DeepSeek’s newly-released V4 model charges double during Beijing office hours and half price at night. Surge pricing has come for artificial intelligence, and even at peak it humiliates Western prices.
Anthropic spent $1.97M lobbying Washington in Q2, more than Nvidia, while OpenAI closed in too. The ‘safety-first’ labs are breaking lobbying records to write their own rulebook. What that money is actually buying.
Meta’s automated moderation is wrongly deleting years-old accounts, sometimes branding people with its worst accusation, and the ‘appeal’ is handled by the same AI. The only reliable way back? Be famous enough for a journalist to intervene.
Tesla’s operating profit fell 57% and its subsidy crutch collapsed while Musk chases robots and robotaxis. Alphabet’s profit rose 298% on an 82% cloud surge, then its stock sank anyway on a $205B capex bill. The AI-spend verdict, split down the middle.
Neill Blomkamp released ‘Nightborne’, a 13-minute AI-generated sci-fi short made with ByteDance’s Seedance 2.0. It’s a genuine technical milestone (13 coherent minutes) and, per critics, a mixed watch, plus 32 licensed faces and a real labour question.
Alphabet and Tesla report today, and the number that matters is whether Big Tech’s vast AI spending is turning into returns. Alphabet is even selling stock to fund data centres, the tell. Watch the capex guidance, not just the profit.
Nvidia disclosed a 9.3% passive stake in cloud firm Nebius, which promptly jumped ~6%. It’s a chipmaker buying a slice of a customer whose whole business is buying its chips, the circular ‘vendor financing’ pattern worth watching, calmly.
A joint AWS, Nvidia, Berkeley Lab and NASA study built a model (the communication-to-computation ratio) for when a quantum processor must sit next to a supercomputer versus when the cloud will do. Unsexy, and exactly the discipline the field needed.
Google is reportedly developing ‘Frozen v2’, a chip that hardcodes Gemini’s architecture into silicon for a claimed 6-10x efficiency gain over its current TPUs. Real direction, unproven magnitude, and not due until 2028.
Meta’s Muse Spark 1.1 quietly tops the tool-use benchmarks that measure whether an AI can finish real multi-step work: computer use across desktop, browser and mobile, a 1M-token context, and parallel subagents. The most useful model of the month.