Google was losing the AI game. That was the narrative in 2023 and 2024. Every tech journalist, every analyst, every investor seemed certain that OpenAI was the future. Google released Bard, and it was embarrassing.
Meanwhile, ChatGPT exploded across the internet. OpenAI was the disruptor. Google was the dinosaur. The story seemed written.
But stories written by people who do not understand patience rarely end the way they predict. In November 2025, Google released Gemini 3. Within days, the conversation had inverted completely.
Gemini 3 is now considered the strongest large language model on the planet. Not by a small margin. By a meaningful one. Google’s stock surged 13% in a week. Nvidia dropped $115 billion in value.
The narrative flipped so completely that it seemed like it happened overnight. But Google did not build this position overnight. The company built it systematically, over a decade, by playing a game that nobody else had the patience to play.
The Hidden Decade That Changed Everything
The story of Google’s AI dominance starts not in 2023, but in 2015. That is when Google released the first Tensor Processing Unit. Most people did not pay attention.
TPUs were internal tools. Google used them to power search, to train their own models, to run their own systems. They were not sold to customers. They did not have splashy marketing campaigns.
They just worked, quietly, in the background, year after year after year. This was patient capital at work. Google had the resources to bet on custom silicon when everyone else was buying GPUs off the shelf.
The company could afford to wait. The company could afford to iterate. The company could afford to optimise. While Nvidia was selling chips to every company with a budget, Google was building an integrated stack.
Model, hardware, software, cloud infrastructure. Everything connected. Everything is optimised for each other. For nearly a decade, this looked like a waste.
Why would a search company spend billions building chips that nobody could buy? Why not just use Nvidia like everyone else? The answer is that Google was not trying to compete in the chip market.
Google was trying to build an entire ecosystem where Nvidia did not matter.
Vertical Integration As Long-Term Strategy
What Google did was almost the opposite of what the tech industry celebrates. While everyone was talking about specialisation, about focusing on one thing and doing it better than anyone else, Google was integrating vertically.
Google built models. Google built chips. Google built the software to run on those chips. Google built the cloud infrastructure to host them. Google built the frameworks and tools for developers.
Each piece is connected to the others. Each piece made the others better. The fundamental advantage of vertical integration is cost efficiency. When you own the entire stack, you do not have to negotiate with suppliers.
You do not pay markups. You do not have supply chain friction. You build for your specific workloads instead of building for every possible workload. Nvidia GPUs are general-purpose.
They need to work with any model, any framework, any software. Google TPUs were built for Google’s specific needs. They are optimised for tensor operations. They strip away everything that does not matter.
The economic difference is staggering. Nvidia’s gross margins on data centre chips are in the 80% range. That means hyperscalers buying Nvidia chips pay roughly 5 times what the hardware costs to manufacture.
Google, making its own chips, pays closer to manufacturing cost. No middleman. No markup. Just hardware at cost. This arithmetic compounds at scale.
Meta is spending $72 billion on AI infrastructure this year. If half of that went to compute hardware, and if switching to Google TPUs reduces costs by 75%, that is $13.5 billion in annual savings.
The Model Becomes The Proof
For all the elegance of the long-term strategy, none of it mattered until Google proved that TPUs could actually train world-class AI models. This is where Gemini 3 enters the story.
Google did not hire the best minds in the world and tell them to build the best model on TPUs and hope it worked out. Google actually did it. Gemini 3 was trained entirely on Google’s custom silicon.
The model is multi-modal, running text, image, audio, and video all at once. The reasoning is superior. The context window is massive. The model understands nuance.
The model works. This is the moment where the narrative shifts from “Google built an interesting internal tool” to “Google built a genuine alternative to Nvidia.”
Gemini 3 proves that you do not need Nvidia to build state-of-the-art AI. The proof point exists. Other companies can see it. Other companies can evaluate it.
Other companies can decide that they want it. The market knew this moment was coming. Analysts had been predicting that custom silicon would challenge Nvidia for years.
But predictions and proof are different things. Gemini 3 is proof. And proof changes behaviour.
Building The Ecosystem Prison Without Looking Like One
Here is where the long-term thinking becomes really sophisticated. Google is not trying to sell TPUs to everyone. Google is building a complete ecosystem where using Google TPUs makes sense for every layer of the AI stack.
Start with the model. Google has Gemini, and it is good. Companies want access to it. Google rents access through Google Cloud. But to use Gemini optimally, you want your workloads running on Google infrastructure.
The model works best on the cloud where it lives. But you also need a framework. Google has JAX and XLA. These are the software tools that make it easy to develop on Google hardware.
Use JAX, and everything just works better. Nvidia CUDA still works, but JAX is simpler. JAX is optimised for TPUs. JAX is the path of least resistance.
Then add the development tools. Google recently launched Antigravity, a platform for agentic development. This is where developers build applications that use AI.
Developers want to build on the platform that makes development easiest. That is increasingly Google. Add analytics and observability. Add machine learning operations tools. Add orchestration and deployment infrastructure.
Each piece of the stack matters, and each piece works best when every other piece is also Google’s. You are not locked in by contract. You are locked in by convenience.
You are choosing Google because its stack is better, not because you have no alternative.
The Pricing Power That Breaks Economics
The final layer is pricing. Because Google owns the entire stack, and because the cost of compute is so much lower than Nvidia’s, Google can price far more aggressively than Nvidia and still be more profitable.
This is the leverage that breaks Nvidia’s model. Nvidia charges what the market will bear because it is the only option. Google charges less and still wins because it does not have Nvidia’s cost structure.
Suddenly, every company evaluating its AI budget thinks differently. The math changes when an option exists that costs a quarter as much for a similar capability. Google’s Gemini API is priced at fractions of what OpenAI charges.
Gemini 2.5 Pro is $10 per 1 million output tokens. OpenAI’s o3 is $40. The difference is not that Google is dumping. The difference is that Google’s costs are a fraction of OpenAI’s.
Google can sustain these prices while maintaining high margins. Nvidia, with its markup-heavy model, cannot.
Playing The Long Game When Everyone Else Burns Cash
The fundamental difference between Google’s approach and Nvidia’s is time. Nvidia plays for quarterly returns. The company extracts maximum value from its market position right now.
Google plays for the next decade. When Meta spends $72 billion on AI infrastructure, Nvidia captures most of it today. That feels like a win.
This happens as the AI market grows. Google is not trying to take 100% of the AI chip market. Google does not need to. Google is building an infrastructure business where AI is one component.
Patience Rewarded
Google took a decade to build this position. The company invested billions in infrastructure that did not immediately generate revenue. The company took criticism for Bard and held its ground.
And then, in November 2025, the market finally caught up to what Google had already built. Gemini 3 launched. Meta noticed. The narrative flipped.
Nvidia lost $115 billion in market value. Google’s stock surged. All because Google played a different game than everyone else. Google played the patient game.
The game where you build for the long term and let compounding work in your favour. The game where integration creates value that quarterly returns cannot explain. The game that Nvidia thought it was winning.
Right up until the moment it realized that the rules had changed. This is how empires transition. Not with dramatic crashes but with patient building by those willing to wait.
Google waited. Now Google is collecting.