Dual Perspectives on AI's Economic Impact: Jim Covello vs Joseph Briggs
1. Profiles and Core Positions
Jim Covello — Head of Global Equity Research at Goldman Sachs, a 30+ year veteran who rose from semiconductor equipment analyst. He is widely characterized as Wall Street's leading AI skeptic. While he recommends continued exposure to AI infrastructure beneficiaries (e.g., NVIDIA, utilities), he is highly cautious on AI application-layer profitability. His June 2024 report *Gen AI: Too Much Spend, Too Little Benefit?* became one of Goldman's most-read publications in 12 years.
Joseph Briggs — joint-lead of global economics investment research at Goldman, a macroeconomic modeler specializing in quantifying structural shifts. His March 2023 report with Devesh Kodnani, *The Potentially Large Effects of Artificial Intelligence on Economic Growth*, decomposed 900+ US occupations and 2,000+ euro-area occupations into work tasks, providing the empirical foundation for downstream AI macro forecasts.
2. The Core Disagreement: Near-Term Economic Viability
#### 2.1 Covello's "Cost Paradox"
Covello's central thesis: current AI is using an expensive solution to replace relatively cheap labor — the inverse of every successful prior technology revolution. Internet-era companies like Amazon and Uber started with structural cost advantages; AI shows the opposite dynamic.
Goldman internal benchmark case: AI-automated spreadsheet updates for analyst workflows saved ~20 minutes per company but cost 6x more than manual processing — even with Goldman's scale economies. Covello reports similar cost-benefit failures in coding assistance (productivity can actually drop 19% when developers over-rely on AI beyond their competence) and customer service chatbots (escalation friction negates savings).
He argues LLMs are not designed for complex problem-solving — even basic summarization frequently produces "indecipherable and meaningless" results — and that NVIDIA's near-monopoly on AI GPUs removes the competitive price discipline that drove Moore's Law cost declines.
#### 2.2 "Workslop" and the Application Layer
MIT's 2025 research finds that despite $30–40B in generative AI spend, 95% of organizations captured no measurable business return. Only ~9.3% of US firms use generative AI in production; only ~5% see measurable P&L impact. Profit is highly concentrated at the chip layer (NVIDIA revenue +200%+ in 2023–2024, gross margins 70%+) while model and application layers report universal losses and rising "acquihire" activity.
#### 2.3 Briggs' J-Curve Rebuttal
Briggs counters with historical adoption curves (electricity ~40-year lag to measurable productivity, computers ~50 years, internet ~10–15 years), invoking the Solow "productivity paradox" precedent. His structural pillars:
- Task automation scale: ~25% of work tasks eventually automatable (US up to 30%, Europe ~20%)
- Cost decline mechanics: algorithmic efficiency, hardware specialization (TPU/NPU/ASIC), software optimization, and competitive open-source pressure could drive inference costs down 1–2 orders of magnitude over 5–10 years
- Complementary investment lag: enterprise process redesign, training, and data accumulation take time — the current low-return phase is the bottom of the J curve, not its terminus
- Valuations: Mag 7 forward P/E 25–30x vs. ~50x at the 2000 dot-com peak — a ~46% discount, but earnings-bubble mechanics already visible (NVIDIA stock flat in late 2024–2025 despite further +37% upward revisions to forward earnings)
- Adoption: Only 5% of firms report measurable P&L impact
- Capex: Hyperscaler capex growth decelerating from +75% (Q3 2025) → +49% (Q4 2025) → +25% by end-2026 guidance
- Physics: Data center electricity demand projected +165% by 2030
3. Long-Term Projections and Labor Market Impact
Briggs' quantitative framework:
| Metric | Projection | |---|---| | Global annual value creation | ~$4.5T | | US labor productivity lift over 10 years | ~15% (annualized ~1.5 pp) | | US GDP uplift | ~6.1% cumulative | | Global annual GDP uplift | ~7% (~$7T) | | Global workforce exposure | ~300M full-time jobs | | Actual replacement rate | 6–7% globally (US 8–10%, eurozone 6–8%, China 4–6%, India 3–5%) |
A 2025 finding flags disproportionate impact on workers aged 20–30 in AI-exposed entry-level roles (junior coding, data analysis, content creation, customer service), described by Goldman's George Lee as the cohort "a bit being sacrificed" during restructuring.
Covello's rebuttals center on: absence of a killer app, fast commoditization eroding 6–12 month competitive moats, and a fundamental ceiling on LLM-based architectures for causal reasoning and complex problem-solving.
4. The 2026 Inflection Point
Multiple converging signals:
5. Dual Bubble Structure
| Bubble type | Signal | |---|---| | Public-market earnings bubble | Stock prices decoupled from fundamentals; NVIDIA flat despite +37% EPS revisions | | Private-market valuation bubble | OpenAI ~$150B (~40–50x revenue), Anthropic ~$18B (~36x); rising acquihires (Stability AI, Inflection, Adept) |
Covello's defensive playbook: stay in infrastructure beneficiaries where revenue has already been booked; avoid application-layer exposure. Briggs' constructive stance: history favors long-term holders, and pullbacks create better entry points rather than signaling bubble collapse.
6. Comparative Valuation vs. Dot-Com Era
| Dimension | AI 2025 | Dot-com 1999–2000 | |---|---|---| | Mag 7 forward P/E | 25–30x | ~50x | | PEG ratio | ~1.6x | ~3.7x | | Implied long-term growth | ~10% | ~16% | | IPO activity (YTD) | 51 | 388 | | Current capex / operating cash flow | ~85% | Lower | | Today leaders' balance sheets | Strong cash flow, buybacks | Fragile, cash-burning |
The 46% valuation discount vs. 2000 is significant — but the comparison cuts both ways: today's leaders are financially more robust, while the tech architecture may lack the cost logic that ultimately validated internet survivors.