I guess I have to mention Oracle.

Right? I mean, you can’t harp on potential AI profitability disappointments and then fail to mention a story which hints at just that.

Oracle, you’re reminded, grabbed every market headline early last month when the shares stormed 36% higher in a single session thanks to a superlative-laden earnings release which showed rapidly expanding demand for the company’s cloud infrastructure and teased a “large upward revision” to the financial roadmap.

The shares have gone nowhere since that bonanza, but they’re up 70% YTD thanks in large part to the September 10 surge. A month on from the stock’s best day in 30 years and The Information had the audacity to suggest that renting out Nvidia chips isn’t necessarily a one-way ticket to easy profits.

The result was a selloff which at one juncture neared 7%. The stock later trimmed losses.

The linked article cited internal documents, which I can only assume means the numbers are accurate. Those numbers show Oracle generated $900 million in revenue from renting Nvidia-powered servers in the three-month period through August, but only $125 million in gross profit.

Is that problematic? Well, you’d have to ask someone who covers Oracle for a living. It’s during times like these when company analysts actually are useful. But I’ll note that Oracle’s gross margin was nearly 70% in fiscal Q1. $125 million on $900 million in sales works out to… well, a helluva lot less than 70%, let’s just put it that way.

Maybe (probably) that’s apples to oranges. Or doesn’t paint an accurate picture. Or focuses too narrowly on one new line of business. Or is just plain stupid. I really don’t know. But I do know that generally speaking, it’s not great if margins are thin in the business where the growth’s concentrated.

That said, you could argue that stupendous growth compensates for lower margins. If you’re only selling a few, your margins need to be very high, whereas if you’re selling a bunch, maybe you can make up for thin margins with volume. But when taken to extremes, that argument risks venturing into “First CityWide Change Bank” territory. If you’re unfamiliar, that means you’re too young to remember late-1980s SNL. Here’s the reference:

You’re welcome. And yeah, that’s how good SNL used to be.

The point, obviously, is that you can have as much “volume” (sales) as you want, but if you’re not keeping anything (enough) from it, then the model makes no sense.

Already, the AI buildup’s costing Oracle margin. As Bloomberg noted, 67.3%, the company’s gross margin last quarter, was “the lowest in more than a year,” and The Information article said Oracle actually lost $100 million from renting out servers powered by Jensen Huang’s next generation Blackwell chip.

I don’t think this is a big deal, but at a time when bubble whispers aren’t whispers anymore, and amid a transition from the “show me” revenue phase of the cycle to the “show me” profit phase, this is well worth a mention, even if it doesn’t turn out to be a canary or any kind of harbinger.

Meanwhile — and relatedly — the self-referential character of sundry strategic tie-ups like last month’s OpenAI-Nvidia deal and this week’s AMD-OpenAI initiative is getting the wrong kind of attention in the most mainstream of all mainstream press.

The title of Tuesday’s DealBook mailer (from Andrew Ross Sorkin) was “The circular economy.” He mentioned “the vendor financing-ish quality” of Nvidia’s pledged investment in OpenAI on the way to telling his many millions of followers that the AMD deal “also has a circular element to it.”