Re‑map your AI exposure by layer
- Investors are increasingly sorting artificial intelligence holdings by layer — chips, cloud, models, tools, apps and distribution — instead of treating “AI” as one trade with one set of economics. - The sharpest dividing line is revenue quality: Goldman Sachs says investors are rewarding companies that show a clear link between AI spending and sales, while rotating away from debt-funded infrastructure exposure. - Sequoia’s 2024 “$600B question” and Goldman’s 2026 capex outlook both pushed the market toward business-model analysis over model-version hype. (sequoiacap.com) (goldmansachs.com)
Artificial intelligence investing is shifting from one big theme to a layer-by-layer map of who sells chips, who rents compute, who builds models, and who keeps customers. (goldmansachs.com) (sequoiacap.com) That shift followed a surge in spending below the surface. Goldman Sachs Research said Wall Street’s consensus for 2026 capital spending by AI hyperscalers reached $527 billion, up from $465 billion earlier in the earnings season. (goldmansachs.com) Sequoia partner David Cahn framed the mismatch a year earlier as “AI’s $600B question,” arguing that infrastructure spending implied far more end-user revenue than the market had actually proven. He wrote in June 2024 that the earlier “$125B hole” had widened as Nvidia revenue and data-center buildouts climbed. (sequoiacap.com) A layer map starts with semiconductors and servers, then moves up to cloud platforms, foundation models, developer tooling, application software, and the channels that distribute AI to users. Different layers carry different margins, capital needs, and customer lock-in. (goldmansachs.com) (docs.nvidia.com) The practical question is not whether a company says “AI” on an earnings call. It is whether the company turns AI spending into recurring revenue, operating profit, and customers who are costly to replace. (goldmansachs.com) Goldman said investors have rotated away from AI infrastructure companies where operating-earnings growth is under pressure and capital spending is funded with debt. It also said investors have rewarded companies that can show a clearer link between capex and revenue. (goldmansachs.com) That makes the middle and top of the stack more important to analyze. An application company with proprietary data, embedded workflows, and renewal revenue can keep pricing power even if the underlying model gets cheaper. (menlovc.com) (goldmansachs.com) Menlo Ventures said enterprises spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024, and that $19 billion of that went to user-facing products in the application layer. Coding tools alone accounted for $4 billion, the firm said. (menlovc.com) Distribution is its own layer because the company that already owns the customer relationship can absorb model improvements faster than a standalone model vendor can. Goldman’s 2026 outlook also pointed investors toward data governance, security, and consumer platforms beyond the original hyperscaler trade. (goldmansachs.com) The result is a more selective market. Goldman said the average stock-price correlation across large public AI hyperscalers fell from 80% to 20% after June, a sign that investors are no longer treating AI exposure as one uniform bet. (goldmansachs.com) That is why the cleaner framework is to re-map AI exposure by layer, then ask which layer captures durable economics after the next model release. The companies that keep the customer, not just the compute bill, are the ones the market is starting to separate out. (goldmansachs.com) (sequoiacap.com)
