A single robots.txt token now carries the AI-training preference
On 30 October Kevin Kelley of Anthropic submitted an IETF Internet-Draft, Machine-Readable Training Preference Signal, which binds usage preferences to paths using an ordinary robots.txt group keyed to the policy token __usage_ai-training. A site writes User-Agent: __usage_ai-training and then its Disallow and Allow lines; the draft adds no new directives and relies on RFC 9309 path matching. Google, Microsoft and OpenAI said the same day that they would honour it. Google and OpenAI keep their existing per-crawler controls working alongside the token; Microsoft says the new preferences replace its own NoArchive and NoCache tags. The draft is an individual submission that carries no IETF endorsement and expires 3 May 2027.
The dates differ by company. Microsoft honours the preferences from 1 February 2027 and stops reading the old tags after 1 March 2027, when a site that has recorded no temporary preference defaults to Allow for grounding. Google aims for early 2027, OpenAI for "the coming months", and Anthropic has begun with no date given. The token states a preference about use, not an instruction to stop crawling; a wildcard User-Agent: * group does not count as expressing one. European publishers should check whether a blanket crawler rule they already run would register as no preference at all. The draft's acknowledgements thank staff from the Financial Times, the BBC, Bertelsmann and the European Publishers Council.
Big Tech's AI capital spending has overtaken its cash generation
A Reuters analysis of LSEG data published 30 October puts combined capital spending at Microsoft, Alphabet, Amazon, Meta and Oracle at $556.9bn for the first nine months of 2026, roughly double the same period last year and above the $542.9bn of operating cash flow the five generated, which itself rose 29% to a record. It is the first time their capital spending has exceeded their cash generation in data going back to 2011. Free cash flow across the group was a $14bn deficit, against a $143.5bn surplus a year earlier. Buybacks fell 83% to $13.6bn from $81.8bn.
Shareholders received 9.3% of the cash generated, down from 27.4%, and the $50.6bn paid out came from reserves, borrowing or share sales. A European buyer signing a multi-year inference commitment is dealing with suppliers whose capacity build has outrun their own cash, so ask what fixes a quoted price and for how long. Amazon's third-quarter capital spending was $75.9bn net against $42.4bn of operating cash flow, and it raised $48.1bn of long-term debt in the quarter.
Quick Hits
- AWS grew 39% in the third quarter, to $46bn of revenue and $17.8bn of operating income on a $184bn run rate. Andy Jassy said AWS "can grow north of 40% in each of 2027 and in 2028".
- Artificial Analysis rebuilt its Intelligence Index. Version 5, published 27 October, runs nine evaluations, 60% on private or held-out test sets. Claude Opus 5.5 at max effort leads at 58; the best open-weight models with released weights score 32.
- Project ProtoStar launched on 29 October, with five US universities, the Astera Institute and Hugging Face as infrastructure partner, to publish publicly funded datasets, models and code in machine-actionable formats. Its first challenge, on climate risk, opens today.
- Pope Leo XIV urged states to move toward a ban on lethal autonomous weapons. On 30 October he encouraged "the States meeting in Geneva in November to take concrete steps", before the CCW review conference of 16-20 November.