Goldman's AI Debate: Jim Covello's Skepticism vs. Joseph Briggs' Long-Term Optimism

Key points

• The debate's protagonists: Jim Covello (Head of Global Equity Research at Goldman Sachs, a veteran analyst known for contrarian calls during the dot-com era) argues AI investment returns are deeply questionable. Joseph Briggs (co-lead of Goldman's global economics research) argues generative AI will drive a historic productivity transformation. Their exchange took place in Goldman's June 2024 report *Gen AI: Too Much Spend, Too Little Benefit?* and follow-up research through 2025–2026.
• Covello's 'cost paradox': Most transformative technologies replaced expensive solutions with cheaper ones (the internet, Amazon replacing physical bookstores, Uber). AI reverses this—replacing relatively cheap human labor with an extremely expensive technology. Estimated AI development/operating costs are around $1 trillion, yet no clearly cost-effective application has emerged to justify it.
• Evidence cited by Covello: In a Goldman internal test, an AI service that updated analyst spreadsheets saved ~20 minutes per company but cost six times more than doing it manually. He argues LLMs produce frequent 'illegible and nonsensical' output even on basic summarization tasks, and questions whether AI—built as statistical pattern matching—can solve the complex problems needed to justify its cost.
• 'Workslop' and weak enterprise ROI: An MIT survey found 95% of organizations saw no business return despite $30–40 billion in generative AI spending; only ~5% saw measurable P&L impact. Only 9.3% of US companies used generative AI in production in a given two-week period. Coding assistants and customer-service chatbots show efficiency gains that are largely offset by review, debugging, escalation, and quality costs.
• Structural profit imbalance: Most AI value accrues to the infrastructure layer (notably Nvidia, with 200%+ data-center revenue growth and 70%+ gross margins), while model companies (OpenAI, Anthropic) and application startups run large losses. Covello advises holding infrastructure beneficiaries even if skeptical about applications.
• Briggs' rebuttal: High early costs and weak returns are typical of the 'J-curve' phase of technology adoption. Electricity took ~40 years to show productivity effects, computing ~50 years, the internet ~10–15 years. Robert Solow's 1980s 'productivity paradox' complaint was eventually refuted by the 1990s IT boom. Cost declines from algorithmic efficiency, specialized chips (TPUs, ASICs), and scale economics could cut AI inference costs by 1–2 orders of magnitude within 5–10 years.
• Briggs' quantitative forecasts: Based on task-level analysis of 900+ US and 2,000+ euro-area occupations, generative AI can ultimately automate ~25% of work tasks (~30% in the US, ~20% in Europe). Projected impacts: ~15% cumulative US labor-productivity gain over 10 years (~1.5 percentage points annually), +6.1% US GDP, +7% global GDP (~$7 trillion annually), and ~$4.5 trillion in annual economic value.
• Labor-market effects: ~300 million full-time jobs globally are exposed to automation, but the eventual replacement rate is estimated at 6–7% (8–10% in the US). Younger tech workers (20–30) in junior programming, data, and content roles are most exposed. AI will also create new roles (AI engineering, workflow design, governance), though the 'natural sellers' labor-transition mechanism may be weaker than in past technological shifts.
• Covello's counter-arguments on the long term: No 'killer app' has emerged comparable to spreadsheets for PCs or browsers for the internet; competition erodes efficiency gains within 6–12 months (weak moats, fast diffusion via open-source models and APIs); and fundamental limits of LLMs (causal reasoning, creative synthesis, value judgment) may be structural rather than temporary.
• Bubble diagnosis: Goldman identifies a public-market 'earnings bubble' (Mag 7 forward P/E of 25–30x vs. ~50x at the 2000 peak—roughly a 46% discount) and a more severe private-market 'valuation bubble' (e.g., OpenAI at ~$150 billion valuation with ~$3–4 billion revenue and heavy losses). Hyperscaler capex (~$381 billion in 2025, +68% YoY, growth slowing toward ~25% by end-2026) is the key trigger: any cut in the 'arms race' could cascade.
• 2026 inflection signals: Mag 7 market-cap gains slowed from $4.8T (2023) and $5.4T (2024) to $3.5T (2025); MIT's 5% measurable-ROI finding; LLM capability plateau; and projected 165% growth in data-center power demand by 2030. Notably, Briggs and chief economist Jan Hatzius stated AI investment contributed 'essentially zero' to 2025 US GDP growth—attributed to GDP accounting (imported equipment offsetting investment) rather than a reversal of long-term optimism.
• The core disagreement restated: Briggs believes the AI J-curve will ultimately turn upward; Covello questions whether the J-curve exists at all or is steep enough to justify the investment. Covello warns 'the day of reckoning will come'—perhaps not this year or next, but eventually—while noting markets can sustain bubbles far longer than rational expectations suggest.