Over the last week, I caught the recent Exchanges at Goldman Sachs podcast - The AI Boom: When Will it Pay Off?

It's a follow-up debate two years on from GS Research's Covello's original sceptical note, framed around a new report he calls a "mark to market" two years later. The podcast is a conversation with GS's Alison Nathan (host) and George Lee.

I've posted this now as an observation following the reaction on Thursday/Friday on the NASDAQ to Broadcom's numbers and the US employment numbers. The simply huge magnitude of the capex going into AI means that different parts of the value chain are (from time to time) going to experience big valuation "adjustments" regardless of how transformational AI proves to be and the timescale over which that happens It is simply inconceivable to me that such vast sums of capex can be deployed with perfect foresight as to where in the value chain the returns will arise and when that happens. As well as the winners, there are going to be plenty of losers investing in this journey.

The summary that follows was produced by Claude.ai from the podcast transcript. But I recommend listening to the whole

The AI investment landscape: key facts cited

The central tension is that capex has accelerated rather than moderated. Covello had predicted that if hyperscaler stocks underperformed (which they have, given the free-cash-flow hit from spending), the companies would pull back; instead they raised capex, which he thinks sharpens the economic question. Lee frames the eventual payback bar against roughly $7–8 trillion of projected spend, arguing that summing up disruption of existing profit pools alone wouldn't justify it.

The most striking structural fact both return to: essentially all the economic value is accruing to the semiconductor layer, while the players above them in the chain are absorbing large losses. Covello (a 16-year semis analyst) says this is historically unusual — normally chipmakers thrive when their customers thrive, whereas here they're thriving at the expense of everyone upstream.

Other facts highlighted: consumer adoption has far exceeded Covello's expectations (a point he credits Lee for calling correctly), but most consumers are on free tiers, so the real economic test is enterprise. Enterprise adoption has been slower than hoped. The independent model companies are described as the fastest-growing companies in US history on the top line — Lee notes they've reached revenue levels in ~3 years that took cloud companies ~15–17 years — though the profits still flow to semis. Coding is flagged as the standout application because it's a verifiable domain, with agentic coding only really hitting product-market fit around late last year. A recurring blocker is data readiness ("putting agents on top of data that isn't ready to be agented"), plus orchestration and SLM-vs-LLM questions. They also discuss a measured C-suite-versus-line-worker productivity gap (surveys consistently show frontline gains lagging executive expectations) and rising US-specific populist hostility to AI as a potential drag.

Jim Covello: the sceptic on ROI

He's as or more sceptical on the economics than two years ago, despite acknowledging the technology's progress has been genuinely impressive. His core test: do enterprises actually make or save money deploying AI? If yes, the technology fulfils its promise; if the same "it's early" debate is being had in another two years, that's a problem, because the short term eventually has to become the long term and upstream companies can't lose money forever. He notes the hill is now steeper precisely because more has been spent. Tactically, he's shifted to favouring hyperscaler stocks over semiconductor stocks — he outlines three scenarios (enterprises start generating returns; hyperscalers modestly moderate capex to reclaim free cash flow; or status quo persists) and sees hyperscalers outperforming in two of the three. He also notes scale is becoming a decisive advantage, as periodic surprise model spend becomes "more defence than offence" that smaller firms can't afford.

George Lee: the optimist on ROI

Lee agrees the payback hill is high but believes the real prize isn't disrupting existing profit pools — it's net-new economic activity and entirely new TAMs, consistent with how prior technology waves played out. He's optimistic about enterprise transformation but on a longer timeline, stressing the technology is paradigmatically novel (probabilistic vs deterministic), needs a new deployment stack and control planes, and that agentic tools are newer still. He raises two ROI complications that genuinely cut against his own optimism: first-mover advantages may prove fleeting as competitors catch up and margin gains get competed away, and surplus may ultimately flow to consumers rather than be captured by the deploying firm — making ROI measurement "elusive." He points to AI-native startups as evidence of extraordinary productivity gains that incumbents, weighed down by legacy workflows, haven't yet realised. The offsetting force keeping spend high is game-theoretic: not investing risks a permanent margin disadvantage.

Where they land

Nathan notes there's not much daylight between them. Both agree the number of unanswered questions is large and the debate will run for years. The neat framing: Covello is good at accurately pricing where the technology is today, while Lee's role is to look further downfield, which requires a leap of technological and economic faith — and Covello's caution is that investors should neither assume it definitely will nor definitely won't pay off, but watch the specific markers along the way.

(mikebrisy note: I found the Covello's references to upstream and downstream confusing. I assumed downstream is closer to the customer, so the order is: chips, hardware, LLMs, applications incl. agents, enterprise, customer,... whereas I think he is refering to layers in the technology stack, where I'm not used to hearing the terms upstream and downstream. It confused me for a moment, but it is clear if you listen to the podcast.)