InvestAnswers: AI capex of $835B dwarfed by $2.4T revenue backlog, trade "not over"

The bear case: credit stress, backlash and a sand-in-the-gears IPO delay

James opens with the bear argument, citing Bank of America warning of "signs of boom and bust" reminiscent of the dot-com era, calling it "classic capital cycle risk." He notes data centers are now, per a Tech Radar stat he cites, more unpopular in the US than the Iran war, with "nearly 70% of the US population" opposed, which he attributes to organized FUD rather than organic opposition.

He flags physical bottlenecks — power hookups, permits, skilled trades, land — as the real constraint, not chips, and says plans to add "43 gigawatt of new AI compute next year" face real difficulty. He points to the trade credit market getting "completely smashed," naming Oracle and CoreWeave as names he has personally doubted for 18 months, with CDS spreads at "record-wide spreads." The Fed raising rates 25 basis points and the 10-year Treasury nearing 6% is, he says, raising the discount rate and making funding harder.

On the news front, James says Anthropic is "weighing delaying their IPO beyond November" with its valuation "rumored down to 1.5 trillion from 2 trillion." He tallies big-four 2026 capex at roughly $735-750 billion, rising to $835 billion once Oracle's $100 billion is added, asking: "How are they going to pay for it? Can they even pay for it?"

The bull case: inference margins, sold-out compute and a $2.4 trillion backlog

Turning to the bull side, James leans on the Jevons paradox argument — cheaper intelligence drives more demand — and cites training costs falling from $20 million for the first GPT model to "100 bucks today." He cites Jensen Huang's breakdown of Nvidia's GPU market: a third classic computing, a third frontier labs, and a third spread across startups, neoclouds and "sovereign AI" players like Palantir.

He highlights inference economics, saying margins for frontier models have risen "from 40% to over 70% just in a few months despite prices falling," and that eventually "99% of all compute will be inference." Cloud revenue backs this up: Azure up 43%, Google Cloud up 82% year-to-date, Amazon up 37%. He also says compute "out 2027 or 2028" is already sold out.

James cites a planned $3-4 trillion per year AI infrastructure spend by 2030, and says Nvidia expects to "double growth next year" with "infinite demand," while rumoring that Elon Musk alone has locked up "40% of the supply" of Vera Rubin chips, with units reportedly trading at "5 million each" on the black market versus a $2 million retail price.

Will they get to 2.7 gawatts by year end? Elon thinks so.

— James

Verdict: capex covered almost three times by backlog, casualties expected along the way

James walks through OpenAI and Anthropic's growth, noting Anthropic's quarterly revenue moving "4.7 billion in Q1, 11.6 billion in Q2... 16.8 billion" in Q3, annualizing toward "110 billion," while OpenAI could reach "70 billion annualized by year end." He is skeptical of a semianalysis estimate that combined revenue could hit "over 500 billion in the next 12 months," saying "I do not see that happening."

He also flags a memory crunch, saying Michael Dell expects it to worsen and that Google is reportedly trying to secure high-bandwidth memory supply "out to 2035." On financing, James concludes most big players — Microsoft, Alphabet, Meta — remain free-cash-flow positive and can fund capex, with only Amazon "slightly negative" for one quarter.

His final tally: against roughly $835 billion in 2026 capex, bond issuance up 79%, and weak CDS on Oracle and CoreWeave, he sets a $2.4 trillion revenue backlog, 82% cloud growth, and planned $3-4 trillion annual infrastructure spend by 2030. His conclusion: "Are we going to go bust? No. Are we going to go boom? Yes. Will there be casualties along the way? Yes."