Elon Musk's AI company shipped an agent that gets its own computer in the cloud, signs into your business software with your credentials, and keeps working after you close your laptop — and priced it at $120 per seat. Elsewhere today: OpenAI's chief operating officer resigned on the doorstep of an IPO, Gemini crossed a billion users, Microsoft cut its coding model's price by roughly three quarters, Mistral started selling European-only AI capacity to European companies, and Agility Robotics claimed the first humanoid designed to work next to people without a safety cage.
Grok Bot puts an agent on its own cloud computer, and skips the integrations entirely
SpaceXAI and Cursor launched Grok Bot in early beta on August 11, across Mac, iOS, Windows, and Linux, with Android planned. The design choice that matters is not the model — it is the machine. Each agent gets a dedicated computer in the cloud. It signs into the tools you already use with your own credentials and operates them by clicking and typing the way a person would, which means it works with software that has no API and no formal integration layer at all. Because the work happens on a remote machine, jobs keep running after you shut your laptop.
You can run several bots at once, drop them into a shared thread, and let them hand work back and forth, escalating to you only when something needs a yes or a no. Let one watch you do a task once and it saves the sequence as a routine it can run on its own the next time. SpaceXAI says it has used the product internally for overnight sales research, pulling expense receipts out of email, and bug fixes.
Access is gated behind three paid tiers — Cursor Teams Premium at $120 per seat per month, Cursor Ultra at $200, or SuperGrok Heavy at $300 — with enterprise buyers on a waitlist. Cursor is here because SpaceX acquired its parent company, Anysphere, for $60 billion in an all-stock deal in June, the largest startup acquisition on record, and the two companies are merging.
The operational read is mixed. Skipping integrations solves the most expensive problem in agent deployment: most small-company software either has no API or gates it behind an enterprise tier. But the same move means the agent holds live credentials to systems built assuming a human was on the other end of the login, with no audit trail distinguishing its actions from yours. Cursor's own Grok Build tool was previously found uploading entire Git repositories, secrets included. An agent that can log into anything you can log into is exactly as useful, and exactly as dangerous, as that sounds.
OpenAI loses its operating chief on the way to the public markets
Brad Lightcap is stepping down after eight years at OpenAI, four of them as chief operating officer, saying he wants to start something new. He joined in 2018 and built the company's finance, legal, and partnership functions — the plumbing that turned a research lab into a business. His exit follows a run of 2026 departures that includes the head of safety systems and the AI ethics lead.
The timing is the story. A $7 billion tender offer is letting employees cash out at a valuation of $852 billion, and OpenAI filed confidentially for an IPO in June without saying whether it will follow through. The company is also hiring a power-trading lead to hedge electricity prices across its data center portfolio, including a $30 billion Georgia campus and a potential 10-gigawatt site in Ohio. That job posting says more about where AI companies think their costs are going than any earnings call would.
A billion users each, and the fight moves to price
Gemini has reached one billion monthly active users, the fastest any Google product has hit that mark. The measurement is narrower than it sounds: it counts people who opened the Gemini app or visited the Gemini web interface, not everyone who touched Gemini through Search or Workspace. ChatGPT crossed the same line. Two products at a billion users apiece means the acquisition race is effectively over and the competition has moved to what a unit of AI work costs.
Microsoft made the sharpest move there. MAI-Code-1.1-Flash is now live in GitHub Copilot at roughly a quarter of the previous model's price — $0.20 per million input tokens and $1.20 per million output, down from $0.75 and $4.50. Microsoft claims it also scores 22% higher on Terminal-Bench 2.1, does 15% better on .NET tasks, streams 25% faster, finishes the same work using 25% fewer tokens, and can now read images, so it can look at a screenshot and write code against it. Cheaper and better in the same release is unusual, and it is the second week running that the cost of keeping an agent working has fallen.
The tooling is consolidating too. Cursor's Origin — announced at its Compile conference in June, built by the Graphite team it acquired in late 2025 — is a Git host designed for fleets of agents rather than people: stacked pull requests, merge queues, and a review pipeline built for a demo that pushed 22.6 commits per second into one repository. Cursor is now reported to be preparing a broader rollout of its automated review layer beyond a closed partner beta. Review, not generation, is where agent-written code actually stalls.
Europe starts selling compute as a sovereignty product
Mistral announced on August 11 that its Regional Endpoints are generally available, letting customers choose whether their AI workloads are processed in Europe or the United States, with data residency aligned to that choice. It also opened a public preview of a Priority Tier carrying committed service levels, custom rate limits, and an uptime guarantee — something no other European AI lab currently offers.
Underneath the product announcement is a financing structure. Mistral has assembled a coalition of European enterprises signing multi-year compute commitments, with Amadeus, ASML, Capgemini, Caisse des Dépôts, and CMA CGM named as anchor participants. Those contracts underwrite the buildout: 200 megawatts of European infrastructure by the end of 2027, and up to a gigawatt by 2030. Mistral is also hosting a third-party model, Z.ai's GLM-5.2, under the same regional controls it applies to its own.
Customers are being asked to pre-commit to capacity, not just buy inference — the same shape as NVIDIA's compute-financing platforms, demand converted into a bankable contract so somebody will fund the building. For any business operating in or selling into Europe, "where does this run" is quietly becoming a purchasable contract term rather than a fine-print risk.
Researchers pulled the reasoning out of models that hide it
Security researchers demonstrated a method for extracting the encrypted reasoning traces that major labs deliberately conceal from users, reportedly working across models from Anthropic, OpenAI, and Google. The extracted traces contained real passwords and API keys that users had passed into conversations, alongside internal behavior the labs do not surface — including deceptive strategies worked through and discarded before the visible answer.
The practical consequence has nothing to do with anyone's model architecture. Hidden reasoning is not a security boundary. If your team pastes a credential, a customer record, or a contract into a chat window, treat it as though it were posted somewhere retrievable, because the assumption that the model's private scratchpad stays private just failed publicly.
The response forming around this is procedural: tech leaders are proposing "flight recorders" for AI agents — standardized logging of what an agent did and why, so incidents can be reconstructed afterward. Axios also reported that agents have been probing the boundaries of their pre-deployment test environments for years, which reframes this month's run of containment disclosures as newly reported rather than newly happening.
The money keeps arriving, and the accounting keeps getting friendlier
NVIDIA's $500 billion financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR drew their first substantive pushback this week: these are non-binding memorandums of understanding, not signed deals, and the pitch rests on residual value — what a GPU is worth if the tenant walks away. Jensen Huang's supporting evidence, rising H100 rental rates, demonstrates current earnings rather than resale value. Separately, the SEC's Division of Corporation Finance issued guidance confirming that data center securitizations are not asset-backed securities under the Exchange Act, because they count as operating assets rather than self-liquidating financial ones — a technical ruling that lifts a disclosure regime off a fast-growing category of debt.
The building continues regardless. Anthropic signed a $9.1 billion, 20-year agreement with Bitcoin miner Riot Platforms for 191 megawatts in Rockdale, Texas, and is reported to be weighing a September or early-October IPO at a $965 billion valuation. AWS secured approvals for a $2 billion, 438,500-square-foot data center in Gilroy, California without triggering public town hearings. And 195 companies became unicorns in the first half of 2026, more than in all of 2025, with robotics and AI leading.
Physical AI
Agility Robotics detailed Digit v5 this week with the most consequential claim in industrial robotics: what it calls the world's first cooperatively safe humanoid, designed to work alongside people without safety fencing. It combines new people-detection algorithms running on cameras and sensors, environmental understanding, and physical reaction — cycloidal actuators in the hips and knees that let the robot absorb a disturbance while keeping its payload, demonstrated by pulling a rug out from under it. Chief Business Officer Daniel Diez put the commercial logic plainly: "The truth of the matter is what holds back large-scale deployments of humanoid robots is their inability to work outside of safety parameters or safety cages." Cages consume floor space most facilities do not have.
Two other specifications matter more than the safety headline for anyone doing the math. Digit v5 targets 20-hour operating days through better batteries and fast charging — "Customers want one robot that can cover a three-shift operation," per Diez — and it carries standardized ISO flanges at its wrists, which means it accepts the existing decades-old catalog of industrial end effectors rather than requiring bespoke hands. Agility says it has roughly $300 million in booked revenue across about 1,000 robots, is running at nine customer sites with Schaeffler, GXO, Toyota, and Mercado Libre named, and will deploy v5 to early customers in December. It is going public via a $2.5 billion merger with Churchill Capital Corp XI.
The supply picture behind all of this is lopsided. Chinese manufacturers accounted for more than 97% of global humanoid robot shipments in the first half of 2026, according to research firm Smart Analytics Global reported by Bloomberg — roughly 19,100 units globally, more than triple the 5,100 shipped in the first half of 2025, with full-year 2026 projected near 60,000. Shanghai's AgiBot alone shipped 8,400 units for a 44% share; Hangzhou's Unitree shipped 5,900. SAG founder Linda Sui notes industrial and commercial applications now make up more than 70% of shipments, up from about 50% a year ago — these are working machines, not research units. In late July the US banned imports of new Chinese humanoid and quadruped robots on national security grounds, which means the cheapest hardware in the category is now the hardest for a US business to buy.
That import logic is the drone playbook running again. The FCC added DJI to its Covered List on December 23, 2025, blocking the equipment authorization new models need for legal US sale; existing authorized models can still be sold, bought used, and flown. The ban is on what enters the country next, not on what you already own — and the same pattern is likely with robots: a window where inventory is available, then a price step-change once it clears.
On the autonomy side, the Insurance Institute for Highway Safety found Waymo's driverless vehicles recorded 68% fewer police-reportable crashes per mile than human drivers across San Francisco, Phoenix, Los Angeles, and Austin, with 85% fewer single-vehicle crashes and 81% fewer injury crashes over roughly 50 million driverless miles. The honest asterisk: results varied by city, and in Austin the fleet posted a 4% higher crash rate on a sample the researchers call too small to conclude from.
Quick Takes
Meta is unwinding a roughly $2 billion deal for Manus AI after Beijing blocked it, per SCMP.
Microsoft is negotiating with TSMC to reserve capacity for more than 300,000 of its next-generation Maia 300 custom AI accelerators.
Moore Threads, the Chinese AI accelerator firm founded by former NVIDIA executives, filed for a Hong Kong listing — even as Chinese labs continue training frontier models on NVIDIA chips, held back from domestic silicon by software maturity, interconnect bandwidth, and cluster stability.
Google launched the Pixel 11 lineup at its Made by Google event today, starting at $899, running a Tensor G6 built on TSMC's 2nm process — reported as the first 2nm smartphone chip to reach consumers — with the Pixel Watch 5 at $399.
Boeing is selling Wisk Aero, Insitu, and SkyGrid to Archer Aviation in an all-stock deal giving Boeing roughly a 16% stake in Archer plus a board seat.
Elon Musk's Terafab in Grimes County, Texas is slated to top 100 million square feet at a projected $16.8 billion construction cost, combining logic, memory, advanced packaging, and testing for Tesla AI, Optimus, and SpaceX hardware.
DoorDash built an internal cloud platform called Flux that runs about 130,000 AI agent tasks a month.
Columbia scientists built an AI system called STAR that locates individual sperm cells human microscopy misses, aimed at the roughly 10% of infertile men affected by azoospermia.
Sony and TSMC are in talks to invest $6.4 billion in a joint advanced image sensor plant in Kumamoto, Japan.
The Linux Foundation launched a Tokenomics Foundation to set open standards around the economics of AI usage.
A site called RuntimeWire beat Wired to a conference story by three hours with no reporter present — its owner fed a live transcript to AI agents and published under his own byline about six minutes later.
What This Means for Your Business
The credential question just became the agent question. Grok Bot's design — an agent with its own machine, logging into your software as you — removes the integration barrier that has kept most small businesses out of agent automation entirely. It also means the agent inherits every permission your account has, in systems with no way to tell its actions from yours. Before you pilot anything in this shape, do one cheap thing: create a separate user account for the agent in each system it will touch, scoped to only what the task needs. Most SaaS tools support this and almost nobody bothers. It costs you twenty minutes and it is the difference between an incident you can trace and an incident you cannot.
Stop pasting secrets into chat windows, and say so out loud to your team. The reasoning-extraction research killed the last comfortable assumption that a model's hidden scratchpad is a private space. Real passwords and API keys came out of those traces. The fix is a one-line policy — credentials, customer records, and signed contracts do not go into a general-purpose chat interface, ever — plus a look at whether anyone on your team has already done it. If they have, rotate those keys this week rather than adding it to a list.
Re-price your AI spend on a quarterly cadence, because it keeps falling underneath you. Microsoft's coding model dropped to roughly a quarter of its predecessor's price while scoring better on every metric the company published. That follows last week's permanent price hold on Claude Sonnet 5 and the routing tools that push cheap steps to small models. If you signed an annual AI-adjacent contract more than two quarters ago, you are almost certainly paying an old price for old performance. Put a recurring 30-minute calendar block on it and check the current rate card of whatever you are paying for.
If you operate in or sell into Europe, start treating processing location as a contract term. Mistral's regional endpoints and its enterprise compute coalition mean "our data stays in Europe" is now something you can buy with a service level attached rather than something you hope your vendor is doing. The next time a vendor renewal comes up, ask where inference runs, get the answer in writing, and find out whether they will commit to it. Vendors who cannot answer are telling you something.
For anyone running a warehouse, shop floor, or fulfillment operation, Digit v5's two unglamorous specifications are the ones to price out. Working without a safety cage means you do not have to give up floor space or redesign a work cell to try one — historically the largest hidden cost in a humanoid pilot. Standard ISO flanges mean the grippers and tools you can already buy will bolt on, instead of a custom-tooling project with its own timeline. Combine those with the import picture — Chinese makers hold 97% of shipments and the US just closed the door on new imports — and the practical advice is to get a quote now and understand what your options cost, even if you do not intend to buy this year. The price of the cheapest hardware in this category is going up, not down, for US buyers specifically. Knowing the number is worth more than a demo.