Anjney “Anj” Midha, a general partner at Andreessen Horowitz and a board member at Mistral, first took notice of DeepSeek’s impressive capabilities six months ago, as he shared with TechCrunch.
It was during this time that DeepSeek launched Coder V2, which competed head-to-head with OpenAI’s GPT-4 Turbo for coding-related tasks, according to a study published last year. This development set DeepSeek on a trajectory to roll out enhanced models every few months, culminating in R1, their recently unveiled open-source reasoning model that is disrupting the tech landscape by delivering industry-standard performance at a significantly lower cost.
Despite the slump in Nvidia’s stock, Midha asserts that the advent of R1 doesn’t signify a halt in the immense investment into AI foundational models which continue to clamor for GPU resources and aspire to rapidly expand their data center infrastructures.
Instead, it indicates that they will maximize the computational capabilities at their disposal.
“When people question, ‘Anj, does Mistral’s billion-dollar fundraising render DeepSeek obsolete?’ I respond, ‘No, it’s actually incredibly advantageous for them to examine DeepSeek’s efficiency gains, adapt them, and then invest significantly,’” he stated.
He elaborated, “Now we can achieve ten times the output using the same computational resources.”
That doesn’t imply Mistral is lagging behind its competitors like OpenAI and Anthropic, who have amassed much greater sums of investment. Reports suggest that OpenAI is currently negotiating to secure an astonishing $40 billion.
Mistral remains a formidable contender due to its open-source model, according to Midha. This argument holds water; open-source projects can leverage essentially free technical contributions from users invested in the project. In contrast, closed-source competitors shield their proprietary information and must pay for both labor and computational resources.
“You don’t need $20 billion, just more computational power than any other open-source model can offer. Mistral is well-positioned in this aspect, boasting the highest computational abilities among open-source providers,” Midha remarked regarding his portfolio company.
Facebook’s Llama, the largest open-source AI model in the West pitted against Mistral, is also set to attract considerable investment. CEO Mark Zuckerberg announced plans to allocate “hundreds of billions of dollars” towards AI, including a substantial $60 billion in capital expenditures focused on data centers in 2025.
a16z’s Oxygen GPU sharing initiative is “overbooked”
Midha, who also serves on the boards of AI imaging startup Black Forest Labs and 3D model creator Luma (and is an angel investor in AI companies like Anthropic and ElevenLabs), has another reason to believe that the demand for GPUs in AI will not diminish anytime soon.
As the head of a16z’s Oxygen program, Midha explained that GPUs, especially Nvidia’s cutting-edge H100s, have turned into a scarce resource. To address this, the venture capital firm took proactive measures about a year and a half ago by acquiring numerous GPUs for its portfolio companies.
“Oxygen is currently ‘overbooked’. I simply can’t allocate enough resources,” Midha chuckled. It’s not just that his startups need GPUs to train AI models; they require even more to maintain their operational AI services for clients.
“At this point, there is an unquenchable demand for inference capabilities,” he clarified.
He also suggested that DeepSeek’s advancements won’t impact Stargate, the recent $500 billion partnership between OpenAI, SoftBank, and Oracle intended for AI data centers.
A significant outcome of DeepSeek’s innovations is the growing acknowledgment by nations that AI represents the next critical infrastructure, akin to electricity and the internet. Midha advocates for what he describes as “infrastructure independence.” Should nations depend on Chinese models riddled with censorship and potential data risks, or should they opt for Western models that adhere to Western laws and moral standards and comply with NATO agreements?
Clearly, he is championing the use of Western models, such as his Paris-headquartered Mistral. Numerous companies share this sentiment and have taken steps to restrict access to DeepSeek, both as a customer service and an open-source model.
However, not everyone shares the apprehension regarding Chinese open-source models. Organizations can implement these locally within their own data centers. DeepSeek is also offered as a secure cloud service through American companies like Microsoft Azure Foundry, alleviating the need to utilize DeepSeek’s cloud solution directly.
In fact, Pat Gelsinger, the former CEO of Intel — who is well aware of the situation in China — informed TechCrunch that his startup Gloo is developing AI chat solutions based on their own iteration of DeepSeek R1, opting against alternatives like Llama or OpenAI.
Should any company be inclined to abandon their data center plans due to DeepSeek’s emergence, Midha humorously puts forth a request: “If you have surplus GPUs, kindly send them my way.”
Don’t miss out! TechCrunch has a newsletter focused on AI. Sign up here to have it delivered to your inbox every Wednesday.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
