Amidst all the hype surrounding Gemini 3, Google DeepMind CEO Demis Hassabis shared his thoughts on what he believes is the secret to the model’s success.
In a recent post on X, Hassabis responded to Oriol Vinyals, VP of Research and Deep Learning Lead at Google DeepMind, and unraveled his winning formula. “In fact, if you want to know what the real ‘secret’ is, it’s that world-class research, world-class engineering, and world-class infrastructure all work closely together with relentless focus and concentration…” he posted.
This statement highlights a key advantage that Google has cultivated for more than a decade: It’s not just about having the best algorithms or the best computing power. It’s about controlling the entire AI stack, from silicon to software, in a way that few competitors can match.
Gemini 3 technological advancements
Vinyals’ original post provides technical context for the model’s excellent performance and highlights two key areas of improvement: pre-training and post-training. On the pre-training front, he refuted the notion that scaling has hit a wall, pointing out that “the difference between 2.5 and 3.0 is the biggest I’ve ever seen.” This suggests that Google has found a new way to extract value from its increased scale, even as some in the industry question whether profits on larger models are shrinking.
Vinyals described the post-training situation as “completely greenfield,” showing that there is more room for algorithmic innovation than simply training a large base model. The combination of breakthroughs in both pre-training scale and post-training improvements appears to have resulted in incremental changes in functionality that have placed Gemini 3 at the top of multiple benchmarks.
Google’s first mover advantage
To understand Google’s current position, it’s worth looking back at the company’s pioneering role in modern AI. Google has been at the forefront of this space for over a decade, and many of the foundational technologies that power today’s AI revolution were born in-house.
Most notably, researchers at Google invented the transformer architecture, which is the backbone of all today’s major large-scale language models, from GPT to Claude to Gemini. The landmark 2017 paper “Attending Is All You Need” introduced Transformers and fundamentally changed the way the industry approaches natural language processing, and by extension, most other AI domains. Although Google published this research openly, allowing competitors to build on it, the company retained deep institutional knowledge of the architecture and its potential applications.
Beyond Transformers, Google Research (now part of Google DeepMind) has contributed to numerous other breakthroughs in machine learning, from advances in neural architecture search to innovations in efficiently training large-scale models. This research stream has given Google a continued advantage in understanding what is possible and what is practical in AI development.
DeepMind Acquisition: Combining Research Excellence with Engineering Scale
Google’s acquisition of DeepMind in 2014, reportedly for about $500 million, looks like one of the most visionary moves in technology history. At the time, DeepMind was a relatively small AI research institute in London, but it brought world-class research talent and groundbreaking achievements in reinforcement learning.
In 2023, DeepMind and Google Brain merged to form Google DeepMind, creating a powerful company that combines pure research excellence with Google’s engineering and infrastructure capabilities. DeepMind’s culture of basic research, exemplified by achievements like AlphaGo and AlphaFold, has merged with Google’s ability to deploy systems on a global scale.
This combination is exactly what Hassabis has hinted at in recent comments. It is not enough to have “world-class research.” You need the engineering capacity to turn research ideas into production systems and the infrastructure to train and deliver models that push the limits of scale.
TPU: Hardware Advantages
One of the often overlooked elements of Google’s AI stack is custom hardware. Google recognized early on that general-purpose GPUs, while useful, were not optimized for the specific mathematical operations required for neural networks, and began developing Tensor Processing Units (TPUs) internally in 2013.
TPUs are application-specific integrated circuits (ASICs) designed specifically for machine learning workloads. They excel at matrix multiplication and tensor operations, which dominate both neural network training and inference. By designing its own chips, Google is able to optimize the exact behavior its models need, potentially delivering better performance per watt and per dollar than competitors that rely on off-the-shelf hardware.
The company is currently developing its sixth generation of TPUs, and these chips power both Google’s internal AI services and are available to cloud customers through Google Cloud Platform. This means that while competitors like OpenAI and Anthropic have to rely on NVIDIA GPUs (which are constantly in short supply), Google supplies its own dedicated AI hardware. Notably, Gemini 3 was trained entirely on Google’s own TPUs rather than NVIDIA’s GPUs.
Google has deep hardware and software integration. The company’s machine learning framework can be optimized specifically for TPU architectures, allowing you to tailor your TPU design to the specific needs of the model you are developing. This vertical integration creates a flywheel effect where hardware improvements enable better models and learnings from model development inform the next generation of hardware.
Control over the entire stack
When Hassabis talks about “world-class infrastructure,” he’s not just talking about chips. The benefits of Google’s infrastructure span multiple layers.
data center and network: Google operates one of the world’s largest and most advanced data center networks, with private fiber connections between data centers. Training large models requires moving large amounts of data between thousands of chips, and using a dedicated high-bandwidth network makes this dramatically more efficient.
software framework: Google developed TensorFlow, one of the most widely used machine learning frameworks, and more recently JAX, which is particularly suited for research. Control over the entire software stack, from low-level kernel optimizations to high-level model APIs, enables optimizations that would otherwise not be possible.
data: As the company behind search, YouTube, Gmail, Maps, and countless other services, Google has access to vast and diverse data sets. While much of Gemini’s training may use publicly available data like its competitors, Google’s own data sources may offer unique advantages for certain features.
operational expertise: Google has been running massive machine learning systems for years on everything from search rankings to YouTube recommendations to Gmail spam filtering. This operational experience in deploying and maintaining AI systems is invaluable when launching a product like Gemini.
In contrast, most of Google’s AI competitors control only part of this stack. OpenAI relies on Microsoft’s Azure infrastructure and NVIDIA GPUs. Anthropic similarly relies on cloud providers (both AWS and Google Cloud) and NVIDIA hardware. Even well-resourced competitors like Meta, which designs its own training chips, don’t have the same depth of vertical integration across research, engineering, custom hardware, and global infrastructure.
The challenge of integration
Of course, having all of these factors does not automatically guarantee success. These must work together effectively. It is important here that Hassabis emphasized that these teams will “work closely together with relentless focus and intensity.”
Large organizations often struggle with coordination between different departments. Researchers may develop algorithms that are impractical to implement at scale. Engineers may optimize for the wrong metrics. The infrastructure may not evolve in sync with the needs of the model. The fact that Google has clearly succeeded in aligning its research, engineering, and infrastructure around a common goal for Gemini 3 is as much an organizational achievement as it is a technical one.
The Google Brain and DeepMind merger was likely motivated by a desire to eliminate silos and achieve tighter integration. Google now has a unified organization that can coordinate on priorities and execution, rather than two separate AI labs that may work together for multiple purposes.
Looking to the future
The combination of factors Mr. Hassabis describes—cutting-edge research, strong engineering, and robust infrastructure—represents a formidable competitive advantage. Competitors can certainly do a better job in individual areas, but few companies can match Google’s strengths in all three dimensions at the same time.
However, AI advances quickly and its benefits may be temporary. The open release of models like Llama signals the rapid dissemination of research breakthroughs. Competitors are raising billions of dollars to build their own infrastructure. Additionally, other technology giants such as Microsoft and Amazon have deep resources and strong engineering cultures of their own.
The success of Gemini 3 shows that in the current era of AI development, having a full stack of research talent, engineering excellence, custom hardware, and extensive infrastructure provides meaningful advantages. As Vinyals pointed out, “the walls are invisible” when it comes to scaling, and there is still huge potential for innovation post-training. At least for now, Google appears to have the pieces in place to take advantage of both opportunities.
Whether this integrated approach continues to deliver industry-leading results depends on its execution. But Hassabis’ comments suggest that Google DeepMind is betting that the secret to AI leadership is not a single breakthrough, but a constant combination of research, engineering, and infrastructure working in tandem.
