THE WEEKLY
Issue of 26 September 2026
This week's facts and what changed on the site, each one still linked to its source.
Every advanced AI chip needs 2.5D packaging before it ships, and TSMC's CoWoS lines have been the industry's tightest bottleneck. TrendForce expects the shortfall between what customers want and what TSMC can supply to ease as capacity ramps toward 200,000 wafers a month industry-wide, narrowing by the end of 2026 from TSMC's CoWoS supply-demand gap narrowing from about 20% to roughly 10% by the end of 2026
See the receipt →US cloud providers signed multi-year supply agreements that stop memory makers raising their prices, so TrendForce expects the third-quarter 2026 increase to fall on everyone else: server DRAM contract prices up another 13–18% in one quarter
See the receipt →
Terminal-Bench 2.0 tests whether AI agents can chain together the kind of multi-step, unsupervised terminal work a developer does in a day — compiling code, training models, setting up servers. Stanford's 2026 AI Index reports agent accuracy on it nearly quadrupled in a year, reaching 77.3% by early 2026, up from 20% in February 2025
See the receipt →
AI chip performance per dollar has been improving by approximately 1.37× per year
See the receipt →
Tesla reports its AI training compute the way it reports factory capacity — in installed megawatts. At the end of the second quarter of 2026 both Texas clusters, Cortex 1 and Cortex 2, were in production, and the compute installed there had more than doubled in six months, to over 205 MW of AI training capacity
See the receipt →
TrendForce's tally of the industry's biggest cloud spenders keeps climbing. Combining Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba and Baidu, it now forecasts their combined 2026 capital expenditure will exceed $886.7 billion, with the five North American hyperscalers accounting for nearly 90% of it
See the receipt →
Broadcom AVGO
Accelerators · The engine
Designs the custom accelerators hyperscalers use to avoid buying merchant GPUs, and reports that work as AI semiconductor revenue.
See it→Data-centre revenue of $89bn in the quarter to 26 July 2026 — and, in the same filings, a guarantee of up to $105bn on OpenAI's leases and $36bn of commitments to rent capacity back. For members; free in Nvidia's week, 19 October.
See it →Revenue of $1.935bn in the quarter to 30 June 2026, up 93%. The filings use the word AI 65 times and never as a revenue line — the money is split only by who pays. For members; free in Palantir's week, 12 October.
See it →The value-chain layer, the signal that stands in for an undisclosed AI share, and the open card's own prose were all still English on /de, /es and /zh. All three now follow the reader. Filed figures stay exactly as filed.
See it →The set went from 17 companies to 27 — and not one of the ten new ones discloses an AI share of revenue. Each shows a cited signal instead: capex the company calls AI-driven, megawatts contracted, orders booked.
See it →Read on 24 September 2026
THE $70 BILLION KITCHEN: How to Read a Company Whose AI Is All Cost
See it→