The week ahead: Hot Chips talks first, NVIDIA reports second

The week of 23-29 August is a hardware week. Three days of chip architecture talks at Stanford run straight into NVIDIA's quarterly results on Wednesday and Marvell's on Thursday. Everything listed below has a date the organiser or the company published itself. Nothing from a frontier lab does: no model release, price change or API deprecation has an announced date inside this window.

Sunday to Tuesday: Hot Chips 2026

Hot Chips 2026 runs Sunday 23 August through Tuesday 25 August at Memorial Auditorium, Stanford, with tutorials on the Sunday and two conference days after it. The published program is dense with AI accelerators, and it is one of the rare venues where the companies describe the parts rather than the products.

The Monday GPU session has NVIDIA presenting its Rubin GPU, AMD giving two talks on the Instinct MI400 series - one on the GPU architecture, one on the system around it - and Intel presenting Crescent Island, a GPU it describes as designed for agentic AI inference. Earlier in the day, NVIDIA's Vera CPU appears in the first CPU session, with Intel's next Xeon, Diamond Rapids, and an Arm chiplet-based server SoC in the second.

Tuesday is the denser day for anyone tracking who is building their own silicon. The two AI sessions list Meta on its custom accelerators, Cerebras on rack-scale wafer-scale systems, SambaNova on the SN50 RDU, Microsoft on the MAIA 200 accelerator, and Google on what it calls the eighth-generation TPU family - two chips, one optimised for training and one for serving. OpenAI has a slot of its own, presented by Richard Ho, Ravi Narayanaswami and Chris Leary. The networking session carries Broadcom's Thor Ultra Ethernet NIC and NVIDIA's BlueField-4 and Spectrum-X.

Two caveats worth holding. These are vendor presentations: architecture disclosures and vendor-run numbers, not independent evaluation, and the program page itself is marked subject to change. And Hot Chips describes hardware that mostly has not shipped, so nothing said this week changes what an API costs today. It changes the shape of what is available to price in 2027.

Slides are posted at 7:00 AM PDT on the day of each talk, so the material becomes public in three tranches rather than all at once.

Wednesday: NVIDIA's quarter

NVIDIA reports second-quarter fiscal 2027 results on Wednesday 26 August, for the quarter that ended 26 July. The company said on 29 July that results go out at approximately 1:20 p.m. PT, with written commentary from CFO Colette Kress posted immediately after, and a call at 2 p.m. PT whose Q&A is limited to financial analysts and institutional investors.

The context has shifted since the last report. On 10 August NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish compute financing platforms intended to mobilise over $500 billion of third-party capital for AI infrastructure over time, and on 17 August it said it had secured land, power and shell capacity in Ohio through SB Energy, providing credit support on the buildout, with OpenAI named as the customer. Those are commitments about how the buildout gets funded, not revenue. The interesting part of Wednesday is less the quarter than what gets asked about the financing.

Salesforce releases second-quarter fiscal 2027 results the same day after the close, with a webcast at 2 p.m. PT, per its 5 August announcement. It is the closest thing on the calendar to a read on whether enterprise agent products are converting into revenue rather than pilots.

Thursday: Marvell

Marvell reports second-quarter fiscal 2027 results on Thursday 27 August, with a call at 1:45 p.m. PT, announced on 3 August. The same release scheduled an investor day for 6 October. Marvell describes itself as a data infrastructure semiconductor company, and its business is weighted toward custom accelerators rather than merchant GPUs, so its guidance reads as a second, partly independent signal about how much of the buildout is going into designs that are not NVIDIA's - the same question Hot Chips spends Tuesday afternoon on.

What has no date

No frontier lab has announced a dated model launch, pricing change or deprecation inside 23-29 August. Expectations circulate; none of them carry a confirmed date, so none of them appear here.

The nearest dated items land just past the window. Google's Gemini API release notes put the shutdown of gemini-robotics-er-1.6-preview at 31 August, announced on 30 July when the ER 2 preview endpoints launched. Broadcom's VMware Explore runs 31 August to 3 September in Las Vegas, with the private AI cloud as its stated theme.

Other technology companies also report during the week. We list only the ones whose dates we could open and read in the company's own announcement; a few investor-relations sites were unreachable from here, and an unverified date is not worth printing.