DeepSeek AI and GPU Demand: Insights from Anjney Midha

DeepSeek's Impact on the AI Landscape

Anjney “Anj” Midha, a general partner at Andreessen Horowitz and board member of Mistral, initially observed DeepSeek’s remarkable capabilities approximately six months ago, as he shared with TechCrunch.

The introduction of Coder V2 by DeepSeek marked a turning point. It demonstrated performance comparable to OpenAI’s GPT4-Turbo specifically for coding tasks, as detailed in their published research last year.

Rapid Model Development

Following Coder V2, DeepSeek consistently released enhanced models at intervals of a few months, culminating in the release of R1. R1 is a new open source reasoning model that has significantly impacted the tech industry by delivering performance on par with industry standards at a considerably lower cost.

Compute Efficiency, Not Reduced Spending

Despite recent declines in Nvidia’s stock value, Midha emphasizes that R1 does not signal a decrease in spending on GPUs and data center infrastructure by AI foundational model developers.

Instead, it suggests an increased efficiency in utilizing existing computational resources. They will achieve more with the compute power already available to them.

Leveraging Efficiency Improvements

“The question arises, with Mistral having secured a billion dollars in funding, does DeepSeek render that investment unnecessary?” Midha poses. “The answer is no. It’s actually incredibly beneficial for them to analyze DeepSeek’s efficiency gains, integrate those learnings, and then apply their billion-dollar investment.”

He elaborates, “This allows us to generate ten times the output from the same computational capacity.”

Mistral's Competitive Position

Midha contends that Mistral is not lagging behind competitors like OpenAI and Anthropic. These rivals have secured substantially larger funding rounds. OpenAI is reportedly exploring another funding round potentially reaching $40 billion.

Mistral’s competitive advantage lies in its open source nature. This approach provides access to a wealth of essentially free technical expertise from individuals motivated to contribute to the project.

Closed source competitors, conversely, must safeguard their proprietary information and bear the full cost of both labor and computational power.

Compute as a Key Differentiator

“A $20 billion budget isn’t required; simply possessing more compute than any other open source application is sufficient,” Midha stated regarding his portfolio company. “Mistral is favorably positioned, currently holding the largest compute capacity among open source providers.”

Continued Investment in AI

Facebook’s Llama, a major Western open source AI model competing with Mistral, is also expected to receive significant further investment. CEO Mark Zuckerberg recently affirmed plans to invest “hundreds of billions of dollars” in AI overall.

This includes a projected $60 billion in capital expenditures for 2025, primarily allocated to data centers.

a16z’s Oxygen GPU Sharing Program Faces High Demand

Midha, who also serves on the boards of AI image generation company Black Forest Labs and 3D model creator Luma, alongside being an angel investor in AI ventures like Anthropic and ElevenLabs, believes the demand for GPUs will remain strong for the foreseeable future.

As the head of a16z’s Oxygen program, he’s directly experiencing the scarcity of GPUs, especially Nvidia’s advanced H100s. Approximately a year and a half ago, the venture capital firm proactively addressed this issue by purchasing a substantial quantity of GPUs for use by its portfolio companies.

Currently, Oxygen is experiencing a state of being fully allocated. Midha notes that he is unable to provide sufficient GPU resources to meet the needs of the startups within the program.

These companies require GPUs not only for the initial AI model training process but also for the continuous operation of their AI-powered products serving customers.

“There is now an unrelenting demand for inference and ongoing consumption,” he clarifies.

This ongoing demand is a key reason why he doesn’t anticipate DeepSeek’s technological advancements to significantly impact the Stargate initiative. Stargate refers to OpenAI’s recently announced $500 billion partnership with SoftBank and Oracle, focused on establishing AI data centers.

The emergence of DeepSeek highlights a growing awareness among governments that AI represents the next essential infrastructure, comparable to electricity or the internet. Midha urges consideration of “infrastructure independence.” He questions whether nations wish to depend on Chinese models, potentially subject to censorship and data access concerns.

He advocates for Western nations to utilize Western-developed models, such as Mistral, based in Paris. Numerous companies share this concern and have already restricted access to DeepSeek, which functions as both a consumer application and an open-source model.

However, not all parties share the apprehension regarding Chinese open-source models. These models can be deployed locally within a company’s own data centers. Furthermore, DeepSeek is already offered as a secure cloud service by American providers like Microsoft Azure Foundry, eliminating the need to utilize DeepSeek’s native cloud platform.

Notably, Pat Gelsinger, the former CEO of Intel, with extensive experience in China, revealed that his startup, Gloo, is developing AI chat services based on their customized version of DeepSeek R1, rather than alternatives like Llama or OpenAI’s offerings.

Should any organization reconsider their data center plans in response to DeepSeek, Midha jokingly requests: “If you have surplus GPUs, please direct them to Anj.”

Stay informed with TechCrunch’s AI newsletter! Subscribe here to receive it weekly in your inbox on Wednesdays.