More than 1,200 staff at OpenAI, Anthropic, DeepMind and Meta signed a letter asking Washington for tools to slow AI down, and OpenAI and Anthropic endorsed it as companies. Microsoft's Azure passed $100 billion a year and Copilot reached 30 million paid seats. The FCC banned foreign-made humanoid robots outright. Three of today's stories share a shape, an announced number that does not survive the detail, and here is the Pulse with sources and what to do about each.

More than 1,200 employees across OpenAI, Anthropic, Google DeepMind and Meta signed "Pacing the Frontier" on 29 July, asking the US to back an international effort to build the technical and governance tools needed to deliberately pace automated AI development. Signatories include Anthropic CEO Dario Amodei, OpenAI chief scientist Jakub Pachocki, Meta AI chief scientist Shengjia Zhao and DeepMind's head of AI safety Anca Dragan. The letter does not ask for a pause now. It asks for the machinery to slow things later, if AI systems start designing their own successors faster than anyone can follow what they are doing.

Why it matters: Read the signature block before the headline. OpenAI and Anthropic endorsed this as companies. Google DeepMind and Meta did not, though senior people at both signed in a personal capacity, including Meta's chief AI scientist, whose CEO published the opposite argument the same day. For anyone building on frontier models, the signal is that senior scientists inside all four labs expect a moment when pacing is needed, and the machinery to do it does not exist yet.

What to do:

Source: Fortune

Microsoft reported FY26 Q4 revenue of $90 billion on 29 July, ahead of the $87.7 billion analysts expected. Azure crossed $100 billion in annual revenue for the first time, growing 43% in the quarter, and Microsoft 365 Copilot passed 30 million paid seats. Commercial remaining performance obligations, meaning contracted revenue not yet recognised, hit $678 billion, up 84%. Microsoft also booked a $3.2 billion gain on its Anthropic stake and roughly $5 billion on OpenAI.

Why it matters: Paid seats are the number worth watching. Pilot counts and usage stats have carried the AI adoption story for two years, and 30 million people whose employers renew a line item every month is harder evidence than either. It also sets the benchmark your own AI spend gets measured against internally, because once seats are demonstrably converting to renewals, "we are still evaluating" gets harder to defend to a board.

What to do:

Source: Fortune

The FCC voted on 28 July to bar imports and sales of foreign-made humanoid and quadruped robots, alongside certain power inverters, citing confirmed backdoors in hardware and firmware capable of undocumented outbound communication. In practice the rule targets Chinese manufacturers, and Beijing has formally objected. It extends the same equipment-authorisation regime already used against Huawei and ZTE networking gear to robotics.

Why it matters: Chinese manufacturers make most of the affordable humanoid and quadruped hardware on the market, so the cheap end of robotics has effectively left the US catalogue. If you have a warehouse, inspection or site-security robotics pilot running on that hardware, the procurement question moved from cost to legality this week. It also marks where restrictions go next: physical AI, not only models.

What to do:

Source: The Washington Post

A 404 Media investigation found AI companies placing bulk book orders through middlemen, ranging from 1,000 copies to as many as one million in a single transaction. ISBNdb pivoted its business to bulk-buying books for AI labs, and one anonymous bookseller's weekly volume went from around 20 books to several hundred. The books are cut apart, run through high-speed scanners, and destroyed, because destructive scanning is cheaper than the careful alternative.

Why it's interesting: Courts have ruled that training on legally purchased books counts as fair use, so the cheapest lawful route to clean, pre-AI human text now runs through a paper shredder, and buyers use middlemen specifically to keep their names off the receipts.

Key takeaway: Human-written text that predates AI is scarce enough to be worth industrial-scale acquisition, which is worth remembering before you give your own archive away.

Source: Tom's Hardware

OpenAI published results on 30 July showing GPT-5.6 Sol scoring 38.3% on ARC-AGI-3, up from the 7.8% recorded on the official harness, after enabling two API settings: retained reasoning, which preserves chain of thought between steps, and compaction, which summarises old context instead of truncating it. The higher score puts it ahead of Claude Opus 5's 30.2%. ARC Prize co-founder François Chollet accepted the settings as legitimate because they are available to every API user, while flagging a parity problem in how the benchmark gets run.

Why it's interesting: Same model, same benchmark, a near-fivefold swing from two configuration flags, which suggests the harness around a model can matter more than the model.

Key takeaway: Before switching models over a disappointing result, check whether you are running the one you have on default settings.

Source: The Decoder

OpenAI research published on 27 July, drawn from roughly 800,000 work-related ChatGPT messages, found 43.5% of job-specific AI use falls outside the user's own occupation. Marketers doing analyst work, engineers drafting legal-adjacent documents, operators building things they would previously have briefed out to someone else.

Why it's interesting: The productivity pitch assumed people would do their existing jobs faster. The data says a lot of them are doing other people's jobs instead.

Key takeaway: Look at what your team stopped outsourcing this year without anyone deciding to, because that is where AI actually moved your cost base.

Source: OpenAI

Context: Microsoft just showed 30 million people paying for AI seats, and that number becomes the internal benchmark your own AI spend gets compared against. Separately, OpenAI's own research found 43.5% of work AI use crosses job boundaries, so the value may not sit in the team you licensed it for. Audit before your next renewal date, not after it.

Prompt: "You are a sceptical CFO reviewing our AI spend. Here is our AI tooling: [list each tool, seat count, monthly cost, and the team it was bought for]. Here is what people report actually using it for: [paste survey answers, notes, or usage exports]. Produce: (1) cost per active user per tool, flagging any tool where under half the seats are used weekly, (2) tasks people are doing with these tools that sit outside their job description, (3) where two tools overlap enough that one could be cut, (4) three questions to ask each vendor at renewal based on our real usage rather than their pricing page."

Do not: Do not accept licence counts as usage. If a tool has no usage data attached, mark it unknown rather than assuming the seats are active.

If/Then: If any tool comes back under 50% weekly active seats, then cut the seat count to match real usage at renewal instead of negotiating a discount on the same volume.

Example: A 40-person agency ran this across six AI subscriptions and found two tools bought for the design team were mostly being used by operations for meeting notes. They cut 22 unused seats, moved that budget into the one tool the whole team actually used, and took roughly £1,100 a month out of the bill without removing anything anyone was using.

Mark Zuckerberg published a Wall Street Journal op-ed on 28 July arguing superintelligence should be widely distributed rather than concentrated in a few institutions, writing that "the defining question of our age isn't whether superintelligence will exist, but who will have access to it." The pacing letter landed the same day, and Meta AI's chief scientist Shengjia Zhao signed it.

Why it's important: Two opposite arguments about AI's biggest risk came out of one company on one day, one from the CEO and one from the chief scientist's signature. Meta is not unusual here. That split runs through the whole industry, and it is why no coherent policy has arrived yet. Open weights are also Meta's commercial strategy, which does not make the argument wrong, but does make it interested.

Business takeaway: Expect AI regulation to stay unsettled for longer than vendor roadmaps assume, and avoid architecture that only works under one policy outcome.

Source: Forbes

The Information reported on 29 July that ChatGPT is approaching one billion weekly active users, seven months later than OpenAI originally projected. The company has more than 50 million paid subscribers, expects around $25 billion in revenue this year against roughly $25 billion in cash burn, and enterprise now accounts for 40% of revenue, heading toward 50% by year end. The Information separately estimates Anthropic is generating at least 35% more revenue than OpenAI.

Why it's important: The headline number is enormous and the trajectory still missed its own target by seven months, while a rival with a fraction of the consumer footprint is out-earning it. Consumer scale and revenue have come apart, and the enterprise contracts are where the money sits.

Business takeaway: Pick AI vendors on enterprise commitment and contract terms rather than on consumer user counts.

Source: PYMNTS

The UAE launched what it describes as the world's first fully integrated AI-powered judicial platform on 28 July. It analyses case files, retrieves relevant legislation and precedent, generates analytical reports and drafts legal documents across the court process. Officials stressed that final judgments remain exclusively with human judges under full oversight.

Why it's important: This is the first national court system to run AI across the whole workflow rather than piloting it in one function. Whatever comes out of it, faster case processing or a documented failure, becomes the reference case every other jurisdiction cites, including ones your business operates in.

Business takeaway: Watch UAE court timelines over the next year, because they are the first real-world data on AI inside high-stakes regulated decisions.

Source: Gulf News

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas