Anthropic Poaches Google TPU Founder Amir Salek to Spearhead Custom AI Chip Development
Anthropic hires Amir Salek, former head of Google's TPU program, to lead its custom AI chip initiative, signaling a major shift in the AI hardware landscape and intensifying competition with NVIDIA.
The race for AI supremacy isn't just about algorithms and models anymore; it's increasingly about the silicon that powers them. In a move that sent ripples across the artificial intelligence and semiconductor industries, Anthropic, the developer behind the Claude AI models, has reportedly hired Amir Salek, the acclaimed founder and former head of Google's Tensor Processing Unit (TPU) program. This significant recruitment, reported on August 21, 2026, signals Anthropic's ambitious push into designing its own custom AI chips, a strategic pivot aimed at gaining greater control over its compute infrastructure and reducing reliance on external providers like NVIDIA.
Salek's arrival marks a pivotal moment in Anthropic's compute strategy, which has been unfolding throughout 2026. This development follows earlier reports of Anthropic establishing an in-house chip design team and engaging in discussions with manufacturers like Samsung. By bringing a veteran architect with a proven track record in specialized AI silicon development, Anthropic is clearly doubling down on vertical integration, a trend increasingly adopted by major AI labs to optimize performance, manage costs, and secure supply chains in a fiercely competitive landscape.
1. The Strategic Imperative: Why Custom AI Chips?
Anthropic's decision to invest heavily in custom silicon is a direct response to the escalating demands and unique characteristics of modern AI workloads, particularly those associated with large language models (LLMs) like Claude. General-purpose GPUs, while powerful, are not always optimally efficient for the highly specific computational patterns of AI training and inference. Custom Application-Specific Integrated Circuits (ASICs), like Google's TPUs, can be tailored precisely to these workloads, offering significant advantages in performance, energy efficiency, and cost-effectiveness.
The current AI hardware market is largely dominated by NVIDIA, which holds an estimated 70% to 95% market share in AI accelerators. While NVIDIA's GPUs and its CUDA software ecosystem have been instrumental in the AI revolution, this dominance creates several challenges for AI developers: supply chain bottlenecks, high operational costs due to hardware vendor margins, and a lack of granular control over hardware-software co-design.
By designing its own chips, Anthropic aims to achieve:
- Optimized Performance: Chips can be designed from the ground up to match Claude's specific architectural requirements, leading to faster processing and lower latency.
- Reduced Costs: In the long term, owning the hardware stack can significantly lower the substantial costs associated with leasing external hardware, especially as AI models scale to serve billions of tokens daily.
- Enhanced Scalability: Greater control over hardware development allows Anthropic to scale its infrastructure more effectively and predictably to meet growing demand.
- Supply Chain Resilience: Diversifying its chip supply reduces vulnerability to market fluctuations and geopolitical risks associated with relying on a single vendor.
- Innovation Control: Vertical integration enables tighter coupling between hardware and software development, fostering more rapid and specialized innovation.
This trend is not unique to Anthropic; other AI giants like Google (with TPUs), Amazon (Trainium, Inferentia), Microsoft (Maia AI Accelerator), Meta (MTIA), and OpenAI (Jalapeño) have all embarked on similar custom silicon initiatives.
2. Amir Salek: A Visionary in Custom Silicon
The recruitment of Amir Salek underscores the seriousness of Anthropic's hardware ambitions. Salek is a highly respected figure in the semiconductor industry, with a career spanning foundational roles at two of the most influential companies in AI hardware.
Before joining Google in 2013, Salek spent approximately eight years at NVIDIA, where he founded and led the company's System-on-a-Chip (SoC) Design organization. During his tenure, he contributed to the development of NVIDIA's GPU and Tegra product lines, gaining invaluable experience in high-performance chip design.
His most notable achievement, however, came at Google. Recruited to establish custom silicon development for Google's data centers, Salek became the founder and head of the Google TPU program. From 2013 to 2022, he directed the development and deployment of multiple generations of specialized processors, including TPUv1 through TPUv4 and Edge TPU – the first seven generations of Google's Tensor Processing Units. He is widely credited as "the brains behind Google's TPU chip," playing a crucial role in developing an architecture specifically designed for AI tasks.
Salek's expertise lies not just in chip design but also in understanding the intricate relationship between hardware and software for optimal AI performance. His experience in building out a custom silicon capability from scratch at Google makes him an ideal leader for Anthropic's nascent, but critical, hardware division. He will report to James Bradbury, Anthropic's head of compute.
3. The Broader AI Hardware Landscape and Future Implications
Anthropic's move, spearheaded by Salek, is a clear signal that the AI industry is entering a new phase of hardware competition. While Anthropic maintains a "multi-chip approach," continuing partnerships with existing providers like AWS, Google, NVIDIA, and AMD, the in-house effort aims to complement these relationships by providing tailored solutions for its specific needs. This strategy acknowledges that developing an advanced AI chip is a costly and complex endeavor, potentially requiring hundreds of millions of dollars.
The shift towards custom silicon is expected to drive down per-token costs and reduce latency over the medium term, making sophisticated enterprise AI use cases, such as multi-step autonomous agents and real-time big data processing, more cost-effective and scalable. For developers, this could translate into more accessible and powerful AI models, enabling new applications and pushing the boundaries of what AI can achieve. As AI models become more complex and widespread, the demand for specialized hardware capable of handling intensive computational tasks efficiently will only increase.
The long-term implications are profound: a more diversified AI hardware ecosystem, increased innovation in chip architectures, and potentially a rebalancing of power in the AI supply chain. The coming years will undoubtedly see intense competition and collaboration as AI companies strive to build the most efficient and powerful compute infrastructure to support the next generation of artificial intelligence.
Comparison Overview
| Company | Custom Chip Initiative | Primary Motivation | Status/Notes |
|---|---|---|---|
| Anthropic | In-house custom AI chips (led by Amir Salek) | Optimize Claude performance, reduce costs, secure supply | Actively hiring, strategic leadership in place. Multi-chip approach with existing partners. |
| Tensor Processing Units (TPUs) | Accelerate AI workloads for Google products and Cloud customers | Multiple generations deployed, backbone of Google's AI infrastructure. | |
| NVIDIA | Dominant GPU Architectures (e.g., A100, H100) | General-purpose AI acceleration, strong software ecosystem (CUDA) | Market leader, but facing increasing competition from custom silicon efforts. |
| OpenAI | Jalapeño chip (in partnership with Broadcom) | Optimize inference for large language models | Unveiled recently, slated for deployment later this year. |
| Microsoft | Azure Maia AI Accelerator, Azure Cobalt CPU | Power Azure data centers, reduce reliance on third-party suppliers | Designed for LLMs, enhancing efficiency and cost-effectiveness in Azure. |
| Meta | Meta Training and Inference Accelerator (MTIA) | Efficiently run and train Meta's AI models | Investing billions to reduce operational costs and improve performance. |
| Amazon Web Services (AWS) | Trainium, Inferentia | Provide optimized AI training and inference alternatives for AWS customers | Offering specialized chips as alternatives to GPUs for cloud clients. |
Frequently Asked Questions (FAQ)
Q: Who is Amir Salek and what is his significance?
Amir Salek is an electrical engineer and technology executive, renowned for founding and leading Google's Tensor Processing Unit (TPU) program from 2013 to 2022. He oversaw the development of multiple generations of TPUs, which are specialized AI chips. His recruitment by Anthropic to lead their custom AI chip initiative is highly significant as it brings deep, proven expertise in building AI-specific hardware from scratch to a leading AI model developer.
Q: Why is Anthropic developing its own AI chips?
Anthropic is developing its own AI chips to gain greater control over its compute infrastructure, optimize performance for its Claude AI models, reduce long-term operational costs associated with leasing external hardware, enhance scalability, and diversify its supply chain. This strategy aims to create hardware specifically tailored to the unique computational demands of its AI workloads.
Q: How does this impact NVIDIA's position in the AI hardware market?
NVIDIA currently holds a dominant position in the AI chip market. Anthropic's move, along with similar initiatives from other tech giants, signals an intensifying competition in the AI hardware space. While Anthropic will continue to partner with NVIDIA and other vendors, its in-house chip development aims to reduce its overall reliance on external suppliers, potentially leading to a more diversified market and increased innovation across the industry.
Q: What are the benefits of custom AI chips for developers?
For developers, the proliferation of custom AI chips can lead to several benefits: more efficient and powerful AI models, lower inference costs, reduced latency for AI-powered applications, and potentially greater accessibility to advanced AI capabilities. These advancements could enable the creation of more sophisticated AI agents and real-time AI solutions across various industries.
Try Our Developer Utilities
Simplify your engineering workflows with our free browser-native tools: