A massive $100bn deal between **Nvidia** and **OpenAI** has reportedly evaporated, raising alarms about the sustainability of AI's circular funding models, where chipmakers fund developers who then use that money to buy the chipmakers' own products.

Everyone is staring at Nvidia's GPU architecture. But the real reason for Nvidia's dominance lies in its control over physical limitations. To understand whether this rally is sustainable, one must look beyond the chips themselves and understand the bottlenecks that Nvidia is currently solving for the entire industry. 1. The "Stacking Dilemma" (Advanced Packaging) Nvidia's performance leaps no longer come solely from smaller transistors, but from 3D stacking. HBM memory and logic dies are stacked on top of each other in towers. The strategy: The more complex these towers become, the more crucial the "assembly" process (OSAT - Outsourced Semiconductor Assembly and Test) becomes. The logic: Whoever possesses the machines and the capacity to precisely stack these chips holds the key to production. Without this packaging, Nvidia's designs remain purely theoretical. 2. Zero-Defect Tolerance (Metrology & Inspection) Stacking chips creates a huge risk: If an internal layer is defective, the entire module is scrap. With a Blackwell chip, we're talking about enormous sums of money. The strategy: Optical and acoustic chip testing has gone from a peripheral issue to the heart of the factory. The logic: I pay close attention to the companies that supply the measuring instruments needed to find defects in stacked chips at the nanometer scale. Without their "eyes," TSMC cannot guarantee the yield for Nvidia. 3. The Energy Constraint in the Rack Nvidia no longer sells components, but complete power systems. A modern rack now requires the power supply of an entire small town. The strategy: The switch to direct liquid cooling and high-precision power conversion systems is vital for Nvidia's survival. The logic: If the power supply in the rack fluctuates, the AI ​​model crashes. Nvidia is working with partners who provide critical components for voltage control and thermal monitoring. Conclusion for Part 1: Nvidia is the architect, but the blueprint only works if the specialists in packaging, testing, and power infrastructure deliver. Identifying these "enablers" reveals the true foundation of Nvidia's success. In Part 2, we'll look at why institutional investors like Vanguard are building precisely these infrastructure assets behind the scenes, while the general public is still debating the software.

OpenAI was looking to alternatives to Nvidia GPUs, because it was disappointed in the inference performance. Both Nvidia and OpenAI have since made deals with chip makers to overcome this deficit. GPUs rely on external memories, while more on chip memory helps with inference. This is where Gemini has a lead with TPUs. One will have to see how Nvidia adapts its products, and how long it will take. In the meanwhile OpenAI might try other chips: it has already made deals. But as a monopoly Nvidia shouldn't be allowed to acquire chip startups. Give startups a chance to grow and monetize through long term contracts, technology licensing, or IPOs. Hopefully there will be more competition from AMD, and smaller startups. Reference: reuters.com/.../openai-is-unsatisfied-wi...