In partnership with
In Today’s Issue:
🤖 Grok Bot lands, and the $60 billion Cursor deal starts paying off
⚡ Anthropic rents 191 megawatts from a Bitcoin miner
🏎️ Nvidia's Nemotron 3.5 Lightning trades IQ for speed
🎬 LTX ships a film-grade video model as open weights
🩺 An AI that learned medicine by doing a residency
✨ And more AI goodness…
⚡ The Signal
The AI industry has quietly stopped selling assistants and started selling staff.
xAI's Grok Bot, opened in beta on Tuesday, does not wait to be prompted: each bot gets its own computer in the cloud, works inside the tools you already use, and keeps going after you close the laptop. It is also the first product out of SpaceX's $60 billion Cursor purchase, which finally answers what that money was for. Three more of today's stories point the same way from different directions. Anthropic signed a 20-year, $9.1 billion lease for 191 megawatts of data center capacity in Texas, because staff need somewhere to live. Nvidia shipped a small, very fast open model built for the repetitive steps agents grind through all day. And Terminal-Bench 3.0, the new scoreboard for professional computer work, has the best model and agent pairing resolving 43.5% of tasks. The hiring is running well ahead of the competence, and everyone is billing for it already.
All the best,
Kim Isenberg
Riot Platforms' data center operations, where Anthropic has just signed a 20-year lease (Riot Platforms)
⚡ Anthropic Rents 191 Megawatts From a Bitcoin Miner
Anthropic signed a $9.1 billion, 20-year lease for AI data center capacity at Riot Platforms' Rockdale campus in Texas. The deal covers 191 megawatts of IT capacity and runs through June 2048, with two five-year options that could lift the total to $16.1 billion. Riot will phase the site in, reaching 96 megawatts by December 2027 and the full build by June 2028; its shares jumped more than 25% after hours, and CEO Jason Les said the company now holds 241 megawatts and roughly $9.8 billion in contracted revenue.
👉 tl;dr: Bitcoin miners already own the two things AI labs cannot buy quickly, land and grid connections, so they are turning into landlords.
Output speed and output tokens per task; Nemotron 3.5 Lightning leads on speed at 669 tokens per second (Artificial Analysis via The Decoder)
🏎️ Nvidia's New Open Model Trades IQ for Speed
Nvidia released Nemotron 3.5 Lightning on August 11, an open-weights model built to be fast rather than clever. It is a 31.6-billion-parameter mixture-of-experts model that activates only 3.6 billion parameters per step, and it runs at about 669 output tokens per second, the quickest in Artificial Analysis's comparison group. It scores 24 on the Artificial Analysis Intelligence Index, level with OpenAI's gpt-oss-120b at a quarter of the size, and Nvidia is pointing it at the repetitive execution layer of agent pipelines rather than at frontier reasoning.
👉 tl;dr: As agents run longer, most of the cost sits in the thousands of small steps rather than in the hard thinking.
LTX-2.5's launch image, from LTX's own announcement (LTX)
🎬 LTX Ships a Film-Grade Video Model as Open Weights
LTX-2.5 landed on Tuesday, an open-weights video model aimed squarely at film and television work. It adds native multishot generation, which holds a character consistent across a whole sequence instead of a single clip, a new rendering stage that recovers fine detail such as faces, and an HDR ACES pipeline that survives a professional finish. Prompts run through a custom Gemma 4 backbone. The weights are on Hugging Face and native in ComfyUI from day one, and VentureBeat clocked a 10-second 720p clip from an image in 6.8 seconds self-hosted on two Nvidia GB200 chips, against 23.7 seconds through LTX's own managed API.
👉 tl;dr: Open weights let a studio run this on its own hardware, which matters more to a production team than another cloud API.
🎬 Watch This
Jeff Dean, Google's (ex-)Chief Scientist, sat down with Y Combinator's Diana Hu at Startup School 2026 and told two stories that both began as napkin math. In 2001 he and Sanjay Ghemawat worked out that Google's entire search index would fit in RAM, shipped it within days, and search got fast. In 2013 a similar calculation showed that three minutes of daily speech recognition per user would have required doubling Google's server fleet, and that one became the TPU. The rest of the hour covers why inference hardware is the next thing to specialize, why long-running agents fail, and where two or three people in a room can still beat Google.
"Sitting here today, mission success feels within sight."
– Brad Lightcap, outgoing COO of OpenAI, in the farewell note he posted on X (TechCrunch)
Lightcap joined OpenAI in 2018 and built the first versions of its finance, legal, people and go-to-market teams. He told staff on Tuesday that he is leaving to start something new and will stay on for a few weeks. OpenAI's mission is framed entirely around AGI, so an operator on his way out calling success "within sight" is a heavier claim than a goodbye note usually carries.
Tibo Sottiaux, who works on Codex at OpenAI, had promised a usage-limit reset for every extra million Codex users, up to 10 million. On Wednesday morning he said the team blew past that and then went quiet, before teasing: "Little surprise for you tomorrow."
Grok Bot Is What the $60 Billion Cursor Deal Was For
The Takeaway
👉 xAI, now inside SpaceX and presenting itself as SpaceXAI, opened the Grok Bot beta on August 11: agents that each run on their own computer in the cloud and keep working unattended.
👉 It is the first consumer product to come out of SpaceX's $60 billion purchase of Cursor's parent Anysphere, agreed on June 16.
👉 Distribution runs straight through Cursor's paid tiers: Cursor Premium Teams at $120 per seat, Cursor Ultra at $200, SuperGrok Heavy at $300 a month.
👉 Show a bot a task once and it saves the sequence as a routine, runs it on a schedule, and hands work to other bots in shared threads.
Grok Bot is the first real answer to what SpaceX actually bought when it paid $60 billion for Cursor. On August 11, xAI opened an early beta of Grok Bot on desktop and iOS. You create a bot, hand it a job and access to the tools you already use, and it works on its own computer in the cloud, carrying on after your laptop is shut and coming back when it needs approval or when the job is done. In xAI's own words: "Bots are AI teammates that do real work for you. They sign in to your tools, use them just like you do, and come back with finished work."
The opening frame of xAI's Grok Bot launch film (xAI)
The corporate history explains the product: SpaceX finished absorbing xAI on May 6, listed on Nasdaq on June 12, and four days later agreed to buy Cursor's parent Anysphere for $60 billion in stock, the largest acquisition of a venture-backed startup on record and still awaiting regulatory clearance. Cursor brought a coding agent that already ran in the cloud, a paying developer base, and an engineering bench xAI badly needed after all eleven of its co-founders left by the end of March. Grok Bot is the first consumer product to carry both sets of fingerprints: agent machinery pointed at ordinary office work rather than at code, sold through Cursor's own subscription tiers.
Musk on screen at SpaceX's Nasdaq listing, days before the Cursor deal was signed (Forbes)
What is genuinely new here is persistence. Walk a bot through a task once and it keeps the sequence as a routine it can run again on a schedule, in parallel with other bots that pass work between them. Browser agents of the last two years mostly drove one session and forgot it. The open questions are real enough: xAI has not said which models the router picks, VentureBeat's hands-on found the router underwhelming with no way to pin one, and the beta only exists behind the priciest tier of each subscription. But the direction is the story. SpaceX did not spend $60 billion to sell a better autocomplete.
Why it matters: The Cursor acquisition was widely read as a data and talent play to feed Grok's training pipeline. Grok Bot suggests SpaceX also bought a product team and a paying distribution channel, and that it means to sell agents as staff rather than as a feature inside a code editor.
Warmly Ran GTM With No Sales Team. Here's How.
That's what Warmly proved. They defined ICP, scored buying intent, and surfaced the right accounts before a human ever touched a lead. HubSpot noticed.
On August 12, Max and Keegan are rebuilding it live in HubSpot — and showing you how to replicate it this week. HubSpot Credits included when you join HubSpot for Startups.
Source: Terminal bench 3 shipped
The chart: Terminal-Bench 3.0 is the successor to Terminal-Bench 2.1, built by the same team under the working name Frontier-Bench, which is why the leaderboard still lives on the old Frontier-Bench domain. A new version was needed because 2.1 had been solved: the best agents were scoring 78 to 83% on it, too high to tell frontier models apart. On 3.0, Claude Opus 5 (max) running the mini-SWE-agent scaffold leads at 43.5%, ahead of GPT-5.6 Sol (max) with Codex at 34.6% and Claude Fable 5 (max) with Claude Code at 34.1%. Grok 4.5 (xhigh) with Cursor CLI sits sixth at 15.7%, having spent $960 and 1.2 billion tokens against the leader's $6.0k.
The lesson: The scaffold and the budget belong inside the score itself. Sonnet 5 with Claude Code burned the most on the board, $7.9k and 17.9 billion tokens, and landed at 14.6%, while Grok 4.5's run cost about an eighth as much and scored 15.7%. A model ranking that does not name the agent and the spend alongside it describes very little.
The caveat: These are max and xhigh effort settings rather than the defaults most people run, and the error bars sit around 1.5 points, so second and third place are effectively tied. The first release holds 74 tasks across 7 domains and is built to keep growing, so the absolute numbers will move.
🩺 The AI That Went Through Residency
⚡ Bottom line
Google put a Gemini model through nearly 50,000 simulated patient encounters, and it got measurably better at consulting.
💡 Why it matters
Clinical skill is procedural, and procedure can be drilled in simulation far faster than in a hospital.
🔎 What it means
Habits learned on primary care carried over to cancer cases the model had never seen in training.
Doctors do not become doctors by reading. They leave medical school knowing the facts, then spend years in residency seeing thousands of patients before anyone calls them competent. A team at Google and Google DeepMind asked whether a model could be put through the same thing, and published the answer on August 7 as ResidencyRL.
They started with Gemini 3.5 Flash and had it run consultations against a second model playing the patient, across 49,870 simulated telehealth scenarios covering 81 conditions. The simulated patients were built to be awkward on purpose: some hide symptoms, some resist advice, some push for treatment they should not get. Each encounter could run up to 60 turns, where a real consultation averages about 21. An automated grader scored every encounter on diagnosis, management, communication, documentation and safety, and those scores became the training signal.
The ResidencyRL loop, from generated patient scenarios through reinforcement learning to the blinded clinician evaluation (ResidencyRL, arXiv)
The trained agent turned noticeably more thorough. Under adversarial conditions its diagnostic accuracy rose from 81.0% to 88.0%, and missed red flags fell by 31%. In a blinded head to head on 97 cases, board-certified clinicians preferred it in 87.6% of comparisons overall, and in 90.7% on completeness of information gathering; on management plan safety it was preferred or tied in 96.9%. The result the team did not train for is the more interesting one: the same habits carried into 300 oncology scenarios the model had never seen, and into the outside benchmarks AgentClinic and CRAFT-MD.
Out-of-domain telehealth scores, baseline Gemini 3.5 Flash in blue against the ResidencyRL-trained model in red, across 299 simulations (ResidencyRL, arXiv)
The obvious objection is that the patients were played by the same base model, and that most of the grading was automated. The human check was real but small, at 97 cases. What the paper does show is that a model can be trained to interview better, which is an underrated part of clinical skill, and that the training transferred to work it had never seen. Whether it survives contact with a frightened person in an actual clinic is the next study.
Rethink the blueprint for customer financial support
See how Fin and Plaid are helping AI take action in financial services, while keeping businesses in control. Join us August 13th to explore faster onboarding, proactive support, and real-time issue resolution.