The AI Investment Boom: When Will It Pay Off?

Investment Stakes and Economic Reality

• Jim Cabello (Head of Global Equity Research) and George Lee (Co-head of Goldman Sachs Global Institute) agree that the AI investment landscape has moved further from realized returns over the last two years rather than closer.
• Cabello asserts that at some point, businesses must generate profit from AI investments; currently, "the stakes are higher" because the technology has advanced while the return on investment (ROI) remains unproven at scale.
• The hosts note that companies are spending "well in advance" of economic justification, driven by Fear of Missing Out (FOMO) across the entire supply chain.

Re-evaluation of Pre-2024 Forecasts (What Was Wrong)

• Cabello acknowledges being wrong about consumer adoption, which has exceeded expectations and is occurring significantly via free versions of AI.
• Cabello admits being wrong about hyperscaler behavior: when their stocks underperformed due to heavy capital expenditure (CapEx) and negative free cash flow, he predicted a CapEx reduction, but instead, hyperscalers have massively increased investment.
• Cabello confirms that the technology itself has progressed at a pace consistent with George Lee's earlier predictions, describing the progress as "incredible."

The "Semiconductor Squeeze" and Value Distribution

• A unique deviation from historical tech cycles (e.g., the internet era) is that economic value is currently accruing exclusively to semiconductor companies, while upstream layers (hyperscalers and software firms) are generating significant losses.
• Cabello warns that this dynamic "cannot go on forever" and must eventually rectify, either through upstream profitability or a scaling back of semiconductor spending.
• Cabello forecasts that hyperscaler stocks will outperform semiconductor stocks in two of three future scenarios:
  • Enterprises begin generating profits from AI, distributing value across the chain.
  • Hyperscalers moderate CapEx spending to recover free cash flow, even if not eliminating spending entirely.
• Semiconductors are projected to outperform only if the current status quo (only semis profiting) persists.

Enterprise Adoption and Productivity Gaps

• Enterprise adoption is slower than expected due to the need for a new technology stack, control planes for compliance, and probabilistic (vs. deterministic) infrastructure.
• Perception Gap: C-suite executives expect significant productivity gains, whereas line workers report little to no benefit; surveys confirm this divergence.
• Data Readiness: The "data isn't ready to be agented," causing economic challenges as companies deploy agents on top of unstructured or siloed data.
• Productivity Variance: Significant productivity leaps are observed in "AI-native" startups built for the technology, whereas legacy enterprises face a "drag coefficient" from retrofitting.

Forward-Looking Statements and Market Dynamics

• Scale as a Barrier to Entry: The necessity of spending approximately $50 million to deploy new models creates a defensive requirement, making scale an absolute advantage and potentially reducing the number of viable competitors.
• Competition of Margins: There is a risk that initial margin advantages will be competed away as adoption matures, or that surplus value may vanish into consumer pockets (lower prices) rather than corporate profits.
• Revenue Velocity: Generative AI model companies have reached revenue levels in three years that took cloud companies 15–17 years to achieve.
• Populist Backlash: George Lee notes a unique, sudden rise in populist resentment toward AI in the U.S. (e.g., protests against data centers), which could impede progress compared to other regions.
• Timeline: Cabello suggests that if the debate on profitability remains unresolved in two years, it may signal a fundamental challenge, as the "short term" cannot become the "long term" indefinitely.

Specific Challenges and "Blocking and Tackling"

• The market for "agentic" AI is nascent, with product-market fit for agentic coding only occurring at the end of the previous year.
• Significant technical hurdles remain regarding the "SLM vs. LLM dynamic," model orchestration, and the need for robust guardrails for powerful agents.
• Cabello advises investors to avoid blind assumptions and instead track specific markers: whether enterprises are making or saving money.

Event Details and Context

• Event: Goldman Sachs Exchanges episode exploring AI.
• Hosts: Alison Nathan, George Lee, and Guest Jim Cabello.
• Recorded Date: Tuesday, May 26, 2026.
• Duration of Dialogue: The speakers note this represents 3.5 years of continuous debate regarding AI economics and adoption.