Why Is Anthropic Looking to Build Its Own AI Chips?
AI companies need more than powerful AI models. They also need a huge amount of computing power to train and run those models.
Anthropic, the company behind Claude, is now reportedly exploring another part of the AI infrastructure: custom AI chips. The company discussed a potential acquisition of MatX, an AI chip startup, in a deal reportedly worth around $7 billion. The acquisition talks have since been put on hold, but the discussions highlight an important trend in the AI industry.
So why would an AI company want to build its own chips?
For Anthropic, having more control over its hardware could help address the growing cost and demand for computing power. It could also reduce some of the company's reliance on outside chip suppliers such as NVIDIA.
In this article, we'll look at what happened between Anthropic and MatX, why custom AI chips are becoming important, and why major AI companies are investing more in their own computing infrastructure.
Who Is Anthropic?
Anthropic is an AI safety and research company that develops Claude, a family of AI models and AI assistant services.
→ Source: Anthropic — Company overview
Claude can help with tasks such as writing, coding, analysis, and answering questions. Like other advanced AI models, Claude requires a large amount of computing power to train and run.
What Is MatX?
MatX is a U.S.-based AI chip startup founded by engineers with experience working on Google's Tensor Processing Unit (TPU), a type of chip designed for AI workloads.
→ Source: Reuters — Anthropic and MatX acquisition talks
The company focuses on designing specialized chips for AI workloads, rather than general-purpose processors.
This makes MatX interesting to companies like Anthropic, which need more computing power as their AI models become larger and more widely used.
In simple terms, MatX is a startup that designs specialized chips for running AI models.
What Happened Between Anthropic and MatX?
Anthropic reportedly explored acquiring AI chip startup MatX for about $7 billion. The goal was to accelerate Anthropic's development of its own custom AI hardware and reduce its reliance on outside chip suppliers.
→ Source: Reuters — Anthropic reportedly explored a $7 billion acquisition of MatX
However, the acquisition talks are no longer moving forward. According to Reuters, discussions between the two companies have shifted toward a potential partnership instead.
→ Source: Reuters — Anthropic and MatX partnership discussions
Reuters did not report why the acquisition talks were abandoned. Neither Anthropic nor MatX commented on the reported acquisition discussions.
So, Anthropic has not acquired MatX. The reported $7 billion deal was only a proposed acquisition that did not move forward.
Why Does Anthropic Need So Much Computing Power?
AI models like Claude require a huge amount of computing power to work.
First, Anthropic needs computing resources to train its AI models. Training involves processing large amounts of data and performing countless calculations to improve the model.
Computing power is also needed for inference, which is the process of generating an answer when a user interacts with Claude. Every question, request, or task requires the model to perform calculations before producing a response.
As more people use Claude, Anthropic also needs more computing power to serve many users at the same time and keep the service fast and reliable.
In simple terms:
More AI users → More computing → More chips
As Anthropic's AI services grow, having enough computing power becomes increasingly important. This helps explain why the company is looking beyond simply buying more existing hardware and is interested in developing its own AI chips.
Why Does Anthropic Rely on NVIDIA GPUs?
NVIDIA GPUs play an important role in modern AI computing. GPUs are well suited to AI because they can perform many calculations at the same time. This makes them useful for training and running large AI models.
→ Source: NVIDIA — Accelerated Computing and AI
As a result, companies such as Anthropic and other major AI companies use NVIDIA GPUs as part of their AI infrastructure. NVIDIA also has a mature software ecosystem that makes it easier for developers to build and run AI applications on its hardware.
→ Source: NVIDIA — CUDA Platform
But relying heavily on one type of hardware also creates challenges.
Why Can't Anthropic Rely Only on NVIDIA?
Buying NVIDIA GPUs is an important part of building large-scale AI infrastructure, but relying entirely on one type of hardware can create challenges as AI demand grows.
Supply: Advanced AI chips are in high demand. If chip supply becomes limited, AI companies may have difficulty expanding their computing capacity. Reuters reported that Anthropic's interest in custom chips is partly related to the constrained supply of NVIDIA hardware.
→ Source: Reuters — Anthropic and MatX
Cost: Running large AI models requires a huge amount of computing power, which can make hardware and infrastructure a major expense. Custom chips could potentially reduce costs for specific AI workloads over the long term. However, building custom chips is also expensive and takes time, so it is not automatically cheaper than using NVIDIA GPUs.
Efficiency: NVIDIA GPUs are designed to handle many types of workloads. A custom AI chip can instead be designed around specific tasks, which could make it more efficient for certain AI workloads.
Control: Developing custom hardware can give an AI company more control over how its AI models, software, and hardware work together. This can allow the company to optimize its infrastructure for its own needs.
For Anthropic, the goal is not necessarily to stop using NVIDIA GPUs. Instead, custom chips could give the company another source of computing power and more control over its AI infrastructure.
→ Source: Reuters — Anthropic and MatX
What Is a Custom AI Chip?
A custom AI chip is a processor designed for specific AI workloads rather than a wide range of computing tasks.
For example, a general-purpose GPU such as an NVIDIA GPU can be used for many different AI and computing tasks. A custom AI chip can be designed around the specific needs of a company's AI models and workloads.
You can think of them this way:
CPU → General-purpose computing
GPU → Highly parallel computing and widely used for AI
Custom AI chip → Designed for specific AI workloads
Because the hardware can be designed around particular workloads, a custom chip may improve efficiency, performance, or cost for certain tasks.
In simple terms, a custom AI chip is hardware designed to fit specific AI needs rather than trying to do everything.
→ Source: IBM — What are AI chips?
 |
| Figure 1. General-purpose GPU and specialized processor comparison |
How Does a Custom AI Chip Work?
A custom AI chip does not think or understand things like an AI model does. The AI model is the software, while the chip is the hardware that performs the calculations the model needs.
When you ask an AI model a question, the model performs many mathematical operations to generate a response. The chip processes these calculations and sends the results back to the software.
The basic process looks like this:
AI model → Mathematical operations → AI chip → Result
A custom chip can be designed around the specific calculations needed by certain AI workloads. You can think of it like building a special highway for the calculations an AI system uses most often.
By focusing on specific workloads instead of supporting every possible task, a custom chip may achieve better speed or energy efficiency for those workloads.
In simple terms, the AI model does the work of generating the answer, while the AI chip provides the computing power needed to make it happen.
→ Source: IBM — What are AI chips?
Why Is MatX Attractive to Anthropic?
MatX could be valuable to Anthropic because the startup already has experience in AI chip design.
MatX was founded by engineers with experience working on Google's TPU, giving the company relevant expertise in designing specialized AI hardware. This could help Anthropic avoid building an entire chip design team from scratch.
There is also a potential speed advantage. Working with an experienced chip startup could help Anthropic move faster in developing custom hardware than starting the process entirely on its own.
For Anthropic, the reported interest in MatX therefore appears to be about more than simply buying a company. It could provide access to hardware expertise, experienced engineers, and a faster path toward custom AI hardware.
In simple terms, MatX could offer Anthropic a combination of technology, talent, and speed.
→ Source: Reuters — Anthropic's reported MatX acquisition talks
How Does This Connect to Anthropic's $45 Billion Computing Deal?
Anthropic's interest in custom AI chips also makes more sense when viewed alongside its growing demand for computing power.
As Anthropic secures access to large amounts of computing infrastructure, the company also needs to think about the cost, availability, and efficiency of the hardware behind its AI models.
This helps explain why Anthropic is looking beyond simply buying existing hardware. Custom AI chips could give the company more options as its computing needs continue to grow.
The reported interest in MatX fits into this bigger picture. Anthropic reportedly explored acquiring MatX for about $7 billion, potentially giving it access to chip design expertise and experienced engineers.
→ Source: Reuters — Anthropic's reported MatX acquisition talks
In other words, Anthropic's growing investment in computing power and its interest in custom chips are part of the same larger strategy: building the infrastructure needed to support its AI models at scale.
Reuters reported that Anthropic agreed to spend about $45 billion to rent cloud computing power from Nscale over six years.
→ Source: Reuters — Anthropic to rent AI computing power from Nscale for $45 billion
Are Other AI Companies Building Their Own Chips?
Anthropic is not the only AI company looking beyond NVIDIA GPUs. Several major technology companies have been developing their own AI chips to support their growing computing needs.
Google has developed its Tensor Processing Units (TPUs), custom accelerators designed specifically for AI workloads. Google uses TPUs for systems including its Gemini models.
→ Source: Google Cloud — Tensor Processing Units (TPUs)
Amazon has developed Trainium for AI model training and Inferentia for AI inference. Both are purpose-built AI chips designed to work with AWS services.
→ Source: AWS — Trainium and Inferentia
Microsoft has developed its own AI accelerators, including Maia 200, which is designed for large-scale AI inference in Azure.
→ Source: Microsoft — Maia 200 AI Accelerator
Meta has developed its Meta Training and Inference Accelerator (MTIA) family of custom AI chips for its own AI workloads.
→ Source: Meta — Four MTIA Chips in Two Years
OpenAI has also moved into custom silicon. In 2026, the company unveiled its first custom AI chip, Jalapeño, developed with Broadcom and designed initially for inference workloads.
→ Source: Reuters — OpenAI unveils custom AI chip developed with Broadcom
 |
| Figure 2. AI infrastructure powering multiple industries |
These companies are taking different approaches, but the broader trend is clear: AI companies are investing more heavily in hardware designed around their own workloads.
This does not mean they are all abandoning NVIDIA. Instead, many are building custom chips alongside GPUs to gain more options as their AI infrastructure grows.
What Does This Mean for Anthropic and NVIDIA?
Anthropic's interest in custom AI chips does not mean that it is replacing NVIDIA.
NVIDIA GPUs remain an important part of the AI computing ecosystem, and Anthropic can continue to use them while exploring other hardware options.
At the same time, developing custom chips could give Anthropic more flexibility and control over its computing infrastructure. Instead of relying on a single type of hardware, the company could use a combination of NVIDIA GPUs and custom AI chips, depending on the workload.
This is part of a broader shift in the AI industry. Companies are not necessarily trying to replace NVIDIA overnight. They are looking for more choices as their AI computing needs continue to grow.
For Anthropic, the relationship with NVIDIA is therefore better described as continued reliance alongside a growing interest in alternative and custom hardware.
→ Source: NVIDIA — Accelerated Computing and AI
→ Source: Reuters — Anthropic's reported MatX acquisition talks
What Are the Challenges of Building Custom AI Chips?
Building a custom AI chip can give an AI company more control over its hardware, but it also comes with significant challenges.
High development cost: Designing a new chip requires a large investment in engineering, research, and development. This can make custom hardware a major financial commitment.
Long development cycle: Chip development takes much longer than developing or updating software. Designing, testing, and improving new hardware can take years.
Manufacturing: Designing a chip is only part of the process. The design must eventually be manufactured, packaged, and tested before the chip can be used at scale.
→ Source: IBM — What are AI chips?
Software ecosystem: A powerful chip is not enough on its own. Developers also need software, libraries, and development tools that work well with the hardware. This is one reason established GPU platforms remain valuable.
→ Source: NVIDIA — CUDA Platform
In other words, successfully building custom AI hardware requires more than a good chip design. Hardware, software, and developer tools all need to work together.
Does Anthropic Really Want to Manufacture Chips?
When Anthropic talks about developing its own AI chips, it does not necessarily mean that the company wants to build and operate its own semiconductor factories.
There is an important difference between chip design and chip manufacturing.
Chip design means designing the structure and functions of a chip for specific workloads.
Chip manufacturing means physically producing those chip designs in a semiconductor factory.
For Anthropic, the focus appears to be on custom chip design and in-house silicon development, rather than becoming a chip manufacturer like Intel or Samsung.
→ Source: Reuters — Anthropic's reported MatX acquisition talks
In simple terms, Anthropic wants more control over the design of the hardware used for its AI systems, but that does not mean it needs to build its own factories to manufacture the chips.
Simple Summary
Anthropic is exploring ways to gain more control over the computing infrastructure behind Claude. Its reported interest in MatX shows why custom AI chips are becoming increasingly important as AI models and user demand grow.
Custom chips could offer benefits such as better efficiency and more control, but developing them also requires significant time, money, and technical expertise. Anthropic is therefore not necessarily replacing NVIDIA GPUs, but exploring additional hardware options for the future.
Key Takeaways
- Anthropic reportedly explored acquiring MatX for about $7 billion, but the acquisition talks did not move forward.
- MatX specializes in custom AI chip design and has engineers with experience working on Google's TPU.
- AI models like Claude require large amounts of computing power for training, inference, and serving users.
- Custom AI chips could provide more efficiency, flexibility, and control for specific AI workloads.
- Developing custom chips is difficult and expensive, requiring hardware, software, and developer tools to work together.
- Anthropic's interest in custom hardware does not mean it is abandoning NVIDIA GPUs.
FAQ
Q. Does Anthropic own MatX?
A. No. Anthropic reportedly explored acquiring MatX for about $7 billion, but the acquisition talks did not move forward. The two companies have reportedly discussed a potential partnership instead.
Q. Why does Anthropic want its own AI chips?
A. Anthropic needs a large amount of computing power to train and run Claude. Custom chips could give the company more control and potentially improve efficiency for certain AI workloads.
Q. Will Anthropic stop using NVIDIA GPUs?
A. There is no indication that Anthropic is planning to completely stop using NVIDIA GPUs. Custom chips could instead be used alongside NVIDIA hardware.
Q. What is a custom AI chip?
A custom AI chip is hardware designed for specific AI workloads. Unlike a general-purpose processor, it can be optimized around the calculations needed for particular AI tasks.
Q. Does Anthropic plan to build its own chip factories?
A. Not based on the information discussed here. Chip design and chip manufacturing are different processes. Anthropic's reported focus is on developing custom hardware rather than becoming a semiconductor manufacturer.
Q. Why are other AI companies building their own chips?
A. Companies such as Google, Amazon, Microsoft, Meta, and OpenAI are also developing custom AI hardware. This gives them more options as their AI computing needs continue to grow.
Final Thought
The AI race is no longer just about building better models.
As companies like Anthropic build larger AI systems and serve more users, the hardware running those systems is becoming just as important.
Anthropic's reported interest in MatX is one example of this larger shift. The company may continue to rely on NVIDIA while developing other hardware options, showing that the future of AI computing could involve a mix of GPUs and specialized AI chips.