0:00 openai just cut its latest funding round target by $10,000,000,000 dropping from 30,000,000,000 to 20,000,000,000 while locking in barely 11,000,000,000 so far in our last episode we looked at where capital translates into real capability this weekend private markets drew a firm line on price across the rest of the board finiti labs emerged from stealth with $5,200,000 backed by0:25 uncore capital to generate chip architectures directly from text prompts context labs launched its ai distillery in abu dhabi offering enterprises governed data packages to decouple internal company records from whichever frontier or open weight model they run and in optimization research the jobs framework demonstrated that predictive scaling curves can steer hyperparameter tuning cutting costly full training iterations out0:52 of the loop entirely now look at where the actual technical leverage is moving the headline isn't another monolithic parameter count it's brute efficiency at the task level samaya ai published results for samaya sterling a specialized financial agent post trained on nvidia's nemotron 3.5 lightning architecture the base model is a 30,000,000,000 parameter mixture of experts but sterling activates1:19 only 3,000,000,000 parameters during generation on the company's frontier finance benchmark sterling scored 58.9% topping claude fable 5.1 at 55.9% benchmark delta aside the decisive weapon here is the inference invoice samaya reports sterling answers each financial query for 61¢ claude fable 5.1 costs $6.89 for the identical prompt that is an 11 fold1:49 drop in operational cost when you run 25,000 multi step audit traces a day that gap is the difference between a functional enterprise software business and a venture subsidy burning through cash this connects directly to the financing squeeze hitting openai its target valuation dropped from $1,400,000,000,000 down to $1,250,000,000,000 as credit markets and institutional allocators bulk at the sheer2:15 debt required to build out raw compute the easy narrative says the frontier model simply eats every vertical underneath it the balance sheets say otherwise when a 3,000,000,000 active parameter specialized agent matches frontier accuracy while slashing inference costs by 90% the moat belongs to architecture and targeted tuning not raw spend
