Goldman Sachs Exchanges
2026-06-08 · Hosted by Allison Nathan · Goldman Sachs
Executive Summary
Goldman Sachs Exchanges hosts a multi-year debate between George Lee (co-head of the Goldman Sachs Global Institute) and Jim Covello (head of Global Equity Research) on whether the AI capex boom will pay off. Both agree consumer adoption has exceeded expectations and the technology has made incredible progress, but Covello remains as or more skeptical on the economics than two years ago, noting hyperscalers raised rather than cut capex despite underperforming stocks. Nearly all the economic value continues accruing to semiconductor companies at the expense of every layer above them, a dynamic the panel agrees cannot persist forever. Lee estimates $7–8 trillion will be spent and argues the payoff requires net-new economic activity, not just disruption of existing profit pools, while Covello reframes the entire question to one test: do enterprises make or save money implementing AI?
Key Stories & Changes
1. Two Years Later: Where the Skeptic Was Wrong
Covello revisited his original skeptical AI report with a new “two years later” piece, opening with where he was wrong
Consumer adoption of AI has been “magnificent,” much greater than he expected — George Lee predicted this accurately
Most consumers still use a free version of AI, so the core economic question remains the enterprise
He predicted that if hyperscaler stocks underperformed, companies would scale back capex; instead they raised capex despite underperformance, which he says calls the economics into question even more
The technology itself has made “incredible progress,” consistent with Lee’s predictions
2. Value Concentration in Semiconductors
All economic value has continued accruing to semiconductor companies — “incredible economic value”
Covello covered semis directly for 16 years; says in every prior cycle semis thrived when customers thrived, but in this cycle semis are thriving “at the economic expense of everybody above them in the chain”
This must eventually rectify: either upstream companies start generating profit, or semiconductor spending gets scaled back
Covello now favors hyperscaler stocks over semiconductor stocks (a change from two years ago), arguing hyperscalers outperform semis in two of three plausible scenarios
3. The Scale of the Payback Hill
Lee’s Goldman Sachs Global Institute paper sized the investment and how high the bar is for sufficient payback
Summing the opportunity profit pool by profit pool (e.g., advertising disruption) still falls short of justifying $7–8 trillion in spend
Lee argues the imperative is to create net new economic activity and new TAMs, consistent with past tech waves (agricultural, industrial, computer revolutions)
4. Coding as the Standout Use Case
Independent model companies’ summed revenue reached a level in ~3 years that took cloud companies 15–17 years to reach
Coding is the best application — a verifiable domain with strong product-market fit; agentic coding takeoff really occurred at the end of last year
Open question whether that product-market fit expands to other, less verifiable domains
5. The Enterprise ROI and Productivity Gap
“In a lot of ways, companies are losing more money today implementing this technology than two years ago”
Data often “isn’t ready to be agented” — agents deployed on unready data create economic challenges
Third-party surveys show a persistent gap between C-suite expectations and line-worker experience of productivity
Lee notes biggest productivity leaps appear at young, AI-native companies built from the jump for the technology; retrofitting legacy companies carries a “drag coefficient”
Trends Identified
1. FOMO-Driven, Game-Theoretic Capex
Covello attributes continued spending to “a tremendous amount of FOMO at every level of the supply chain” — enterprise, model layer, and hyperscaler. The game theory is symmetric: deploy and risk that margin advantages get competed away, or don’t deploy and accept a permanent margin disadvantage. This dynamic keeps spending running well ahead of demonstrated economics.
2. Scale as the Decisive Competitive Advantage
As models demand unplanned, non-revenue-generating outlays (e.g., a sudden $50 million model spend), AI is becoming “the cost of doing business” rather than a durable edge. This makes scale a massive advantage and shrinks the set of companies that can afford to compete, reinforcing concentration at the top of the economy.
3. Fleeting Advantage and Value Leakage
Lee raises that enterprise advantage from AI may be temporal — competitors catch up and margin gains get competed away — while surplus may “vanish into consumers’ pockets,” as in prior technology waves. The elusiveness of measuring ROI in hard terms adds to the uncertainty even for AI believers.
4. Rising Populist Resentment of AI
Lee flags a “somewhat sudden turn towards populist resentment of AI,” from booed commencement addresses to data-center-related violence, noting it is almost uniquely intense in the US. He links it to the midterm elections and warns it could contribute to slower US progress. —-
Sentiment Analysis
Overall Market Sentiment: Cautiously Skeptical
The dominant mood is a balanced, long-running debate: genuine optimism about the technology paired with hard skepticism about whether and when the economics justify the spend.
Risk Factors Highlighted
Economics never close the gap: Enterprises may not make or save enough money to justify $7–8 trillion in spend.
Capex rising despite weak returns: Hyperscalers raised capex even as stocks underperformed, increasing free-cash-flow strain.
Value stuck in semiconductors: Concentration of all profit in semis while upstream layers lose money is unsustainable.
Data not ready for agents: Deploying agents on unprepared enterprise data adds cost and undermines ROI.
Productivity gap: Line-worker productivity gains lag C-suite expectations across surveys.
Fleeting advantage / value leakage: AI margin gains may be competed away or pass to consumers.
A rough patch could trigger pullback: A market or economic downturn could cause hyperscalers to scale back spending.
Populist backlash: Rising US resentment of AI and data centers could slow deployment, especially around the midterms.
This episode was covered in today’s The Market Signal — 2026-06-08, a cross-source synthesis of multiple podcast reports.