Elon Musk Steals, Meta Goes Mad Max, and Anthropic Gives Sunday Sermons and Hacks for the NSA on Monday

It's All Stolen: Elon Musk Trained on Anthropic's Outputs for Months

According to The Information, Elon Musk's xAI trained its own coding model for months directly on the outputs of Anthropic Claude—a classic case of distillation. When Anthropic blocked official access in January, xAI engineers continued working through private accounts and the intermediary service Blackbox AI. Musk had already admitted in court that xAI had “partially” used OpenAI models to train Grok, calling it the industry standard. Internally, the picture is bleak: the pre-training team has shrunk to fewer than five people, four Grok code leads have left within a few months, along with many co-founders. An employee accidentally deleted critical training data, costing the team two to three weeks of work. And all that computing power Musk bought? He's now renting it out to Anthropic and Google via SpaceX instead of training his own models. Supposedly, it's just a temporary solution. → Techpresso

Synthszr Take: When you train your coding model on a competitor's outputs, it's an admission that you can't dig the moat yourself. Musk has invested billions in GPUs, built Colossus from the ground up, established the largest compute position in the market, and in the end, he's leasing that computing power to the very company whose model he secretly siphoned. In May, we wrote about the “Drama-Rama” between Musk and Amodei, the enemy-of-my-enemy pact over SpaceX compute. Now we know the other side of the story. Compute buys you oxygen, but not a product: a pre-training team with five people and four departed code leads isn't a scaling problem, it's a substance problem. Let's remember that at Anthropic, 80 percent of the codebase comes from Claude Code. Then you see who's really getting recursively better here: perfect timing for the IPO of SpaceX and xAI...

Meta is Turning Its Data Centers into 'Mad Max' Sets

Meta is setting up tent-like structures across the U.S. to house AI servers. Tom's Hardware describes the scene as looking like something “out of the movie Mad Max”: the structures take only three months to become operational and are partly powered by jet engines. In parallel, an industry coalition is warning the Trump administration of the urgent need for action because the extreme memory appetite of AI data centers is starving other industries. The AI-driven memory bottleneck could drive up prices in the automotive, medical, and telecommunications sectors. At the same time, Anthropic reports that Claude is improving faster than expected, and Nvidia is still planning the RTX-50 Super series for 2026. The pace of expansion is enormous, and the side effects are just now becoming visible. → Tom's Hardware

Synthszr Take: A three-month construction time for a tent with a jet engine—that's the most honest statement about the state of AI expansion we've seen in months. There's no planning here, just provisional patching, because every month of delay is counted as a lost advantage. In January, I wrote about Meta's reality check with the metaverse; now the same company is burning kerosene to feed memory chips. The interesting part is the math behind it: when AI buys out the DRAM market, car manufacturers, clinics, and network operators pay the surcharge without ever having trained a model. This is the Jevons paradox in physical form: compute gets cheaper per token and more expensive for everyone else. Anyone planning in these sectors should already be factoring memory prices into their procurement as a strategic risk today, instead of waiting until the next quarterly report. Speed wins this market, but someone has to put the external costs on the bill, and it won't be Meta.

Anthropic Gives Sunday Sermons and Hacks for the NSA at the Same Time

According to the Financial Times, Anthropic has deployed about half a dozen engineers at the NSA to use its most powerful model, Mythos, for offensive cyber operations. Mythos is the same model that Anthropic does not release publicly, citing risks of misuse; it's available only through the closed project Glasswing with Microsoft, Apple, and Amazon. On the same day, the company's in-house institute published the study “When AI Builds Itself” and called for a globally coordinated moratorium, comparing it to the nuclear disarmament treaties of the Cold War. The reasoning: Claude now writes over 80% of the code that goes into production (in early 2025, the figure was in the low single digits), and engineers are shipping about eight times as much code per day as in 2024. In parallel, Anthropic is suing the Pentagon, which classified the company as a supply chain risk after a $200 million contract fell through—a label otherwise applied to Huawei. The point of contention: Anthropic refused to release Claude for fully autonomous weapon systems and domestic mass surveillance, but the NSA contract was exempt from this block. And all this is happening while Anthropic prepares for an IPO that could value the company at over a trillion dollars. → Techpresso

Synthszr Take: Anthropic is selling a hard brake to everyone else while stepping on the gas itself. The moratorium paper reads like an attempt to build a moat with moral concrete: whoever is already ahead and has the NSA as a client benefits most when the rest stand still. In April, we wrote here that Mythos is the AI that detects software vulnerabilities in a flash; now we see what that same capability is being used for, and it's not just defense. The 80% figure is the real bombshell: when a model writes the majority of its own successor, the race has become recursive, and no treaty with Cold War charm will voluntarily bring it to a halt. The stance against the Pentagon on autonomous weapons and domestic surveillance remains noteworthy, even if the NSA exception shows where the red line is drawn. Anyone aiming for a trillion-dollar valuation while calling on the industry to pause shouldn't be surprised when credibility is the first thing to go. You build trust with consistent behavior, not with two press releases on the same day.

Google Rents Compute from SpaceX for $920 Million a Month

One week before the biggest IPO in history, SpaceX has brokered its next compute deal, this time with Google. According to an SEC filing on Friday, Google will pay around $920 million per month from October 2026 to June 2029 for access to approximately 110,000 NVIDIA GPUs, along with CPUs and memory. The deal is similar to the agreement with Anthropic from late May, under which SpaceX will earn $1.25 billion monthly for the entire Colossus-1 data center in Memphis, originally built by xAI for its own purposes. This gives Google roughly half the computing power that Anthropic uses there, while Musk is likely reserving Colossus 2 for xAI. Notably, Google is considered the world's largest single owner of AI compute, yet it justifies the rental with unexpectedly high demand for Gemini Enterprise. Parent company Alphabet has already committed over $180 billion in capex this year and recently raised $80 billion through a stock issuance. SpaceX aims to raise around $75 billion in its Nasdaq IPO at a valuation of $1.75 trillion, and both companies are already discussing data centers in orbit. → Techpresso

Synthszr Take: When Google, despite $180 billion in capex and the title of the world's largest compute owner, still has to pay SpaceX $920 million a month, then computing power is the scarcest raw material of this decade. This is the Jevons paradox in its purest form: the more powerful Gemini Enterprise becomes, the more demand it creates, and efficiency gains evaporate into even more workloads. The money loop behind this is neat, as Alphabet will hold over $100 billion in SpaceX after the IPO and is now paying rent for its own investment. The 90-day termination clause starting at the end of 2026 reveals that even Google sees this only as a stopgap until its own TPU capacity catches up. For everyone playing at a smaller scale, this makes compute discipline basic arithmetic. If you don't design your model and vendor architecture for interchangeable capacity, you'll end up paying Memphis prices for something that should be a commodity.

Meta is Charging for AI for the First Time: Hatch to Cost Up to $200 a Month

Meta is building its first paid AI product, “Hatch,” with a price tag of up to $200 per month. Hatch is designed as a user-friendly version of the open-source tool OpenClaw: you describe what you need in simple language, and the agent builds a functional tool, schedules appointments, or sends emails. Internal documents show a free version as well as a “Hatch Plus” subscription with five to ten times higher usage limits, positioning Meta to compete directly with OpenAI and Anthropic, who charge $100 to $200 for their top-tier subscriptions. A broader U.S. launch is planned for July. Hatch is also intended to power Meta's planned AI hardware, including new Smart Glasses with “supersensing” and an AI pendant, with internal testing scheduled for spring 2027. Mark Zuckerberg sees AI agents as a path to revenue beyond advertising, which is necessary to recoup the massive infrastructure investments of around $600 billion that have already led to layoffs. Microsoft with Scout and Google with Gemini Spark recently introduced similar systems. → Techpresso

Synthszr Take: For two decades, Meta lived off the fact that users were the product and advertising paid the bill. Now, for the first time, there's a price tag on it, and it's $200 a month. That's the more honest part of the story: if you sink $600 billion into data centers and are already laying people off for it, you can't push the refinancing onto the ad business forever. What's interesting is that Hatch is built on OpenClaw, the same open-source tool that appeared in Microsoft Teams in June as a 24/7 colleague (we wrote about it in early June). Three corporations, one open tool, three price tags: the real competition is shifting from “who builds the best agent” to “who integrates it most deeply into hardware and daily life.” That's exactly where Meta's bet lies with its glasses and pendant, and that's also where it can fail if the $200 feels more like a mandatory subscription than a tenfold value-add. If you're charging $200 now, you have to deliver what wasn't possible for free before, otherwise the Plus button will remain unclicked.

China's Overlooked Model Builders: Xiaomi, Meituan, and StepFun

Tech Buzz China shifts the focus away from the usual labs and onto three names that often slip through the cracks in the AI debate. Xiaomi's MiMo model is directly linked to smartphones, cars, HyperOS, IoT devices, and its own distribution channels. Meituan operates its models, LongCat and Xiaomei, right in the middle of its offline service business, where demand, supply, routing, payment, and complaint data already converge. StepFun still appears to be an independent lab from the outside, but it is also directing its agent push towards smartphones. The analysis is part of the new China AI Atlas, a living field guide to labs, talent flows, investors, and parent platforms. In a companion piece, Weijin Research separates the question of where Chinese models are used globally from the question of where token production and value creation actually occur. → Tech Buzz China

Synthszr Take: The most interesting model builders in China aren't on the leaderboard; they're in the supply chain. A lab without distribution first has to find a home for its tokens; Xiaomi, Meituan, and StepFun already have the channel before the model is even finished. Meituan feeds LongCat with real transactions from local commerce—every order, every complaint, every route optimization becomes a training signal. That's a moat a pure research lab can't replicate without a funding round. In May, we wrote about Xiaomi's price discounts of up to 99 percent, and that only makes sense when inference happens within your own ecosystem instead of on a foreign platform. To understand Chinese AI, you have to look at the data flows in operation, not at benchmarks. The competition is decided by who has the touchpoint with the user.

The Skeptic's Guide to Viral Robot Videos

The timeline is full of humanoids sorting packages, folding laundry, and walking around in real-time streams. Dipam Patel, a PhD candidate in computer science at Purdue University and a Research Assistant at the US Army DevCom Army Research Lab, provides a sobering checklist in an Ars Technica piece. His first point: many of these demos run via teleoperation, with a human directly controlling the robot. As long as a paper or company doesn't explicitly talk about full autonomy, the highest level of suspicion is warranted. Patel also advises checking whether the robot is mastering a new test environment or just repeating a previously trained task in the exact setup where it learned it. And the classic: playback speed. Robots are slow for safety reasons; some videos run at 2x or 4x speed, meaning the robot actually takes twice or four times as long as a human. → Techpresso

Synthszr Take: The same principle applies here as with Vibe Coding: a demo that runs in the browser is miles away from production readiness. A robot grabbing a cup in its trained kitchen at 4x speed has about the same indicative value as code that works on localhost but crashes with a thousand concurrent users. The real progress is in generalization—the ability to autonomously handle an unfamiliar environment on the first try. That's precisely what can't be squeezed into a 15-second clip for the timeline, which is why we rarely see it. Anyone investing or buying should ask the two questions Patel suggests: fully autonomous or teleoperated, and new environment or repetition. Back in May, we wrote that AI is leaving the lab and real value is replacing demos; for humanoids, this dividing line is even sharper. Checking the playback speed takes ten seconds and saves a lot of false euphoria.

GEO, SEO, Content: Stop thinking in channels

On June 9, an online event called “BreakingSilos 2026” will take place, hosted by SE Ranking along with Planable and a long list of well-known marketing minds. The question behind it: How can brands secure visibility when artificial intelligence decides what gets seen at all, across SEO, social, content, and PR? Speakers include people like Rand Fishkin (SparkToro), Eli Schwartz (author of “Product-Led SEO”), Ross Simmonds (Foundation Marketing), and Crystal Carter, Head of AI Search & SEO at Wix. The core message is evident in the session titles: artificially generated brand signals will be filtered out or even penalized by AI systems, while brands with high citation rates in large language models are doing something else that is hard to copy. The thesis of the day is to stop managing SEO, social, and content as competing individual disciplines and instead treat them as one cohesive visibility system. The promise: actionable frameworks from practitioners, applicable that same week. → TAAFT - There's An AI For That

Synthszr Take: The title hits the mark, even if the event is ultimately a lead-gen funnel for SE Ranking. The point is still valid: in 2026, if you still have one team for “socials,” one for SEO, and one for PR, you're building silos that no longer map to the discovery model of ChatGPT and others. This is exactly the pipeline layer I write about in Code Crash. Everyone buys the tools—HubSpot-AI, Klaviyo, a few Copilot seats—that will be a commodity in 24 months. What makes the difference is the delivery logic behind it: brand voice as a pattern library, content as a cohesive asset rather than platform-specific output, fed by people with real domain knowledge. We already wrote in mid-May that Google is killing scaled AI content; now we're seeing the other side of the same coin, which is that authentic brand signals are the only currency a language model will still cite. If you organize your marketing department along channels instead of the outcome pipeline, you're optimizing a dead horse.