Microsoft has cut its internal Claude spending by more than a third, and Meta has halved its internal Anthropic AI users, The Information reported Monday in an exclusive by Aaron Holmes, Jyoti Mann, and Kevin McLaughlin, read in full via the subscriber session. Microsoft had been on track to spend at least $1 billion a year on Anthropic technology internally; it cut Claude Code subscriptions, told staff to access Claude through GitHub Copilot (whose auto mode defaults to cheaper OpenAI models), and in its roughly 60,000-person cloud-and-AI group cut the per-employee monthly AI budget from $100,000 to about $10,000. Total Microsoft payments to Anthropic are roughly flat, because customer-facing Copilot Claude spend, on pace for $2 billion a year, keeps rising. At Meta, Claude Code users fell from about 60,000 earlier this year to about 30,000 as the company pushed its own Muse Code (over 6,000 users, built on Meta's Muse Spark models) and MetaCode (over 30,000); Meta still spent more than $105 million on Claude Code in a recent 28-day period, and it built its Muse consumer agent on Claude before switching to its own models for the shipped product. The backdrop: Anthropic passed $65 billion in annualized revenue in July, up more than 7x since January, while OpenAI surged about 70% to nearly $70 billion, and both OpenAI and Google have been undercutting Anthropic's prices.
Why it matters
This is the hyperscaler playbook made visible: adopt the best model, learn from the usage, then steer employees onto your own models and pocket the margin. Microsoft even keeps total Anthropic payments flat by shifting spend to customer-facing features while starving internal seats. For Anthropic, the signal is sharp: its fastest-growing revenue came from enterprises that are also its competitors, and those competitors are now the price-setters. Founders should price for the switching moment, not the honeymoon.
OpenAI announced textGrain, a system that embeds an invisible statistical signal in model output through word-choice distributions. API customers worldwide can now opt in to watermarking on select models, off by default; in the coming weeks, eligible ChatGPT and Codex text output in the EU will carry the invisible mark, answering the EU AI Act's machine-readable marking rule. OpenAI published unusually honest fragility numbers: at a 1% false-positive target, the detector caught about 80% of 200-token passages and 95% of 400-token passages for flexible writing like psychology, much less for constrained content like math; swapping 10% of words for synonyms dropped detection from about 92% to 66%, and 25% swapping dropped it to 17%. The company says the watermark costs essentially nothing in output quality on GPT-6 Astra and plans to open-source the technology; detector access stays limited to approved researchers and expert organizations.
Why it matters
This is the first time a watermark has shipped at ChatGPT scale, and the published failure modes matter more than the launch. A watermark that survives paraphrase only when the text is long and free-form cannot be the basis for content authentication, and OpenAI is saying so itself. For founders: treat watermark detection as a weak signal, not proof. For everyone else: the EU's marking rule just became real, and every lab will now ship one of these.
OpenAI introduced a visual ad format for ChatGPT that uses images for product inspiration, usage, and experiences, with testing starting later this month in the US inside image generation; ads are clearly labeled and kept separate from generated content. The company also stood up a full measurement stack: conversion integrations with Hightouch, Tealium, and LiveRamp; attribution through AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge, and Tenjin; and incrementality partnerships with Haus, Measured, and WorkMagic. Early numbers: WeightWatchers' attributed cost per acquisition on ChatGPT Ads was 15.3% below its blended paid-search benchmark (measured by DV Rockerbox); 67% of incremental purchases for wellness brand Dose came from net-new customers (WorkMagic); 93% of Portland Leather's visitors from ChatGPT Ads were new (Triple Whale). Brand-safety pilots with DoubleVerify and Integral Ad Science are underway, and businesses can sign up at ads.openai.com. OpenAI says ChatGPT reaches 1.2 billion people each week.
Why it matters
OpenAI is not testing ads; it is building an ad-tech company inside ChatGPT. A full attribution and incrementality stack on day one means it wants performance budgets, not brand budgets, and the early CPA numbers are aimed straight at Google's paid search. Growth teams should get into the pilot: a channel with this much intent and this little competition does not stay cheap.
a16z published the seventh edition of its Top 100 Consumer AI Apps, adding for the first time a ranking of observed US consumer card spending via YipitData alongside the Similarweb web and Sensor Tower mobile lists. The money picture: only 4.5% of US consumers had an active paid personal subscription to ChatGPT, Gemini, or Claude as of August, up from 2.1% a year earlier, and spending is extremely concentrated: the top 1% of payers average $903 a month against a $25 median, with the top 10% of spenders accounting for roughly half of all observed spend. ChatGPT has about three times the US paid subscribers of Claude or Gemini; Claude passed Gemini on US consumer subscribers earlier this year, and 7.3% of Claude's payers are on its $100-a-month Max plan, versus about 1% for ChatGPT and Gemini. Other signals: 29 of the top 50 vendors by spend appear on neither traffic list; only 11 products debuted this edition, the fewest ever; ChatGPT is the only product topping all three lists, joined on all three by Canva, Notion, Perplexity, Photoroom, and Suno; NSFW products were excluded but would have taken more than 20% of the web list. And the business-model note that matters most: OpenAI's ChatGPT advertising hit a $1 billion annualized run rate in August on 1.2 billion weekly actives, while among the top web products 84% use subscriptions, 64% use usage credits, only 14% use ads, and 2% take transaction fees.
Why it matters
Consumer AI has a monetization shape problem: usage is wide, but real money comes from a tiny slice of power users. The $903-a-month top percentile is not a consumer market; it is a prosumer market wearing a consumer costume. The white space is the other end: ads and transaction fees are barely used across the top 100, which is exactly why OpenAI is racing to build the ad stack. Founders should stop copying the subscription playbook and ask what their product can sell besides seats.
Uber is acquiring ezCater, the platform that lets businesses order catering from restaurants nationwide for meetings and events, in an all-cash $2.3 billion deal announced Tuesday morning, per The Information's AM Briefing. The move pushes Uber Eats beyond single meals into high-value group and workplace catering, its second major food-delivery expansion this year.
Why it matters
Uber keeps buying its way up the order-value ladder: single meals, then groceries, now corporate catering. Workplace food is recurring, high-ticket, and relationship-driven, which is exactly the kind of demand a delivery network can defend. Founders in food and logistics should read this as the template: the platform with the densest driver network wins each adjacent category by acquisition, not by building.
Salesforce signed a definitive agreement on September 29 to acquire Listen Labs, the AI market-research company, Sequoia announced. Founders Alfred Wahlforss and Florian Juengermann will join Salesforce AI Labs, and the product will sit alongside Marketing Cloud and Service Cloud. Sequoia, which led Listen Labs' 2023 seed before there was a demo, led the Series A, and followed through the Series B, frames it as perfect market-technology fit for LLMs: market research was a roughly $100 billion services market built almost entirely on language and untouched by technology, and LLM-led interviews plus LLM-led synthesis turned customer studies from annual events into real-time, continuous ones. Within months of launch, Google, Levi's, and Microsoft were customers.
Why it matters
This is the clearest acquisition signal yet that language-native services are the fastest LLM disruption targets: research, legal, support, any market where the work product is words. Salesforce is buying its way into that disruption rather than building it, which tells you the build-vs-buy math has flipped for incumbents. Founders selling into services markets should expect the acquirers to come knocking; founders competing with incumbents should expect them to buy a startup like yours instead of building.
Anthropic committed $100 million to train 10,000 Frontier Deployed Engineers by the end of 2027 through the Claude Frontier Academy, announced October 2. The first program, the Frontier Deployed Engineer Residency, is modeled on medical training: a multi-day in-person session with Anthropic engineers and licensed instructors, a simulated enterprise deployment ending in a graded practical that earns the Claude Resident Engineer badge, then a 12-week residency leading a real Claude use case at the trainee's own organization, with a final assessment earning the Claude Frontier Deployed Engineer badge; the first credentials are expected in early 2027. Cohorts are running now in San Francisco, New York, and London with engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk, with participation by nomination through Anthropic account teams. The Academy builds on the Claude Partner Network, where 46,000 firms have earned more than 175,000 certifications and about 4,000 completed Basecamp.
Why it matters
Anthropic has decided the scarcest resource in enterprise AI is not models but people who can take them from pilot to production, and it is minting those people inside its biggest customers. A vendor-owned credential pipeline is a classic lock-in play: the certified engineers become the reason accounts stay. Watch whether OpenAI and Google copy the credential model; whoever owns the deployment talent owns the renewal.
DayOne Data Centers filed for a $4.6 billion Nasdaq IPO, The Information's AM Briefing reported Monday. The filing lands amid market uncertainty about how much of the AI buildout is real demand and how much is vendor-financed.
Why it matters
A $4.6B IPO filing from a data-center company is the public-market test of the AI buildout thesis: do public investors still want to fund the picks-and-shovels layer at these valuations? If DayOne prices well, the neocloud funding window stays open; if it struggles, expect every private data-center raise to get harder. Watch the book-building, not the filing.
Kuaishou's Kling, the AI video-generation unit, has picked banks including Goldman Sachs, JPMorgan, and CICC for a Hong Kong IPO of more than $1 billion, The Information's AM Briefing reported Tuesday.
Why it matters
The first major IPO of a Chinese AI application company, and it is video generation, the category where Chinese labs have led. A $1B-plus Hong Kong listing tests whether global capital will pay up for Chinese AI apps the way it has for Chinese EVs. For US founders: Kling's public comps will set the valuation bar for every AI video startup raising its next round.
On October 1, SpaceX flew three orbital missions in a single day, its launches page shows: the Crew-13 astronaut mission to the ISS, now on orbit with return listed for April 2027; the Transporter-18 Falcon 9 rideshare from SLC-4E in California; and the NROL-97 Falcon Heavy national-security mission from LC-39A in Florida.
Why it matters
Three orbital launches in one day, spanning crew, rideshare, and heavy national-security payloads, is not a PR stat; it is the visible output of a launch cadence no competitor is close to matching. Cadence is the moat: every launch refines the machine that makes the next launch cheaper. For investors, the question is no longer whether SpaceX leads but how far the cost curve can keep falling.
