DeepSeek, the viral AI sensation, has introduced a new family of multimodal AI models called Janus-Pro, claiming they surpass OpenAI’s DALL-E 3 in performance.
The Janus-Pro models, available on the AI development platform Hugging Face, range from 1 billion to 7 billion parameters—smaller yet powerful. Parameters are a key measure of a model’s problem-solving capacity, with larger parameter counts often indicating better performance. Despite their compact size, Janus-Pro models reportedly outperform not only DALL-E 3 but also rivals like PixArt-alpha, Emu3-Gen, and Stability AI’s Stable Diffusion XL in evaluations such as GenEval and DPG-Bench.
Operating under an MIT license, Janus-Pro is free for commercial use. Its “novel autoregressive framework” allows for both image analysis and creation. However, the models currently support a maximum resolution of 384 x 384 pixels. DeepSeek describes the series as a game-changer for unified multimodal models, balancing simplicity, flexibility, and effectiveness.
The DeepSeek Revolution: A Threat to the U.S. AI Leadership?
DeepSeek, backed by Chinese investment from High-Flyer Capital Management, made headlines after its chatbot app soared to the top of the Apple App Store charts. Its innovative language models, developed with cost-efficient training techniques, have raised concerns among analysts about the U.S.’s ability to maintain its AI lead.
The debut of DeepSeek-R1 trained on NVIDIA’s H800 GPUs with a modest $6 million budget, has drawn significant attention. Analysts question whether AI chip demand will be sustained as efficient training reduces the cost of creating advanced models.
NVIDIA’s Tumultuous Ride with DeepSeek
Despite DeepSeek’s reliance on NVIDIA’s GPUs, its latest moves have shaken the market. On January 22, NVIDIA’s stock plummeted by 17%, erasing $600 billion in market value—the most significant single-day loss in Wall Street history. This downturn was linked to DeepSeek’s success, even though NVIDIA GPUs remain integral to DeepSeek’s operations.
DeepSeek’s recent pivot to Huawei’s 910C chips for inference has added complexity to the situation. While NVIDIA reiterated its dominance in supplying GPUs to China, it faces rising competition from Huawei’s Ascend 910C processor, touted as a rival to NVIDIA’s H100 chip.
The Jevons Paradox of AI Compute
The demand for computing power paradoxically grows as AI models become more efficient. This phenomenon, known as the Jevons Paradox, suggests that improved efficiency often leads to greater consumption. Microsoft CEO Satya Nadella echoed this sentiment, predicting that as AI becomes more accessible, the demand for GPUs like NVIDIA’s will only expand.
Meanwhile, critics argue that smaller language models and mobile-friendly solutions will soon reduce reliance on GPUs. Companies like Meta are already ramping up AI infrastructure investments, with plans for $65 billion in 2025 capital expenditure. NVIDIA has also unveiled Project DIGITS, a $3,000 supercomputer aimed at researchers and students, signaling its intent to dominate the local AI market.
NVIDIA’s Complicated Relationship with China
China remains a critical market for NVIDIA despite U.S. restrictions on AI chip exports. To maintain its foothold, the company has created export-compliant versions of its GPUs, such as the H20 and L20. However, investigations have revealed smuggling networks and the unauthorized use of NVIDIA chips in Chinese servers, prompting scrutiny from U.S. regulators.
Huawei’s domestic processors, like the Ascend 910C, are poised to challenge NVIDIA’s dominance. DeepSeek, having trained its models on NVIDIA hardware, is now leveraging Huawei chips for inference, further fueling the narrative of a shifting global AI landscape.
The Bottom Line
DeepSeek’s rise underscores the rapidly changing AI industry, where cost efficiency, open-source development, and geopolitical tensions are reshaping the competitive landscape. While NVIDIA remains a key player, challengers like DeepSeek and Huawei are redefining the rules, ensuring the AI race is far from over.
