Anthropic is recruiting a former Google TPU pioneer to lead its internal silicon accelerator program, marking a strategic shift toward heterogeneous computing for inference.
Anthropic PBC has hired Amir Salek, founder of Alphabet Inc.’s Google custom AI chip project, to join its compute team. In his new role, Salek will report directly to James Bradbury, Anthropic’s Head of AI Compute. Prior to 2022, Salek oversaw Google’s Tensor Processing Unit (TPU) business and helped design the first seven generations of TPU chips.
This executive appointment reflects a broader transformation in AI computing infrastructure. Alongside Microsoft (MSFT.US) planning to soon launch its latest self-developed AI accelerator, Maia, and a new agreement between Google and fabless chip company Marvell Technology covering customized chips including AI accelerators, storage, and memory controllers, the industry is comprehensively upgrading from a reliance on collective NVIDIA GPU procurement to a heterogeneous computing system. This new architecture features the large-scale coordinated operation of NVIDIA GPUs, AMD GPUs, self-developed AI Application-Specific Integrated Circuits (AI ASICs)/XPUs, and CPUs/DPUs. As AI large model architectures stabilize and global token call volumes exhibit exponential growth, high-concurrency AI inference workloads are increasingly well-suited for reducing cost per token through customized silicon.
Consequently, AI training operator processes and frontier workloads with the most complex architectures and fastest rates of change continue to rely heavily on AI GPU clusters. Frontier model pre-training, reinforcement learning, and rapidly evolving new operators remain more dependent on GPU programmability, the CUDA ecosystem, and NVLink/NVSwitch clustering capabilities. Meanwhile, large-scale AI inference workloads surrounding mature and open-source AI large models, Copilot, and AI agents are increasingly suited for specialized self-developed AI chips.
According to institutions such as Morgan Stanley and Wedbush Securities, which remain bullish on AI compute infrastructure investments, the seemingly boundless frontier of cutting-edge compute demand in the AI inference era—and the growing resource requirements driven by AI agents—positions AI ASICs as a critical component of a second trillion-dollar compute ecosystem without cannibalizing GPU demand. This dynamic further reinforces the investment thesis that the AI semiconductor supercycle is not solely a GPU cycle, but rather a broader cycle of increasing silicon content across data centers.
Anthropic currently procures AI chips and accelerators at scale from multiple channels, including NVIDIA, Google, and Amazon. However, the company has recently signaled its intent to establish an internal self-developed silicon chip business. The San Francisco-based firm has begun recruiting specialized talent and posting related job openings. Media reports also indicate that Anthropic has initiated negotiations with Taiwan Semiconductor, the leading global chip foundry, regarding long-term cooperation for advanced manufacturing processes and advanced packaging.
Anthropic has not yet disclosed the architecture or tape-out timeline for its latest generation of self-developed chips, nor has it announced formal external fabless chip co-design partners such as Broadcom or Marvell Technology. Broadcom previously participated in Anthropic’s approximately $35 billion computing capacity expansion project, providing customized AI ASIC chips, network infrastructure solutions, and financing support. However, it has not been confirmed as a long-term partner for Anthropic’s proprietary chip business to date.
Like other super-giants in the AI application sector, Anthropic is racing to secure sufficient data center infrastructure to support its ambitious goals. Customized AI chip clusters could help the company address the global shortage of AI chip supply more comprehensively, optimize chip design and overall performance according to its specific workload requirements, and significantly reduce AI inference costs on the data center operational side.
OpenAI, the developer behind ChatGPT and Anthropic’s strongest competitor, is pursuing a similar trajectory. OpenAI has released a prototype architecture for its self-developed AI chip, named “Jalapeno,” designed in collaboration with chip giant Broadcom, and plans to begin deploying it later this year. Concurrently, Anthropic is accelerating its infrastructure footprint, signing agreements for AI chips and ultra-large-capacity data centers at a rapid pace while engaging in multi-year collaborations with established suppliers and emerging cloud computing providers. Recently, the company reached an agreement with UK chip startup Fractile, placing an initial chip order valued at approximately $250 million, with plans to expand the contract in the future. Anthropic has also secured large-scale data center capacity supply agreements with Riot Platforms Inc. and Volta Infra Holdings Ltd.
Furthermore, leadership within the Claude AI super platform is actively conducting relevant valuations and IPO fundraising projections. Management is preparing to publicly file the necessary documents for a potential mega-IPO as early as the end of this month. Citing informed s