Keep pulling the thread on Ben Thompson.
There are rumors that NVIDIA's Blackwell GPUs are unreliable and that training runs using them are difficult due to their cutting-edge nature.
The speaker speculates that OpenAI's next-generation model may be delayed in part due to reliability issues with NVIDIA's Blackwell GPUs during training.
The SpaceX Starship is expected to reduce satellite launch costs by an order of magnitude.
A viable space-based data center would require a complete, first-principles redesign of every component and process, rather than adapting terrestrial data center architectures.
A primary challenge for data centers in space is cooling, as the vacuum prevents convective cooling and forces reliance on less efficient radiative cooling.
Space-based data centers are fundamentally not serviceable or maintainable once deployed in orbit.
There is a years-long lead time for large electrical transformers, creating a significant bottleneck for new electrical generation projects in the US.
The speaker speculates that China could develop a semiconductor manufacturing process that is less reliable but more scalable and cost-effective than ASML's, following an "Elon Musk" style of engineering.
SpaceX is likely the only entity capable of building data centers in space due to its vertically integrated stack, low-cost launch capabilities, and operational experience with Starlink.
The economic case for space-based data centers is primarily driven by terrestrial constraints like land use permitting and energy shortages, not by inherent advantages of operating in space.
Google's Gemini 3 model's primary advantage is its larger size and parameter count, while OpenAI's GPT-4 level models still maintain higher quality on a per-parameter basis.
GPUs frequently fail, which would require any space-based data center to have a massively fault-tolerant architecture.