🛎️ Musk Sells To Musk

Good Morning, AI Enthusiasts!

The boom is eating labor, sleep, benchmarks, GPUs, and apparently its own expense policy.

TOKEN

Musk Sells To Musk

👀 What's happening: Tesla told staff it will cap employee AI spending at $200 a week starting July 6, months after gamifying token use with leaderboards that ranked engineers by consumption. It worked too well: software engineers were burning thousands of dollars in tokens weekly. But the cap carves out one exception, beta versions of xAI products, Musk's own company. So spending is throttled everywhere except the tools that pay Musk. Grok isn't popular internally.

🌍 How this hits reality: When you need an expense policy to win internal market share, that isn't product strength, it's a captive audience dressed as adoption. And the mechanism scales past Tesla: SpaceX is buying Cursor's parent for $60 billion, so so Musk company owns the tool his engineers actually want, closing the loop where Musk sells to Musk.

🛎️ Key takeaway: A trillion-dollar AI thesis that can't win its own engineers without a spending cap isn't selling a product, it's mandating one and calling the captive traffic demand.

TOGETHER WITH BLOOME

Meet Bloome — Where Your AI Agents Actually Work Together

Most AI tools work alone. Bloome is a messaging platform where agents collaborate — one drafts, another challenges, a third catches what's missing. You get the sharpest answer, not just the first.

• All your agents, one place: Claude, ChatGPT, DeepSeek, any agent you choose — all in one conversation. No tab-switching, no lost context.
• Your whole team, always in context: Teammates and agents share the same thread. Anyone who joins later catches up instantly.
• Built for people who work through communication: For PMs, marketers, founders — people who move work forward. No code. No setup.

AGENTS

Half The Valley Can't Clock In, The Rest Can't Clock Out

👀 What's happening: The same AI wave is splitting the workforce two ways at once. On one side, agents quietly absorb the tasks that used to be entry jobs, leaving people with no shift to show up for. On the other, the builders can't log off. Wispr AI's CEO slept in the office for three straight weeks and bought beds for staff. A Gradient Ventures partner worked AI deals from the hospital sofa beside his wife's delivery room.

🌍 How this hits reality: This is the absurd core of the agent economy. AI does not just lengthen the workday, it raises the default stake for everyone in the game. Your agents keep shipping while you sleep, so does every competitor's, which turns rest from a shared rhythm into personal exposure. One group loses the job entirely; the other loses the right to stop. Bloomberg found AI's own architects can't quit building. Even MyClaw has started emailing users to go to bed, a tool built to run forever now negotiating with humans over sleep.

🛎️ Key takeaway: The scarce resource in an age of infinite output is not intelligence, it's the will to say there is no next step.

COMPUTE

SoftBank Joins The GPU Sale

👀 What's happening: SoftBank is entering the neocloud business in America, a new company called SB Neo renting GPU compute to hyperscalers for AI training, starting in fiscal 2027. The move gives it something to do with the 10-gigawatt US server farm it is building. Masayoshi Son called it "world-class AI infrastructure." But the timing tells another story: SoftBank Group just went back to lenders for a $10 billion loan against its OpenAI stake, and had to personally guarantee repayment after banks balked at the collateral.

🌍 How this hits reality: SoftBank isn't leading anything here. SpaceX is selling spare compute, Meta is opening its clusters, and now Son is scrambling to rent out silicon he already committed to buy. The pattern is the same: everyone who overbuilt for AI is now looking for someone to sell the excess to pay back its AI bill. Which raises the question the whole boom is built on avoiding. If this many giants are suddenly to offload compute, was compute ever as scarce as the trillion-dollar spending assumed? And that is before counting the incumbent neoclouds like CoreWeave already sitting on their own mountains of rentable silicon.

🛎️ Key takeaway: The moment everyone rushes to rent out their shovels, you know the gold rush priced the shovels for a scarcity that may not exist.

AGENT

Every AI Benchmark Is Lying To You

👀 What's happening: The UK's AI Security Institute tested frontier models across seven benchmarks and found the scores everyone trusts are broken. The problem is the fixed compute budget baked into every test. Give an agent more tokens to work with and it just keeps getting better: up to 25% higher on software tasks, 22% on math. One cybersecurity challenge that takes a human expert 20 hours went unsolved until models were handed over 30 million tokens, far past any standard benchmark's cutoff.

🌍 How this hits reality: Capability is a curve that climbs with spend, and every benchmark chops it off partway up. So every leaderboard, every "model X scores Y" headline, every safety evaluation regulators lean on is measuring the same thing: not what the agent can do, but what it can do on a budget nobody bothered to justify. The scores aren't conservative, they're wrong. AISI's own cyber progress rate went from doubling every 67 days to every 40 the moment the cap came off. We have been grading the frontier through a keyhole and calling it the whole room.

🛎️ Key takeaway: Every number used to rank, sell, or regulate these models is an artifact of a budget, not a measure of intelligence, and the entire industry has been quoting receipts as if they were IQ scores.

DAILY TL;DR

• Anthropic is facing a demand from more than 100 authors seeking over $75 million over books allegedly used to train AI systems.
• China is tightening rules on humanlike AI agents, including restrictions around emotional manipulation, identity deception, and unsafe advice.
• Singapore filed new charges in an alleged Nvidia GPU smuggling case, with authorities seizing a mansion and freezing bank accounts.
• Google is preparing Gemini 3.5 Pro for July, with a focus on longer context, agents, and higher token capacity.
• AI coding tools are causing workplace paralysis for some engineers, who now spend more time reviewing, debugging, and managing AI-generated code.
• Data center developers are still announcing new AI infrastructure projects, even as local opposition around power and land use keeps growing.

READ MORE

Let the Future Come to Your Inbox

Stay ahead without drowning in information. We turn the most important signals across AI, tech, marketing, and future products into 5-minute reads you can actually finish.

• AI Secret uncovers what really matters in AI
• Bay Area Letters decodes tech and business shifts from Silicon Valley
• Robotics Herald tracks how robots move from labs into daily life
• Marketing Secret breaks down real growth and go-to-market playbooks
• The Hardwire explores hardware, consumer tech, and what’s coming next

TOGETHER WITH US

AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, reaching over 2 million leaders across the global innovation ecosystem, from OpenAI, Anthropic, Google, and Microsoft to top AI labs, VCs, and fast-growing startups.

We've helped promote over 500 Tech Brands. Will yours be the next?

Email our co-founder Mark directly at mark@aisecret.us if the button fails.