THE MAN WHO CO-INVENTED THE TRANSFORMER ARCHITECTURE JUST LEFT GOOGLE FOR OPENAI. AGAIN.
So there is a 2017 paper called “Attention Is All You Need.” Eight people wrote it. One of those eight people is named Noam Shazeer. If you want to understand why that matters, consider this: the Transformer architecture introduced in that paper is the technical foundation for every major AI model that exists today. GPT-4. Gemini. Claude. Llama. All of them. The Transformer is the engine, and Shazeer was one of the people who built it.
He left Google in 2021 to co-found Character.AI, the AI chatbot company. Google watched him walk out the door and then, in 2024, decided it needed him back badly enough to pay $2.7 billion to bring him home. Two. Point. Seven. Billion. Dollars. Google paid that because Shazeer and his team from Character.AI represented some of the best pretraining expertise in the world, and Google needed every advantage it could get against OpenAI and Anthropic. So they opened the vault and brought him back. He became co-lead of Gemini, one of the two or three most important model development programs on Earth.
And then on Wednesday he announced he is leaving Google to join OpenAI.
This is the kind of news that makes very smart people in AI look at their phones and do a double take. Shazeer is not just a famous name on a famous paper. He is still actively working at the frontier of model research. Google was counting on him to help build the next generation of Gemini. That work continues now without one of the most important people who was doing it.
Sundar Pichai said at Google I/O in May that Google is “a bit behind” in agentic coding. That comment was widely read as strategic honesty, a CEO getting ahead of a narrative before the narrative got ahead of him. In the context of losing Shazeer, that comment reads a little differently now.
For OpenAI, this is a significant moment. The company has been on an aggressive hiring campaign in the weeks before its IPO filing. The same week Shazeer announced his move, OpenAI also brought on Dean Ball, a former Trump administration AI policy official. The message OpenAI is sending to Wall Street is clear enough: we are the destination for the best people in this field.
The question nobody can answer yet is what Shazeer will actually work on at OpenAI. He was co-leading one of the most important model programs in the world. What does that translate to at a company simultaneously trying to build the best AI on Earth, file for an IPO, figure out its corporate governance, and compete with half a dozen very well-funded rivals? Something big, presumably. The man does not do small things.
Google spent $2.7 billion to solve this problem two years ago. They are back to zero, and the guy is on the other team. That is a rough week in Mountain View.
Source: CNBC | Axios | TechCrunch
AMAZON IS GOING TO SELL ITS AI CHIPS TO EVERYBODY. NVIDIA SHOULD PROBABLY PAY ATTENTION.
Andy Jassy wrote something interesting in his annual shareholder letter back in April. He said Amazon’s homegrown AI chips, called Trainium, had gotten so good and so in-demand that he was thinking about selling them externally to companies outside of Amazon’s own cloud. He estimated that if the chip operation were a standalone business selling to AWS and third parties at current rates, it would run at about $50 billion a year. That is not a side project. That is a company.
This week, Amazon’s AI chief Peter DeSantis confirmed that AWS is in active talks to sell Trainium chips to third-party data center operators. He declined to name potential customers. The talks are described as early stage.
Here is why this matters. Nvidia has controlled the market for AI training chips in a way that borders on monopoly for the last three years. The H100. The H200. The B200. Jensen Huang has been the gatekeeper of the hardware that every AI company in the world needs, and the waitlists for Nvidia chips have been a running joke in the industry. Everybody needs more chips and Nvidia is the only serious game in town.
For years, potential competitors have been making progress without making a dent. Google has TPUs. AMD has MI300. But none of them have gotten chips out the door, to enough customers, with enough software support, to genuinely threaten Nvidia’s position. The software ecosystem Nvidia has built around CUDA is a moat that most rivals cannot cross.
Selling Trainium externally changes the calculus, at least potentially. Amazon has real manufacturing relationships with TSMC, the same foundry that makes Nvidia’s chips. Amazon knows how to ship hardware at massive scale because that is the foundation of what AWS is built on. If they sell Trainium to third parties, those companies get a real, production-grade alternative to Nvidia’s products, made by one of the most capable infrastructure companies on Earth.
The obvious catch: every chip Amazon sells externally is a chip it is not using internally. AWS customers are on waiting lists too. Expanding external sales means someone in the existing queue waits longer. Amazon acknowledged this tension and suggested that solving the production constraints is part of the plan.
Nvidia’s stock barely moved on the news, which tells you what the market thinks of the short-term threat. And the market is probably right. Closing Nvidia’s software ecosystem advantage takes years, not press releases. But watch the trajectory. The first credible external challenger to Nvidia’s AI chip dominance is not going to come from a startup. It is going to come from a company with manufacturing relationships, distribution reach, existing customer bases, and a balance sheet to absorb losses while the market develops.
Amazon just raised its hand.
Source: TechCrunch | Bloomberg
AMAZON EMPLOYEES TESTIFIED AT CITY HALL ABOUT AI DATA CENTERS. THEN AMAZON CALLED THEM INTO HR.
A few weeks ago, five Amazon engineers decided to do something that people in a democracy are supposed to be able to do. They went to their local city council and testified at a public hearing about AI data centers. Seattle was considering a one-year moratorium on new large-scale data center construction, and the city wanted to hear from residents. These engineers showed up and said their piece.
Seattle passed the moratorium in a unanimous vote on June 9. That part worked out fine.
What happened next is where it gets interesting. Three of the Amazon employees were separately called into Zoom meetings with a human resources representative, who informed each of them that an investigation had been opened about a concern raised by their testimony. The employees were told the investigation could result in disciplinary action. One was told the potential consequences could go all the way to termination.
The workers filed a complaint with the Seattle Office for Civil Rights on June 18, arguing that Amazon violated a Seattle city ordinance that specifically prohibits employers from discriminating against employees based on political ideology. That law exists precisely for situations like this one.
Amazon’s response is worth reading carefully. The company says employees must follow certain internal procedures before speaking in a way that could be perceived as representing Amazon. Amazon looked at how the employees presented themselves and how their comments were received, and determined they may have been speaking “in their capacity as Amazonians and not as private citizens.”
The employees were testifying at a city council hearing, as individuals, about local public policy. They were not at an Amazon press event speaking on behalf of the company. They were people with opinions about data centers who showed up to a democratic process and said those opinions out loud.
The AI industry’s relationship with the people building it has always been complicated. Companies have been laying off tens of thousands of employees while simultaneously spending hundreds of billions on AI infrastructure. Amazon is a prominent example of exactly that pattern, announcing large layoffs while committing to enormous data center expansion. Some of the engineers doing that work have started speaking publicly about the pace and direction of the spending, raising concerns about environmental impact, community effects, and power grid implications.
The response from at least one of those companies appears to be: you can think whatever you want, but be careful about saying it where people can hear you.
The complaint is filed. The investigation is open. Amazon has not backed down. The employees are still employed, for now. This story is not finished.
WHILE WASHINGTON BANNED ANTHROPIC FROM SELLING AI ABROAD, MICROSOFT WAS QUIETLY COLLECTING A BILLION DOLLARS FROM CHINA’S TIKTOK COMPANY FOR OPENAI MODELS
Two weeks ago, the US government did something pretty dramatic. It ordered Anthropic to immediately shut off access to its two most powerful AI models, Fable 5 and Mythos 5, for every non-US national on Earth, regardless of where they were located. The official reason was national security. A company had apparently claimed it could jailbreak Mythos in a way that could enable dangerous capabilities. The ban was immediate, sweeping, and sent alarms through the AI industry about what export control policy could look like going forward.
Meanwhile, Bloomberg published a quiet story last week that has been getting considerably less attention than it deserves.
ByteDance, the Chinese company that owns TikTok, is Microsoft’s single largest AI customer. ByteDance has been buying OpenAI models through Microsoft’s Azure cloud platform for years, and it is currently on track to spend more than $1 billion on Microsoft AI and cloud services in 2026 alone.
Hold that next to the Anthropic story for a moment.
The US government emergency-banned an American AI company from selling its models to anyone outside the United States, citing national security. At the same time, an American tech giant is generating more than a billion dollars a year selling a different American company’s AI models to one of China’s biggest technology firms.
To be fair: these situations are not legally equivalent. The Anthropic action was triggered by a specific national security determination about specific models deemed dangerous enough to restrict under export control law. Microsoft’s sales to ByteDance are going through established cloud contracts under existing legal frameworks. No export control has been applied to the OpenAI models ByteDance is using through Azure.
But the optics are a genuine problem for anyone who wants to argue that US AI policy is coherent and consistent. The US has spent years treating ByteDance as a national security concern. Congressional hearings. Forced divestiture demands. Threats to ban TikTok entirely. And through all of it, ByteDance has been quietly the largest customer of an American company that was selling it products built by one of America’s most prominent AI startups.
The Bloomberg report landed on a Tuesday. Nobody in Washington said anything about it. The focus stayed entirely on Anthropic. The $1 billion ByteDance deal remained a footnote.
The AI policy debate in the US has been dominated by questions about which models are too powerful to export, who counts as an adversary, and what technology they should be allowed to access. The Microsoft-ByteDance relationship sits directly in the middle of all those questions and has somehow not become the conversation everyone is having.
It probably should be.
Source: Bloomberg