CHINA BUILT A SECRET ARMY OF 25,000 FAKE CLAUDE USERS TO STEAL ANTHROPIC’S BRAIN
Source: CNBC | Tom’s Hardware
So here is what happened. Between April 22 and June 5 of this year, someone at Alibaba’s Qwen AI lab sat down and thought: why build the most advanced reasoning AI from scratch when you can just steal it from the Americans? They set up 25,000 fake Claude accounts. Twenty-five thousand. That is not a hack, that is a military operation. Those accounts then had 28.8 million conversations with Claude, carefully engineered to extract the most valuable capabilities Anthropic had spent years building. Then they fed all of that data straight into Qwen to make it smarter, faster, and better at the exact tasks Claude was trained to do.
Anthropic found out, blew a gasket, and wrote a letter to the Senate Banking Committee on June 10 calling it the largest distillation attack the company had ever seen, dwarfing everything that came before it combined. Distillation attacks, for the uninitiated, are when you query a competitor’s AI at industrial scale and use its answers to train your own model. It is essentially cloning someone’s brain by making them answer millions of test questions without telling them why. Legal? Grey area. Ethical? Absolutely not. Effective? Very much so.
The response from Washington came fast, and Anthropic probably did not love how it went. The Commerce Department slapped export controls on Anthropic’s Fable 5 and Mythos 5 models, meaning foreign nationals anywhere in the world could no longer access them. Since Anthropic has no reliable way to screen users by nationality, they had to shut both models off entirely for everyone. Their own best products, pulled from the market because the government panicked. Anthropic publicly disagreed with the decision, pointing out that the same jailbreaks the government cited also work on OpenAI’s GPT-5.5, which somehow did not get banned. Nobody could explain that one with a straight face.
The bigger story here is that the AI race between the US and China has moved from abstract rhetoric into actual industrial espionage at a scale that would embarrass a Cold War spy novel. Alibaba did not try to hire away Anthropic’s engineers or reverse-engineer the weights. They just made 25,000 fake accounts and had polite conversations with Claude until they had everything they needed. The playbook is so simple it is almost insulting. And the answer from Washington, instead of going after Alibaba directly, was to restrict Anthropic’s ability to sell its own product. The company that got robbed got punished. That is the kind of policy that makes you want to order another round.
ANTHROPIC IS ABOUT TO MAKE MONEY FOR THE FIRST TIME. OPENAI IS ABOUT TO LOSE $14 BILLION. SAME INDUSTRY, VERY DIFFERENT STORY.
Let us talk money, because the numbers coming out of the AI industry right now are genuinely insane in opposite directions depending on which company you are looking at. Anthropic, the company that most people a couple of years ago thought was the scrappy underdog that would get crushed by Google and OpenAI with their practically infinite resources, is on track to post its first operating profit in Q2 of this year. The projected number is $559 million in operating income for the quarter, on a revenue run rate that has crossed $47 billion annually as of this spring. A year ago, that number was $10 billion. They have grown their revenue nearly five times in twelve months.
OpenAI, in the meantime, is projected to lose $14 billion in 2026. Their infrastructure alone costs $25 billion a year to run. They owe Microsoft roughly $6 billion annually as part of their partnership agreement. They are burning cash at a pace that would make a Las Vegas casino blush. They generated about $5.7 billion in Q1, which is real money by any normal measure, but their costs are so catastrophically high that none of it matters yet. They are running what amounts to the most expensive charity in tech history, giving away access to their best AI for free or near-free and hoping the market share justifies it eventually.
What is interesting about this comparison is not just the numbers but what they say about strategy. Anthropic focused relentlessly on enterprise and API customers, meaning companies paying real money for reliable, high-quality outputs. OpenAI chased consumer scale, which means hundreds of millions of free users and ChatGPT becoming a household name. One of these strategies is producing operating profit in 2026. The other is producing the most famous AI product on the planet while bleeding out. You can decide which tradeoff you think is smarter, but investors looking at upcoming IPOs from both companies are going to have a very interesting conversation about which one they actually want to own.
The irony is that for months, the narrative was that Anthropic needed to catch up to OpenAI on consumer awareness. It turns out the company quietly building the boring business infrastructure was the one that figured out how to actually make money at this thing. Sometimes the nerd who stays in and does homework beats the popular kid who throws the best parties. At least in the balance sheets.
QUALCOMM DROPS $14 BILLION TRYING TO BECOME THE COMPANY THAT DETHRONES NVIDIA
Source: Bloomberg | CNBC | TechTimes
Qualcomm held its Investor Day on June 24 and spent the whole presentation making one very loud argument: we are coming for Nvidia. They announced the acquisition of Modular, an AI inference software startup, in an all-stock deal worth $3.9 billion. Modular makes software that lets you write AI applications once and deploy them on any hardware, whether that is an Nvidia GPU, an AMD chip, an Intel processor, or Qualcomm’s own silicon. The CUDA killer, in other words. The tool that makes the thing everyone hates about the AI industry, which is that everything is locked to Nvidia’s proprietary software stack, suddenly optional.
And then there is the Tenstorrent situation. Qualcomm is reportedly in advanced talks to acquire Tenstorrent, the AI chip startup built by Jim Keller, for somewhere between $8 billion and $10 billion. Keller is not a household name outside of chip nerd circles, but among chip nerd circles he is basically the Michael Jordan of processor design, having led major architecture work at AMD, Apple, Tesla, and Intel. Tenstorrent builds AI chips using the open RISC-V standard instead of proprietary architectures, which means they are betting that open source hardware eventually beats closed proprietary hardware the same way Linux beat Windows in the server room. That bet is not crazy.
Add these together and you get Qualcomm spending upward of $14 billion in a single week to build a full-stack alternative to Nvidia. Custom AI accelerators. Open instruction set architecture. Software that runs anywhere. Meta has already signed a multi-generational agreement to use Qualcomm’s new Dragonfly C1000 server processors, which is a meaningful endorsement from the company that runs some of the largest AI infrastructure in the world. If the Tenstorrent deal closes, Qualcomm would have the hardware, the compiler technology, and the software stack to credibly tell any AI company that they do not have to rent Nvidia’s kingdom anymore.
Whether that pitch works is a different question. Nvidia’s moat is not just the hardware, it is the decade of software investments, the ecosystem of tools, and the fact that every AI researcher on earth learned to code on CUDA. You do not displace that overnight. But every industry has its IBM moment, that inflection where the incumbent is so dominant that everyone assumes they are permanent, right up until they are not. Qualcomm is placing a very expensive bet that the AI chip market is about to have its IBM moment. The next two years will tell us if they are early visionaries or just very expensive optimists.
THE ERA OF THROWING TOKENS AT PROBLEMS IS OVER. EVERYONE JUST TRIPLED THEIR PRICES TO PROVE IT.
Source: CNBC
For the past two years, the unofficial strategy of every ambitious AI user was what CNBC this week called tokenmaxxing, which is a term that sounds like something a finance bro invented but actually describes something real. Tokenmaxxing means throwing as many tokens as possible at every problem, using the most expensive models for every task, turning on extended thinking for everything, and generally treating AI compute as if it were free because the models were so impressive you wanted to squeeze every drop out of them. Companies were burning money on AI inference the way startups burned money on office ping pong tables in 2018, which is to say enthusiastically and without a clear return on investment calculation.
That era is ending, and the price changes hitting the industry simultaneously are the clearest evidence. GPT-5.5 launched at double the per-token price of GPT-5.4. Gemini 3.5 Flash, which is supposed to be the cheap fast option, now costs roughly three times more than its predecessor Gemini 3.1 Flash. Claude Fable 5 came in at $10 per million input tokens and $50 per million output tokens, which is double what Opus 4.8 cost. Across the board, the message from every major AI lab is the same: the introductory pricing party is over, adults are drinking now, and the adults pay full price.
What is driving this is not greed, or not only greed. The labs genuinely cannot keep running at introductory pricing while building the next generation of models, which cost billions to train and billions more to run. OpenAI’s infrastructure bill alone is $25 billion annually. They cannot keep subsidizing that with venture money forever. The shift from tokenmaxxing to efficiency also reflects a maturing market. The companies that are getting real value from AI are the ones building specific, targeted workflows that use the right model for the right task rather than the most expensive model for every task. GitHub Copilot now explicitly routes users to the cheapest model that will do the job adequately. That is a grown-up approach.
The funny thing is that this shift probably helps the industry in the long run. When everything was cheap, nobody really had to think hard about whether the AI was actually doing something valuable. Now that prices are real, companies are going to have to justify their AI spending with actual numbers. That is going to kill a lot of pilot programs that were just vibes-based investments. It is also going to surface the use cases where AI genuinely delivers returns that justify the cost. The gold rush phase required cheap tools. The industrial phase requires real economics. We are entering the industrial phase, whether everyone is ready for it or not.