Overview
Oracle and OpenAI abandoned one planned Texas expansion while preserving Stargate's larger capacity target. Alibaba responded to departures from Qwen with a foundation-model task force and a promise to continue open releases. FinSheet-Bench supplied an immediate deployment check: leading models still made frequent errors on complex financial spreadsheets.
Developments
1. Stargate reallocates a planned 600-megawatt expansion
Oracle and OpenAI ended talks to add 600 megawatts near the flagship Abilene site after financing discussions dragged and OpenAI's needs changed, Reuters reported. A separate 4.5-gigawatt expansion remained planned, with the displaced capacity expected at another campus.
Two of Abilene's eight buildings were already operating, according to Reuters, which makes the dropped expansion a site decision rather than a reversal of the wider program. The episode separates a 10-gigawatt ambition from usable capacity: financing, site allocation, construction, power delivery, and server commissioning can each change on a different schedule.
Sources: Reuters
2. Alibaba creates a foundation-model task force after Qwen exits
Qwen technical lead Junyang Lin and two colleagues announced their departures after Alibaba released its Qwen3.5 small-model series. In a March 5 internal message published by VentureBeat, Alibaba CEO Eddie Wu accepted Lin's resignation, said Jingren Zhou would continue leading Tongyi Laboratory, and created a three-person foundation-model task force.
Wu also said Alibaba would maintain its open-source model strategy, increase AI research investment, and accelerate hiring. Existing Apache-licensed weights are unaffected; the observable continuity tests are who maintains future releases, whether artifacts and training details remain open, and whether the new task force restores a stable release cadence.
Sources: VentureBeat · Qwen
3. FinSheet-Bench finds spreadsheet accuracy falls with complexity
A March 7 preprint evaluates ten model configurations on 24 synthetic financial portfolio spreadsheets modeled on private-equity data rooms. Gemini 3.1 Pro led at 82.4% overall accuracy, equivalent to roughly one wrong answer in six, while the average across models fell to 48.6% on the largest sheet.
Synthetic data protects confidential source material and permits exact grading, but cannot capture every formatting and semantic irregularity in a live diligence file. The size gradient is the actionable result: document understanding and deterministic calculation remain separate control points once a spreadsheet grows beyond simple lookup questions.
Sources: FinSheet-Bench preprint