I’m Allyson Klein from the TechArena, and I’m delighted to join the Cosmos community to share view and discuss what’s the latest in AI with the community. Here are my most recent thoughts on the DeepSeek news from last week and what it means to AI model development.
And just like that, the world of artificial intelligence (AI) model training was thrown into turmoil. In case you were vacationing or otherwise occupied late last week, China introduced DeepSeek, a free-to-the-public generative AI model that is outperforming Chat GPT.
While the performance is stunning, what’s most notable about DeepSeek is how this elegant solution was purportedly trained with a fraction of the resources required by leading models today. And this news emerged literally days after OpenAI announced a $500 billion investment with Microsoft, Arm, NVIDIA and others to drive U.S. AI superiority and the U.S. delivered the first meaningful threat to China’s TikTok access to U.S. markets.
Chatter across social networks dubbed the makers of DeepSeek geniuses, heralding an end to U.S. AI leadership and calling NVIDIA stock deeply overvalued. To unpack what we know and what we suspect as the dust settles on this massive shock to the AI world, we need to dig deeper.
DeepSeek appears to have taken some innovative approaches to training its algorithm, including utilizing less precise math (eight decimal places vs 32) and processing larger groups of words, driving down precision. But it’s also created more efficient multi-token ingest and allocated training across multiple experts like a group of smaller, smart models working in tandem. All of these examples are innovations that will likely get attention from across the AI developer community and come at a time when developers were seeking disruptive approaches to delivering AI models more efficiently.
What is raising questions among many experts whom I chatted with over the weekend is the full transparency in the cost of training the model. One pegged the true cost at 1.5-to-2 orders of magnitude higher than what the startup has stated, with disclosed costs solely focused on knowledge distillation and fine-tuning of the algorithm. They point to the fact that this version of the model has had the benefit of training of previous iterations, similar to the investment alternatives like ChatGPT, Llama, Gemini and others have based on iteration of versions of models over time. The truth is likely that the cost of the model we’re looking at today is much higher than the $5 million that promoters are claiming. Yet, this should not discount the value of a competitive model and the overall performance it’s delivering.
Of course, the timing of DeepSeek’s release does give credence to this lack of transparency being a shot across the bow of the Stargate announcement, continued tensions on TikTok restriction from the U.S. market and U.S.’s focus on AI as a central policy imperative. And the well-timed emergence of DeepSeek to access U.S. data sources can’t be overlooked, given that China has historically used TikTok to collect data from Americans and this new generative AI model provides a much more powerful way to collect that data and deliver potential misinformation through content.
While the coming weeks will provide more clarity on the full truth of this model’s efficiency, one thing is clear: DeepSeek has absolutely captured the community’s attention, seizing leadership on AppStore downloads vs Chat GPT.
So, what’s my take?
Isee the arrival of DeepSeek as a reminder that while the U.S. hyperscalers grab the headlines for AI model advancement, their Chinese counterparts have been dedicated to developing their own AI solutions for years. Bytedance alone has committed $20 billion in AI investment in 2025 ($12 billion earmarked for U.S. spending), and others such as Alibaba, Baidu, and Tencent have similar large-scale operations established. In this high-stakes realm of the race for LLM superiority, I expect to see both dizzying announcements of innovation and overstated differentiation at both a corporate and geo-political scale. And while I applaud any advancement to drive efficiency into AI training and are seeking further clarity of the full veracity of what’s been delivered with DeepSeek, I’m not yet ready to call an end to the Blackwell era before it’s really even begun.
We also see the public’s zeal for utilization of all of these LLM models igniting without a lot of thought about what happens to the data provided to the model owner, whether that be an enormous tech conglomerate in the U.S. or a Chinese startup with ties to the government. In this world of rapid adoption of LLMs, we wonder what defines truth in the future and will there be multiple definitions of truth depending on who controls the algorithm.
Share your thoughts!
