The week ahead: Broadcom reports, and almost nothing else is scheduled

The 30 August to 5 September window has almost nothing scheduled that an AI lab controls. No model release carries a published date, no price change takes effect, no API endpoint shuts down and no regulatory deadline falls inside it. What the week does have is earnings — four of them — and one product change that quietly alters how a Google tool meters your work. The single most informative event is Broadcom's third quarter on Wednesday, which is the clearest public read on custom AI accelerator demand outside NVIDIA.

Wednesday: Broadcom reports the quarter it guided to $16 billion of AI silicon

Broadcom said on 3 August that it will report third-quarter fiscal 2026 results after the close on Wednesday 2 September, with a conference call at 2:00 p.m. Pacific (5:00 p.m. Eastern).

The reason to care is the guide it is reporting against. In its June quarter release, Broadcom reported $10.8 billion of AI semiconductor revenue, up 143% year over year, and told investors to expect that line to grow more than 200% year over year to about $16.0 billion in the quarter ending 2 August.

Consolidated revenue guidance was approximately $29.4 billion, an 84% increase, with non-GAAP operating income at roughly 67% of revenue. Those are Broadcom's own forecasts, made three months ago; Wednesday is when they meet an actual result.

This lands a week after NVIDIA reported $96.2 billion for the quarter ended 26 July. The two companies read the same demand from opposite ends: NVIDIA sells merchant GPUs, Broadcom builds the custom accelerators that large buyers design instead of buying them. A Broadcom number that clears $16.0 billion says the custom path is still widening; one that misses says the buyers pulled back, or that the schedule slipped. The forward guide for the October quarter will matter as much as the print itself.

Wednesday: Gemini Notebook starts counting compute instead of prompts

Google said on 28 August that Gemini Notebook's new usage limits begin rolling out to consumer accounts on web and mobile on 2 September. The old daily prompt cap is replaced by a compute budget that refreshes every five hours, and what draws it down is broader than the number of questions you ask: Google says the limit factors in prompt complexity, chat length, the number of sources in the notebook, and which features you use.

If you hit the ceiling, outputs like Video Overviews and Slide Decks can be deferred and generated later.

Google published no numbers alongside this, so nobody outside the company can say what any given account will actually get. The structural change is clear enough, though, and it is the kind that changes how you work rather than what you pay: in a source-count-weighted system, a large reference library is no longer free to keep loaded, and a long-running chat costs more than the same questions asked in fresh sessions. Anyone who has been treating Notebook as a place to park every document they own has a reason to prune before Wednesday.

The rest of the earnings

• Tuesday 1 September — GitLab reports second-quarter fiscal 2027 results, for the quarter ended 31 July, after US markets close, with a call at 4:30 p.m. Eastern. The company describes itself as "the intelligent orchestration platform for DevSecOps"; it is one of the few public quarterly reads on whether AI features are pulling developer-tooling budgets up.
• Wednesday 2 September — HPE holds its fiscal 2026 third-quarter call at 3:30 p.m. Central (4:30 p.m. Eastern), announced on 12 August. In its June quarter HPE raised full-year revenue growth guidance to 29–33%, put Networking growth at 72–75%, and lifted free cash flow guidance to at least $3.5 billion.
• Thursday 3 September — Zscaler releases fourth-quarter fiscal 2026 results after the close, with a call at 1:30 p.m. Pacific, per its 6 August notice.

What is not scheduled

This is worth stating plainly, because a week-ahead is easy to pad with things people expect rather than things anyone has committed to. Nothing on the following list has a date this week.

• No model launch. Neither OpenAI's newsroom feed, its API changelog nor Anthropic's newsroom carries a dated announcement landing between 30 August and 5 September. New frontier models are widely anticipated; none has an announced date, so none appears here.
• No API retirements. OpenAI's deprecations page lists nothing shutting down in this window. The next dates are 24 September, when the Videos API and Sora 2 variants are removed, and 28 September for legacy snapshots including gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106.
• No regulatory deadline. A Federal Register query for documents mentioning artificial intelligence with effective dates in this window returns a single result, a Health and Human Services rule on children and family programs that is not an AI rule.

What it adds up to

For anyone choosing models or budgeting API spend, this is a week where nothing moves under you except one consumer product's meter. Broadcom's Wednesday number is the week's only genuine signal, and it is a supply-side one: it says how fast inference capacity outside NVIDIA is actually being built, which sits upstream of every eventual argument about token prices and rate limits. It will not change anything you can act on by Friday. It is the better indicator of what the next two quarters cost.