Mistral Board Member and a16z VC Anjney Midha on DeepSeek

Andreessen Horowitz general partner and Mistral AI board member Anjney “Anj” Midha first noticed DeepSeek’s groundbreaking performance six months ago. That was when DeepSeek introduced Coder V2, a model that rivaled OpenAI’s GPT-4 Turbo for coding-specific tasks, according to a research paper it released last year. This breakthrough put DeepSeek on a trajectory of rapid model improvements, culminating in the launch of R1—its latest open-source AI model, which has significantly disrupted the AI industry by offering cutting-edge performance at a fraction of the usual cost.

AI Compute Efficiency: DeepSeek R1’s Impact on the Industry

Despite Nvidia’s stock sell-off, Midha asserts that the emergence of R1 does not mean AI foundational models will cease their billion-dollar investments in GPUs and data centers. Instead, it signals that they will achieve more with the compute power they acquire.

“When people ask, ‘Anj, Mistral AI has raised a billion dollars. Does DeepSeek mean that all that money is unnecessary?’ No, actually, it’s incredibly valuable. DeepSeek’s efficiency improvements can be internalized and amplified with further investment,” Midha explains.

DeepSeek’s advancements mean AI companies can now achieve 10 times more output from the same compute resources. However, this does not imply that Mistral AI is lagging behind competitors like OpenAI or Anthropic. While OpenAI is reportedly in talks to raise another $40 billion, Mistral AI remains competitive due to its open-source strategy.

Open-Source vs. Proprietary AI Models: The Competitive Edge

Mistral AI’s open-source nature grants it access to vast technical contributions from the developer community. In contrast, proprietary AI companies like OpenAI and Anthropic must fund all research and compute power themselves. Midha highlights this advantage:

“You don’t need $20 billion. You just need more compute than any other open-source model. Mistral AI is well-positioned—they have the most compute among open-source providers.”

Facebook’s Llama, the largest Western open-source AI model competing with Mistral, is also attracting significant investment. Mark Zuckerberg recently announced plans to spend “hundreds of billions of dollars” on AI, including $60 billion in capital expenditures for data centers in 2025.

a16z’s Oxygen GPU Sharing Program: Addressing Compute Shortages

Midha, who also serves on the boards of AI-focused startups like Black Forest Labs and Luma, sees no slowdown in the demand for GPUs. He leads a16z’s Oxygen program, which addresses the scarcity of high-performance GPUs, particularly Nvidia’s H100s, by acquiring them for its portfolio companies.

“Oxygen is overbooked right now. I can’t allocate enough,” Midha admits. The demand for AI model training GPUs is soaring, and even after models are built, AI inference—running these models for real-world applications—requires even more compute power.

The Future of AI Infrastructure: DeepSeek’s Role

While DeepSeek R1 has revolutionized AI model efficiency, it won’t disrupt large-scale AI infrastructure projects like OpenAI’s recently announced $500 billion partnership with SoftBank and Oracle. However, it has highlighted a critical geopolitical issue: the importance of AI infrastructure independence.

Midha raises a thought-provoking question for national governments: “Do they want to rely on Chinese AI models, which come with censorship risks and potential data security concerns? Or do they prefer Western AI models that adhere to democratic laws and NATO agreements?”

Western nations are already taking steps to safeguard their AI ecosystems. Hundreds of companies have blocked DeepSeek, despite its availability as both an open-source model and a consumer AI service. However, skeptics argue that concerns over Chinese AI models may be overblown. Companies can deploy DeepSeek R1 locally in their own data centers or access it via American cloud providers like Microsoft Azure Foundry.

The Industry’s Next Moves: Navigating AI Compute Challenges

Intel’s former CEO, Pat Gelsinger, recently revealed that his AI startup, Gloo, is developing AI chat services using DeepSeek R1 instead of proprietary options like Llama or OpenAI’s models. This highlights the increasing appeal of efficient, open-source AI models in the evolving AI landscape.

Yet, as AI companies continue their relentless pursuit of compute power, Midha humorously concludes with a request: “If you have extra GPUs, please send them to Anj.”

The Future of AI Compute and Open-Source Models

DeepSeek R1 marks a significant shift in AI compute efficiency, but it does not negate the need for substantial investment in GPUs and AI infrastructure. While open-source models like Mistral AI and DeepSeek R1 provide cost-effective alternatives to proprietary models, the hunger for compute power continues to grow.

As AI competition intensifies, the balance between open-source innovation and proprietary AI investments will shape the industry’s trajectory. Companies and governments must decide how to secure their AI future—whether through open-source collaboration or billion-dollar AI infrastructure investments.

One thing is certain: the race for AI supremacy is far from over.