Why reasoning models like DeepSeek-R1 are increasing (not decreasing) demand for GPUs

The rise of advanced AI reasoning models like DeepSeek-R1 has transformed the computational landscape for AI infrastructure providers. Together AI has secured significant funding to meet growing enterprise demands for high-performance AI model deployment and reasoning capabilities.

Recent funding and growth: Together AI has raised $305 million in Series B funding led by General Catalyst and Prosperity7, marking a significant expansion of its AI deployment platform.

• The company has experienced 6X year-over-year growth with over 450,000 registered developers
• Notable customers include AI startups like Krea AI, Captions, and Pika Labs
• The platform now supports multiple AI modalities including language, reasoning, images, audio, and video

Infrastructure demands: Contrary to initial expectations, DeepSeek-R1 and similar reasoning models are driving increased infrastructure requirements.

• DeepSeek-R1’s 671 billion parameters require distribution across multiple servers
• User requests can last up to 2-3 minutes, necessitating dedicated computational resources
• Together AI has introduced “reasoning clusters” offering 128 to 2,000 chips for optimal performance

Key applications: Reasoning models are finding practical applications across various use cases.

• Enhanced coding capabilities through problem decomposition
• Reduced model hallucinations through verification processes
• Quality improvements in non-reasoning models
• Self-improvement capabilities using reinforcement learning

Technical innovations: Together AI has made strategic moves to enhance its platform capabilities.

• Acquisition of CodeSandbox for lightweight VM deployment
• Implementation of Nvidia’s Blackwell GPUs, offering 2X performance improvement
• Optimization of inference speeds, achieving 85 tokens per second compared to Azure’s 7 tokens per second

Competitive landscape: Together AI operates in an increasingly crowded market for AI infrastructure.

• Major cloud providers Microsoft, AWS, and Google offer competing platforms
• Specialized AI infrastructure providers like Groq and Samba Nova target similar markets
• Together AI differentiates through full-stack offerings and superior performance metrics

Market implications: The increasing demand for AI infrastructure, particularly for reasoning models, suggests a continuing trend toward more sophisticated and resource-intensive AI applications, potentially creating new opportunities and challenges for infrastructure providers and enterprises alike.