Key Takeaways
- OpenAI discontinued Sora on March 24, 2026, generating only $1.4M in revenue versus ChatGPT's $1.9B — a 1,357x revenue gap reflecting unviable compute economics for video generation
- Video generation requires 100-1000x more compute per second of output than text generation, making consumer product economics structurally impossible even at scale
- TSMC's 2nm capacity meets only one-third of what major AI customers require for accelerators — the entire 2026 production is fully booked with demand exceeding capacity by 25-30% through 2027
- PCB lead times stretched from 6 weeks to 6 months for optical transceivers; laser device capacity is constrained — the bottleneck is the entire electronics ecosystem, not just chips
- TSMC advanced node pricing is rising 3-10% in 2026 — the cost deflation assumption underlying every AI business plan is wrong at the physical layer
Sora: From Hype to Discontinuation
OpenAI's Sora shutdown on March 24, 2026 is the first high-profile casualty of compute economics. The product generated approximately $1.4 million in revenue versus $1.9 billion for ChatGPT over the same period — a 1,357x revenue gap that made the project a structural black hole for computational resources.
The economics failure is not a product mistake; it is a fundamental modality problem. Video generation requires 100-1000x more compute per second of output than text generation, while user willingness-to-pay does not scale proportionally. A user will pay $20/month for unlimited ChatGPT access but will not pay $200-2000/month for equivalent video generation access. The demand-side constraint is absolute.
Disney's planned $1 billion investment (structured as stock warrants, never finalized in cash) dissolved simultaneously. The entertainment industry's strategic bet on Sora as a content generation tool evaporated not because the technology was inadequate but because the unit economics could not work under any plausible scenario.
OpenAI's explicit pivot of Sora techniques to world simulation research to advance robotics signals the company has internally concluded that video generation capabilities are worth more as training substrate for physical-world models than as consumer products. This is rational: the same 3D scene understanding and physics modeling that powers video generation can train robotic foundation models, where the value capture occurs through physical automation rather than content generation.
The TSMC Capacity Wall
The supply side amplifies the demand-side failure. TSMC's entire 2nm production capacity for 2026 is fully booked. Current capacity meets 'about one-third' of what major customers require for AI accelerators. International Business Strategies projects wafer demand at 5nm and below will exceed capacity by 25-30% through 2026 with shortages extending into 2027.
This is not a temporary constraint. TSMC is responding with record $52-56 billion in 2026 capex and three additional Taiwanese facilities at $28.6 billion. But the Arizona fab targeting mass production around 2030 is already 'largely reserved before construction begins,' demonstrating that demand has permanently decoupled from capacity planning cycles.
The Compute Economics Wall: Key Metrics
Revenue gap, supply deficit, and price trajectory data showing compute economics moving against AI expansion
Source: Bloomberg / EE Times / TrendForce / Broadcom
The Cascading Ecosystem Bottleneck
The constraint extends far beyond chips. PCB lead times for optical transceiver applications have stretched from 6 weeks to 6 months. Laser device capacity is constrained. Broadcom's Senior Director stated TSMC capacity has 'reached its limits' after being 'virtually unlimited' previously. This is not a chip shortage — it is an electronics ecosystem shortage.
Advanced node pricing is rising 3-10% in 2026 — the cost deflation assumption underlying every AI business plan is wrong. Compute cost curves will not decline; they will increase through 2027 minimum. This invalidates the financial models of companies that depend on inference cost decline for unit economics viability.
Structural Competitive Advantages
Companies that secured multi-year TSMC capacity agreements (Apple, Nvidia, Google, Amazon) have a structural competitive moat that cannot be replicated by algorithmic innovation alone. Nvidia's next-generation Feynman AI chip platform may be forced to redesign due to 2nm constraints. Every AI startup's unit economics model that assumes declining compute costs needs immediate revision.
The implications cascade through the AI value chain. Pure-play video generation companies like Runway, Pika, and Kling face the same compute economics headwinds that killed Sora. Text and image generation economics remain viable because the compute-to-value ratio is fundamentally different. Embodied AI applications (robotics, autonomous systems) face the same constraint but have higher value-per-computation since physical outputs create lasting economic value.
What This Means for Practitioners
ML engineers building compute-intensive applications (video, 3D, simulation) should immediately model rising rather than declining inference costs through 2027. Startups depending on inference cost decline for unit economics viability must revise financial models today — not next year.
For teams selecting which inference modalities to target, the hierarchy is now clear: text and image generation have viable product economics. Video generation does not. Pure-play synthetic media companies should pivot to either (1) vertical applications with very high per-unit willingness-to-pay, or (2) use video as training substrate (as OpenAI is doing) rather than end product.
Plan for 18-24 months of elevated compute costs regardless of software optimization. This is a hardware constraint, not an engineering problem that innovation will solve. Budget accordingly.