Founded in May 2023, DeepSeek was little more than a footnote in the annals of AI history—then it released DeepSeek R1 on January 20, 2025, and suddenly, everyone sat up like they'd just found out their house was on fire.
The reason? DeepSeek R1 didn’t just match OpenAI’s flagship o1 model; it trounced it across nearly every major LLM benchmark. Have a look at the results for yourself:
And DeepSeek R1 did so while charging users prices a fraction of OpenAI’s prices. DeepSeek’s pricing: $0.55 per million input tokens, $2.19 per million output tokens. OpenAI’s o1? $15 per million input tokens, $60 per million output tokens—roughly 96% more expensive, or to put it another way, OpenAI is charging first-class airfare while DeepSeek is offering budget flight prices with first-class features.
And then came the real kicker. DeepSeek claimed they trained R1 for a mere $12 million. For added fun, they revealed their other model, DeepSeek V3 (a GPT-4o competitor), cost just $6 million to train. By comparison, the Wall Street Journal reported that OpenAI is burning through $500 million every six months to train its upcoming GPT-5, codenamed “Orion.”
The tech industry promptly went into speculative overdrive. Was this the beginning of the end for exorbitant GPU spending? Had DeepSeek cracked the code to high-performance AI at bargain prices? Should Nvidia be nervous? Maybe. But before anyone starts shorting GPU stocks, a little skepticism is in order.
DeepSeek: Too Good to Be True?
While some have hailed DeepSeek as the AI messiah, others—particularly those who have been building LLMs —have raised an eyebrow, or in some cases, both. The skepticism boils down to one question: Can DeepSeek’s claims really be trusted?
A few notable doubters include:
Elon Musk finds it highly unlikely that DeepSeek trained its models with just 10,000 Nvidia A100 GPUs, suspecting the real number is significantly higher.
Alexander Wang, CEO of Scale AI, speculates that DeepSeek somehow got its hands on 50,000 Nvidia H100 GPUs, despite U.S. export restrictions making that technically impossible.
SemiAnalysis, a research firm, clarifies that DeepSeek’s 10,000 A100 GPUs were acquired before restrictions in 2021 and claims the company has since spent at least $500 million more on hardware. They estimate DeepSeek actually has:
10,000 H800s (China-friendly versions of H100s with lower network bandwidth)
10,000 H100s
Pending orders for many more H20s (another China-specific Nvidia chip, of which over 1 million have been produced in the last 9 months)
DeepSeek’s total server CAPEX is closer to $1.3 billion, with operating costs at $715 million. Oh, and that $6 million training cost for DeepSeek V3? That, they say, was just the pre-training bill—final expenses were likely much higher.
That said, even with these revised estimates, DeepSeek still built DeepSeek R1 at a fraction of the cost compared to OpenAI, Google, and Anthropic. Which, no matter how you slice it, is a remarkable achievement.
Now, the real test begins: Can DeepSeek R1 actually rival OpenAI’s o1 in reasoning and comprehension? Or was this just a masterclass in AI hype?
DeepSeek R1 Strong on Paper but Loses to o1 in Logic Puzzle
To test DeepSeek R1’s actual reasoning ability, we posed a logic puzzle that OpenAI’s o1 (and Claude 3 Opus) had previously cracked with ease:
Q: A cable of 80 meters (m) is hanging from the top of two poles that are both 50 m from the ground. What is the distance between the two poles, to one decimal place, if the center of the cable is 10m above the ground?
A: 0 meters. The poles are standing side by side.
Chat GPT o1 Answer
DeepSeek R1 Answer
DeepSeek R1, unfortunately, did not arrive at this answer.
This highlights an important reality: while DeepSeek R1 shines on paper—outscoring o1 in benchmark tests—real-world reasoning is another matter. If a model is optimized primarily to ace benchmarks, that doesn’t necessarily mean it excels at logic, chain-of-thought reasoning, or common sense.
That said, DeepSeek has proved one thing decisively: a high-performing LLM can be built at a much lower cost than previously thought. Maybe not for the miraculous $6 million claimed for DeepSeek V3, but even SemiAnalysis’ estimate of $1.3 billion in CAPEX is still a bargain compared to the multi-billion-dollar investments of OpenAI, Meta, Google, and Anthropic. The pressure is now on AI incumbents to prove that their relentless spending on GPU clusters will yield performance gains substantial enough to justify the cost—especially in light of DeepSeek’s far more economical approach.
Comments