The race for artificial intelligence is not won only with better algorithms: it's won with more compute. The deal between Anthropic and xAI —through which Anthropic accesses 300 MW of power and more than 220,000 NVIDIA GPUs from the Colossus 1 datacenter in Tennessee— is the clearest sign so far that infrastructure has become the industry's new competitive moat.

According to TechCrunch on May 10, the deal gives Anthropic massive access to computing capacity without having to build its own datacenter. Here we analyze what changes, why it matters, and what it means for developers in LATAM who use the Claude API.

The details of the deal

The deal between Anthropic and xAI is for access to computing capacity, not an acquisition. Anthropic will use the infrastructure of xAI's Colossus 1 datacenter, located in Memphis, Tennessee, to train and serve its Claude models. The numbers are striking:

• 220,000+ NVIDIA GPUs (predominantly H100 and H200)

• 300 MW of energy capacity, equivalent to supplying a city of ~300,000 inhabitants

• Colossus 1 was built by xAI in record time (less than 6 months in 2024) to train the Grok models

The deal allows Anthropic to scale its training capacity without the time and capital required to build its own infrastructure, while xAI monetizes its datacenter in the moments when it doesn't use it for its own models.

Why Colossus 1 is important

Colossus 1 is not just big: it was designed specifically for the type of workloads that LLMs require. Its high-speed network architecture (InfiniBand) and its cooling capacity are optimized for training massive models with thousands of GPUs working in parallel. When xAI built it to train Grok, it set a benchmark for infrastructure deployment speed that the rest of the industry noticed.

Compute: the new competitive advantage in AI

For years, the advantage in AI was in the data and the algorithms. Today, with pretraining datasets relatively democratized and fine-tuning techniques accessible to everyone, compute has become the differential factor. More GPUs = larger models = better capabilities = more users = more revenue to reinvest in compute. It's a cycle that favors those who can scale fast.

OpenAI has the infrastructure deal with Microsoft Azure. Google has its own cloud and its TPUs. Meta has one of the largest GPU clusters in the world. Anthropic, until now, depended mainly on AWS (Amazon is one of its main investors) and Google Cloud. This deal with xAI diversifies that dependence and gives it additional levers to scale the training of Claude.

To better understand how the economics of compute in AI work —tokens, API costs, and model architectures— you can read this article about what a token is in AI and how to optimize API costs.

How the Anthropic vs OpenAI vs Google dynamic changes

The deal with xAI has strategic implications beyond compute itself:

• Vs OpenAI: OpenAI has an advantage in scale and product (ChatGPT has hundreds of millions of users), but Anthropic built a solid reputation in the enterprise market and in the developer community with Claude. More compute means potentially better models and better API latency.

• Vs Google: Google is both a competitor (Gemini) and an infrastructure provider for Anthropic (Google Cloud). The diversification toward xAI reduces that dependence and gives Anthropic more negotiating power.

• The most important signal: xAI and Anthropic are competitors in the AI model market. That xAI decides to monetize its infrastructure with Anthropic indicates that the economic logic (revenue from idle compute) outweighs the model rivalry, at least for now.

What it means for developers in LATAM who use Claude

The most direct impact for developers who integrate the Claude API into their projects can be seen in three dimensions:

1. Greater inference capacity: with more compute available, Anthropic can serve more requests in parallel, which translates into lower latency and better API availability during demand peaks.

2. More powerful models in the future: access to more training capacity accelerates the development cycle of the next Claude models. More parameters, better reasoning, a larger context window.

3. Service continuity: diversifying the infrastructure reduces the risk of downtime in the face of problems at a specific provider.

If you're building applications with the Claude API or want to start, you can see a concrete example in this article about how to create your own AI agent with the Anthropic API.

The hardware war: energy and GPUs as strategic assets

The Anthropic-xAI deal also puts the spotlight on something that goes beyond the companies involved: energy infrastructure has become a strategic asset in the AI industry. AI datacenters are mega-consumers of energy, and the availability of cheap and reliable energy is limiting the speed of expansion of all players.

This has implications for LATAM: countries with cheap renewable energy (Paraguay with hydroelectric, Chile with solar, Colombia with geothermal) are starting to appear in conversations about where to build the next generation of AI datacenters. The region could have a more active role in this race than was anticipated just two years ago.

Coderhouse courses to take advantage of the Claude ecosystem

If you want to work with the Anthropic API, build AI agents, or understand how to scale applications on models like Claude, these Coderhouse courses are the starting point:

• AI Engineering Course: learn to integrate LLM APIs (Claude, GPT, Gemini) into real applications, with production architectures.

• AI Agents Course: build autonomous agents with tools like LangChain, LlamaIndex, and the Anthropic API.

• Introduction to Artificial Intelligence Course: to understand the fundamentals of how the models behind this deal work and the AI ecosystem in general.

Frequently asked questions

Did Anthropic buy xAI or just access its infrastructure?

It just accesses the infrastructure. There is no acquisition or merger. It's a commercial deal for computing capacity: Anthropic pays to use the GPUs and the energy of xAI's Colossus 1 datacenter during the periods when xAI doesn't need them for its own Grok models.

Why does xAI rent compute to a direct competitor?

The answer is economic. A datacenter of 220,000 GPUs has enormous operating costs (energy, cooling, maintenance) regardless of whether it's at 100% utilization or 40%. Monetizing the idle capacity with a client like Anthropic is smarter than leaving the racks empty. In the cloud industry, this is called "capacity arbitrage" and is a common practice.

Does this affect the privacy of the data I process with the Claude API?

It shouldn't. The deal is about infrastructure: the data you send to the Anthropic API is still processed under Anthropic's privacy terms, not xAI's. The underlying infrastructure doesn't change the API provider's data policies. If you have specific doubts about data handling in the Claude API, the correct source is Anthropic's official terms of service.

When will the improvements in the Claude API be seen thanks to this deal?

The impact on inference capacity (latency, availability) can be relatively fast. The impact on more powerful models takes longer: the complete training cycle of a frontier model with that infrastructure can take months. It's expected that the next Claude models will benefit from this additional training capacity.

Does it make sense to learn to use the Claude API with all these infrastructure changes?

Totally. The Claude API is one of the most used in the world for enterprise applications and agent development, and this deal reinforces Anthropic's ability to maintain and scale the service. Learning to integrate Claude into applications today is a skill with high demand and low risk of obsolescence in the short term.