Anthropic is developing proprietary inference accelerators while Nvidia expands its software stack to directly compete in the generative AI model training and deployment space.
The boundaries between semiconductors and AI models—the two foundational pillars of the artificial intelligence industry—are rapidly blurring. Model developers such as Anthropic and OpenAI are designing custom silicon, while semiconductor leaders like NVIDIA and AMD are expanding into AI model development. This strategic pivot seeks to reduce dependence on third-party vendors, cut costs, mitigate supply chain vulnerabilities, and capture direct control over the AI ecosystem. Industry analysts observe that competition has moved beyond benchmarking isolated chips or models, escalating into a “full-stack” race to secure every component necessary for end-to-end AI operations.
On August 21, Bloomberg reported that Anthropic, creator of the Claude AI model, has hired Amir Salek, who spearheaded Google’s Tensor Processing Unit (TPU) development from 2012 to 2022. Anthropic has since formed a dedicated engineering group to design custom semiconductors tailored for Claude, actively recruiting professionals with expertise spanning both hardware and software. The initiative aims to co-engineer Claude models alongside proprietary silicon, optimizing both computational performance and power efficiency.
In parallel, Anthropic maintains a multi-chip strategy, deploying NVIDIA graphics processing units (GPUs), Google TPUs, and Amazon Trainium accelerators. Developing in-house silicon is intended to diversify supplier dependencies and scale efficiently amid surging compute demand.
Similarly, OpenAI unveiled Habanero, a custom inference chip co-developed with Broadcom in June. Optimized for OpenAI’s GPT family, Habanero targets faster inference speeds, improved energy efficiency, and reduced operational costs, with data center deployment scheduled for later this year.
Meanwhile, NVIDIA is pivoting upstream by developing Nemotron-4, an open-source AI model designed to compete with leading Chinese systems such as Moonshot AI and Zhipu AI. To accelerate progress, NVIDIA invested $6 billion to license Fullstack AI’s model-development technology and recruited more than 100 of its core engineers. Anticipating that model developers will increasingly rely on custom silicon rather than NVIDIA GPUs, the company plans to offer its high-performance models to startups at no cost—a move designed to drive direct demand for its underlying hardware.
The strategic convergence is underscored by recent industry milestones, including OpenAI CEO Sam Altman and Broadcom CEO Hok Tan showcasing a prototype wafer for the Jalapeno AI chip. Companies are crossing traditional boundaries because controlling a single tier of the AI value chain no longer guarantees sustained market leadership. Performance and unit economics now hinge on the seamless integration of models, silicon, software stacks, and data center infrastructure. End-to-end ownership accelerates iteration cycles, compresses costs, and anchors users within closed ecosystems. While giants like Google, Amazon, and Meta have long pursued vertical integration, specialized model and semiconductor firms are now racing to build comparable full-stack capabilities.
Despite this territorial overlap, a dynamic of competitive coexistence is taking hold. Anthropic continues to procure accelerators from AWS, Google, NVIDIA, and AMD alongside its internal silicon efforts. OpenAI’s inference-focused Habanero still relies on NVIDIA GPUs for large-scale model training. Likewise, NVIDIA remains a critical hardware partner for both OpenAI and Anthropic even as it builds its own generative models.
The evolving landscape was highlighted at the San Francisco AI Summit on the 24th of last month at The Midway, where NVIDIA CEO Jensen Huang and OpenAI CEO Sam Altman were joined by South Korean President Lee Jae Myung. Reflecting the industry’s pragmatic trajectory, one AI sector analyst noted, “Ultimately, companies will prioritize their own chips and models but adopt a ‘multi-chip and multi-model’ strategy, leveraging rivals’ products based on specific use cases.”