The honeymoon is over.
Silicon Valley is bleeding cash. Wall Street is finally asking for the receipt. The euphoria that defined the last twenty four months has curdled into a cold, hard audit of return on investment. On February 23, Goldman Sachs portfolio strategist Ryan Hammond broke the silence. He signaled a pivot in how institutional investors view the artificial intelligence trade. The focus is no longer on potential. It is on disruption risk and the thinning margins of the hyperscalers.
The trillion dollar infrastructure trap
Capital expenditures have reached a terminal velocity that defies traditional accounting logic. Microsoft, Alphabet, and Meta have committed hundreds of billions to data centers that may never see full capacity. The technical bottleneck is no longer compute power. It is power itself. Electrical grids in Northern Virginia and Ireland are at a breaking point. This physical limitation is creating a massive lag between spending and earning. According to recent Reuters market reports, the gap between AI infrastructure investment and realized enterprise revenue has widened by 40 percent since the start of last year.
The unit economics of inference remain the industry’s dirty secret. Training a model is a one time sunk cost. Running that model for millions of users is an ongoing operational nightmare. Every query costs money. If the query does not lead to a high value transaction, it is a net loss for the provider. Most enterprises are still in the testing phase. They are using free credits or subsidized tiers. They are not yet paying the premium prices required to make these chips profitable for the cloud providers.
Goldman Sachs and the disruption pivot
Ryan Hammond’s recent commentary via Bloomberg terminal data suggests that the market is beginning to price in the losers of AI disruption rather than just the winners. This is a fundamental shift. In 2024, everything touched by AI went up. In February 2026, we are seeing the opposite. Legacy software providers are being cannibalized by custom internal agents. Why pay for a thousand seats of a CRM when a single localized LLM can manage the database for a fraction of the cost? This is the disruption risk Hammond warned about. It is not just about who builds the AI. It is about whose business model the AI destroys.
Visualizing the Expenditure Gap
The following data represents the estimated quarterly capital expenditure versus AI-attributed revenue for the top four hyperscalers as of late February 2026. The divergence is the primary driver of current market volatility.
Quarterly Capex vs AI Revenue Gap (Billions USD)
The enterprise adoption friction
Corporate boards are hesitant. The initial promise of 30 percent productivity gains has failed to materialize in the broader workforce. Most employees use AI for trivial tasks like drafting emails or summarizing meetings. These use cases do not justify a $30 per month per user seat price. To reach the next level of ROI, AI must take over complex, multi step workflows. This requires a level of data cleanliness that most Fortune 500 companies simply do not possess. They are spending millions on data governance before they can even begin to see a return on their AI spend.
Security remains the final hurdle. The fear of proprietary data leaking into public training sets has paralyzed the financial services and healthcare sectors. While private cloud solutions exist, they are significantly more expensive to maintain. This adds another layer of cost to an already bloated balance sheet. Investors are now looking at SEC filings for signs of “AI Fatigue” in enterprise sales cycles. The data suggests that the sales cycle for AI tools has lengthened from three months to nine months over the last year.
Market Concentration and the Liquidity Trap
The S&P 500 is top heavy. A handful of names dictate the direction of the entire market. If the AI ROI narrative continues to weaken, the indices have nowhere to hide. We are seeing a rotation into defensive sectors that were ignored during the 2024 rally. Utilities and consumer staples are outperforming tech for the first time in eighteen months. This is a defensive crouch. It suggests that the smart money is preparing for a period of stagnation in the tech sector while the infrastructure catches up to the reality of the software.
| Metric | Q1 2025 Actual | Q1 2026 Projected | Change (%) |
|---|---|---|---|
| NVIDIA H200/B100 Backlog | 4 Months | 7 Months | +75% |
| Enterprise AI Seat Price (Avg) | $28.50 | $19.20 | -32% |
| Hyperscaler Capex Intensity | 14.2% | 18.9% | +33% |
| AI Revenue Contribution (SaaS) | 4.1% | 6.8% | +65% |
The table above highlights the fundamental tension. While revenue contribution is growing, it is not keeping pace with the intensity of the capital expenditure. The backlog for high end chips is growing, but the price customers are willing to pay for the software is falling. This is a classic squeeze. The manufacturers are winning while the software integrators are losing. Eventually, the manufacturers will run out of buyers if the integrators cannot find a way to turn a profit.
The next major data point arrives on March 12. The Bureau of Labor Statistics will release the latest productivity numbers. If AI is truly the generational shift it claims to be, we should see a significant spike in output per hour. If those numbers remain flat, the AI trade will face its most existential challenge yet. Watch the 10 year Treasury yield. If it stays elevated, the cost of financing this AI build out will become unsustainable by the third quarter.
