months later. That's a startling number for a drug class people mostly associate with vanity metrics and Ozempic jokes. It's a quiet reminder that the biggest labor-market shocks of this decade might not all come wearing an AI label.

Put together, these four numbers describe an economy in the middle of several overlapping bets — on AI infrastructure, on cheap fast talent, on biotech reshaping who shows up to work. None of them are settled yet, which is exactly why they're worth watching closely rather than believing the headline version of any one of them.