Google is on track to become the new leader in artificial intelligence.

After ChatGPT 3.5 was released in November 2022, it seemed that Google had fallen behind in the race for artificial intelligence. Google’s response to OpenAI’s actions, the Bard chatbot, was simply poor, while AI Overviews (i.e. summaries of web pages displayed in search results) advised people, for example, to add glue to pizza.

How poorly Google performed came as quite a surprise to many people. After all, it is a company whose existence has been based from the very beginning on algorithms, machine learning, and artificial intelligence. In 2017, Google researchers created the legendary paper “Attention Is All You Need,” which became the foundation of the transformer architecture and LLMs (i.e. language models). There were even voices claiming that chatbots would replace Google Search, and that SEO itself (optimizing a company’s website for Google) would no longer be enough, requiring the development of GEO (Generative Engine Optimization) or AISEO—in other words, making sure that chatbots recommend a given company. Many people assumed that Google was too large and too rigid a company to respond appropriately to the situation that had arisen.

A few months ago, I wrote about how Google is finally making good AI. Gemini, the successor to Bard, worked much better, and Notebook LM deserves particular attention—a new type of AI product based on language models but configured to process data provided by the user.

It is now becoming clear that Google has not only caught up with the frontrunners but has taken the lead in the race and is well on its way to becoming the undisputed leader. The launch of the Gemini 3 model was crucial; it proved to be a better model than ChatGPT 5, which had been released only a short time earlier.

What Is Google’s Advantage? An Ecosystem OpenAI Doesn’t Have.

Google did not win thanks to a single chatbot or flashy launches. Its advantage is structural and multi-layered. The company controls the entire AI stack, from designing its own TPU chips (based on tensors rather than graphics cards), through data centers and cloud services, to language models and ready-made products used daily by billions of people. This allows it to reduce costs where competitors have to pay external providers, optimize the system as a whole rather than individual components, and roll out new features globally faster.

On top of that, there is the vast ecosystem—Gmail, Search, YouTube, Android, and Maps—which makes AI a natural addition to the tools already in use, rather than a separate application that requires an additional payment.

The mechanism works as follows: if someone is already paying for Google Workspace (that is, a more professional version of Google Drive) and immediately has access to Gemini in office applications as well, why would they pay separately for ChatGPT?

Google Search, YouTube, Maps, Waymo autonomous cars provide Google with enormous amounts of data that can be used to train models.

Combined with its own infrastructure, Google Cloud’s growing profitability, research and financial resources, and freedom from the most burdensome antitrust restrictions, this creates an advantage that no other company can easily challenge today.

Gemini 3, Veo 3.2, and Nano Banana Pro showed that Google not only matches ChatGPT, but can surpass it. Admittedly, various models regularly top the benchmarks, and Claude and Grok have achieved this as well, but in this case the impact was much greater, and the shift more lasting. Gemini 3 is particularly effective for designing the visual appearance of applications (UX), and such prototyping is one of the better uses of vibecoding.

Google’s AI is above all faster and cheaper than solutions developed by OpenAI.

Hardware Designed for AI

Google’s AI processors, known as TPUs (tensor processing units), differ from graphics cards in that they are designed exclusively for one task: the massive tensor computations used in AI models. They do not handle graphics or a broad range of applications like GPUs do; instead, they perform precisely the operations that dominate model training and deployment.

GPUs are flexible and excellent for experimentation and developing new models, which is why they are the standard in startups and research. TPUs come into play when AI operates at an industrial scale: it needs to be affordable, fast, energy-efficient, and stable. Thanks to their close integration with Google’s data centers, they make it possible to run models more cheaply than on cards purchased from external suppliers.

A GPU is a versatile tool for work and testing, while a TPU is specialized infrastructure for large-scale use of AI in finished products.

How ChatGPT Saved Google

Paradoxically, ChatGPT’s popularity helped Google. In 2020, the company faced a federal antitrust lawsuit, accused of illegally monopolizing the search market. The threat of breaking up the conglomerate was real; Chrome, Android, and the possible separation of key assets could have permanently changed the company’s structure.

Paradoxically, the emergence of ChatGPT in November 2022 became an unexpected lifeline for Google. For the first time in years, a real competitor emerged for its core search business, which generates over $100 billion annually from advertising. When the judge considered remedies in 2024, the competitive landscape was radically different from what it had been two years earlier. Ultimately, the company was not broken up, as it was determined that AI had introduced sufficient competition.

The result? Google emerged from the process stronger, retaining full vertical integration and the strategic freedom of maneuver that competitors can only envy.

Artificial intelligence stack (AI stack): Five key components of language models.

To understand why some companies dominate in the era of artificial intelligence while others fall behind, AI must be viewed not as a single product, but as a complex technology stack consisting of five fundamental layers.

Infographic created in NanoBanana Pro – a tool available in Google Gemini.

Layer 1: Chip Design (Silicon)

At the very bottom are specialized integrated circuits designed for AI computations. These are not ordinary processors—they are chips optimized for specific mathematical operations (matrix multiplication, tensor computations), which form the core of machine learning. Nvidia dominated this space for years with its A100 and H100 GPUs, but alternatives are emerging, such as Google’s TPUs, Amazon’s Trainium, and chips designed by Apple.

Whoever controls chip design controls the economics of the entire stack, and can reduce costs and optimize performance in ways unavailable to companies purchasing hardware externally.

Layer 2: Data Centers (Infrastructure)

Thousands of these chips need to be installed in enormous halls and provided with power (often hundreds of megawatts per facility), cooling, and connectivity. This is physical infrastructure requiring billion-dollar investments and years of construction. Microsoft Azure, Amazon AWS, Google Cloud, and Oracle are companies that rent out this computing power to others.

Layer 3: Foundation Models

These are the true “brains” of AI—large language models such as GPT, Claude, Gemini, and Llama. Training them costs tens or hundreds of millions of dollars and requires months of computation on thousands of chips. A model is not just code, but “encoded knowledge” extracted from terabytes of training data.

Training the best models requires access to layers 1 and 2, vast data resources, and the best research teams. The barrier to entry grows every year.

Layer 4: Products and Applications (Applications)

This is the level seen by the end user: chatbots, assistants, content creation tools, and AI features built into applications. ChatGPT, Claude, Gemini, and Copilot are the most popular examples, but there are also thousands of specialized applications using the APIs of large models.

The product determines adoption. The best model in the world is worthless if no one can use it conveniently. UX, integrations, and the ecosystem are of fundamental importance here.

Layer 5: Consumer Equipment (Devices)

Physical devices through which people access AI—smartphones, laptops, tablets, future AR glasses, or other wearable devices. Apple with iPhones, Samsung with Galaxy devices, Google with Pixels: controlling the hardware gives control over how and when users interact with AI.

Layers 0: Energy and Talents

When we talk about AI infrastructure, we must not forget two key components: Energy and Scientists. For large technology companies, providing energy for their AI factories is not a major challenge. For example, in order to launch Grok quickly, Elon Musk rapidly built a power plant based on gas turbines. Although the costs of such temporary measures are higher, they give companies time to build more efficient, cheaper power sources for their factories.

A much greater challenge is attracting talent—scientists capable of creating artificial intelligence. There are relatively few of them because, over the past dozen or so years, this field was niche compared with other areas of computer science.

Five Fundamental Google Advantages in the AI Era

While the world focuses on the rivalry between ChatGPT and Gemini, Google’s true advantage lies much deeper—in the infrastructure it has built over the past two decades. This is not a story about a better chatbot, but about systematic dominance at every level of the AI technology stack.

Advantage 1: Full Control Over Computing Infrastructure

Google is the only major technology company that has designed and deployed its own AI chips at industrial scale. TPUs (Tensor Processing Units) are not an experiment—they are the foundation of all AI operations at Google. While OpenAI pays Microsoft (which pays Nvidia) for every second of training, Google uses its own integrated circuits, eliminating intermediaries and their margins. The economics are brutal: where competitors spend billions on hardware, Google invests in R&D and expanding its own capacity.

Advantage 2: A Global Data Center Network Built Over Years

Physical infrastructure is something that cannot be copied quickly. Google has dozens of data centers across six continents, connected by its own fiber-optic cables running along the ocean floors. This network was built over 25 years and has cost hundreds of billions of dollars. OpenAI may have brilliant models, but it does not have facilities where it could train them efficiently without renting someone else’s infrastructure. Microsoft and Amazon have data centers, but they do not have their own chips. Google has both.

Advantage 3: Decades of Experience in Machine Learning

It was Google that invented the Transformer architecture in the famous 2017 paper “Attention Is All You Need” — the technology powering all modern large language models, including GPT. While OpenAI had to build its expertise from scratch after 2015, Google has teams working on machine learning since the early 2000s. Google Brain and DeepMind are institutions employing Turing Award winners, authors of the most cited AI publications, and creators of AlphaGo and AlphaFold. This intellectual and institutional capital is an advantage that cannot be quickly overcome.

Advantage 4: Access to training data on an unprecedented scale

Every day, more data flows through Google’s servers than through any other company in the world. YouTube is the largest video database on the planet. Google Search processes 8.5 billion queries a day, revealing people’s intentions, questions, and needs in real time. Gmail, Maps, Chrome—each of these products generates signals about human behavior on an unprecedented scale. Data is the fuel of AI, and Google is sitting on the world’s largest oil reserve. What’s more, most of this data is unique—no one outside Google has access to the search histories of billions of users.

Advantage 5: Financial Power Enabling Hundreds of Billions in Investment

Google generates over $300 billion in revenue annually, with operating profit exceeding $80 billion. This gives the company the luxury of thinking long term and the ability to invest tens of billions of dollars annually in AI without pressure to achieve immediate profitability. Despite its spectacular success, OpenAI still loses billions of dollars a year and depends on subsequent funding rounds. Anthropic is in a similar situation. Google can invest calmly, experiment, and outlast the competition—an advantage that cannot be overstated in a race requiring capital on an unprecedented scale.

These five advantages form a system of mutually reinforcing elements. In-house chips reduce infrastructure costs, enabling greater investment in research, which leads to better models, attracts more users, generates more data, and further improves the models. It is a classic flywheel—and Google already has it in full motion, while competitors are only beginning to get theirs turning.

How Google Is Undermining Nvidia’s Empire.

For the past decade, Nvidia has been the undisputed king of AI hardware. Its H100 and A100 GPUs power nearly every major language model in the world. Margins reaching 70 percent made it the most profitable semiconductor company in history.

Google is methodically undermining this position. TPU v5 and v6 not only match Nvidia GPUs in performance, but also offer better energy efficiency and integration with Google’s infrastructure. The fact that Gemini 3 was trained exclusively on TPUs is a signal to the entire industry: the alternative not only exists, but operates at the highest level.

What’s more, Google is entering Nvidia’s traditional territory by selling TPUs to external customers. The deal for one million chips with Anthropic is just the beginning. Rumors suggest talks with Meta and even OpenAI—strategic customers of Nvidia. If Google captures a significant share of this market, Nvidia’s margins will inevitably decline, and its dominance will erode.

Google Cloud: The Quiet Money-Making Machine

While the world watches the chatbot showdown, the real battle is taking place in the cloud. Training large language models requires enormous computing power; GPT-4 probably cost more than $100 million, and future models will be even more expensive.

Companies such as Microsoft Azure and Oracle Cloud rent computing power to AI companies. The problem is that they have to buy expensive chips from Nvidia, pay for energy, cooling, and staff, and only then add their own margin. The economics are brutal, cloud margins are notoriously thin, and price competition is merciless.

Google plays by different rules. By designing its own TPUs, it eliminates Nvidia’s margin. By controlling the entire stack, it optimizes each layer to work with the others, achieving efficiencies impossible for competitors that purchase components externally. The result? Google Cloud can offer AI compute more cheaply than anyone else while maintaining higher margins.

The numbers speak for themselves: Google Cloud revenue is growing at a rate of 30 percent annually and already accounts for nearly 20 percent of the company’s total revenue.

The Google ecosystem, the system that gives them an advantage.

ChatGPT may have 900 million weekly users, and Gemini 650 million monthly users. The numbers are impressive, but focusing solely on chatbots misses the essence of Google’s strategy.

The real advantage does not lie in a separate app, but in a habit repeated billions of times. According to 2024 data, Chrome holds more than 60 percent of the browser market, Gmail serves 1.8 billion users, Android runs on 70 percent of the world’s smartphones, YouTube attracts 2.5 billion monthly users, and Google Search processes 8.5 billion queries a day.

This is an integrated ecosystem in which each service strengthens the others. When Google weaves AI into Gmail (help with writing emails), Docs (automatic formatting), Search (direct answers), and Maps (context-aware route suggestions), the product becomes “sticky.” The cost of switching to a competitor is not limited to changing chatbots; it would require leaving an entire, carefully constructed digital world, which is practically impossible for most users.

OpenAI may have a better chatbot. However, it has no chance of replicating decades of investment in an ecosystem used by literally half of humanity.

A Critical Perspective: Did Google Really Win?

Despite the compelling case for Google’s dominance, it is worth maintaining a healthy skepticism. The history of technology is full of examples of giants that once seemed unassailable—IBM in mainframes (hardware), Microsoft in operating systems, and Nokia in phones—yet were ultimately dethroned by more agile rivals.

Potential Weaknesses of Google’s Strategy:

The innovation paradox: Large companies with well-established revenue streams often struggle with revolutionary innovation. Google earns over $100 billion annually from search advertising. Does it really want to replace that with a conversational model that could reduce ad click-through rates? This conflict of interest may hinder radical innovation.

Corporate culture: Google has a reputation for launching hundreds of projects only to abandon them later (Google+, Google Wave, Stadia, countless messaging apps). Is this culture of experimentation appropriate for an era that requires long-term commitment to a single AI vision?

Regulatory challenges: Despite the easing of the antitrust case in the US, Google continues to face regulators in Europe, lawsuits concerning Google Play and Android, as well as its dominance in online advertising. Vertical integration, which is their strength, may become the target of further regulatory interventions.

OpenAI and Microsoft: OpenAI, despite lacking its own infrastructure, has one crucial asset: developers’ attention. The GPT API supports millions of applications. Microsoft, as a partner, offers distribution through Office 365 (400+ million paying users) and Azure.

On the other hand, it is also possible that OpenAI will create its own office ecosystem.

Anthropic and Claude: Although smaller, Anthropic (the creator of Claude) attracts top talent and stands out for its emphasis on AI safety, which may be an advantage in an era of growing concerns about alignment. Claude enjoys an exceptionally positive reputation among advanced users and developers.

Is there any AI company that can truly compete at Google’s level?

The answer depends on how we define “the game.” When it comes to controlling the entire technology stack, Apple is probably the only company with similar capabilities, although it has not yet leveraged this potential in AI. Apple’s AI, grandly named “Apple Intelligence,” turned out to be a disaster. What’s more, Apple intends to integrate Google’s chatbots into its products. And perhaps this would require moving away from a fairly closed ecosystem, where some of the restrictions were deliberately imposed by Apple.

Meta has hardware and a huge social network, but it lacks a layer of utility products beyond Facebook and Instagram. Microsoft has Azure and Office, but it does not control hardware or chips (on such a scale).

However, if we are asking about the possibility of creating a better language model or building a more useful AI tool, history shows that breakthroughs often come from unexpected places. Google itself was once a small startup that defeated giants such as Yahoo and Microsoft.

Perhaps the real question is not “can anyone compete with Google,” but rather “is vertical integration truly the optimal strategy for the AI era?” Open models (Meta’s Llama, Mistral), specialization (Anthropic in safety, Perplexity in search), or entirely new interaction paradigms may prove more suitable than a universal ecosystem controlled by a single company.

Conclusions. Google has a structural advantage, useful in the long game

Regardless of future twists and turns, Google’s current position is impressive and difficult to challenge. Five-layer vertical integration creates a self-reinforcing system: proprietary chips lower cloud costs, making model training cheaper, which enables better products, which attract more users, which generate more data, which improve the models. This is a classic strategic flywheel.

Add to that decades of experience in machine learning (Google invented the Transformer architecture that powers all modern LLMs), world-class research teams (Google DeepMind), and virtually unlimited financial resources.

Has the AI race come to an end? Probably not. But Google undoubtedly has pole position, the best car, and the largest team of mechanics. In the long game for dominance in the AI era, Mountain View holds all the cards, while the competition is only just discovering that the game is being played on five different fronts simultaneously.

The question remains whether this is good for innovation, users, and society as a whole. The history of technology suggests that the concentration of power, even in the hands of an efficient company, rarely leads to the best long-term outcomes. But that is a topic for an entirely different article.

The text is largely based on this video, although I added some of my own thoughts.