On December 25th, 2024, Deepseek released Deepseek-V3, a new large language model (LLM) that it claimed to be the strongest open-weights model to date. Accompanying this release was a report detailing Deepseek’s innovative training process. This was the latest in a series of Deepseek models, following the launch of Deepseek-LLM in November 2023.
Shortly after, on January 20th, 2025, Deepseek introduced Deepseek-R1, a reasoning-focused model built on the Deepseek-V3 base. It was designed to rival OpenAI’s o1 series of reasoning models, again claiming to be the strongest open-weights model ever released.
A relatively normal week as far as things go in the world of open-source LLMs.
Within days, the Deepseek Assistant app topped the US Apple Store charts, $1 trillion in market value was wiped from the stock market, and AI conversations exploded across social and professional circles. But what’s the big deal?
What is the big deal
The Models
Deepseek-V3 is a general-purpose, instruction-tuned model designed to handle a wide range of queries—similar to the models behind ChatGPT. Deepseek-R1, on the other hand, is a specialized reasoning model akin to OpenAI’s o1 series. These models excel at tasks requiring structured thinking, such as math, coding, and visual reasoning. They leverage techniques like multi-step problem-solving, multiple approach testing, and self-correction.
Performance Comparison
So how do the Deepseek models stack up against OpenAI and other leading providers? Standard benchmarks provide a useful comparison:
Artificial Analysis: A site tracking AI performance shows that Deepseek-R1 matches OpenAI’s o1 model across various tasks—despite being 100x cheaper ($0.14 per 1M input tokens vs. OpenAI’s $15 per 1M input tokens).
LMSys Arena: This ranking, based on large-scale human evaluations, places Deepseek-R1 in 4th place overall, trailing only OpenAI’s GPT-4o and two Google Gemini models. Deepseek-V3 also makes the top 10—the only open-source models to do so.
The takeaway? Deepseek models are extremely capable, dramatically cheaper, and competitive with the best proprietary models. But their technical prowess was only part of the story.
The Crash
A week after Deepseek-R1’s release, investors started scrutinizing Deepseek’s report more closely. One line in particular set off alarm bells:
"Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training."
H800 GPUs are a GPU designed by NVIDIA to be exported to China, and are similar to state of the art data centre GPUs like H100s. Using an estimated cost of $2 per hour per GPU, this translates to a total training cost of approximately $5 million—a fraction of the $100 million OpenAI reportedly spent on GPT-4 and significantly less than Meta’s 30M H100 GPU hours for Llama 3.2 405B.
This revelation sent shockwaves through the market. Investors had long assumed that training ever-more-powerful AI models would require ever-larger investments in GPUs, fueling massive demand for NVIDIA hardware. But Deepseek’s efficiency raised a troubling question: What if you don’t need tens of thousands of GPUs to train a state-of-the-art model anymore?
The market reaction was swift:
- NVIDIA’s stock plunged 20%, wiping out hundreds of billions in market value.
- The NASDAQ dropped nearly 4%, reflecting broader concerns about AI infrastructure spending.
With companies like xAI recently purchasing a 100,000-GPU cluster and Meta planning nuclear power plants to support AI growth, the idea that top-tier models could be trained more efficiently threatened to upend long-held assumptions about AI economics.
What Comes Next? The Future of AI and the Market
Stock Market Implications
Does this signal the beginning of NVIDIA’s decline, or just a temporary adjustment before continued growth? The answer depends on how close we are to the LLM performance ceiling.
- Scenario 1: Close to the Ceiling If AI models are approaching their fundamental limits, then reducing training costs means fewer GPUs will be needed. Large-scale GPU purchases might decline, and NVIDIA’s growth could slow.
- Scenario 2: Far from the Ceiling If major AI breakthroughs are still ahead, cheaper training could accelerate progress—leading to even more demand for GPUs. More companies may now be able to afford to train state-of-the-art models, driving broader adoption and usage. The potential to produce more capabilities with the same GPUs, and the expectation of many more capabilities to come, means purchasing as many GPUs as possible is still the right strategy. If similar economics to the DeepSeek team can be achieved, the return on investment for your GPU cluster will be higher than previously expected.
At TitanML, we believe there is still much more to unlock in AI. The ability to train powerful models more efficiently should fuel further advancements, not slow them down.
The Future of Open-Source LLMs
Deepseek-V3’s reported $5 million budget is a game-changer for open-source AI. It puts state-of-the-art model training within reach of well-funded research teams, open-source projects, and even enterprises looking to develop specialized models.
We expect:
- More open-weight models matching Deepseek’s performance in the coming months.
- Faster innovation in training methodologies and inference efficiency.
- Greater accessibility of powerful AI tools to businesses outside of Big Tech.
But why wait? With TitanML, you can easily deploy any open-source model in your own environment. For instance, if you have an idle 8xH200 setup, you can deploy the Deepseek model today (or 8xH100 in 4-bit mode on the TitanML platform). If you're working with smaller hardware, consider one of the distilled model versions or explore the thousands of models available on Hugging Face.
Whether you have on-premise clusters or GPUs from cloud providers like AWS and GCP, TitanML offers a centrally managed, customizable AI Model Hub that integrates seamlessly into your environment.
So, whether you're looking to add Deepseek, LLaMA models, or the latest open-source innovations, start your journey today by connecting with the TitanML team!