Just a few weeks ago, we took a comprehensive look at the cutting-edge large language models (LLMs) that were dominating the AI space: Google’s Gemini series, OpenAI’s o1, Anthropic’s Claude, and a handful of promising open-source contenders like Mistral. Since then, one name in particular has roared into the global conversation: DeepSeek.
By now, you’ve likely heard about DeepSeek, the name at the center of the recent AI stock crash dominating the headlines on January 27th—even though news of the company had already circulated widely for a week. Over that same period, Artificial Analysis Leaderboards have recorded a noteworthy surge in performance across multiple domains—from coding to scientific reasoning to multilingual capabilities. Yet the numbers only scratch the surface of what’s happening.
The bigger story is that DeepSeek is quickly emerging as the “Robin Hood” of AI, offering near top-tier performance at a fraction of the usual cost—and making it open source to boot. It’s a development that’s challenging established narratives and leaving both investors and industry titans wondering if their spending spree will be rewarded.
Even more eye-opening? DeepSeek has quickly become the talk of Silicon Valley, with many calling it a potential Android moment for AI.
Updated Benchmark Highlights
Thanks to fresh data from artificialanalysis.ai and other sources, we now have a clearer sense of where each major LLM stands. The chart below reflects one of the multiple performance categories:
Scientific Reasoning & Knowledge (GPQA Diamond)
Quantitative Reasoning (MATH-500)
Coding (HumanEval)
Multilingual Performance (Artificial Analysis Multilingual Index
Deepseek R-1 is ranking amongst top three in most categories, and tops even some of them such as Quantitative Reasoning.
Clearly Deepseek is a new AI Titan to reckon with, or maybe more appropriately, the Robin Hood of AI.
Indeed, whilst most Titans are racing to deploy hundreds of billion to scale compute capabilities to deliver nextgen LLMs, Deepseek allegedly trained R1 on the incredibly low budget of 5.6M$!
Why Everyone in AI is Freaking Out About DeepSeek
Seven days ago, on January 20th—immediately after his inauguration as President of the United States—Donald Trump announced Project Stargate, a 500-billion-dollar initiative to build the compute infrastructure needed to train and host OpenAI’s large language models (LLMs). This bold move was touted as a way to reaffirm the United States’ dominance in Artificial Intelligence—“AI Made in the US!”
Just two days later, on January 22nd, Deepseek (dubbed the “Chinese Robinhood”) unveiled R1 in direct response to Stargate. In the span of a single week, this new contender casted doubts on the previously unquestioned metric that has defined AI competition so far: compute capacity.
The current narrative assumes that larger investments in compute yield superior, proprietary LLMs—such as OpenAI’s “O1.” However, R1 appears to challenge that assumption by achieving impressive capabilities with far fewer resources. As a result, organizations may start to question whether they should pay a premium for OpenAI API tokens when they could simply deploy Deepseek R1 on their own GPU cloud—public or private.
The Legend of Robinhood
Could it be that, like all good bandits, Deepseek’s “Robinhood” pulled off a daring heist of banned H100 GPUs to train R1? Esteemed insiders Dylan Patel from SemiAnalysis and Scale AI CEO Alexandr Wang certainly seem to suggest so, estimating a massive, billion-dollar-scale GPU deployment. Deepseek, meanwhile, insists it spent a modest $5.6 million on its model—conveniently matching the narrative that no smuggled U.S.-restricted hardware ever made it into their racks. But if their cluster is anywhere near the scale of America’s “Titans,” these claims may be more myth than reality. After all, running models on that level typically comes with a much bigger price tag than a few million dollars—unless, of course, you’re the legendary Robinhood, who knows a thing or two about acquiring resources by unconventional means.
A third possibility is that the company could have trained a larger model on a hidden H100 cluster, then used model distillation to create a smaller version for deployment on the H800 system.
Meanwhile, Meta’s Chief AI Scientist, Yann LeCun, offers another perspective: that these models may have been developed so inexpensively thanks to open-source innovations—a principle once central to OpenAI’s mission but seemingly abandoned in recent years.
Alternative Explanation: The Role of Distillation
A plausible reason behind DeepSeek’s rapid performance gains could be model distillation—a technique for extracting knowledge from a larger “teacher” model (like GPT-4) to train a smaller, more efficient “student” model. Typically, distillation is straightforward if you own both teacher and student, but it’s still feasible to do this indirectly (and more awkwardly) via APIs or even chat interfaces. While distillation can violate the terms of service of leading-edge models—and providers might try to stop it with IP bans or rate limits—it’s widely believed to be an under-the-radar practice. This could be why so many new models are converging on GPT-4-level quality.
If DeepSeek did take this route (and it wouldn’t be surprising if they did), it sheds light on how they might offer near state-of-the-art results without the enormous training budgets of American tech giants. Ironically, this situation hurts cutting-edge model creators like OpenAI, Anthropic, and Google, who bear the brunt of R&D costs that others effectively piggyback on. It may also explain why Microsoft is increasingly cautious about investing tens of billions into data centers for next-generation models that might become commoditized faster than their infrastructure can be depreciated.
Even in this scenario, the resemblance with robin hood figure's becomes even more striking
Robin Hood steals from the rich and gives to the poor
Deepseek metaphorically “stole from the rich” as AI stocks in the United States tumbled dramatically today, with chipmaker NVIDIA alone shedding nearly $600 billion in market value. Continuing the Robin Hood allegory, one might wonder: who are the “poor” in this scenario?
It’s the end users who stand to benefit. With competitive, open-source models available at a fraction of the cost of premium American solutions, subscription prices could fall substantially, making advanced AI more accessible to a broader audience. Meanwhile, GPU cloud providers win as well, offering the crucial infrastructure needed to run these open-source alternatives.
By providing affordable, high-performance open-source models, Deepseek and similar efforts are helping democratize AI, loosening the grip of the few hyper-scale companies that have traditionally dominated the market.
Hyperscaler Slowdown and the Rise of Independent Operators
We’re likely to see a slowdown in large-scale capacity deployment by major hyperscalers, as they become more cautious about their forecasts for cluster growth. This shift will offer independent operators—which can often deliver infrastructure at a fraction of the cost and the time—an opportunity to step in and fill the gap, fostered by the increased demand from indipendent GPU Clouds and LLM API Providers. The result? A more distributed and stable industry landscape, fueled by fresh competition and innovation. We believe this shift will foster enterprise adoption particularly in the training/fine-tuning hypo-scale dimension (training with 5-10k GPUs) and straight on inference. We believe that his event will favour the drift of NVIDIA away from hyperscalers' demand - as they will slow down deployments. The Neoclouds (or GPU Clouds) will be the major beneficiaries from this shift, proving that NVIDIA diversification strategy in supporting these NEO Clouds was a strategic play.
There will be though reluctance from enterprise clients to adopt made-in-china LLMs in NATO Countries.
Therefore, for this opportunity to fully materialize for indipendent Neoclouds and AI Labs (all NATO based), we will need to see first NATO-based Open-source players like Meta and Mistral to be able to leverage DeepSeek unquestionable breakthroughs such as Reinforcement Learning. We actually believe that Mistral and others alike are the secret winners of this change as they will be able to adopt Deepseek open weights for the model (first time model weights have been made publicly available) and RL technology (see Hugging face community already working on this).
Most western enterprise clients will prefer adopting anyway a reasoning model from Mistral or Meta rather from Deepseek, so we will likely see significant adoption of these western copies of Deepseek breakthrough (who would have ever thought West copying China!). If that happens, we expect to see a surge in orders from enterprise clients, where the main beneficiaries will indeed be indi AI Labs, Neo Clouds, Indi Co-Locators and surprisingly for the stock market: NVIDIA.
Even though already mentioned, it's important to stress out that all the major hyperscalers have plans to reduce reliance on NVIDIA by developing their own AI chips. Google already did, Amazon is trying with limited success, Microsoft and Meta have a plans to do it. So in the long run it's essential for NVIDIA that its NeoCloud children grow their sales and market share to contrast the reduction in orders from hyperscalers building their own AI-chip value chain.
China’s AI Renaissance: Beyond DeepSeek
DeepSeek is merely one shining example of a larger Chinese AI surge across several sectors—from text-to-image and image-to-video, to text-to-video. Until recently, American giants maintained their dominance in text generation. However, they’ve struggled to roll out “Sora,” promised since March 2024 but released only recently to few users. During this gap, smaller Chinese contenders like Flux1 have filled the market need throughout most of 2024, and they continue to hold a commanding position.
Conclusion: A Taste of What to Expect in 2025?
As we head into 2025, the global AI landscape appears more dynamic—and more uncertain—than ever. The once-unquestioned link between massive compute investment and superior model performance seems to have been challenged by Chinese players like Deepseek and Flux1. Meanwhile, in the United States, Project Stargate is poised to supercharge infrastructure investment, yet competition from open-source and independent operators may undercut the dominance of established giants.
With hyperscalers showing signs of caution in their expansion plans, there’s growing room for smaller, agile players to enter the market. The result is likely to be a more distributed AI ecosystem—one in which cost-competitive, open-source models drive down prices and give end users more choices.
In essence, 2025 is set to be a year of recalibration: potentially a shift from monolithic, capital-intensive AI models toward a more inclusive environment that rewards efficiency, collaboration, and innovation. The winners of tomorrow will be those who can adapt swiftly to changing compute economics and user demands—whether they’re giant, well-funded hyperscalers or lean, independent upstarts.
コメント