With all the hype around Gemini 3, Google DeepMind CEO Demis Hassabis has chimed in on what he believes has been the secret behind the success of the model.
In a recent post on X, Hassabis responded to Oriol Vinyals, VP of Research and Deep Learning Lead at Google DeepMind, breaking down what he sees as the winning formula. “Actually if you want to know what the real ‘secret’ is it’s world-class research AND world-class engineering AND world-class infra all working closely together with relentless focus and intensity…” he posted.
The statement underscores a crucial advantage that Google has cultivated over more than a decade—one that goes far beyond simply having the best algorithms or the most compute power. It’s about controlling the entire AI stack, from silicon to software, in a way that few competitors can match.
The Technical Leap Behind Gemini 3
Vinyals’ original post provided technical context for the model’s impressive performance, highlighting two key areas of improvement: pre-training and post-training. On the pre-training front, he pushed back against the narrative that scaling has hit a wall, noting that “the delta between 2.5 and 3.0 is as big as we’ve ever seen.” This suggests that Google has found new pathways to extract value from increased scale, even as some industry voices have questioned whether the returns on larger models are diminishing.
On post-training, Vinyals described the landscape as “a total greenfield,” indicating substantial room for algorithmic innovation beyond simply training larger base models. This combination of breakthroughs in both pre-training scale and post-training refinement appears to have delivered the step-change in capabilities that has positioned Gemini 3 at the top of multiple benchmarks.
Google’s Early 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 the field for over a decade, and many of the foundational technologies powering today’s AI revolution originated within the company’s walls.
Most notably, the transformer architecture—the backbone of every major large language model today, from GPT to Claude to Gemini—was invented by Google researchers. The landmark 2017 paper “Attention Is All You Need” introduced the transformer, fundamentally changing how the industry approached natural language processing and eventually most other AI domains. While Google published this research openly, allowing competitors to build on it, the company retained deep institutional knowledge about the architecture and its potential applications.
Beyond the transformer, Google Research (now part of Google DeepMind) has been responsible for numerous other breakthroughs in machine learning, from advances in neural architecture search to innovations in training large-scale models efficiently. This research pedigree has given Google a sustained advantage in understanding both what’s possible and what’s practical in AI development.
The DeepMind Acquisition: Combining Research Excellence with Engineering Scale
Google’s 2014 acquisition of DeepMind, reportedly for around $500 million, now looks like one of the most prescient moves in tech history. At the time, DeepMind was a relatively small AI research lab in London, but it brought world-class research talent and a track record of breakthroughs in reinforcement learning.
The merger of DeepMind with Google Brain in 2023 to form Google DeepMind created a powerhouse that combines pure research excellence with Google’s engineering and infrastructure capabilities. DeepMind’s culture of fundamental research—exemplified by achievements like AlphaGo and AlphaFold—merged with Google’s ability to deploy systems at planetary scale.
This combination is exactly what Hassabis is alluding to in his recent comments. Having “world-class research” isn’t enough on its own; you need the engineering prowess to turn research ideas into production systems, and the infrastructure to train and serve models that push the boundaries of scale.
TPUs: The Hardware Advantage
One often-overlooked element of Google’s AI stack is its custom hardware. Google began developing Tensor Processing Units (TPUs) internally in 2013, recognizing early that general-purpose GPUs, while useful, weren’t optimized for the specific mathematical operations required by neural networks.
TPUs are application-specific integrated circuits (ASICs) designed specifically for machine learning workloads. They excel at the matrix multiplications and tensor operations that dominate both training and inference for neural networks. By designing its own chips, Google has been able to optimize for the exact operations its models require, potentially achieving better performance per watt and per dollar than competitors relying on off-the-shelf hardware.
The company is currently on 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 must rely on NVIDIA GPUs (which are in perpetually short supply), Google has its own dedicated supply of specialized AI hardware. Gemini 3, in particular, was fully trained on Google’s own TPUs as opposed to NVIDIA’s GPUs.
The integration between hardware and software at Google runs deep. The company’s machine learning frameworks can be optimized specifically for TPU architectures, and TPU designs can be informed by the specific needs of models in development. This vertical integration creates a flywheel effect where improvements in hardware enable better models, and learnings from model development inform the next generation of hardware.
Controlling the Entire Stack
When Hassabis talks about “world-class infra,” he’s referring to much more than just chips. Google’s infrastructure advantage spans multiple layers:
Data centers and networking: Google operates one of the world’s largest and most advanced networks of data centers, with private fiber connections between them. Training large models requires moving massive amounts of data between thousands of chips, and having a purpose-built, high-bandwidth network makes this dramatically more efficient.
Software frameworks: Google developed TensorFlow, one of the most widely used machine learning frameworks, and more recently JAX, which is particularly well-suited for research. Having control over the entire software stack—from low-level kernel optimization to high-level model APIs—allows for optimizations that wouldn’t be possible otherwise.
Data: As a company that operates Search, YouTube, Gmail, Maps, and countless other services, Google has access to vast, diverse datasets. While much of Gemini’s training likely uses publicly available data like competitors do, Google’s proprietary data sources can provide unique advantages for certain capabilities.
Operational expertise: Google has been running machine learning systems at massive scale for years—from search ranking to YouTube recommendations to Gmail spam filtering. This operational experience in deploying and maintaining AI systems is invaluable when it comes to launching products like Gemini.
By contrast, most of Google’s AI competitors control only parts of this stack. OpenAI relies on Microsoft’s Azure infrastructure and NVIDIA GPUs. Anthropic similarly depends on cloud providers (both AWS and Google Cloud) and NVIDIA hardware. Even well-resourced competitors like Meta, which does design its own training chips, don’t have the same depth of vertical integration across research, engineering, custom hardware, and global infrastructure.
The Integration Challenge
Of course, having all these pieces doesn’t automatically translate to success—they need to work together effectively. This is where Hassabis’ emphasis on these teams “working closely together with relentless focus and intensity” becomes critical.
Large organizations often struggle with coordination across different functions. Researchers might develop algorithms that are impractical to implement at scale. Engineers might optimize for the wrong metrics. Infrastructure might not evolve in sync with model needs. The fact that Google has apparently managed to align research, engineering, and infrastructure around a common goal for Gemini 3 is an organizational achievement as much as a technical one.
The merger of Google Brain and DeepMind was likely motivated in part by a desire to eliminate silos and create tighter integration. Rather than having two separate AI labs potentially working at cross-purposes, Google now has a unified organization that can align on priorities and execution.
Looking Ahead
The combination of factors Hassabis describes—cutting-edge research, strong engineering, and robust infrastructure—represents a formidable competitive advantage. While competitors are certainly capable of excellent work in individual areas, few can match Google’s strength across all three dimensions simultaneously.
That said, AI is moving fast, and advantages can be fleeting. The open release of models like Llama has shown that research breakthroughs diffuse quickly. Competitors are raising billions in capital to build out their own infrastructure. And other tech giants like Microsoft and Amazon have deep resources and strong engineering cultures of their own.
What Gemini 3’s success demonstrates is that in the current era of AI development, having the full stack—research talent, engineering excellence, custom hardware, and massive infrastructure—provides meaningful advantages. As Vinyals noted, there are “no walls in sight” when it comes to scaling, and post-training remains wide open for innovation. For now, at least, Google appears to have the pieces in place to capitalize on both opportunities.
Whether this integrated approach continues to deliver industry-leading results will depend on execution. But Hassabis’ comments suggest that Google DeepMind is betting that the secret to AI leadership isn’t any single breakthrough—it’s the relentless combination of research, engineering, and infrastructure working in concert.