Who Is Noam Shazeer: From Transformer to Gemini
If you write a prompt into ChatGPT, Claude, or Gemini today, you are using technology that Shazeer helped create. In 2017, he co-authored the scientific paper "Attention Is All You Need" — the document that introduced the Transformer architecture to the world. It has become the foundation of virtually all modern large language models. Without exaggeration, it is one of the most cited research papers in the history of computer science.
Shazeer's career at Google began as early as 2000, when he joined the company as one of its first few hundred employees. Over more than two decades, he rose to become one of the most influential artificial intelligence researchers in the world. In recent years, he served as vice president of engineering and co-led the Gemini project — Google's flagship in the field of large language models, which today powers hundreds of millions of users across the Google Workspace ecosystem.
A $2.7 Billion Journey: How Google Bought One Man Twice
Shazeer's relationship with Google was not always straightforward. In 2021, he left the company and co-founded the startup Character.AI — a platform that allowed users to create and communicate with personalized AI characters. The startup quickly gained popularity, especially among younger users, and was among the first to show that conversational AI could also be an entertainment platform, not just a productivity tool.
Google was well aware of what it had lost. In 2024, it therefore paid $2.7 billion to acquire Character.AI — primarily with the goal of bringing Shazeer and his team back under its roof. It was one of the most expensive acqui-hire transactions in tech history. The transaction drew attention from regulators, as it was part of a broader trend of tech giants buying AI talent through acquisitions of entire startups.
Now, less than two years later, Shazeer is leaving Google again — this time heading to OpenAI. For Google, this means not only the loss of a key researcher but also a return on investment that, from a talent retention perspective, proved to be short-lived.
OpenAI Before a Historic IPO: Trillion-Dollar Ambitions
Shazeer's arrival is not an isolated personnel announcement — it is part of a much larger story. OpenAI is preparing for its initial public offering (IPO), which could go down in history as one of the largest ever. According to sources at CNBC and BeInCrypto, the company has confidentially filed its S-1 application with the SEC, with investment banks Goldman Sachs and Morgan Stanley serving as the lead advisors.
The numbers being discussed around the IPO are staggering:
- OpenAI Valuation: private investors already value the company at more than $850 billion. At the public offering, the valuation could exceed the one trillion dollar mark, bringing OpenAI into the exclusive club of companies like Apple, Microsoft, or Nvidia.
- Capital Raised: OpenAI has so far raised over $180 billion from investors, making it one of the most heavily funded startups in human history.
- Competition: Anthropic, OpenAI's main rival and the creator of the Claude model, has reached a valuation of $900 billion and, according to available information, generates $30 billion in annualized revenue. In recent months, Anthropic has dominated particularly in the enterprise and developer segments.
- SpaceX (with xAI): Elon Musk's company, which is also preparing to go public after merging with xAI, creates additional competitive pressure. The result is an exceptionally competitive IPO season in the AI sector, attracting investor attention from around the world.
Shazeer's arrival at OpenAI in this context sends a strong signal to the market: despite growing competition from Anthropic and SpaceXAI, OpenAI can still attract the absolute top research talent. For investors considering participation in the IPO, this is a significant signal of confidence.
What Shazeer's Move Means for AI Development
The move of such a pivotal figure as Shazeer has several layers of impact:
For Google: the loss is all the more painful because Gemini faces fierce competition from both Claude and ChatGPT. Google recently launched Gemini 3.5 Flash with significantly improved capabilities and lowered subscription prices — but Shazeer's departure comes at a time when the company needs every research capacity.
For OpenAI: gaining Shazeer represents a strengthening of the research team at a critical moment. With the IPO approaching, the company needs to demonstrate not only financial results but also the ability to continue attracting and retaining top talent. Shazeer's expertise in the Transformer architecture — the very foundation of the GPT models — is invaluable in this context.
For the AI labor market: Shazeer's move is further evidence of the unprecedented mobility of top AI researchers. Recall Andrej Karpathy, a co-founder of OpenAI, who recently moved to Anthropic. In an industry where a few hundred people determine the technological direction of the planet, every such move is geopolitically significant.
The Czech and European Dimension: What It Means for Us
For the Czech reader, the whole situation around trillion-dollar valuations and the movement of research stars may seem distant. However, it has concrete implications:
First, the concentration of AI talent in a few American companies means that European — and even more so Czech — startups and research institutions compete for talent on an extremely uneven playing field. OpenAI and Anthropic offer compensation packages in the order of millions of dollars per year, which is unattainable for European entities.
Second, OpenAI's IPO will impact European investors. European pension funds, insurance companies, and institutional investors will have to consider whether and to what extent to participate in the AI IPO. While the EU AI Act regulates the deployment of models on the European market, the ownership structure of the companies behind the models remains beyond its reach.
Third, Shazeer's story is a reminder that Europe has so far failed to create an environment that would prevent brain drain to the U.S. The Czech Republic has a strong technical tradition — recall the successes of the IT Faculty at CTU or the activities of the Czech AI Factory in Ostrava, which was created as part of the European network of AI centers. Yet the gap between a lab in Prague or Ostrava and a research team in San Francisco remains enormous.
Summary: The Man Who Created the Transformer Bets on OpenAI
Noam Shazeer's departure from Google to OpenAI is an event that connects three key stories of the current AI landscape: the ruthless battle for talent, historic IPO ambitions, and a return to fundamentals — Shazeer is returning to the architecture he helped create and on which OpenAI built its entire business. Whether this will be enough for OpenAI to stand up to Anthropic and SpaceXAI in the eyes of investors will be revealed in the coming months. One thing is certain: the AI labor market is more fluid than ever before — and that is precisely the sign of an industry still at the beginning of its trajectory.
Why did Google pay $2.7 billion for Character.AI if Shazeer ended up leaving less than two years later?
Through the acquisition of Character.AI, Google gained not only Shazeer and his team but also the technology and user base of the platform. Although Shazeer's departure is a loss, Character.AI's technology has been integrated into the Google ecosystem and part of the team remains at the company. In retrospect, however, it shows that money alone is not enough to retain top talent — vision and the company's direction are equally crucial.
When exactly is the OpenAI IPO expected and can Czech citizens invest in it?
The exact IPO date has not yet been set, as the company filed its S-1 confidentially. Typically, 3 to 4 months elapse between the S-1 filing and the first day of trading. Czech citizens can participate in the IPO through brokers with access to U.S. exchanges (e.g., Interactive Brokers, XTB, Fio), once the shares become publicly tradable.
What is the difference between the Transformer architecture by Shazeer and today's models like GPT-5?
The Transformer architecture from 2017 is the foundational building block — the principle of how a model processes text in parallel and "pays attention" to different parts of the input simultaneously. Today's models like GPT-5 build on this foundation but add orders of magnitude more parameters, training data, better safety mechanisms, and multimodal capabilities (working with images, audio, video). It's a similar relationship as between the internal combustion engine and a modern supercar — the principle remains, but everything else has radically advanced.