Where Ed Zitron's AI-spending essay refuses to round its numbers
/Lorna Petrakis reads Ed Zitron's AI capex essay for craft tells rather than its conclusion, and finds the discipline in numbers that never soften into estimates.
TL;DR: Ed Zitron’s essay on Where’s Your Ed At argues hyperscaler AI spending, $1.3 trillion by the end of 2026 alone, is building a debt spiral rather than paying for itself. Lorna Petrakis isn’t here to referee that argument. She’s reading the piece the way she reads galley proofs after nineteen years at a wire service, for whether the specificity holds, whether the restraint cracks, whether a human actually sat with the source filings long enough to hate them properly. The numbers never round themselves into something comfortable, and neither does the prose.
Nineteen years reading galley proofs for a wire service teaches you to stop reading for the argument first. You read for the hand. Ed Zitron’s essay on Where’s Your Ed At makes a case about AI capital spending that plenty of people will fight over on the merits. What I want to talk about is what the sentences are doing while that argument happens, because that’s where you catch whether a human sat with the material long enough to actually hate it.
The tell people miss with financial writing that might be AI-assisted sits in the rounding, not the vocabulary. Machine-smoothed prose tends to flatten numbers into something rhetorically clean: over a trillion dollars, nearly half, the vast majority. Zitron won’t do it. Hyperscaler AI capex sits at $1.3 trillion through the end of 2026, not “over a trillion.” Capex eats 24.4 to 43.4 percent of those companies’ revenue, a range, not a summary. Amazon’s July bond sale pulled 1.6 times demand against a normal 4 times for investment-grade debt, two decimal-adjacent numbers sitting side by side, daring a reader to notice the gap. Somebody pulled the filings for that and refused to let the prose flatten what the filings actually said.
The second tell is structural, and it took me two reads to catch because on the first pass it looks like a stylistic quirk. Every section of the piece builds the same way: a hyperscaler buys more chips, memory suppliers raise prices past 60 percent, servers cost more, more debt gets raised, bond spreads widen, the next round costs more than the last. He runs that loop four separate times against four different pieces of the industry, GPUs, memory, the CoreWeave financing, the Oracle build-out, and never once shortcuts it back to a summary line. He makes the reader sit through the mechanism again each time, with new numbers plugged in. A CoreWeave that raised $23 billion in two years and lost $740 million last quarter gets walked through the same wheel Oracle gets walked through: downgraded by S&P, sitting on 5.6 percent of the capacity it’s contracted to deliver, cutting 21,000 jobs while building it anyway.
That kind of repetition reads like a habit until you count how many times it holds under pressure without once cracking into shorthand. He had a hundred chances to compress. He made his readers earn the pattern by living through it four times instead of being told about it once.
The other place a human hand shows is in what doesn’t get explained. The NVIDIA-SoftBank-OpenAI financing chain, $250 billion committed to a subsidiary that has never built a data center, gets one historical comparison: the 1998 Lucent-Winstar deal, which cost Lucent $488 million when its own circular financing collapsed. Then Zitron moves on. He trusts the comparison to do its work rather than spelling out the moral underneath it. A summarizer explains the analogy. A writer drops it and walks.
None of this settles whether Zitron is correct that the whole structure amounts to what he calls kayfabe, wrestling’s term for a worked outcome everyone quietly agrees to perform as real. Fitch’s own language, quoted in the piece, hedges harder than he does anywhere: “medium- and long-term potential of the underlying technology is highly uncertain.” That’s a ratings agency covering itself. Zitron skips the hedge, but he earned the right to, because he built the case number by number before he let himself reach for the word.
Nineteen years of proofs don’t train you to check for agreement. They train you to check whether the specificity is load-bearing or decorative, whether the discipline holds when a writer clearly wants to swing harder than the sentence allows, whether the piece would survive someone checking every number against the source filing. This one, on the sourcing alone, does the work. Whether the argument survives contact with 2027’s actual numbers is a different question, and the prose can’t answer that one for him.
Comments are open to members. Sign up free →