The week ahead: Imagen 4 switches off, and China's two AI platforms report

The confirmed calendar for 16-22 August is short, and what is on it is not about new capability. One image-model generation switches off inside Google's API, and the two largest Chinese AI platforms report quarters that now turn mostly on cloud revenue and capital spending. Everything below carries a date stated by a source we opened. Nothing here is a prediction.

Monday 17 August: Imagen 4 stops answering

Google's Gemini API deprecations table lists three models with a shutdown date of 17 August 2026: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001. All three carry a release date of 24 June 2025, so this closes a roughly fourteen-month lifecycle. The recommended replacement Google names for all three is gemini-3.1-flash-image.

That replacement is worth testing rather than pasting in. It is a Gemini image model, not an Imagen point upgrade, so prompt behaviour, aspect-ratio handling and safety filtering are not guaranteed to match what your current calls return. If you have Imagen 4 model IDs pinned anywhere - a batch job, a CMS integration, an internal tool nobody has opened since last year - Monday is a hard stop rather than a warning banner. The same table shows embedding-2-preview already went dark on 10 August and gemini-robotics-er-1.6-preview follows on 31 August, which makes this a reasonable week to audit every model string in your codebase, not just this one.

Tuesday 18 August: Baidu reports

Baidu told the SEC on 31 July that it will report second-quarter 2026 results, for the quarter ended 30 June, before the US market opens on 18 August, with a management call at 8:00 a.m. US Eastern time - 8:00 p.m. Beijing time the same day.

We are not forecasting the numbers. The lines worth reading when they land are AI Cloud revenue and the capital expenditure behind it, because quarterly disclosure from a listed company is one of the few audited reads on how much compute a major non-US lab is actually buying, as opposed to what its model announcements imply.

Wednesday to Thursday, 19-20 August: Alibaba

Alibaba filed a board-meeting notice on 7 August stating that its directors meet on 19 August to discuss unaudited results for the three months ended 30 June 2026, and that the results announcement will be approved and published on 20 August. Several financial outlets also report a 7:30 a.m. Eastern call that day; Alibaba's own investor-relations site is unreachable from the environment we verify in, so we are citing only the dates the filing itself states.

As with Baidu, the disclosure that matters here is infrastructure rather than models: cloud growth, the capital expenditure trajectory, and whatever the company says about commercial pull-through for Qwen.

Just outside the window, and heavier

The centre of gravity is the following week. Both items below fall after 22 August, but each already has published detail, so they are worth putting in the calendar now.

Hot Chips 2026 runs Sunday 23 to Tuesday 25 August at Stanford, and the advance program is already up. Scheduled talks include:

• NVIDIA on its Rubin GPU, in a session titled "Driving the Era of Agentic AI", plus a separate talk on its Vera CPU
• AMD on the Instinct MI400 series, split across an architecture talk and a system-architecture talk
• Intel on Crescent Island, described on the program as a GPU designed for agentic AI inference, and on Diamond Rapids, its next-generation Xeon
• Microsoft on MAIA 200, presented as a data-centre-scale AI system
• Meta on its custom AI silicon, and Cerebras on rack-scale wafer-scale architecture
• Broadcom on Thor Ultra, an Ethernet NIC aimed at AI and HPC, and Samsung on an LPDDR5X processing-in-memory part for inference
• A keynote from Waymo's Daniel Rosenband on the compute demands of autonomous driving

This is the layer that eventually decides what an API call costs, which is why it matters to people who will never buy a chip. One in-window detail if you want to attend in person: Stanford dorm housing registration closes today, 16 August.

NVIDIA reports second-quarter fiscal 2027 results on Wednesday 26 August. Per its 29 July announcement, results are published at approximately 1:20 p.m. Pacific alongside written commentary from CFO Colette Kress, with the analyst call at 2:00 p.m. Pacific. The quarter ended 26 July 2026.

What is not scheduled

No major lab has announced a dated model launch inside this window. We checked OpenAI's news feed and API changelog, Anthropic's newsroom, and Google's Gemini deprecation and changelog pages; the most recent entries across all of them stop at 14 August. Google's table lists gemini-3.7-flash with an August 2026 release and "no shutdown date announced", which is a lifecycle status rather than a roadmap. If you read a confident claim this week that a particular frontier model ships on a particular day, no company has published that date.