Nvidia has agreed to backstop as much as $105 billion for the PORTS-Pike Technology Campus in Pike County, Ohio — a site that SB Energy will build and own, and that OpenAI will lease for 20 years. The important nuance, and the one most headlines missed: this is not a direct investment. It is a guarantee. If OpenAI cannot pay its lease and power bills, Nvidia covers them. Nvidia is separately putting $1.5 billion of real equity into SB Energy. In exchange, OpenAI has committed to using Nvidia chips exclusively at the site — potentially 1.5 million GPUs, worth $150–200 billion in revenue to Nvidia through 2030. OpenAI is contracting for roughly 8 gigawatts of computing capacity, with the first 800 megawatts due online in 2028. Fortune noted on Wednesday that the final figure came in $145 billion below earlier reports, which is feeding the argument that chip demand is being propped up by circular financing between suppliers and customers.
OpenAI released a teen-specific version of ChatGPT on Tuesday, built around learning rather than open-ended chat, with stronger default protections, “healthy use” features designed to discourage over-reliance, and a new layer of parental controls. On the same day it announced a partnership with CodeAI aimed at building AI literacy in students — teaching them to think critically about the technology rather than simply operate it. Read together, the two announcements are a deliberate repositioning: OpenAI wants to be seen as the responsible option for minors at a moment when child safety is the single most politically dangerous issue facing consumer AI. For communications teams, this is the template competitors will now be measured against.
Anthropic has begun embedding hidden, machine-readable watermarks into text written by Claude. You cannot see them; software can. The mechanism is based on Google DeepMind’s SynthID-Text approach, and it survives copy-and-paste — it will even appear when Claude has only corrected your spelling. Coverage extends across Claude, the Claude API, Claude Code, Claude Cowork and Claude Tag. The trigger is Article 50 of the EU AI Act, enforceable since 2 August, which requires AI-generated content to carry machine-readable marks; Anthropic has chosen to apply it globally rather than only in Europe. David Gerard’s Pivot To AI captured the pushback bluntly this week, and the practical concern for any comms function is real: if a press release, statement or briefing note passes through Claude at any point, it may now carry a detectable signature. That is a workflow question worth answering before a journalist asks it.
OpenRouter is plumbing most people have never heard of: a single connection point that lets a developer reach more than 400 different AI models, automatically picking whichever is cheapest, fastest or most reliable for a given request. Stripe — the payments company — has bought it for north of $7 billion, a 5.4x markup on the $1.3 billion valuation OpenRouter carried in its Series B just three months ago. Stripe now owns both the toll booth and the meter for a large slice of AI traffic, gaining eight million developers in the process. There is a geopolitical footnote worth filing: a CNBC investigation in July found Chinese-origin models accounted for 46% of US enterprise token usage flowing through OpenRouter.
In a post published Tuesday, OpenAI said it is strengthening monitoring, alignment and security work around frontier models, and that these safeguards will now govern the pace at which it ships. In plain terms: the company is saying it will hold models back if their hacking-related capabilities outrun its ability to contain them. It arrived alongside a companion piece, “The Defender’s Window,” arguing that AI is reshaping cybersecurity for attackers and defenders simultaneously. The timing is notable — it lands the same week OpenAI is reported to have reduced its AI safety team, a contradiction reporters have already begun probing.
Google ships Gemini 3.7 Flash three weeks after 3.6Coding benchmark FrontierCode 1.1 jumps from 34.4% to 43.6%; DeepSWE from 49% to 65.3%. Introductory pricing of $0.75 per million input tokens holds until 31 December 2026, then doubles.
Perplexity’s Comet browser goes enterprise, adds a confirmation layerComet — the browser that can click, fill forms and buy things for you — is now deployable across a company via mobile device management, and now demands explicit human approval before sensitive actions. It also shipped inside Samsung Internet.
Perplexity · Aug 2026
05
OpenAI adds opt-in desktop activity logging to ChatGPTThe desktop app can now watch what you are doing on your machine to build context, if you let it. AI Breakfast flagged the obvious enterprise question: what happens when that machine holds embargoed material.
Anthropic reportedly bids $6 billion for DecartThe Neuron reports Anthropic is chasing the real-time video generation startup, alongside claims the company is holding a model it has chosen not to release. Treat both as unconfirmed.
EU AI Act transparency rules now being actively enforcedSince 2 August, chatbots must identify themselves, deepfakes must be labelled, and AI-made content must carry machine-readable marks. Penalties reach €15 million or 3% of global turnover.
OpenAI funds 14 independent AI policy projectsGrants aimed at economic opportunity and societal resilience — an attempt to seed a friendly research ecosystem ahead of the next regulatory cycle.
OpenAI · Aug 17
14
Google lets users remove its visible AI watermarkA striking divergence: Anthropic hardens its invisible marks the same week Google makes its visible one optional. The two approaches will not stay compatible for long.
Data centres get too hot for Texas politiciansGovernor Greg Abbott spent years courting power-hungry industries; Dallas-Fort Worth is now the second-largest data centre market in the US and the ERCOT grid is straining. The politics have turned.
Pivot To AI · Aug 18
17
Voice-cloning scams are emptying bank accountsThe Automated details the current attack pattern: a cloned voice of someone you know, used live on a phone call. Worth briefing your team on — executives are the primary target.
The curious economics of a $6 AI agentAzeem Azhar on what happens when an autonomous agent costs less than a coffee — and on Amazon’s “catastrophically expensive” AI spending.
The dedicated Substack tools were not available to this run. The newsletters below were captured through their Inoreader / kill-the-newsletter feeds instead, which covers your Substack subscriptions but not any subscriber-only sections that sit behind a login.
The core argument: frontier models have become expensive enough, and open-weight models good enough, that almost nobody should be sending every request to one model any more.
“Routing” means automatically sending each request to whichever model handles it best for the price — a cheap open model for summarising, an expensive frontier model for hard reasoning.
This is the exact demand curve that made OpenRouter worth $7 billion to Stripe two days later. Latent.Space effectively explains the thesis behind the deal.
Practical read for you: if your team runs any AI workflow at volume, routing is where the cost savings live.
A reader survey framed around the gap between agent hype and agent habit — most people who have access to autonomous agents still do not use one daily.
Ben’s framing: the blocker is not capability, it is trust and the absence of a repeatable job to hand over.
Pairs neatly with the One More Thing in AI piece below, which argues the fix is to stop asking AI questions and start assigning it jobs.
Azhar is recruiting an economist to extend Exponential View’s tracking of actual AI revenues and spending, following its June state-of-the-AI-economy report.
His Monday data roundup, published the day before, noted AI revenues have continued climbing since that June baseline — worth reading alongside the Nvidia–OpenAI backstop story, which points the other way.
Also from Azhar this week: “The market misread Google’s AI exodus,” arguing the Jeff Dean and Sanjay Ghemawat departures were misread as a capability loss rather than a generational handover.
Explains why OpenAI needed Nvidia to co-sign the Ohio lease at all — a credit signal to the lenders and landowners behind the project, not a cash injection.
Also reports OpenAI has cut its AI safety team, which sits awkwardly against Tuesday’s “pacing model development” post.
Includes Microsoft’s chip math, and in Monday’s edition, the claim that if you use Cursor, Musk’s SpaceX now owns it.
A practical guide to managing Claude’s context window — the amount of material Claude can hold in mind at once before it starts forgetting the beginning of a conversation.
Plus the Amazon book-scanning story, and in Monday’s edition, a clear explainer on how Claude’s watermarks actually work versus Google’s removable one.
Traces Texas from courting bitcoin miners in 2021 to a grid under pressure in 2026, with Dallas-Fort Worth now second only to Northern Virginia for data centre concentration.
Gerard also released a 50-minute interview with the tech commentator Tante on the AI bubble, available now to Patreon and YouTube supporters and publicly on Friday.
Written specifically for business readers: what Anthropic is doing, what software can detect, and what it means for corporate content pipelines.
The line that matters for a comms function — you will not see the mark, but software may.
See also · Top Stories
From Your AI Feeds
Inoreader AI Folder.
17 articles in the last 24 hours from your “AI” folder. The day was dominated by OpenAI, which published five separate posts, and by the Nvidia backstop story rippling through every newsletter you subscribe to. The most distinctive material came from GDELT, which ran three experiments using Gemini on archival and journalistic research.
Kalev Leetaru pointed Gemini Deep Research at a live piece of US sanctions legislation and asked it to find the unintended consequences and loopholes a drafter might have missed. The experiment is a genuinely novel use of a research model — not summarising a document, but stress-testing it adversarially. For anyone who reads regulation for a living, the method transfers directly.
GDELT archives two Iranian broadcasters: Persian-language IRINN, aimed at a domestic audience, and English-language PressTV, aimed abroad. This piece runs a single day of both through Gemini to surface how the same events are framed differently for each audience. This is comparative media analysis at a speed that used to require a research team — and it is a direct template for tracking how a story about your own company diverges across languages and markets.
A follow-up to last month’s experiment in which Gemini spotted and transcribed a handwritten dealer’s note in the margin of a 1539 Holbein volume. This instalment traces the provenance trail from that note through the London booksellers Maggs Bros to Harvard. A small, elegant demonstration of multimodal AI doing archival detective work.
OpenAI commits to letting its safety monitoring, alignment work and security posture set the release pace for frontier models. Published alongside “The Defender’s Window,” a companion argument about AI shifting the balance between cyber attackers and defenders.
A learning-oriented ChatGPT for under-18s with stronger built-in protections, healthy-use features and parental controls. Ran in your feed twice, from two different OpenAI URLs.
An initiative to support the government bodies that oversee AI use in defence and intelligence, offering them tools, training and technical expertise. Notably, this positions OpenAI as an enabler of scrutiny over its own sector — a defensible posture heading into the next round of US policymaking.
Asana used OpenAI’s Codex to replace a legacy testing system in a fortnight, for roughly $12,000 — work the company had scoped at five years. The headline number is doing a lot of work here, but it is the kind of concrete before-and-after that makes an internal business case.
Signal AI walks through five questions communications and risk teams actually brought to its platform: what risks are we missing, how do we prove comms moves the business, and where is the white space competitors have not claimed. Its scale claim — 5 million-plus articles and documents processed daily across 226 markets and 120+ languages — is the relevant benchmark if you are evaluating media intelligence vendors.
A build tutorial for a desktop app that turns meeting audio into structured CRM records automatically, then triggers downstream workflows from them. Written for wealth management, but the pattern — meeting in, structured contact record out — is exactly the gap between a transcription tool and a personal CRM. Worth reading with Mesh in mind.
The technical case for routing requests across many models rather than committing to one — published two days before Stripe validated the thesis with a $7 billion cheque.
David Gerard on the political turn against data centres in Texas, and the ERCOT grid’s difficulty absorbing them. He also posted that his 50-minute interview with Tante on the AI bubble is available now to supporters, publicly on Friday.
Ben Tossell on the gap between people who have access to autonomous agents and people who actually use one. Arrived in your folder twice, via both the Ben’s Bites feed and the aggregated Newsletters on AI feed.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
Claude watermarking now covers Claude Cowork and Claude Code — audit your pipeline
This is the single most operationally relevant item today. Anthropic’s invisible watermark is applied across Claude, the API, Claude Code, Claude Cowork and Claude Tag — meaning anything that passes through your Claude-based workflows now carries a machine-readable signature, even if Claude only touched it lightly. Before this becomes an awkward question, it is worth mapping which of your outputs actually route through Claude: this briefing does, and so would any draft statement, translation check or content summary you generate the same way. The mark is not a problem in itself; being surprised by it would be.
RelevantClaude Cowork · Claude Code · Drafts · Obsidian
Claude Code ships a plugin marketplace, Remote Control, and background code review
Anthropic’s August updates to Claude Code are more useful to a non-coder than they sound. The plugin marketplace means capabilities can now be installed as bundles rather than assembled by hand. “Remote Control” lets you start work on one machine and steer it from another — relevant if you are moving between the Mac mini and the MacBook Pro. And the /code-review command now runs quietly in the background instead of flooding your conversation. Alongside those, a batch of fixes to MCP authentication and Windows paths, which had been the most common source of broken connector setups.
Model routing is now a mainstream cost lever, not a developer curiosity
Latent.Space made the technical argument on Tuesday; Stripe paid $7 billion for OpenRouter to prove it. The idea in plain terms: you do not need your most expensive model for every task. Summarising a press clipping, tagging an email, or translating a routine line can go to a cheap open-weight model; only the hard reasoning needs a frontier model. If any part of your team’s AI usage runs at volume — media monitoring, transcript processing, content classification — routing is where the budget savings are, and OpenRouter is the single API that makes it a configuration choice rather than an engineering project.
RelevantOpenRouter · n8n · Make · MacWhisper pipelines
n8n’s August release adds MCP and agent capabilities to its AI builder
n8n — the automation platform you can self-host — shipped a broad update this month covering its AI workflow builder, editor, and new MCP and agent handling. The most practical addition is a button in the credentials card that sets up connections automatically through the AI Assistant, which removes the fiddliest part of wiring a new service into a workflow. If you have been meaning to build an automated pipeline between Inoreader, DEVONthink and Obsidian, the setup friction just dropped.
Recall.ai published a build guide for a desktop app that listens to a meeting, extracts the substance, and writes structured contact records — then fires downstream workflows off them. It is aimed at wealth managers, but the shape of the problem is identical to yours: you finish a call, and the useful details about who said what live in a transcript nobody re-reads. This is the missing link between MacWhisper or PLAUD capturing the audio and Mesh actually knowing something new about the person you spoke to.
Perplexity Comet adds enterprise deployment and a human-approval gate
Comet is Perplexity’s browser that can act on your behalf — clicking, filling forms, working through multi-step tasks. Two changes matter. First, it can now be deployed across an organisation through mobile device management, which moves it from a personal experiment to something IT can sanction. Second, it now requires you to explicitly approve any sensitive action before it executes, following security patches earlier this year. That approval gate is the difference between an agentic browser you can use on a work machine and one you cannot.
Gemini Deep Research as an adversarial reader for documents
GDELT’s Kalev Leetaru published three experiments in 24 hours, and the useful pattern across all of them is the same: rather than asking a model to summarise a document, ask it to attack one. He used Gemini Deep Research to hunt for loopholes in draft US sanctions legislation, and to compare how the same day’s events were framed by Iran’s domestic and international broadcasters. Both methods transfer straight into comms work — run a draft policy through it looking for the reading you did not intend, or compare how a story about you is framed across two markets.
RelevantGemini Deep Research · media monitoring · issues management
One More Thing in AI’s five techniques all rest on a single reframe: the model performs badly when treated as a search box and well when handed a defined task with a defined output. The practical version is to write the brief you would give a capable junior — here is the material, here is the audience, here is the format, here is what “done” looks like — rather than asking an open question and hoping. It is the cheapest improvement available to anyone who has tried AI and found it disappointing.
The Automated’s report on voice-cloning scams describes a pattern worth circulating internally: a cloned voice of a known colleague, used live on a phone call, to authorise a transfer or extract information. Senior executives are the highest-value target because their voices are publicly available in interviews and on podcasts. The mitigation is procedural rather than technical — a pre-agreed verification step for any voice request involving money or confidential material.
Relevantissues management · executive protection · team comms
No direct Tencent AI coverage in the last 24 hours
Nothing in today’s news sweep, your Inoreader AI folder, or your newsletters mentioned Tencent directly. The items below are the most recent lines a reporter is likely to raise, and the one indirect development worth watching.
Hy3, released on 6 July, is now available internationally through WorkBuddy, Tencent Design Miora and Tencent Cloud TokenHub. The detail most likely to come up in a question: WorkBuddy access is free to users worldwide until 31 August 2026 Pacific Time — which is now under two weeks away, so expect enquiries about what happens on 1 September. The technical claims on record are 90% task completion on autonomous-agent tasks in internal applications, and reduced token consumption per task across software development, office productivity, financial modelling, front-end design and game production.
A CNBC investigation published on 7 July found that Chinese-origin models accounted for 46% of US enterprise token usage flowing through OpenRouter. With Stripe having acquired OpenRouter this week, a single US payments company now controls the gateway, billing and routing decisions for that traffic. This is the kind of structural detail that produces a “will US infrastructure keep carrying Chinese models?” question — worth having a view on before it is asked rather than after.
The 3D asset creation tool remains the most frequently cited Tencent AI product in Western trade press, particularly in games coverage. Useful context if a games or creator-economy reporter comes in cold.