How X AI Replaced Google as The Third Best AI Lab

#tech

Google's failures

As someone who uses Antigravity CLI for coding, I have to say it's not the best model and has a huge amount of issues. With Gemini 3.5 Flash dropping 3 months ago and with no release date on 3.5 Pro, it seems like Google is falling behind. The current models use way too many tool calls for simple changes and don't stop after they have solved the problem. These two issues mean that Gemini models burn more tokens or credit for a task compared to other models.

If we look at benchmarks overuse of tokens seems to be an issue with Google models. In DeepSWE benchmark V1.1 Gemini models are at the bottom right which makes them the least efficient models that don't even score well.¹ The insane thing is the top model GPT 5.6 Sol on Max is cheaper and scores the best of any model.

What you need to make LLMs

To train a good large language model you need three main parts: data, compute and researchers. While that is not everything you need it's the main parts you need. Data is not all created equally. There are two types of data, first is general and world knowledge which is used while doing the main training of the model.

Once you have trained the base model you can start on reinforcement learning at this stage there are two important types of data for a good LLM at this stage. Chat histories don't have to be directly code but even chat's about a semi aquatic mammal are needed. Chat histories about non code may not be need if we are looking at just code performance right? You need to remember that all LLMs are is text prediction machines. This means that it needs to be fine tuned to work better in a chat setting.

Now direct code prompts and answers can be very good because LLMs don't know how to deal with tool calls. This data is where a model learns to make tool calls and with current frontier models call sub agents to investigate and solve small problems. This type of data is also important for teaching how it should write code like comments and style.

Now we have all this data how do we turn it into a large language model? We need computers to put all this data into the models weights but not all compute is equal. CUDA has become the default for training LLMs with most if not all models trained in CUDA. While Nvidia basically has a monopoly with CUDA on training. But for inference there is a bit more variety which I may talk about in the future.

Researchers have become a huge asset to companies with Researchers being given 1M+ packages to join companies. With Researchers becoming more expensive companies have been forced to get them Nvidia GPUs as they are easier to work with.

What Google has

Before the huge AI hype train Google had one of the best AI Researcher talent pool. This huge pool of talent was mostly gained by their acquisition of DeepMind in 2014.² After the acquisition it was merged with Google Brain to form one subsidiary. From an outsiders point it looked like Google left them alone, which for the researchers is probably what they wanted. Now inside of DeepMind, Google probably has its hands all in there trying to get the next model released with it being a high stress place to work. This can be seen with departures of prominent researchers. With this talent leaving it is most likely becoming a more stressful place to work which then makes more researchers leave and then it perpetuates.³

Now let's talk about data while Google is known as the advertising company that wants your data this data is not what you need to train an LLM. For general data I think Google has a good amount of this data due to scraping the internet. For the second type of data Google has strong areas but also weak areas. Chat history is where Google should have a good amount of data with the Gemini app having 750 million active users with training on history on by default.⁴ Now let's talk about the type of data Google has the least of Coding. Google does have Antigravity but due to the worse code quality of the current models there is not enough people using it in order to get the data they need. Now I think Google is trying to solve this because now Gemini can create simple code that can be used to make interactive things in the Gemini App which will help them gather the data they need.

Compute is where Google has an advantage which has disadvantages. No other top AI lab has their own silicon but Google does with their custom tensor processing units. While this provides a way to get compute that does not require Nvidia. But they are still constrained by chip production at TSMC. Now how is this a disadvantage well most researchers prefer working with CUDA which means most training of AI models is done on Nvidia GPUs. This means that if they want to do inference on TPUs, there needs to be changes which can reduce the models performance.

What X AI has

The recent acquisition of Cursor by X AI has really changed the game as Cursor has good strengths where X AI had weaknesses. Let's first talk about researchers while X AI does not have that many prominent researchers they still have a good talent pool. Cursor seems to have had one of the best teams for reinforcement learning as seen with Composer 2.5 where they did reinforcement learning on Kimi K2.5 to massively increase the coding performance.⁵

Now let's talk about compute this is where X AI is doing very well as they have built two massive data centers called Colossus. In a normal market where compute was less constrained the amount X AI built would have been a bad idea, but due to the difficulties of getting GPUs from Nvidia it's been very successful.⁶ If the supply for power and Nvidia GPU shortage and the lack of usage of X AI models building so much capacity would have been a bad idea. But you can see with the deals with Google⁷ and Anthropic.⁸ Both of these deal make up around 84% of X AI's Revenue according to the S-1 filings.⁹

Data is where the recent acquisition of Cursor makes a whole lot of sense as X AI has good General data from Twitter/X, But did not have that much code data. Cursor has this type of data which X AI is lacking because much like Google Cursor Collects users requests and is allowed to use them for training new models. This type of data is the type of data X AI is lacking.

The recent X AI's recent acquisition of Cursor is already paying off. As seen with the release of Grok 4.5 which put it in the third position on artificial analysis intelligence benchmark, until today with the release of Kimi K3 which has displaced it for third place.¹⁰

The future

The future of Google and DeepMind is really riding on Gemini Pro 3.5, as if it's only slightly better than Flash they are going to really struggle to gain market share. If Gemini 3.5 pro has been in the oven longer for a good reason as is a Sol/Fable class model it's going to put Google near the top again.

For X AI the future looks bright with the full acquisition of Cursor and the potential of a Grok 5 which has both teams working alongside each other, they can keep their position in the top three AI players. If they do what Elon said and release a new model every month this will allow them to refine and make incremental improvements on models which should keep them near the top of the charts.