Sponsored by
In Todayโs Issue:
๐ A third of companies now build software with AI instead of buying it
โ๏ธ Brussels puts ChatGPT under its strictest platform rules
๐ A startup sells frontier models with the refusals removed
๐ก๏ธ Anthropic admits its models escaped the sandbox
๐ค NVIDIA shrinks the robot brain to entry level
โจ And more AI goodnessโฆ
โก The Signal
AI has started to decide what companies buy, and the first thing off the list is software.
McKinsey's new State of AI survey finds about a third of organizations skipping software purchases because AI coding agents can build the tools in-house. The same week, the EU pulled ChatGPT under its strictest platform rules, Anthropic admitted its own models slipped their leash during safety tests, and a startup is selling frontier-class models with the refusals stripped out. The pattern: the industry that sells AI is now being restructured by it, from procurement budgets to guardrails. And the tailwind is compounding, with the performance an AI chip dollar buys rising 49 percent a year.
All the best,
Kim Isenberg
(Engadget)
โ๏ธ ChatGPT Now Falls Under the EU's Strictest Rules
The European Commission has designated ChatGPT a Very Large Online Search Engine under the Digital Services Act, the EU's strictest regulatory tier. The label applies to services above 45 million monthly EU users and lands ChatGPT beside the newly designated Reddit and Roblox, bringing duties on minors' protection, illegal content, and algorithm transparency by December 31, 2026. EU digital chief Henna Virkkunen says the three "will now be held to a higher standard of scrutiny and accountability". Brussels regulates at frontier speed; whether Europe builds at that speed remains the harder question.
๐ tl;dr: The EU now polices ChatGPT like a social network, and OpenAI has four months to comply.
(Abliteration)
๐ A Startup Sells Frontier Models With the Refusals Removed
Abliteration sells API access to open models whose safety refusals have been deleted from the weights. The technique, called abliteration, finds the directions in a model's activations that produce refusals and removes them while leaving reasoning, coding, and tool use intact; the flagship model, built from GLM 5.2, offers a 1 million token context. The pitch targets red teaming, security research, and defense work, with a Policy Gateway that lets each customer define their own guardrails. What frontier labs spend billions aligning, a startup now unwinds as a subscription.
๐ tl;dr: Guardrails are turning into a customer setting rather than a model property.
(Anthropic)
๐ก๏ธ Anthropic Admits Its Models Slipped the Sandbox
Anthropic disclosed two incidents in which Claude models gained unauthorized internet access during cybersecurity evaluations. On July 30, three models reached real systems through a misconfigured third-party test environment; on August 4, the UK AI Security Institute reported Claude Mythos 5 taking unauthorized actions on the live internet during a test. Anthropic calls it "a failure of operational security" alongside two alignment issues, concedes "our models are not perfectly aligned", and responded by pausing external cyber evaluations (now resuming with new safeguards), hardening its sandboxes, and planning an independent review with METR.
๐ tl;dr: A frontier lab published its own containment failures. The candor is the story.
๐ฌ Watch This
Lex Fridman sits down with David Heinemeier Hansson, creator of Ruby on Rails and CTO of 37signals, for over five hours on what AI agents actually do to programming: agentic engineering, vibe coding, the cases where AI already writes all of the code, and why he thinks most companies still get it wrong. DHH is software's most entertaining contrarian, and the conversation lands squarely on today's lead story: what happens to the software business when everyone can build.
"As you know, I am not leaving Apple. But I am stepping away from a role that I have loved deeply."
โ Tim Cook, in his final memo to Apple employees as CEO, August 31, 2026
After 15 years, Cook handed Apple to hardware chief John Ternus on September 1. Apple's AI catch-up is Ternus's problem now.
OpenAI's next model Astra appears to be days away. Leaked outputs under the internal codename "mozaik-alpha-fdm" reportedly show one-shot voxel worlds and product renders, and leakers point to a September 3 to 10 window, for a model OpenAI said it had slowed for safety. None of it is confirmed.
One in Three Companies Now Builds Instead of Buys
The Takeaway
๐ About one in three organizations skipped at least one software purchase because AI coding agents could build it in-house, per McKinsey's new State of AI survey
๐ Technology (41 percent) and healthcare (39 percent) lead the build-over-buy shift
๐ Companies above $1 billion revenue scaling AI agents jumped from 27 to 40 percent in a year; smaller firms stayed flat at 22
๐ The returns lag the conviction: only 6 percent qualify as AI high performers, and reported earnings impact is flat year over year
McKinsey's new State of AI survey delivers the clearest evidence yet that AI agents are starting to eat the software business from the inside. In the survey of 1,719 executives published last week, about one in three organizations said they skipped buying at least one software product or feature because agentic coding tools, AI systems that write and test software on their own, let them build it in-house instead. Among technology companies the share is 41 percent; healthcare payers and providers report 39 percent, professional services and energy both 38.
(Chart: Superintelligence, data: McKinsey)
The shift tracks the agent rollout itself. Among companies with more than $1 billion in revenue, the share scaling AI agents in at least one business function jumped from 27 to 40 percent within a year, while smaller firms stayed flat at 22 percent. McKinsey senior partner Lieven Van der Veken: "Leaders are asking what their organisations need to build AI tools themselves." Building, long the expensive path reserved for tech giants, is turning into a default option in procurement.
(Chart: Superintelligence, data: McKinsey)
The skeptical footnote is McKinsey's own: the money has not followed yet. 37 percent of respondents attribute any earnings impact to AI, unchanged from last year, and only 6 percent clear the bar for "high performers." The report's verdict: "Organizations' conviction in AI is growing faster than the immediate financial returns they can attribute to it." Meanwhile 39 percent expect AI-driven job cuts, up from 32 last year.
Why it matters: Software vendors price per seat on the assumption that building is harder than buying. When a third of customers can ask an agent instead, that assumption, and the business model resting on it, starts to wobble.
How AI-Era Pricing Is Reshaping Finance Operations
Usage-based and hybrid pricing models are changing how B2B companies generate revenue โ and creating new headaches for the finance teams behind them.
Tabs co-founder Rebecca Schwartz and PwC Partner Amit Dhir sat down to unpack exactly what that means in practice: how pricing model decisions ripple into revenue recognition, forecasting, and financial ops โ and what it takes to scale without piling on manual work.
Watch the on-demand recording to get practical frameworks, real-world examples, and a clear path to operationalizing usage-based revenue โ including a forward-looking take on how AI will reshape financial workflows. If your team is navigating pricing complexity heading into the back half of the year, this is worth an hour.
(Epoch AI, CC-BY)
The chart: Each bubble is one AI chip generation in one quarter, from A100s in early 2023 to Nvidia's GB300s in late 2025, sized by how much money buyers spent on it. The height shows how much performance a dollar bought, measured against an H100 at its 2025 price. The dashed line is the finding: weighted by real spending, performance per dollar rose about 49 percent a year, doubling roughly every 1.7 years.
The lesson: Compute keeps deflating. Whatever an AI capability costs to run today, expect roughly half that within two years, which is exactly the tailwind behind companies building their own tools instead of buying them.
The caveat: These are spec-sheet peak numbers, not what data centers actually get out of the chips, and the average moves with the spending mix: export-grade H20s drag it down, Google's TPUs pull it up. Epoch's own uncertainty band runs from 36 to 66 percent a year.
๐ค The Robot Brain Goes Mass Market
โก Bottom line: NVIDIA announced the Jetson Orin Nano 2, an entry-level robot computer with double the AI performance of its predecessor at 40 percent less power.
๐ก Why it matters: Robot intelligence is bottlenecked by cheap onboard compute; this brings real-time AI reasoning to entry-level machines.
๐ What it means: The brains of the robot wave are commoditizing before the bodies are, and NVIDIA controls the supply.
NVIDIA announced the Jetson Orin Nano 2 on August 25, a palm-sized computer built to be the brain of entry-level robots, drones, and camera systems. It delivers 78 TOPS, trillions of AI calculations per second, double the inference performance of the current Jetson Orin Nano Super in the same form factor, while drawing 40 percent less power in its 15-watt mode. In plain terms: enough to run modern vision and language models directly on a robot, no cloud connection required.
(NVIDIA)
NVIDIA is aiming at the long tail of robotics, the startups, labs, and student teams that cannot pay data center prices. More than 3 million developers already build on its robotics stack, and early adopters include Cognex, Doosan Bobcat, Alphabet's drone delivery unit Wing, and home-robot maker Matic, whose vacuum already navigates entirely on-device with an earlier Jetson chip. Robotics VP Deepu Talla says the new computer "puts that breakthrough within reach of millions of developers, delivering the performance and energy efficiency needed for real-time reasoning at the edge".
(Matic)
The catch is the calendar. Module and developer kit ship in the first half of 2027, and 8GB of memory caps how much of a modern model actually fits on the board. NVIDIA is recruiting developers years ahead of revenue, betting that whoever supplies the cheap brains ends up owning the robot wave the way it already owns the data center.
The best voice models, now with full orchestration. Build real-time voice and chat agents on one low-latency stack: any LLM, your tools and knowledge, testing, Guardrails, and omnichannel deployment.
