Anthropic’s reported discussions with artificial intelligence chip startup MatX highlight a strategic problem that is becoming increasingly important for leading artificial intelligence companies: access to computing power is no longer simply a purchasing issue. It is becoming a question of how much control an AI developer wants over the hardware on which its models are trained and operated.

Anthropic reportedly explored acquiring MatX for about $7 billion before discussions shifted toward a possible partnership, according to people familiar with the matter. MatX, founded by former Google engineers who worked on the company’s specialized artificial intelligence processors, has developed technology aimed at large-scale AI workloads. The reported talks come shortly after Anthropic confirmed that it was building an internal team to design custom chips for its Claude models.

The significance of the MatX discussions therefore lies less in whether an acquisition eventually occurs and more in what they reveal about Anthropic’s changing hardware strategy. The company is simultaneously committing enormous sums to external computing providers while attempting to develop technology that could eventually give it greater control over performance, supply and costs.

Chip shortages have made hardware a strategic issue

Anthropic’s dependence on computing infrastructure has expanded rapidly alongside demand for Claude. Training increasingly capable models requires enormous quantities of specialized processors, while serving those models to millions of users creates a separate and continuing demand for computing capacity.

For years, companies such as Anthropic could largely treat chips as infrastructure purchased from specialist suppliers. Nvidia became the dominant supplier of the processors used to train and operate many advanced AI systems, while Google, Amazon and other technology companies developed competing specialized processors. Anthropic itself has deliberately used several hardware platforms rather than relying on a single supplier.

That strategy remains in place. Anthropic has publicly described its hardware approach as diversified, using Nvidia graphics processors alongside Google’s tensor processing units and Amazon’s Trainium processors. The company has also committed to very large amounts of additional computing capacity through Amazon and Google.

The problem is that diversification does not eliminate dependence on external suppliers. It simply spreads that dependence across several companies. If demand for advanced processors continues to rise faster than manufacturing capacity, the ability to secure sufficient chips can become a constraint on an AI company’s growth.

Nvidia has itself warned that demand for its data centre products remains exceptionally strong. For Anthropic, developing its own processors could therefore provide another way to manage a bottleneck that cannot be solved simply by signing larger purchasing agreements.

MatX offers expertise that could shorten the development process

The reported interest in MatX is particularly significant because advanced chip design is difficult, expensive and time-consuming. A company starting from scratch has to recruit semiconductor architects, develop processor designs, build software tools, test prototypes and work with manufacturers before a production chip can reach a data centre.

MatX was founded by former Google engineers with direct experience developing tensor processing units, giving it expertise in exactly the area Anthropic is trying to build internally. The company raised $500 million in February 2026, bringing its publicly reported funding to more than $600 million. Its processors are being developed specifically for demanding artificial intelligence workloads, with initial production plans involving Taiwan Semiconductor Manufacturing Company.

An acquisition could therefore give Anthropic access to experienced engineers and an existing technology programme rather than requiring it to develop every capability internally. The reported $7 billion price under discussion would nevertheless represent a substantial premium over MatX’s latest disclosed funding round and would illustrate how valuable specialist semiconductor talent has become within the AI industry.

The fact that discussions reportedly moved toward a partnership is equally revealing. Anthropic does not necessarily need to own a chip company to benefit from its expertise. A partnership could allow it to gain access to design capabilities while maintaining a broader hardware strategy involving Nvidia, Google and Amazon.

That could be particularly useful while Anthropic’s own silicon programme remains at an early stage.

The goal is not to replace Nvidia immediately

It would be misleading to interpret Anthropic’s custom-chip effort as evidence that the company is preparing to abandon Nvidia. Its current strategy points in the opposite direction. Anthropic continues to use Nvidia processors and has committed to enormous amounts of externally supplied computing capacity.

The more immediate objective is flexibility. A processor designed specifically around Claude’s architecture and workloads could potentially improve performance or energy efficiency for particular tasks. It could also allow Anthropic to optimize the relationship between its models, software and hardware instead of relying entirely on processors designed for a broad range of customers.

This is the same strategic logic behind custom processors developed by Google and Amazon. Google’s tensor processing units are designed around the company’s artificial intelligence workloads, while Amazon’s Trainium and Inferentia families target model training and inference. Microsoft and Meta have also pursued specialized silicon.

The competitive significance is therefore broader than Anthropic versus Nvidia. The AI industry is moving toward a model in which the largest model developers increasingly want to control more layers of the computing stack.

Nvidia remains exceptionally strong because its advantage extends beyond individual processors. Its software ecosystem, networking technology, developer tools and integrated data centre systems make it difficult for customers to switch entirely to alternatives. Custom chips therefore do not need to eliminate Nvidia to be strategically useful. They need to perform well enough on selected workloads to reduce dependence and improve economics.

Training and inference require different calculations

Anthropic’s interest in MatX also highlights the distinction between training AI models and running them for users. Training involves processing enormous datasets to develop model capabilities, while inference occurs when a trained model generates an answer or performs a task for a user.

The two workloads have different hardware requirements. A chip optimized for training does not necessarily provide the best economics for inference, and vice versa. This creates opportunities for specialized processors designed around particular workloads.

MatX has focused on chips aimed at training large language models, according to earlier company statements. Anthropic’s own requirements may eventually encompass both training and inference, but the company has not publicly committed to a single architecture or timetable.

That uncertainty explains why Anthropic has been speaking with several chip startups rather than immediately selecting one acquisition target. The company needs to determine which workloads should receive custom hardware and whether developing specialized processors will provide enough performance or cost benefits to justify the enormous investment.

The decision is not merely technical. A custom processor becomes useful only when it can be manufactured in sufficient quantities, supported by software and integrated into data centre systems.

Anthropic is spending heavily before its own chips arrive

The scale of Anthropic’s external infrastructure commitments shows why custom silicon will not provide an immediate solution. The company has signed major agreements with Amazon and Google to secure future computing capacity, while also pursuing additional infrastructure arrangements.

In April, Anthropic announced an agreement with Amazon that could provide up to five gigawatts of computing capacity over time, including Amazon’s custom Trainium processors. Anthropic also announced a major expansion of its relationship with Google and Broadcom involving multiple gigawatts of next-generation Google processors expected to become available from 2027.

These agreements demonstrate that Anthropic is pursuing two strategies simultaneously. It is securing enormous amounts of external computing capacity to meet current and near-term demand while investing in technology that could give it more control over future hardware.

That approach reduces the risk of betting everything on a chip programme that may take years to mature. It also means Anthropic can test its own designs against hardware from companies with much larger semiconductor operations.

The financial stakes are considerable. Advanced chip development can cost hundreds of millions of dollars for a single generation, while successful deployment requires additional investment in manufacturing, packaging, networking, software and data centre infrastructure.

For Anthropic, the attraction is that those costs could eventually be offset if custom processors improve the economics of operating Claude at scale.

The MatX talks reveal a broader shift in AI strategy

The reported discussions ultimately show how the economics of artificial intelligence are changing. Model developers are no longer competing only on algorithms and data. They are increasingly competing over access to electricity, data centres, networking equipment, processors and the engineers capable of designing them.

Anthropic’s decision to build an internal silicon team, recruit experienced chip engineers and explore relationships with specialized startups indicates that hardware is becoming part of the company’s core technological strategy. The MatX discussions fit that broader effort whether they ultimately result in an acquisition, a partnership or no transaction.

The company is unlikely to become independent of external chip suppliers in the near term. Its enormous infrastructure commitments make continued cooperation with Nvidia, Google and Amazon essential. But custom hardware could gradually give Anthropic another source of computing capacity and greater influence over how Claude is optimized.

That is why the reported MatX talks matter even without a completed acquisition. They suggest that Anthropic sees control over chip design as increasingly important to the economics and scalability of its AI business. The immediate objective is not to replace established suppliers, but to build enough internal capability to make hardware a competitive advantage rather than a constraint.

If that strategy succeeds, the effect could extend beyond Anthropic. As major AI companies develop processors tailored to their own models, the industry’s dependence on general-purpose AI hardware could gradually become more fragmented, creating a more competitive but also more expensive race across the entire computing stack.

(Adapted from EuroNext.com)