AI labs are increasingly prioritizing inference over pre-training costs, leading to heightened reliance on specific GPU manufacturers like Nvidia. This shift reflects a growing trend where frontier models are co-designed to optimize performance for particular hardware, such as Nvidia’s GB300 racks. As system architectures diverge, the ability to easily switch models between different types of GPUs is diminishing. This situation particularly benefits companies like Anthropic, which can adapt its models across various chips, whereas other labs face rising switching costs and loss of portability as they refine their models to operate best on specific systems.

Google: Google develops custom TPUs optimized for Gemini models using torus interconnect topologies suited to its AI workloads. Full-stack control enables seamless operation across models, chips, and networking for internal teams. Highlighted for Gemini’s TPU dependency and arrangements allowing Anthropic access to TPU engineers.
Nvidia: Nvidia dominates AI acceleration with GPU architectures like Blackwell and Rubin, emphasizing inference through hardware-software co-design for optimal token efficiency. Recent GTC announcements highlight full-stack systems for agentic AI and large-scale inference. The post underscores Nvidia’s growing power as labs co-design models for its switched scale-up topologies, particularly for MoE inference.
OpenAI: OpenAI advances frontier AI models with inference optimizations on diverse hardware including Cerebras alongside Nvidia GPUs. Recent deployments focus on Cerebras for specialized tasks like code generation to enhance performance. Noted as reliant on Cerebras and Nvidia due to model-specific hardware tuning.
Anthropic: Anthropic builds Claude models trained across multiple platforms including Google TPUs, AWS Trainium, and Nvidia GPUs, enabling rare portability. Strategic partnerships deepen ties with Nvidia’s Blackwell and Rubin for inference economics. Positioned as the sole lab affording switches, yet prioritizing Nvidia for frontier model inference.
Gavin Baker: Gavin Baker serves as Managing Partner and CIO of Atreides Management, focusing on AI and technology investments with prescient market analysis. He shares expertise on AI hardware economics via X and podcasts. Authors the post detailing inference-driven erosion of model portability favoring Nvidia.

Co-Design Trend: Frontier models increasingly co-designed for specific accelerators like Nvidia GB300 racks.
Inference Shift: AI labs prioritize inference over pre-training costs, amplifying hardware-specific optimizations.
Portability Decline: Diverging system topologies between accelerators reduce true model portability across hardware.