Vercel CEO Guillermo Rauch on the fight to split off models from agents

Vercel CEO Guillermo Rauch says the platform now handles 6M deployments per day, with over 50% driven by coding agents, and routes more than 1T tokens daily through its AI gateway—signaling strong production adoption beyond pilots. He highlights Eve (natural-language agent framework) and Vercel Sandbox (policy-controlled “sandbox cage”) to address internal-agent risks around data access, auditing, and tool-call trails. The article also frames intensifying competition with major AI labs as agents become plug-and-play components, but positions Vercel as the “AWS of this generation” via open protocols.

Analysis

The key market mechanism is not “AI adoption,” it’s stack unbundling: as agents move from demos into production, buyers optimize for routing, policy, and deployment control rather than single-model loyalty. That shifts pricing power away from frontier labs and toward infrastructure layers that sit between models and end users. The clearest beneficiary is the platform layer that can arbitrate model choice, identity, and deployment; the clearest loser is any SaaS incumbent whose moat depends on owning the workflow UI and the underlying data exhaust.

For CRM, the risk is less immediate seat destruction than gradual bypass. If agents become the primary interface for internal work, the value of dashboards and manual navigation compresses while the value of secure APIs, audit trails, and permissioning expands. That argues for a longer-term multiple split inside enterprise software: vendors that can become the control plane for agents should hold up, while “systems of record” that mainly monetize trapped usage may see slower net retention and more customer pressure to unbundle modules. AMZN is a secondary beneficiary if the market continues to treat compute and hosting as neutral commodities; however, edge-native platforms can still take share in developer-facing workloads where latency and deployment velocity matter.

Contrarian view: the consensus is probably overestimating near-term lab disintermediation. Production buyers care about price/performance and governance, which favors multi-model routing and open protocols, but that same governance concern also slows enterprise rollout if security or training-leak headlines flare up. The thesis breaks if one lab successfully bundles hosting + models + agent tooling at a discount, or if a sandbox/privacy incident forces a pause in enterprise agent adoption over the next 1-3 months.

Key Decisions for Investors

• Long EVE on pullbacks over the next 1-3 weeks; thesis is that production agent traffic is shifting monetization toward the orchestration/hosting layer. Use a tight stop if the stock loses post-event support or if management commentary implies usage growth is still mostly experimental.
• Pair trade: long EVE / short CRM over 1-3 months. The risk/reward favors the layer that captures agent workflow traffic over the incumbent whose moat depends on UI-driven data access; thesis falsified if CRM shows accelerating agent-native attach or material reacceleration in usage-based metrics.
• Watchlist AMZN for a secondary long entry on AWS commentary tied to model-routing and inference demand. This is a lower-conviction beneficiary, but if cloud consumption data inflects with agent traffic, AWS should capture the compute and networking spend.
• Avoid chasing pure-model exposure here; use any strength in lab-adjacent names to fade if they are trading as if hosting and governance will remain optional. Reassess only if a lab proves it can own both the agent and the deployment layer profitably.
• Set an alert for any enterprise security/data-leak event around code assistants or sandboxing; that would be the most credible near-term catalyst to delay adoption and compress the entire agent-infra trade.