Topic
Cost of serving model tokens, per-token pricing pressure, and the economics of giving inference away.
Rohan Pinto says BitNet b1.58 runs a 100-billion-parameter model at five to seven tokens a second on a single CPU. That is enough for overnight batches and air-gapped servers, and short of what a cloud tier needs.
Reality
The forecast, carried by state broadcaster CCTV, describes compute demand that is billed every time somebody uses a model. Europe is building for the same demand with seven gigafactories on a EUR 30bn budget, of which about EUR 1bn is committed.
Reality
The opposition numbers land on product teams, because an AI feature that fires without being asked bills compute for every user and adds load in towns already organizing against the next build.
Reality
The partnership removes the external API hop for enterprises that cannot let sensitive records leave the perimeter, which also moves the inference bill off someone else's per-token pricing and onto hardware the customer has to buy and keep busy.
Reality
SemiAnalysis says a $200 plan can yield $8,000 of Anthropic API-equivalent usage and $14,000 of OpenAI's. Even the best published caching fix leaves the gap several times wide.
Reality
Atomico led a $100m seed into a London compute-routing startup that raised $10.25m in February. The UK's Sovereign AI Fund made its first-ever investment here, in April.
Perspective Coverage
3 publishersReality
Two Nvidia workstations behind a bar in Zhongguancun self-host DeepSeek V4 Flash and hand out inference like bar snacks. Read it as a pricing signal, not a novelty.
Reality