THE SHORT VERSION
● Siri AI opens to all developers — App Intents, the 27.3 betas, six languages. App Store submissions in the spring.
● Anthropic FlashProve — cryptographic proof that a reply came from the model claimed. Apache 2.0, about 110 tokens/sec on one H100.
● AMD data centre $8.5B — up 97% year over year. Anthropic, Meta, Microsoft, OpenAI and Oracle all on Helios racks.
● ChatGPT at 1.2B weekly users — reported, not published by OpenAI. Up about 200 million in two months.
Apple opens Siri AI to every developer
Apple published a developer preview on 3 November that lets any app bring its content and actions into Siri AI through App Intents. It is live now in the iOS 27.3, iPadOS 27.3, macOS 27.3, watchOS 27.3 and visionOS 27.3 betas. Several actions can be chained into what Apple calls skills.
Launch languages are English, French, Japanese, Korean, Portuguese and Spanish, with Traditional Chinese, Norwegian, Turkish and Vietnamese named as coming. App Store submissions open in the spring, and Apple says the feature ships to customers in a few months, so this is a build-now window rather than a release.
Three constraints worth reading before you plan around it. Siri AI needs an Apple Intelligence-enabled device set to a supported language. It is unavailable in the EU on iOS, iPadOS and watchOS, with Apple saying it is working with regulators on a path forward. And it is not available to users under 13.
The line most coverage skipped. Apple’s own note says server-side Apple Intelligence features, Siri AI included, carry daily usage limits, and that “expanded access to such features will be available for a fee in the future.” If your app depends on Siri AI for a core flow, you are building on a metered surface whose price has not been set.
Anthropic releases FlashProve, proof of which model answered you
Anthropic published FlashProve on 2 November, an open-source prototype for provable inference. It lets a user verify, from a cryptographic proof plus a commitment to the model weights, that a response actually came from the model a provider claims to be running — without the weights themselves being revealed.
The benchmark ran on Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model using about 17 billion parameters per token at 4-bit weights. Proof generation hit roughly 110 tokens per second on a single H100, producing about 19 megabytes of proof per 1,024-token request. The honest number is the overhead: proving costs roughly 70 to 95 times the GPU time of simply running the model on vLLM.
That multiple is why this is a prototype rather than a product, and Anthropic presents it as such. It matters anyway, because the problem it addresses is one nobody currently has a way to check. When a provider says you are talking to the flagship model, there has been no mechanism to confirm you were not quietly routed to a cheaper one. The code, Lean proofs and documentation are released under Apache 2.0, with collaborators at EPFL, MIT, StarkWare and Succinct.
AMD data centre revenue up 97% as Helios racks go live
AMD reported third-quarter revenue of $13.4 billion on 3 November, up 45% year over year, with the data centre segment at a record $8.5 billion, up 97%. Fourth-quarter guidance is approximately $17 billion, plus or minus 3%.
The detail that matters for anyone choosing an inference provider is the customer list. AMD says Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are all running models and workloads on the first Helios rackscale systems, built on Instinct MI450 Series GPUs and 6th Gen EPYC CPUs. That is the five largest buyers of frontier compute named on one press release, on non-Nvidia silicon.
ChatGPT reported at 1.2 billion weekly users
OpenAI’s Adam Fry gave figures at a conference this week putting ChatGPT at 1.2 billion weekly active users, up roughly 200 million in two months, with ChatGPT search at about 20% of Google’s query volume, Voice at 250 million weekly users and 1.5 billion images generated weekly.
Treat these as reported, not published. The numbers come from remarks relayed by Big Technology, not from an OpenAI release or filing. We have not found a primary source and are not treating them as confirmed.
Money keeps moving into inference and incident tooling
Two rounds closed this week. Fal raised $1.3 billion at a $16 billion valuation led by Bessemer Venture Partners, double its $8 billion mark in April, on reported annualised revenue near $800 million. Traversal raised a $120 million Series B at a $1 billion valuation, co-led by Kleiner Perkins, Redpoint and Khosla, taking total funding to $173 million.
Both figures are from press reporting rather than company releases. The pattern is the readable part: the capital is going to the layer underneath the models — serving inference, and working out what broke — rather than to another chat interface.
What to act on today
If you ship an iOS app, the 27.3 developer preview is the item with a deadline attached. App Store submissions open in the spring, which means the apps that feel native to Siri AI on day one are the ones being wired up this month. Start with the one or two actions a user would genuinely ask for out loud, not your whole feature surface.
If you buy inference, FlashProve is worth reading even though you cannot deploy it. It names a problem your contract probably does not cover: there is no clause and no test that tells you which model actually served your request. Asking a provider how they would demonstrate it is a reasonable question now in a way it was not last week.