Will Sour Ed Zitron Have His Bitter Revenge?

Written by Brook Schaaf

Zitron: the word means “lemon” in German, and I can’t get over the Dickensian suitability of his name as a “speaking name” that hints at his personality. He feels like a foreign, male counterpart to Tina Fey’s Liz Lemon on the TV show 30 Rock, who readily recognizes others’ dysfunctions, though not her own. And speaking of speaking, is Zitron ever verbose! Each week, he writes and speaks thousands of thousands of words about the Big Tech “rot economy,” most especially AI, toward which he is the doomiest doomer of them all. Not in the sense that it will kill us, but that we shall have to kill it because it’s impractical and unprofitable (bad combo).

He makes his case succinctly in Bloomberg: “People cannot measure how much an AI task actually costs.” Nor, he argues, “can you measure the return on investment.” Therefore, “lossy” OpenAI and Anthropic shouldn’t be allowed to go public. But, you say, what about Anthropic’s recent claim to Q2 profitability? Zitron makes a convincing argument that this is numbers fudging enabled by SpaceX’s discounted compute costs.  

You have by now seen the Uber and Microsoft examples of the Tokenpocalypse. We have caught glimpses of this at FMTC as well. For example, an AI service we use to parse program newsletters went from basically all-you-can-eat at $25 a month last year to about $400 a month after tokenization. (This is separate from AWS, ChatGPT, GitHub Copilot, and Claude costs.) And this is just for the portion that provides the data to humans, who still correct and confirm it in our work queues. So do costs go down overall? The jury is still out, though speed and thoroughness certainly go up. 

Will inference costs decrease enough in the future? I don’t have any Polymarket bets placed, but I admit the specter of vastly higher AI costs is a new monster in the backrooms of my mind. It does seem to me that if some queries cost a dollar, we’d still run them, but would that mean search rolls back to blue links? Somehow it’s hard to imagine. Perhaps some hybrid to balance costs? Even if costs are unsustainable, it feels like a different course won’t be taken for years because so much money is invested in the system.

And honestly, if costs do go up, it’s basically to our advantage because FMTC represents a cost efficiency for proofing and structuring data. Similarly, we recently reviewed a solution for partial processing of AI that would do about a third of the processing for two-thirds of the cost. It’s not entirely without a use case, but it’s certainly no slam dunk, at least not yet, if ever.

Zitron pooh-poohs comparisons to AWS and dot-com-era fiber infrastructure because there was either cash flow or durability (vs. burned-out GPUs). He predicts a “massacre” for an industry that has “absorbed over a trillion dollars in the last ten years” because hyperscalers are desperate for another hyperscale product, which may simply not be in the cards. (Cf. Meta’s VR ghost world.)

If he is correct, it will be a bitter disappointment for the world, but I suspect an oh-so-sweet vindication for him.