AI digest: Valuations gone mad
Anthropic heads for a $900B valuation whilst the AI infrastructure race hits overdrive.
The AI money machine keeps spinning faster. This week brought eye-watering valuations and some proper technical advances.
Anthropic chases $900B valuation in record time
Sources tell TechCrunch that Anthropic is asking investors to submit allocations within 48 hours for a funding round that could value the company at over $900 billion. That would make it worth more than most countries’ GDP, just three years after launch. The speed suggests either massive FOMO or genuine belief that Claude’s capabilities justify this bonkers number.
OpenAI hits 10 gigawatt compute target years early
OpenAI announced it reached its 10 gigawatt AI compute capacity goal in the US several years ahead of schedule. That’s enough power to run a small city, all dedicated to training models. The infrastructure race is clearly accelerating faster than anyone predicted, which explains why everyone’s scrambling for chips and power deals.
Moonshot AI open-sources FlashKDA for faster attention
Chinese AI lab Moonshot has released FlashKDA, optimised CUDA kernels for their Kimi Delta Attention mechanism. The benchmarks show meaningful speed improvements over existing implementations. Open-sourcing core infrastructure like this signals the Chinese labs are serious about competing on the technical fundamentals, not just throwing compute at problems.
Musk admits xAI trained Grok on OpenAI models
In court testimony, Musk confirmed that xAI used model distillation to train Grok using OpenAI’s models as teachers. This is the first major admission of the practice that everyone suspected was happening. It’s going to make the legal battles around model IP even messier.