openai filed its ipo paperwork.
then reportedly told advisers it would rather wait until 2027.
because $1 trillion or nothing.
anthropic, meanwhile, is already at $965B and didn't blink.
deepseek didn't release a new model.
they released a framework that makes the old one 85% faster.
no new hardware. no retraining. just smarter decoding.
while everyone's racing to build bigger, they're figuring out cheaper.
princeton trained AI to design radio chips from scratch.
no human templates. no prior art. just reinforcement learning and diffusion models.
the chips look like QR codes.
they outperform the human-designed ones.
anthropic hired john jumper — the man who cracked protein folding, won a nobel prize, and then changed labs mid-IPO.
monday he speaks publicly for the first time.
science doesn't usually move this fast.
google capped how much of its own AI meta can use.
rivals competing for the same ad dollars. now one is rationing compute to the other.
the AI arms race turned silicon valley's biggest enemies into each other's customers.
google deepmind expanded its 'ai coding strike team' to include midtraining.
six researchers left for anthropic, openai, and meta the same week.
$270B wiped from alphabet.
the strike team is now rebuilding itself.
the bank that called the 2008 crash compared the AI boom to canal fever, electrification mania, and the dotcom bubble.
each one ended in economy-wide recessions.
five US tech companies spending $1 trillion on AI next year.
the BIS has been right before.
qualcomm confirmed acquiring Modular, the compiler that runs AI models on any chip (nvidia, AMD, intel), for $4B.
also in talks to buy Jim Keller's Tenstorrent for $10B.
$14 billion to crack nvidia's closed loop.
CUDA was a moat. they brought a shovel.
reid hoffman called xAI a 'complete train wreck.'
meanwhile goldman sachs says AI is erasing 11,000 US jobs per month.
the people building this thing can't agree if any of them are winning.
the workers already know the answer.
openai vs anthropic was the wrong rivalry.
both labs are now stuck behind the same informal government approval process that doesn't officially exist yet.
same problem. different logo.
