Anthropic has hired Amir Salek, who led Google's Tensor Processing Unit (TPU) development from 2017 to 2022, to build custom semiconductors for its Claude AI models. The move, reported by Bloomberg on August 21, signals that the AI industry's two pillars - model companies and chip makers - are now competing on each other's turf, with the race shifting from individual component performance to full-stack control.
Model companies are moving into silicon, and chip firms are building models. OpenAI unveiled its custom inference chip, "Habanero," co-developed with Broadcom in June, optimized for its GPT models to improve speed and power efficiency. It is scheduled for deployment in data centers by late this year. Anthropic has established a dedicated team to design custom semiconductors and is recruiting engineers experienced in both hardware and software, aiming to co-design chips and models for better performance and power efficiency.
NVIDIA, the dominant AI chip maker, is moving in the opposite direction. It is developing open-source AI models to rival top-tier Chinese models like Moonshot AI and Zhipu AI. Its "Nemotron-4" model is being accelerated through a $6 billion acquisition of AI startup Fullstack's model development technology, along with more than 100 of its engineers. NVIDIA is also supplying high-performance models to startups for free, a countermove aimed at boosting demand for its chips as model companies reduce their dependency on NVIDIA hardware.
AMD has also released "Instella," an AI model optimized for its GPUs, to expand its developer ecosystem.
Why companies are crossing into each other's markets
Controlling only one segment of the AI market is no longer enough for long-term dominance. AI performance and cost depend on how well models, semiconductors, software, and data centers integrate. Direct control over these elements accelerates product development, lowers costs, and locks customers into an ecosystem.
Google, Amazon, and Meta have already pursued vertical integration. Specialized model and semiconductor companies are now joining that race. This shift aims to reduce reliance on external partners, lower costs, mitigate supply risks, and seize direct control over the AI ecosystem.
Competition and collaboration coexist
As companies encroach on each other's territories, cooperation and competition are becoming intertwined. Anthropic plans to continue using AWS, Google, NVIDIA, and AMD chips even as it develops its own. OpenAI's Habanero is designed for inference, but the company still requires NVIDIA GPUs for advanced model training. NVIDIA, while developing its own models, retains OpenAI and Anthropic as key GPU customers.
"Companies are pushing vertical integration to control costs and supply chains, but semiconductors, models, and cloud services each require decades of specialized expertise," said a source from the AI industry. "Ultimately, a 'multi-chip and multi-model' strategy - centering on in-house technologies while leveraging external products as needed - will spread."
For product development teams, this convergence has practical implications. The full-stack war means AI infrastructure decisions are no longer just about picking the best model or the fastest chip. Teams will need to evaluate how models, hardware, and cloud services work together as a system. The rise of multi-chip and multi-model strategies, as the industry source noted, suggests that flexible integration - rather than single-vendor loyalty - will become a core design principle for AI for Product Development. Understanding the trade-offs between custom silicon and off-the-shelf options, and between in-house models and external APIs, will shape both cost structures and product roadmaps. The Generative AI and LLM landscape is shifting quickly, and product teams that plan for interoperability across chips and models will be better positioned than those that bet on a single vendor.
Why this matters for product development
For product teams, the practical takeaway is that AI infrastructure choices now carry strategic weight. The convergence of chips and models means that a product's performance ceiling depends on how well its underlying hardware and software are co-designed - not just which model you choose. Teams should expect more vendor lock-in pressure as companies build integrated stacks, but also more options as multi-chip and multi-model strategies spread. Product roadmaps should account for the possibility of switching or combining vendors, and cost models should reflect that custom silicon and in-house models are becoming viable alternatives to renting GPUs and calling APIs.
Your membership also unlocks:
