TSMC's dilemma, OpenAI or Oracle, prediction on ambient computing (myriadperspectives.com)

OpenAI confirmed a long-rumored custom AI chip partnership with Broadcom and will buy 10 GW of Broadcom-designed inference systems over four years starting H2 2026—adding Broadcom to its existing NVIDIA and AMD suppliers. The deal is part of OpenAI’s vertical-integration push to become a hyperscaler: co-designing chips, racks, networking and software to optimize inference efficiency at scale. That optimization across the full stack promises big performance and cost gains, but also drives enormous demand for leading-edge wafer capacity measured in gigawatts of data‑center power. That surge creates a strategic dilemma for TSMC and the semiconductor supply chain. TSMC currently dominates advanced-node fabrication but its CapEx (~$42–50B/year) and geopolitically driven U.S. fab investments may not keep pace with projections (OpenAI/counterpart estimates imply tens to hundreds of GW over a decade). Shortfalls could trigger allocation battles, bidding dynamics, or intentional undersupply as risk management. Practically, ambient “always‑on” agents exacerbate the problem: real-time, low-latency inference requires low batch sizes (poor utilization) and much higher per‑token costs (OpenAI’s realtime audio pricing is many× higher than batch GPT pricing), and adding reasoning/context multiplies compute needs further. The net: ambient computing UX is constrained by physical compute economics and fab capacity—only a dramatic drop in real-time marginal cost (and large new fab investment) will enable ubiquitous always‑on agents.