Speaking on the 20VC podcast, Positron AI co-founder Thomas Sohmers argued that the binding constraint on artificial intelligence has shifted from compute to memory bandwidth, noting that memory bandwidth improved 17x per device between 2014 and 2024 versus roughly 120x for compute. He said 96% of tokens in real agentic coding sessions are cached, costing roughly a thousandth of a generated token to serve, and cited Anthropic's reported 80-point gross margin on its API business as evidence frontier labs would be "massively profitable overnight" if they stopped training. Positron AI recently raised $875 million at a $5 billion valuation, betting inference decides AI economics. Sohmers pushed back on the "burning cash" narrative, the on-device shift thesis, and the push to pace the AI frontier, arguing memory capacity per device, not data-center electricity, is decisive. He warned anti-data-center sentiment is "almost entirely a Chinese scion," and predicted local AI agents will drive orders-of-magnitude more cloud token consumption. He said token-serving margins are temporarily obscene but doomed to compress, and Nvidia's memory trajectory is diverging from Positron's.