Another important thing to keep in mind here is the software. Nvidia has been optimizing their software stack for decades to squeeze every clock cycle out of the hardware under any limits. TensorRT alone does straight up magic. Apple is just starting out with MLX, and their hardware is often idling not because it is worse per watt but because the compiler does not know how to optimally load the ALU blocks for specific graphs yet
AI makes it cheaper to create working fragments. It does not automatically make those fragments part of a maintainable system. In practice it may make the canonization step more important not less
Everything depends heavily on the segment. Open-weight models already look good enough for a ton of internal tasks right now but there is always gonna be that small percentage of workloads where the last few percent of quality are totally worth the money.
I feel like the market is just gonna become way more mixed. Not like "everyone is switching to open" or "everyone is staying on frontier" but a mix of multiple models for different scenarios
The weakest part of the article is that the forecasts up to 2029 assume the current market structure will barely change. In three years literally everything can change, like prices, models, hardware, and how we actually use LLMs
That is exactly why big clouds never put all their eggs in one basket, even if it is a super cheap and cold basket. The cost of protecting and backing up network lines for an isolated island quickly eats up any benefits from geothermal energy. Physical security for terabit lines is way more expensive than air conditioning these days
Iceland and Norway are part of the EEA so the AI Act and GDPR will reach them just like Germany or France. Running away there from regulators makes no sense. But running from bureaucracy to get land permits and substation connections - yeah maybe municipalities work faster there