Anthropic has explored acquiring AI chip startup MatX for roughly $7 billion as the Claude maker accelerates efforts to develop its own custom silicon and reduce its reliance on Nvidia processors. The acquisition talks have since evolved into discussions about a potential partnership, according to people familiar with the matter cited by Reuters. MatX, founded by former Google engineers who worked on the company’s Tensor Processing Units (TPUs), is now seeking fresh capital at a valuation of about $4 billion.
The discussions highlight how the economics of generative AI are pushing leading model developers deeper into the semiconductor business. Anthropic has already begun building an in-house silicon team while maintaining a multi-chip strategy involving Nvidia, Google, Amazon and other suppliers. The company is simultaneously committing tens of billions of dollars to external computing capacity, including a planned $36 billion purchase of Google’s AI chips, a $45 billion Nscale cloud-computing agreement and a commitment to pay SpaceX $1.25 billion per month through May 2029 for computing capacity.
Anthropic Explored $7 Billion MatX Acquisition
The potential MatX acquisition would have given Anthropic direct access to a team specializing in AI chip architecture and potentially accelerated its efforts to build custom processors.
Reuters reported that the discussions are no longer active as an acquisition, although the two companies are discussing a possible partnership. The reason the acquisition talks ended was not immediately clear.
MatX is currently seeking new funding at a valuation of approximately $4 billion, meaning the reported $7 billion acquisition discussion would have represented a substantial premium to the startup’s current fundraising target.
Anthropic-MatX Deal At A Glance
| Particular | Details |
|---|---|
| Potential target | MatX |
| Potential buyer | Anthropic |
| Reported acquisition value | ~$7 billion |
| Current MatX fundraising valuation | ~$4 billion |
| Current status | Acquisition talks no longer active |
| Possible next step | Partnership |
| MatX founders | Former Google TPU engineers |
| Primary technology | AI training chips |
| Anthropic objective | Accelerate custom-chip development |
| Anthropic position | Multi-chip strategy |
| Nvidia reliance | Being reduced, not eliminated |
Neither Anthropic nor MatX provided Reuters with a comment on the acquisition discussions.
Why Anthropic Wants Its Own AI Chips
The driving force behind Anthropic’s hardware strategy is the extraordinary amount of computing power required to train and operate Claude.
Frontier AI models require enormous quantities of specialized processors, and Nvidia currently dominates the market for high-performance AI accelerators.
Anthropic wants greater control over the hardware that powers its models.
Claude Demand Rises
↓
Compute Requirements Increase
↓
More AI Chips Needed
↓
Higher Dependence On External Suppliers
↓
Custom Silicon Development
↓
Greater Control Over Cost And Performance
The company has said its in-house silicon effort will allow it to co-design hardware and models so Claude can run faster and more efficiently at scale.
MatX Could Accelerate Anthropic’s Chip Design
Designing an advanced AI processor from scratch is a lengthy and expensive undertaking.
Reuters reported that a viable piece of hardware can take a year or more to produce, while the design costs for a single generation can reach hundreds of millions of dollars.
An acquisition of MatX could therefore have saved Anthropic time by bringing in engineers with direct experience designing specialized AI processors.
Why MatX Matters
| MatX Capability | Potential Anthropic Benefit |
|---|---|
| Former Google TPU engineering talent | Faster access to chip-design expertise |
| AI-specific processor development | Hardware optimized for Anthropic workloads |
| Training-chip focus | Potential support for model training |
| Specialized architecture | Greater performance customization |
| Existing startup team | Faster than building every capability internally |
| Semiconductor expertise | Reduces reliance on external design resources |
The potential acquisition was therefore less about buying a conventional software startup and more about acquiring scarce semiconductor engineering talent.
MatX Was Founded By Former Google TPU Engineers
MatX was founded by Reiner Pope and Mike Gunter, engineers with backgrounds at Google.
Pope spent about a decade working on machine learning at Google, while Gunter worked on Google’s AI chips, known as Tensor Processing Units.
The startup has been developing a server chip designed specifically for artificial intelligence workloads.
Earlier investor materials indicated that MatX was targeting large language model training and believed its architecture could deliver substantially better performance per dollar than future Nvidia processors, although those claims remain company projections rather than independently verified results.
MatX Background
| Metric | Details |
|---|---|
| Founded | 2024-era startup |
| Founders | Reiner Pope, Mike Gunter |
| Founders’ background | Google / TPU engineering |
| Headquarters | Mountain View, California |
| Product focus | AI server chip |
| Primary workload | AI model training |
| Earlier funding | $25 million |
| Current fundraising | $75 million-$100 million reported earlier |
| Current target valuation | ~$4 billion |
MatX previously raised $25 million in a round led by Safe Superintelligence co-founder Daniel Gross and former GitHub CEO Nat Friedman, according to The Information.
Anthropic Is Building An In-House Silicon Team
The MatX discussions are part of a broader hardware push already underway at Anthropic.
The company publicly confirmed earlier this month that it was building an internal silicon team to design custom chips for Claude.
Anthropic has also recruited experienced chip engineers from Google and OpenAI.
This week, it hired Google chip veteran Amir Salek. In June, it hired former OpenAI chip engineer Clive Chan, who worked on the company’s custom processor program.
Anthropic’s Hardware Strategy
In-House Silicon Team
Former Google Engineers
Former OpenAI Chip Engineers
Potential MatX Partnership
Nvidia
Google TPUs
Amazon Chips
↓
Multi-Chip Computing Strategy
Anthropic has emphasized that custom silicon will complement rather than immediately replace its external suppliers.
Anthropic Will Continue Using Nvidia And Other Chips
The push toward custom chips does not mean Anthropic is abandoning Nvidia.
The company has explicitly said it will maintain a multi-chip approach.
Anthropic’s models already run across hardware supplied by Nvidia, Google and Amazon, and the company has agreements with multiple infrastructure providers to secure future capacity.
Anthropic’s Chip Supplier Strategy
| Supplier / Technology | Role |
|---|---|
| Nvidia GPUs | Major existing AI compute source |
| Google TPUs | Large-scale AI compute |
| Amazon Trainium / Inferentia | Alternative AI accelerators |
| Anthropic custom silicon | Future workload-specific hardware |
| MatX technology | Potential training-chip capability |
| Other startups | Being evaluated |
This diversified strategy could reduce supply-chain risk while allowing Anthropic to optimize individual workloads for different processors.
Anthropic Has Committed Billions To External Compute
The scale of Anthropic’s external infrastructure commitments explains why custom silicon has become strategically important.
The company plans to purchase $36 billion worth of Google’s AI chips and has signed a $45 billion agreement with Nscale for AI cloud-computing capacity. It has also committed $1.25 billion per month to SpaceX for computing capacity through May 2029.
Major Anthropic Compute Commitments
| Partner | Reported Commitment | Purpose |
|---|---|---|
| $36 billion | AI chips | |
| Nscale | $45 billion | Cloud computing capacity |
| SpaceX | $1.25 billion/month through May 2029 | Data-center compute |
| Amazon | Up to 5 GW new capacity | AI infrastructure |
| Google/Broadcom | 5 GW next-generation TPU capacity | AI compute |
These commitments demonstrate that Anthropic expects its demand for computing power to remain extremely high even if it successfully develops its own processors.
SpaceX Deal Could Provide Access To 220,000+ Nvidia Chips
Anthropic’s agreement with SpaceX includes access to the Colossus 1 facility, which houses more than 220,000 Nvidia chips, according to Reuters.
The deal illustrates the company’s immediate need for massive amounts of established compute while its internal chip program remains years away from full-scale deployment.
This creates a two-track strategy:
Today
↓
Nvidia + Google + Amazon + Cloud Providers
↓
Transition
↓
Custom Anthropic Silicon
↓
Long Term
↓
Hybrid Multi-Chip Infrastructure
The strategy allows Anthropic to continue scaling Claude without waiting for its own processors to become commercially ready.
Training Chips Could Be The Initial Focus
The MatX discussions provide an indication of the type of processor Anthropic may be interested in developing first.
MatX has been working on a chip intended for AI training, the process through which large models learn from enormous datasets.
Training differs from inference, where an already-trained model generates responses to user prompts.
Training Vs Inference
| Area | Training | Inference |
|---|---|---|
| Purpose | Build / update AI models | Run trained models |
| Compute requirement | Extremely high | High but workload differs |
| Example | Training Claude | Generating Claude response |
| Optimization | Large-scale parallel computation | Latency and efficiency |
| Potential Anthropic chip | MatX-style training processor | Could be developed separately |
Reuters reported that Anthropic could ultimately pursue an inference processor as well.
The distinction matters because the optimal chip architecture can differ substantially depending on whether the workload is training or inference.
OpenAI Is Also Moving Into Custom Chips
Anthropic is not alone.
OpenAI has been developing its own custom AI processor, known as Jalapeno, as it seeks greater control over its computing infrastructure.
At a conference this week, OpenAI executives said the processor outperformed a comparable Nvidia chip in inference calculations and was more energy efficient, according to Reuters.
AI Labs Moving Toward Silicon
| Company | Hardware Strategy |
|---|---|
| Anthropic | Building in-house silicon team |
| OpenAI | Developing custom Jalapeno chip |
| TPU ecosystem | |
| Amazon | Trainium / Inferentia |
| Meta | Custom AI accelerators |
| Nvidia | Full-stack AI hardware and software |
The trend reflects a broader effort by large AI companies to optimize hardware for their own models and workloads.
Nvidia Supply Constraints Are Another Factor
Custom silicon could also provide Anthropic with a hedge against shortages of Nvidia processors.
Nvidia said during its latest earnings discussion that its processors are expected to remain in short supply through 2027.
For an AI company whose business depends on continuous access to compute, supply availability can be as important as chip price.
Why Reduce Nvidia Dependence?
| Issue | Nvidia Dependence | Custom Silicon |
|---|---|---|
| Supply | Subject to market availability | Greater internal control |
| Cost | External pricing | Potential long-term optimization |
| Performance | General-purpose AI platform | Model-specific optimization |
| Energy efficiency | Vendor-defined | Can be customized |
| Software | Mature ecosystem | Development burden |
| Flexibility | Broad workloads | Tailored workloads |
Custom chips do not automatically guarantee lower costs. Developing and manufacturing them requires major upfront investment, specialized engineers and extensive validation.
The strategic value comes from potential optimization at very large scale.
Custom Chips Could Improve Claude’s Economics
Anthropic’s AI business is expanding rapidly, which makes inference efficiency increasingly important.
Every time a user asks Claude a question, the company incurs computing costs.
At enormous scale, even a small improvement in cost per query can translate into billions of dollars of potential savings.
Higher Claude Usage
↓
More Inference Compute
↓
Higher Infrastructure Costs
↓
Custom Hardware Optimization
↓
Lower Cost Per Computation
↓
Potentially Higher Gross Margins
The same principle applies to training, where the cost of developing increasingly capable models can reach enormous levels.
Anthropic’s IPO Plans Add Strategic Pressure
Anthropic is also preparing for a potential public listing.
Reuters reported that the company plans to unveil its IPO prospectus after the U.S. Labor Day holiday, with a potential listing in late September or early October.
The company has reportedly been targeting a valuation of up to $2 trillion based on aggressive future revenue expectations.
This makes infrastructure economics particularly important.
Public-market investors are likely to scrutinize how Anthropic plans to fund the enormous computing requirements associated with future Claude models.
Anthropic’s Pre-IPO Scale
| Metric | Reported Figure |
|---|---|
| Series H funding | $65 billion |
| Post-money valuation at Series H | $965 billion |
| May 2026 revenue run rate | $47 billion |
| Potential IPO valuation target | Up to $2 trillion |
| Planned Google chip purchase | $36 billion |
| Nscale computing agreement | $45 billion |
| SpaceX compute commitment | $1.25 billion/month |
Anthropic’s own May funding announcement confirmed a $65 billion Series H round at a $965 billion post-money valuation and said its revenue run rate had crossed $47 billion.
Anthropic Has Raised $65 Billion This Year
Anthropic’s financial resources give it unusual capacity to pursue semiconductor development.
In May, the company announced a $65 billion Series H financing at a $965 billion post-money valuation. It said the funding would be used to expand compute, advance AI research and scale Claude products and partnerships.
The company also said it had agreements for up to five gigawatts of new Amazon capacity and five gigawatts of next-generation TPU capacity with Google and Broadcom.
This suggests Anthropic is pursuing infrastructure at a scale more commonly associated with the largest technology companies.
Anthropic Is Evaluating Other Chip Startups
The MatX discussions do not appear to be an isolated search.
Reuters reported that Anthropic has recently held meetings with a range of AI chip startups and has not yet decided whether to make an acquisition.
The meetings are intended to help Anthropic’s engineers and executives understand different approaches to AI chip design.
This means MatX may be one option among several rather than the definitive hardware partner.
Anthropic’s Possible Paths
| Option | Advantage | Challenge |
|---|---|---|
| Build internally | Maximum control | Slow and expensive |
| Acquire startup | Immediate talent and IP | High acquisition cost |
| Partner with startup | Shared development risk | Less control |
| Buy Nvidia | Mature ecosystem | Higher dependence |
| Buy Google TPUs | Strong AI performance | External supplier |
| Multi-chip strategy | Supply diversification | Greater software complexity |
The move toward partnership talks with MatX could indicate that Anthropic currently sees collaboration as a more attractive balance between speed, cost and control than a multibillion-dollar acquisition.
The Bigger Picture
Anthropic’s discussions with MatX show that the competition in artificial intelligence is increasingly moving beyond models and software into the semiconductor layer. The reported $7 billion acquisition proposal would have given Anthropic direct access to specialized chip-design talent at a time when the company is building an internal silicon team and spending tens of billions of dollars securing external computing capacity.
The potential shift from acquisition to partnership also reflects the difficulty of building AI chips. Anthropic needs enormous amounts of compute today, while custom silicon may take years and hundreds of millions of dollars to develop. By maintaining relationships with Nvidia, Google and Amazon while exploring MatX and other chip startups, Anthropic can pursue hardware independence without immediately abandoning the infrastructure that powers Claude today.
Looking Ahead
The key question is whether Anthropic ultimately develops a proprietary training chip, an inference chip, or a combination of both. The company’s recent hiring of chip veterans, discussions with MatX and other startups, and continuing investment in external compute suggest that custom silicon is becoming a long-term strategic pillar rather than a short-term experiment.
For MatX, the end of the acquisition discussions could still leave a major opportunity if a partnership with Anthropic materializes. The startup is reportedly seeking funding at a roughly $4 billion valuation, while Anthropic has the capital and commercial scale to become a significant anchor customer. More broadly, the growing number of AI labs designing their own processors could reshape the semiconductor industry by creating new competitors to Nvidia and increasing demand for highly specialized chips optimized for individual AI workloads.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
