Some AI Firms a Little ‘Overvalued,’ Khosla’s Choi Says

Khosla Ventures partner Ethan Choi said the world has only scratched the surface of AI and that some AI model startups may be overvalued but those valuations can be justified by high buildout costs and potential upside. No specific financial figures were cited; the remarks are constructive for long-term AI sentiment but are unlikely to materially move markets or individual stocks.

Analysis

The immediate, measurable winners from a continued large-capex AI buildout are upstream compute and real-estate plays: GPU/IP incumbents, hyperscaler cloud services, and colocation/data‑centre landlords capture the bulk of near-term margin and contract pull‑through. Expect a 12–24 month acceleration in multi‑year hardware orders and colocation leases even if downstream model providers fail to monetize—that creates durable cash flows for NVDA, MSFT, AMZN and EQIX/DLR independent of startup outcomes.

A big second‑order consequence is financing risk concentrated in late‑stage private rounds and crossover public holders: a funding squeeze or modest 20–40% repricing of private model valuations would force markdowns and potential forced exits in 6–18 months, hitting investor NAVs and secondary markets. Algorithmic efficiency gains or open‑source model advances that cut training/serving costs 3x–5x would be the fastest path to destroying the “build cost” justification currently embedded in valuations.

Catalysts that will re‑rate this market are concrete: sustained GPU utilization >80% across hyperscalers (weeks→months), material changes in unit training cost, or regulatory constraints on model deployment (6–24 months). Trade friction lies in timing: hardware and real‑estate winners rerate quickly on utilization data, while model providers’ multiples compress only after funding and revenue signals surface—this asymmetry favors long infra/short late‑stage model exposure for a 6–18 month play.

Key Decisions for Investors

• Long NVDA via a 9–12 month call spread (buy 12‑month ATM calls, sell 25–35% OTM calls) sizing 1–3% NAV: asymmetric upside to capture further GPU tightness if hyperscaler utilization stays >75%; max loss = premium (~100% of position cost), target 50–150% return if NVDA core revenue accelerates.
• Pair trade: Long Equinix (EQIX) or Digital Realty (DLR) 6–18 month exposure (buy stock or 9–12 month LEAPs) vs buy 6–12 month puts on C3.ai (AI) to express infra stability vs model‑provider re‑rating. Expect 20–40% absolute upside on REITs from contracted book while puts capture a 30–60% downside in the event of funding shock; net funding cost of hedge limited to put premium.
• Buy 12‑month MSFT calls (small, 1–2% NAV) to own cloud capture of model deployment and enterprise monetization; rationale: durable ARR growth even if proprietary model valuations compress. Risk = option premium; reward = levered exposure to Azure GPU/RAG adoption with corporate GTM optionality.
• Hedge tail‑risk in private model markets by buying 3–9 month OTM puts on AI pure‑play equities (C3.ai) sized to cover estimated NAV markdown risk from a 20–40% private valuation reset. This limits downside from a financing cliff over the next 6–18 months while retaining upside in infra names.