In the AI industry, ‘agentic’ takes on a life of its own

Eric Schmidt’s admonition to “found an agentic AI company” frames the article, which defines agentic AI as systems that automate entire processes with minimal human input. The piece balances industry hype with skepticism: researchers and former OpenAI staff note prioritization, fidelity and autonomy limitations (Karpathy suggested it may take ~10 years to work through key issues), and OpenAI reportedly de-emphasized agents to focus on ChatGPT. For investors, this signals conceptual hype and early-market talent/venture interest but not immediate revenue or product-readiness signals; governance, execution and alignment risks predominate.

Analysis

Agentic workflows will reprice the marginal economics of compute and enterprise software: one multi-step agent executing a business process creates repeated model calls, state management, and orchestration overhead that could raise cloud/GPU consumption per workflow by a low-single-digit to mid-single-digit multiple versus single-turn prompts. Expect enterprise pilots to lift cloud spend line items first (12–24 months) and labor-replacement effects to be visible only after tooling, governance, and liability solutions mature (3–7 years).

The principal-agent frictions implicit in delegation create an underpriced market: companies that provide monitoring, explainability, access control, and human-in-the-loop gating capture durable annuity economics and pricing power because buyers will pay to avoid litigation and operational outages. Conversely, models and consumer-facing agent apps that emphasize autonomy without enterprise-grade controls will face steep commercial friction and higher churn.

Key risks are idiosyncratic failure modes and regulatory backlash; a single high-profile agent-caused loss or data-exfiltration incident could pause large enterprise rollouts for 6–18 months and force expensive remediation. Catalysts to watch are (1) major cloud providers bundling managed agent orchestration at scale, (2) tier-1 enterprise procurement approvals for agent pilots, and (3) any meaningful GPU supply or pricing shock that reshapes the cost curve for running agents.

Key Decisions for Investors

• Long NVDA (buy 9–12 month calls or a modest call spread) — thematic exposure to sustained GPU demand from agent orchestration. Timeframe: 6–18 months. Risk/reward: asymmetric upside (>30–50%) if enterprise pilots scale; downside limited to premium paid or spread width if hardware demand softens or macro slows.
• Long DDOG or SPLK (buy shares or 6–12 month call options) — observability and telemetry will become mandatory for agent deployments; vendors can expand ARR through higher attach rates and professional services. Timeframe: 6–18 months. Risk/reward: 20–40% upside if enterprise budgets reallocate to monitoring; 20–30% downside if adoption stalls or pricing competition intensifies.
• Long PLTR (buy shares or 12–24 month calls) — take exposure to enterprise orchestration, data lineage and ops that sit between models and business processes. Timeframe: 12–24 months. Risk/reward: binary upside on large government/enterprise contracts (2x+ on success); high idiosyncratic risk if sales cycles lengthen.
• Pair trade: long NVDA / short UPWK (buy NVDA calls and buy 12–24 month put spread on UPWK) — monetize asymmetric tail of compute demand vs. secular headwind to freelance marketplaces if agents begin substituting low-skill gigs. Timeframe: 12–36 months. Risk/reward: captures re-rating of infra/ops leaders while hedging the risk that agents augment rather than replace freelance work (loss limited to put premium/spread).