OpenAI thought it could own AI videos. The reality was too expensive

OpenAI is winding down Sora after roughly six months amid sharply falling engagement (downloads down 70% from November; DAUs down 34%) and estimated running costs cited at about $15M per day. The shutdown signals a strategic retrenchment toward core products as OpenAI faces intensifying competition and a heavy compute-driven cash burn despite roughly $13B revenue last year and a target to triple revenue by 2026.

Analysis

A retrenchment by a leading LLM provider from consumer-facing, compute-heavy multimodal video products crystallizes two durable lessons: consumer attention economics and marginal compute cost matter more than model novelty. Consumer-video that substitutes for serendipitous, real-world moments shows poor long-term retention unless the product either (a) surfaces social proof that can’t be easily synthetic, or (b) embeds monetization that covers very high unit compute cost (e.g., B2B licensing, creative tools for agencies). Expect venture and corporate capex to reallocate away from pure consumer video generators toward enterprise workflows, creative tooling, and robotics where per-session ARPU is higher and retention is stickier.

Strategically, this creates a near-term runway for incumbents who can marry large language models with established enterprise distribution (search, ads, cloud). Firms that own distribution and billing relationships — and who can amortize model costs across high-ARPU customers — gain optionality without having to subsidize viral consumer trials. Conversely, startups chasing viral video faces a longer path to sustainable unit economics unless they secure deep-pocketed strategic partnerships or proprietary, low-cost inference stacks.

Key catalysts that would reverse the current dynamic are technical: a 2-3x drop in inference cost via next-gen chips or model compression within 12-24 months would re-open consumer product economics quickly; likewise, clear regulatory guardrails on deepfakes could either shorten or lengthen time-to-market depending on compliance costs. Watch quarterly disclosures for shifts in enterprise bookings and any vendor commentary on realized inference pricing and client concentration — those numbers will be the earliest hard signals of where capital and product effort will migrate next.

Key Decisions for Investors

• Long Alphabet (GOOG) — 6–12 month horizon. Increase exposure to 2–3% of fund NAV; thesis is durable enterprise distribution and ad/AI bundling will capture monetization as competitors retrench. Upside 15–25% if enterprise adoption accelerates; main risks are regulatory fines or ad cyclicality.
• Bull-call spread on GOOG — 6–9 month expiry. Buy ATM calls and sell out-of-the-money calls to cap premium outlay; targets asymmetric payoff if Google accelerates Gemini/product integration while limiting downside to premium paid.
• Short/put position on SMWB — 3–9 month horizon sized to 0.5–1% NAV risk. Use puts to limit tail risk; rationale: small-cap consumer AI/social names face higher cash-burn-to-retention mismatch and will reprice violently if funding tightens. Reward potential 30–60% if user metrics and financing lines deteriorate.
• Pair trade: long GOOG funded by short SMWB (ratio 1:1 by notional) — 3–12 months. This captures a rotation from speculative consumer AI toward platform-owned enterprise monetization while hedging broader market moves; unwind if market-wide liquidity tightens or if SMWB reports a credible path to sustainable ARPU.