In partnership with
In Today’s Issue:
🛰️ SpaceX and Nvidia design an AI server rack for orbit
🚕 Nevada clears 8,000 robotaxis for Las Vegas
🍎 The Mac mini quietly became an AI machine
⚖️ Taiwan charges nine over smuggled AI servers
🤖 A robot runs the 100 meters faster than Bolt
✨ And more AI goodness…
⚡ The Signal
AI's physical footprint is running out of room on Earth, and the industry has started shopping elsewhere.
Nvidia and SpaceX said on Monday that they have designed a space-optimized Vera Rubin rack for launch into orbit late next year, part of the same announcement that introduced Vera, a processor built to keep AI agents busy rather than to run the models themselves. The same pressure shows up in smaller places. Apple cannot keep the Mac mini in stock partly because buyers turned it into a home AI server and memory chips are scarce. Taiwanese prosecutors charged nine people, among them employees of Nvidia and Supermicro, with routing restricted AI servers to Chinese buyers through Japan and Indonesia. And OpenAI is putting a five-hour limit back on Plus accounts to smooth the load on its own machines. Compute is the thing everyone is now rationing, smuggling, or launching.
All the best,
Kim Isenberg
(Tim Goessman / Bloomberg / Getty Images via TechCrunch)
🚕 Nevada Just Cleared 8,000 Robotaxis for Las Vegas
Nevada's transportation regulator voted unanimously on 20 August to let Tesla, Waymo and Uber run paid driverless rides in Clark County, home to Las Vegas. Tesla's permit covers up to 5,000 vehicles, with 1,000 each for Waymo and Uber, which operates through Motional and Zoox. Counting Zoox's existing permit, the ceiling reaches about 8,000 robotaxis over twelve months. Tesla's own Cybercab chief engineer, Eric Early, played the number down: "I don't think we'll be in a position by this time next year to deploy 5,000 vehicles." Local taxi and livery operators objected that the roads cannot absorb them.
👉 tl;dr: Las Vegas is now America's largest robotaxi market on paper. The fleets to fill it do not exist yet.
(MacRumors)
🍎 The Mac Mini Became an AI Machine, and Apple Cannot Keep It in Stock
Apple is preparing to refresh the Mac mini for the first time in almost two years, and it could arrive within days, Bloomberg's Mark Gurman reported today. The company has tested versions with both the M5 and M6 generations of its processor line. What makes a routine chip bump interesting is why the current model sold out: it became a favorite box for running AI models locally, and that demand ran straight into a memory chip shortage that has stretched deliveries to weeks or months. It would also be the last product launched under Tim Cook, who hands the CEO job to John Ternus on 1 September.
👉 tl;dr: The cheapest Mac quietly turned into a personal AI server, and Apple's supply chain never planned for it.
⚖️ Taiwan Charges Nine Over AI Servers Smuggled to China
Prosecutors in Keelung indicted nine people on Monday over the alleged illegal resale of Supermicro servers carrying Nvidia's B300 chips to buyers in China. The accused include a distribution manager at Nvidia's Taiwan office, two sales managers at Supermicro's Taiwan branch and the chief executive of the distributor Albatron Technology. Investigators say 74 servers reached Chinese customers through Japan and Indonesia on falsified paperwork, and intercepted another 56 still sitting in Taiwan. It is the most detailed public account so far of how restricted American hardware travels through Taiwan's distribution chain.
👉 tl;dr: Export controls rarely fail at the border. They fail at the distributor.
🎬 Watch This
Michael Kratsios directs the White House Office of Science and Technology Policy, which makes him the person most responsible for how Washington thinks about AI. He was COO of Scale AI before that, so he has seen the boom from both sides. In this 40-minute conversation with Y Combinator's head of public policy, Luther Lowe, recorded at Startup School 2026, he explains how technology policy actually gets made, why the administration backs open-source AI, and why he thinks smaller companies need a seat at the table. Useful counterweight to the frontier-lab view of regulation.
"This is necessary as (a) the 5h limit allows us to smoothen the load on our compute, allowing to keep the plan generous in terms of weekly usage."
– Thibault "Tibo" Sottiaux, Codex and ChatGPT lead at OpenAI, posting on X, 25 August 2026
OpenAI is reinstating the five-hour rate limit for Plus subscribers on ChatGPT Work and Codex, after lifting it for several weeks. Pro subscribers on the $100 and $200 tiers keep the limit switched off for the coming months. Rationing by tier is what a compute shortage looks like from the customer's side.
At his first all-hands with Cursor staff since SpaceX closed its $60 billion purchase, Elon Musk reportedly told employees that AI becoming impossible for humans to control is inevitable, so they should build it first. He also conceded that Grok is behind and named Anthropic the current leader.
The Data Center Is Leaving Earth
The Takeaway
👉 SpaceX and Nvidia have designed a space-optimized Vera Rubin NVL72 system for launch to orbit in the fourth quarter of next year, aboard SpaceXAI's first Starmind satellite.
👉 On the ground, SpaceXAI is adopting Vera, which Nvidia pitches as the first CPU built for AI agents rather than for running the models themselves, to scale Grok toward gigawatt capacity.
👉 Vera carries 88 Nvidia-designed Olympus cores and up to 1.2TB/s of memory bandwidth, with a claimed 1.8x faster task completion than x86 on agentic workloads.
👉 Every performance figure here is a vendor number, and the orbital rack still has to survive power, heat, bandwidth and reliability limits that do not apply in a building.
The AI industry has spent three years arguing about how much power it can pull out of the grid. Elon Musk's answer is to stop asking the grid. Nvidia announced on Monday that SpaceXAI will use its new Vera CPU to scale the infrastructure behind Grok toward gigawatt-level capacity. Sitting inside the same announcement is the part that changes the picture: SpaceX intends to put an optimized Vera Rubin NVL72 rack, the same class of system that fills frontier data centers, into orbit aboard its first Starmind satellite. "SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028," Musk wrote on X.
Nvidia's Vera CPU, built around 88 custom Olympus cores. (NVIDIA)
Vera is the less cinematic half of the announcement and probably the more consequential one. It is designed around what AI agents do between model calls: executing code, moving data, calling tools and orchestrating long chains of work. That work has been running on general-purpose x86 processors while the expensive GPUs waited on it. Vera puts 88 Nvidia-designed Olympus cores and up to 1.2TB/s of LPDDR5X memory bandwidth against the bottleneck, and Nvidia claims up to 1.8x faster task completion than x86 on agentic, reinforcement-learning and data-processing jobs. "Vera gives us the CPU performance and memory bandwidth to run enormous amounts of orchestration, code and data processing," said SpaceXAI president Mike Nicolls.
The Vera Rubin rack family. A version of the NVL72 is the system SpaceX plans to fly. (NVIDIA)
The orbital half deserves more caution than the silicon. A rack in space has to manage power, heat, bandwidth and radiation with no technician on site, and Nvidia's own release says these products are offered on a when-and-if-available basis. The speed figures come from the company selling the part. What is harder to argue with is the direction of travel: SpaceX now owns the rockets, a satellite network, a coding company it bought for $60 billion, and an early claim on Nvidia's newest silicon.
Why it matters: Compute has been the binding constraint on frontier AI for three years, and the companies that own the constraint are now integrating vertically around it. Owning the rocket, the satellite and the rack is a different kind of moat than owning the model.
Sources:
Everything GTM. One platform.
Small teams don't have time to stitch together five tools and hope it works.
Apollo gives you everything you need to find leads, reach them, and close deals — all in one place:
230M+ verified contacts
AI-powered outreach
Data enrichment
Inbound lead capture
Meeting scheduler
And more
Stop juggling tools and start building pipeline that scales.
With Apollo, the AI revenue engine powering 4M+ users.
The chart: Arena's Pareto plot for Code Arena: WebDev, the leaderboard where models are ranked by head-to-head human votes on web development tasks. Arena score runs up the vertical axis, blended price per million tokens runs along the horizontal, cheapest on the right. Qwen3.8-27B lands at 1595 points for $2.35 per million, ninth overall, on the frontier line between Zhipu's GLM-5.3-Max and DeepSeek's V4-Flash. Claude Opus 5 Max still tops the board at roughly 1690, from the expensive end of the axis.
The lesson: A 27-billion-parameter model, small enough to run on hardware a developer can own outright, now sits six ranks behind Alibaba's own flagship Qwen3.8-Max and within about a hundred points of the best model on the board. Gemma 4-31B, released in April in the same size class, ranks #80. That is what four months buys right now.
The caveat: Arena scores are human preference votes on web development prompts, not correctness tests, and they reward answers that look good side by side. The price axis is a blended API rate, so it says nothing about what the model costs to run on your own machine.
🤖 A Robot Just Beat Usain Bolt, and It Means Less Than It Sounds
⚡ Bottom line
A humanoid robot ran 100 meters in 9.39 seconds in Beijing, faster than Usain Bolt's human world record.
💡 Why it matters
Running and jumping are the solved part of robotics. Doing useful work in an unfamiliar room is not.
🔎 What it means
China is winning the spectacle and the capital, while the question of what these machines are actually for stays open.
The number is real and it sounds impossible: 9.39 seconds over 100 meters, against Usain Bolt's 9.58. It happened at the second World Humanoid Robot Games, which opened on Saturday in Beijing's National Speed Skating Oval, the venue built for the 2022 Winter Olympics. The machine came from Chinese manufacturer X-Humanoid, which also took the standing high jump at 2.88 meters, comfortably past the 2.45 meters Javier Sotomayor cleared in 1993. In all, 666 teams from 16 countries brought 2,056 robots to compete in 51 events, from table tennis to tug of war.
Booster robots in formation at the Games in Beijing. (AP Photo / Achmad Ibrahim)
Here is what those records do and do not tell you. Sprinting is an engineering problem with one clear objective: apply force, stay upright, go straight. Actuators can be built stronger than muscle and control loops can run faster than a nervous system, so the record was always going to fall eventually. The part that has not fallen is everything else. None of these machines can walk into a kitchen it has never seen and put the groceries away.
Robots and human dancers at the opening ceremony, 22 August. (CGTN)
Which is why the Games read as industrial policy more than sport. The money has arrived well ahead of the use case. Unitree raised about $904 million listing on Shanghai's STAR Market on 19 August and closed its first day up more than 460%, the first humanoid maker to go public on a mainland exchange. XPeng's robotics unit raised more than $900 million this month at a $6.3 billion valuation, targeting mass production before the year ends. A stadium of synchronized robots is an advertisement, and the audience is in Shanghai and Washington.
The honest read is that humanoids sit somewhere between a demo and a product. They are improving quickly at everything you can time with a stopwatch. Whether they can hold down a job is a question the Games are not built to answer, and the companies competing in Beijing have not answered it either.
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 isn’t eSIM compatible.