In a recent explosive segment of The All-In Podcast, industry heavyweights unpacked the latest twists in the global AI arms race. Titled “DeepSeek Panic: What’s Real, Training Innovation, China, Impact on Markets and the AI Industry,” the discussion—featuring insights from Travis Kalanick (Uber & CloudKitchens Co-Founder) and David Sachs (White House AI & Crypto Czar)—shed new light on a Chinese startup’s disruptive approach, the shifting balance between open source and proprietary models, and the broader geopolitical stakes in the burgeoning world of artificial intelligence.
At the heart of the conversation was DeepSeek, a Chinese AI startup that has suddenly thrust itself into the spotlight with its new language model, R1. According to DeepSeek’s claims, R1 rivals the performance of state-of-the-art Western models, yet it was trained on only 2,000 GPUs for a final run costing roughly $6 million. In stark contrast, industry giants reportedly spent tens of millions—and are planning to invest billions—on their next-generation models. This staggering disparity in claimed training costs not only caught the attention of market watchers but also ignited debate over whether these numbers are being misrepresented by comparing apples to oranges.
Central to the controversy is the process known as “distillation,” where a large, proprietary model feeds its knowledge into a smaller, more accessible one. Critics have suggested that DeepSeek’s R1 might be leveraging outputs from models like OpenAI’s “o1” through distillation—a process that blurs the line between genuine innovation and replication. As the conversation unfolded, David Sachs noted that while DeepSeek’s open-source release is surprising in a market dominated by closed platforms, the move fits within a broader trend of challenging established players in the AI space.
The podcast’s discussion quickly moved beyond cost comparisons to highlight the ingenious methods DeepSeek appears to be employing. While Western AI companies often rely on established frameworks like CUDA for GPU programming, DeepSeek reportedly bypassed these norms by working directly with low-level PTX—akin to writing in assembly language—to maximize performance. Such technical agility allowed the startup to navigate the constraints imposed by export controls on Nvidia’s latest GPUs and, in doing so, to innovate under financial and hardware limitations.
Yet, as several panelists pointed out, comparing the “final training run” cost of DeepSeek’s R1 to the fully loaded research and development expenditure of companies like OpenAI is misleading. When one factors in the cost of the extensive compute clusters, including tens of thousands of GPUs acquired over time, the narrative of a “scrappy” company training a world-class model on a shoestring budget loses some of its luster. Nonetheless, the methods and efficiency gains demonstrated by DeepSeek underscore a broader point: necessity is spurring innovation that could reshape how AI models are built and deployed.
One of the most heated debates during the segment centered on the merits—and pitfalls—of open-sourcing AI models. DeepSeek’s decision to make its R1 model available to the public stands in stark contrast to the closed, proprietary approach favored by companies like OpenAI. For many in the community, the open source move is a breath of fresh air that could democratize AI development and spark a wave of downstream innovation. Critics, however, argue that this openness might also undercut the competitive advantages held by U.S. companies, as well as re-sparking accusations that some Western firms might have infringed in intellectual property rights while training their models.
No discussion on cutting-edge AI would be complete without addressing the geopolitical ramifications. DeepSeek’s breakthrough has forced many to reassess how far behind—or ahead—China truly is in the AI race. Just a few months ago, many experts believed that China lagged behind the West by six to twelve months. Now, with R1 emerging as a viable competitor to Western models, that timeline may be halved.
The impact of these developments rippled through the financial markets as well. On the very day DeepSeek captured headlines, Nvidia endured one of its worst trading sessions on record. Concurrently, speculation arose that a surge of advanced chips bound for Asia—potentially routed through clandestine channels in Singapore—might be fueling China's AI ambitions despite the U.S. ban on exporting state-of-the-art GPUs. This situation underscores not only the intricate interdependencies of global supply chains but also the delicate balance between driving innovation and protecting national security.
Panelists noted that the ongoing U.S.-China rivalry in AI is as much about market share as it is about strategic dominance. With giants like OpenAI raising astronomical sums (rumored to be around $40 billion at valuations of over $340 billion) and cloud providers like Microsoft caught in the middle, the stage is set for an all-out technological and economic confrontation.
Bringing a unique perspective to the table, Travis Kalanick—one of the few American entrepreneurs who has successfully scaled a business in China—shared vivid recollections of his days at Uber China. According to Kalanick, the frenetic pace at which Chinese companies can copy and then innovate on new ideas is unmatched. He recounted how local teams could replicate a service in weeks, then swiftly improve upon it, leading to a cycle of relentless innovation that keeps pushing the boundaries.
The lesson is clear: constraints drive creativity. In markets where resources are limited or tightly regulated, the pressure to innovate sparks breakthrough solutions. This dynamic is not only propelling AI development but also reshaping how real value is captured within the tech ecosystem. As AI models become increasingly commoditized, genuine monetization opportunities may well shift to the application layer—where user experience and customized services create lasting differentiation. Yet, pressing questions remain: How will the foundational layer be structured? Will it consist of multiple specialized models that interchange based on the task, or will a single model dominate? And as AI becomes embedded into our daily infrastructure—serving as a pervasive intermediary layer between users and the myriad applications they rely on—the pressing question remains: who will ultimately control the vast amounts of data processed by this ubiquitous AI?
The All-In Podcast segment encapsulates this pivotal moment in AI evolution. DeepSeek’s bold, open-source approach challenges the entrenched, closed systems of traditional players while redefining cost structures and innovation pathways. The convergence of technical ingenuity, market forces, and geopolitical strategy hints that the next wave of AI advancements may emerge from the most unexpected quarters.
