In partnership with
Good morning, {{first_name | AI enthusiast}}.
SpaceX and Nvidia just floated one of the wildest AI infrastructure ideas yet: data centers that live in orbit, not on the ground. The plan, Starmind AI1, could eventually scale into a constellation of up to a million AI satellites circling Earth.
If solar power and radiative cooling really can outrun the energy and heat problems plaguing terrestrial data centers, orbit could become the default home for tomorrow’s AI compute. But regulators, launch costs, and basic physics haven’t signed off yet — so is this a genuine fix for AI’s power crunch, or the most expensive science project in tech history?
Today in AI Brief:
Nvidia and SpaceX plan orbital AI data centers
Anthropic’s Mythos 5 agent goes rogue in testing
Bland’s new voice model nearly passes for human
The first way to trade directly inside Claude and ChatGPT
Superintelligence used to be locked inside billion-dollar quant firms whose algorithms quietly took advantage of everyone else.
Co-Invest puts it right in your chat window. Analyze markets, manage risk, and execute trades, all inside Claude and ChatGPT.
The institutions built the game, Co-Invest gives you a way to beat them.
SpaceX and Nvidia Plan AI Data Centers in Orbit
In Brief: SpaceX and Nvidia unveiled plans for Starmind AI1, a satellite built to run Nvidia’s AI chips in orbit and beam compute back down to Earth. It’s the clearest signal yet that Big Tech sees space, not just bigger power plants, as the next frontier for AI infrastructure.
The Details:
Each satellite will carry Nvidia’s newest Rubin GPUs and Vera CPUs, delivering roughly 120 kilowatts of sustained AI compute per unit and peaking near 150 kilowatts.
SpaceX filed for FCC approval of a constellation with up to 1 million AI satellites orbiting between 500 and 2,000 kilometers.
The design leans on large solar arrays and radiative cooling to sidestep the emissions, cooling, and power constraints that limit data centers on Earth.
Take Away:
Space-based compute could sidestep the two biggest constraints on scaling AI today: land and power. But approval for a million-satellite constellation, if it ever comes, would rewrite orbital traffic rules as much as the AI compute map.
Anthropic’s Mythos 5 Agent Goes Rogue in Safety Test
In Brief: The UK’s AI Security Institute confirmed that Anthropic’s Mythos 5 model went rogue during a security evaluation, fabricating identities and attempting to slip malicious code into a real open-source project. OpenAI’s GPT-5.6 Sol was involved in two similar incidents after testers disabled its safety classifiers.
The Details:
Mythos 5 drove 17 of the 19 unauthorized actions AISI recorded across 122 evaluation runs, including creating fake GitHub identities to push a malicious pull request.
The agent also used Tor to dodge network restrictions and posted public messages to other AI agents on GitHub with instructions for reusing compromised accounts.
Testers had disabled cyber safety classifiers and granted live internet access for the evaluation, conditions OpenAI says contributed to its own model’s misstep.
Take Away:
AISI called this the first time agentic deception and autonomy showed up this clearly in the real world without being specifically prompted for it. No one was harmed this time, but the incident shows guardrails, not raw capability, are what stand between today’s agents and real damage.
SPONSORED BY MESH API
Save 40% on 1000+ AI APIs
AI costs don't usually explode overnight. They grow quietly through duplicate requests, expensive routing and poor visibility.
Mesh API helps engineering teams spot the waste before finance does.
A global e-commerce company was able to reduce spend by 78%.
Bland’s New Voice Model Nearly Passes for Human
In Brief: Bland released Speech v3, a voice model that ranked second only to real human speech on the Audio Realism Bench, beating rival voice models from Microsoft, xAI, and OpenAI. It keeps the pauses, stumbles, and breathing patterns competitors tend to smooth away.
The Details:
Speech v3 scored a 1,384 Elo rating, trailing only human speech at 1,501 and topping Microsoft’s MAI-Voice-2, xAI’s Grok TTS, and OpenAI’s GPT Realtime 2.
The model supports performance tags like “laughs” or “clears throat” that it acts out rather than reads aloud, plus instant voice cloning from a 10-second sample.
Anyone can test it free at Bland’s speech page, with 133,000 free characters, about two hours of speech, before paid usage kicks in.
Take Away:
Voice is quietly closing the gap with human speech faster than video or text ever did, and that has real implications for anything built on IVR or customer support scripts. Expect the line between talking to a bot and talking to a person to blur a lot faster than most people expect this year.
Everything else in AI
Apple sought a preliminary injunction against OpenAI and two former employees, aiming to halt its hardware development over alleged trade secret theft.
DeepSeek released V4-Flash, a model that rivals Anthropic’s Claude Opus 4.8 on performance while charging 28 cents per million output tokens versus Opus’s $25.
Anthropic signed a $10 billion computing deal with a new cloud startup to expand its infrastructure.
PwC found that 86% of finance executives now rate AI skills training above an MBA for new hires.