In partnership with
In Todayโs Issue:
๐ Cursor launches Origin and starts hosting your code
๐ก๏ธ Greg Brockman says the defender's window is open now
โก OpenAI takes about 8 GW on a former uranium site
๐ฐ Anthropic hits a $65B run rate before its IPO
๐ค Nous Research gives Hermes a roster of named bots
โจ And more AI goodnessโฆ
โก The Signal
Every big story today is a company outgrowing the thing it was built on.
Cursor outgrew GitHub and started hosting code itself. OpenAI outgrew shared cloud capacity and took a 20-year lease on about 8 gigawatts at a campus being built for it on a former uranium enrichment site in Ohio. Anthropic outgrew private money and is heading for a listing with a $65 billion run rate in hand. Three companies, three foundations, GitHub, other people's data centers, venture capital, and each is being traded for something the company controls or something far larger. Watch what that does to everyone still standing on the old ground.
All the best,
Kim Isenberg
(Greg Brockman, OpenAI co-founder and president. Photo: Getty Images via TechCrunch)
๐ก๏ธ Brockman: The Defender's Window Is Open Now
OpenAI president Greg Brockman published an essay on August 16 arguing that defenders have a narrow head start in AI security and are mostly not using it. He describes a coordinated group of AI agents that chained vulnerabilities together to get inside OpenAI's own infrastructure, and recounts pointing ChatGPT Work at his personal website, where it found 13 security issues in 15 minutes and fixed them within the hour. His warning carries a date: he expects open-weight models to reach frontier-level cyber capability by the end of August 2026, after which the attacker's toolkit is a free download.
๐ tl;dr: "The defender's window is open now", and the ten-step checklist he attaches is written for a timeline measured in weeks.
(Groundbreaking and site aerial at the PORTS-Pike Technology Campus in Ohio. Image: SB Energy)
โก OpenAI Takes 8 Gigawatts On A Former Uranium Site
A former uranium enrichment plant in Pike County, Ohio, is being rebuilt as the PORTS-Pike Technology Campus, and OpenAI has taken a 20-year lease on roughly 8 gigawatts of IT capacity there. SB Energy builds and operates the site, Nvidia supplies the compute exclusively and is putting $1.5 billion into SB Energy to do it, and the US Department of Energy is a partner. Capacity comes online in phases from 2028, the buildout runs to 2032, and the projections are 35,000 construction jobs and 2,500 permanent ones, alongside a $40 million community fund from OpenAI matching SB Energy's own and up to $84 million in Codex credits for roughly 844,000 eligible Ohio students.
๐ tl;dr: The gigawatt race now runs through Cold War infrastructure, and OpenAI is buying goodwill in Ohio at the same time it buys power.
(Dario Amodei, Anthropic co-founder and CEO. Photo: Getty Images via TechCrunch)
๐ฐ Anthropic Hits A $65 Billion Run Rate Before Its IPO
Anthropic is on track for more than $65 billion in annualized revenue, more than seven times its pace at the end of last year. The run rate crossed that mark by the end of July, according to people familiar with figures the company shared with investors, after topping $9 billion in late 2025 and $47 billion in May. Preliminary numbers put its latest completed quarter above $11.5 billion, against $787 million a year earlier, with positive adjusted operating income, and the company has filed confidentially for a listing that could come as soon as this fall, ahead of OpenAI.
๐ tl;dr: For scale, Bloomberg puts OpenAI's own run rate just above $40 billion, though the two firms may not measure it the same way.
๐ฌ Watch This
Greg Brockman sat down with CNBC's Squawk Box the day after publishing his security essay, and the conversation ranged well beyond security. His headline framing is that compute is becoming the new oil of the AI age, a convenient line for a company that just leased about 8 gigawatts in Ohio. He also fields the awkward questions: the run of executive departures, the capex bill, and where OpenAI actually sits in the race. His answer on the departures, roughly that none of this is unusual, is worth hearing in his own voice.
The timing checks out: OpenAI's DevDay is set for September 29 in San Francisco, six weeks out, which could land it in the same window as Anthropic's planned listing.
(Bot Mode for Hermes Desktop. Image: Nous Research via MarkTechPost)
Nous Research has switched on Bot Mode for Hermes Desktop, turning agent profiles into a roster of named bots that each keep their own model, memory, skills and avatar, and can hand work to one another. It follows Grok Bot into the same territory, and the plugin sits on GitHub under an MIT license.
Cursor Just Made GitHub Optional
The Takeaway
๐ Cursor shipped Origin, its own code hosting, on August 17, putting repositories, pull requests and agents in one place.
๐ Repos live at cursor.com/codebase, and synced GitHub repos sit beside them, with code, pull requests and comments moving both ways.
๐ Agents browse the codebase, answer questions, make changes and push branches without leaving the platform.
๐ It is an early beta for all paid plans, except enterprise organizations whose admins opt out.
Since the first AI coding assistants shipped, they have been guests in somebody else's house. Your code lived on GitHub and the assistant came to visit, authenticating with a token, cloning, opening a pull request, waiting on a webhook. Origin, shipped on Monday, is Cursor's bid to end that arrangement. The shift is quieter than a model launch and probably larger: once most of the code in a repository is written by an agent rather than typed by a person, the repository stops being a place people visit and becomes the runtime the agent lives in. Cursor has decided it would rather own that runtime than rent it.
Origin adds a Codebase tab to Cursor. You create repositories inside it, at cursor.com/codebase, and push to them either with a new Origin CLI or with ordinary git against Cursor's own remote. This is a real code host, what developers call a forge, with the machinery to match: pull requests arrive with diffs, comments, status checks and a merge button, and repository settings carry branch protection rulesets, required approvals, and merge or squash policies. GitHub repositories you already own can be synced in and sit alongside the native ones, with code, pull requests and comments syncing in both directions.
(Branch protection and merge rules inside an Origin repository. Image: Cursor)
Cursor describes Origin as "designed for agent scale", and says that "your code, PRs, and agents are now in the same place." The mechanical argument behind that phrasing is easy to follow. An agent working against an external forge burns a real share of every run on authentication, cloning, polling for build results and translating between two systems that were never designed to talk to each other. Collapse the forge into the platform and that overhead disappears. Cursor has also wired in the surrounding tooling directly, with Vercel for preview deployments and Depot and Buildkite for continuous integration, installable straight from a Codebase settings page.
(Vercel, Depot and Buildkite install directly into a Codebase. Image: Cursor)
The objection writes itself. Cursor is asking companies to move their most sensitive asset onto infrastructure run by a company that, until Monday, sold them an editor, and two-way GitHub sync is exactly the sort of bridge that tends to narrow with each release. It is also an early beta, enterprise administrators can switch it off, and the unglamorous half of a mature forge, issue tracking, organization-wide permissions, compliance attestations, gets no mention in the launch notes. Anyone who has moved a forge before knows the code is the easy part.
Why it matters: If agents end up writing most of the code, then whoever hosts the code sets the rules for what those agents are allowed to do to it. Cursor has just moved from selling an editor to competing for that position, and it is a far bigger business.
Sources:
๐ https://cursor.com/changelog/origin-code-hosting
Your Competitors Just Hired an Entire GTM Team. No Humans.
AI agents are doing the work of full GTM teams. Pipeline, content, customer service. The startups using them aren't waiting for headcount approvals.
Get the free Practical Guide to Agentic GTM for Startups and see exactly where to start.
The chart: DeepSeek's V4-Flash-0731 and V4-Pro-0813 measured against GLM-5.3, Kimi-K3, Claude Opus 4.8 and Fable 5, with each DeepSeek model scored twice: as it ships, and wrapped in J-Space. J-Space is an open-source prompt harness, not a new model and not a fine-tune. It works at inference time only: it sorts each task into a fast, full or looping mode, keeps a running five-field ledger (goal, core, verified, open, next) so the model does not lose its place across tool calls, and forces verification checkpoints instead of one-shot answers. Alone, V4-Pro-0813 scores 42.7 on HLE without tools, 61.5 on NL2Repo and 62.7 on DeepSWE. With J-Space: 48.0, 73.4 and 72.0.
The lesson: The report's claim is that the capability was already in the weights and was leaking out on the way to the answer, through what it calls "capability-realization loss": state drift, mismatched interfaces, long chains nobody verified. The numbers fit that story. Gains are largest on the multi-step agentic tests (NL2Repo +11.9, DeepSWE +9.3) and smallest on closed-book knowledge (HLE without tools +5.3), exactly what you would expect if the harness helps a model stay on task rather than know more.
The caveat: None of this is from DeepSeek. It is one GitHub repository under a pseudonymous account, every number is a single run with no error bars, and nobody has reproduced it, so read each row as a claim. The benchmark selection flatters it too: Fable 5 still leads HLE without tools at 53.3, and GLM-5.3 tops AutomationBench at 48.2 against J-Space's 38.2.
๐ค Where The Robotics Money Actually Went
โก Bottom line
Business Insider asked investors at 14 venture firms to name the robotics startups they rate most highly right now.
๐ก Why it matters
Their 25 picks lean heavily toward software, data and narrow industrial work over general-purpose humanoids.
๐ What it means
The robot that pays for itself first will probably be wheeled, unglamorous and built for exactly one job.
The method is the interesting part. Business Insider asked investors at 14 firms, among them Sequoia, Felicis and Bessemer, to name two companies each: one they had already backed, and one they had not. That second pick is where the useful signal sits, because it is the only one an investor has no reason to talk up.
Of the 25 names, only a handful build the thing most people picture when they hear the word robot. Unitree ($265M) and Noble Machines make humanoids, and Walden Robotics ($300M) builds what its CEO Russ Tedrake calls "a pragmatic version of a humanoid". Much of the rest sits underneath the hardware or beside it: Skild AI ($2.2B) and Generalist ($540M) on cross-robot control, Dexterity ($295M) and FieldAI ($405M) on autonomy, Foxglove ($60M) on data infrastructure, XDOF ($78M) and Mecka AI ($68M) on training data. The applied end is narrower still: Gecko Robotics ($347M) inspects power plants, GrayMatter ($86M) finishes surfaces, Machina Labs ($169M) forms sheet metal with robot arms, Upside Robotics ($7.5M) puts fertilizer on crops.
(Machina Labs' RoboCraftsman forming sheet metal on the factory floor. Image: Machina Labs)
Sunday Robotics ($200M) shows what the pragmatic version looks like at home: its robot Memo rolls on wheels rather than walking on legs, and learns chores from a capture glove worn by a person instead of through teleoperation. Its chief executive Tony Zhao puts the bet plainly: "I think 2026 will be when the timeline for autonomous home robots shrinks in a very drastic way because of this breakthrough."
(Sunday Robotics' Memo working in a test kitchen. Image: SiliconANGLE)
The smallest disclosed raise on the list is the reality check. Avatar Robotics, at $6.5M, puts humanoids on warehouse packing lines with remote human operators still in the loop. That is the open question hanging over the entire category: how much of the autonomy on display is autonomy, and how much is a person somewhere else holding the controls. On why the data companies keep attracting money, Bain Capital Ventures' Amanda Huang offers the oldest rule in the field: "Garbage in, garbage out."
(Avatar Robotics humanoids packing kits alongside warehouse staff. Image: Business Insider)
Never worry about roaming again
Stay connected on every trip with Saily eSIM plans. From beach vacations to business travel, access data in 200+ destinations.
VIP perks available.
Activate instantly upon arrival.
Download SAILY in your app store and use code newsletter15 at checkout to get an exclusive 15% off your first purchase.
Chat support available 24/7. Get a full refund if your device isnโt eSIM compatible.