The recent emergence of DeepSeek has stirred quite a debate in the investment world, drawing attention to its next-generation open-source models that boast impressive low costs coupled with high performanceSome reports have even suggested that DeepSeek could replicate OpenAI's capabilities for a mere $5 million, creating a wave of concern within the AI infrastructure industrySuch statements imply a potential 'doomsday' scenario for established players in the market.

However, prestigious Wall Street investment bank Bernstein, after a thorough examination of DeepSeek's technical documents, has released a report asserting that the prevailing market fear appears to be exaggeratedThey clarify that the claim of "replicating OpenAI for $5 million" does not accurately reflect the broader pictureIn fact, Bernstein argues that while DeepSeek does display significant enhancements in efficiency, these advancements are not astonishing breakthroughs when viewed through a technical lens.

One of the key points raised by Bernstein is that even if DeepSeek has achieved a tenfold increase in efficiency, this merely correlates with the annual cost escalation associated with existing AI models, rather than signaling a radical shift in the landscape

Advertisements

The investment bank emphasizes that the demand for AI computing has not yet hit its ceiling, thus indicating a sustained growth trajectory in the sector.

Delving into the specifics of the "replicating OpenAI for $5 million" narrative, Bernstein provides critical clarificationThey refer to this as a narrow interpretation of the costs associated with training the DeepSeek V3 model, which simply equates GPU rental costs with total investmentThe estimation of $5 million is based solely on the rental price of $2 per GPU hour and overlooks the extensive research and development expenses, data costs, and miscellaneous charges involved in launching such a sophisticated model.

Bernstein further elaborates on the technical prowess of DeepSeek, particularly shedding light on its innovative V3 and R1 modelsThe V3 model is recognized for its revolutionary efficiency, achieving performance levels comparable to mainstream large models using a staggering 2048 NVIDIA H800 GPUs and approximately 2.7 million GPU hours

Advertisements

The architecture of the V3 model incorporates a mixture of experts framework, deliberately designed to reduce both training and operational costsIn addition, it employs multi-head latent attention techniques that significantly lower memory usage and cache size.

Moreover, the utilization of FP8 mixed precision training enhances performance optimization even furtherThe amalgamation of these technologies means that V3 requires just 9% of the computational power of corresponding open-source models to achieve similar, if not superior, outputFor instance, the pre-training phase of V3 necessitates about 2.7 million GPU hours, contrasting starkly with the approximately 30 million GPU hours required by models like LLaMA of the same scale.

While discussing the efficiency of V3, Bernstein posits that the improvements it brings to the table, when contrasted with the typical 3 to 7-fold efficiency boosts seen in the industry, do not represent a seismic shift

Advertisements

The crux of the MoE architecture is its capacity to significantly curtail training and operational costs, operating with only a subset of effective parameters during any given training epochThis approach contrasts with dense models where all parameters are updated continuously.

Comparative studies of other MoE frameworks indicate typical efficiency improvements range from 3 to 7 times; although V3 appears to achieve even better results (over 10 times), branding it as a wholly revolutionary concept feels somewhat overstated, especially when considering the hysteria reflected on social media platforms like Twitter in recent days.

Shifting to DeepSeek's R1 model, it builds upon V3 by harnessing advanced techniques including reinforcement learning, which significantly enhances its inference capabilities to rival OpenAI's o1 modelNotably, DeepSeek also employs a 'model distillation' strategy, utilizing the R1 model as a guiding entity to generate data for fine-tuning smaller models that can compete with counterparts such as OpenAI's o1-mini

• Nvidia's $4.2 Trillion Vanish Sparks 3% Plunge in Nasdaq
• Fed Signals Potential Pause in Rate Cuts
• UAE Deepens Cooperation with BRICS Nations
• DeepSeek: A New Era Begins
• Key Drivers Behind New Productive Forces

This approach not only reduces costs but also introduces an innovative pathway for the mainstream adoption of AI technologies.

Bernstein remains optimistic about the AI sector, postulating that even if DeepSeek has indeed realized tenfold efficiency improvements, this merely aligns with the annual cost increase associated with current AI modelsThe ongoing advancements, driven by innovations such as Mixture of Experts, model distillation, and mixed-precision computing, are pivotal in the realm of AI development, especially as they emerge against the backdrop of the prevailing model size laws driving costs higher.

According to the Jevons Paradox, increased efficiency typically spurs greater demand rather than cutting expendituresBernstein asserts that the current appetite for AI computation remains far from saturated; it is likely that any new processing power will simply accommodate the continually growing demand for AI functionalities.

In conclusion, Bernstein’s analysis underlines that the uncertainty surrounding the perceived threat posed by DeepSeek is more a reaction to digestion of incomplete information rather than an accurate reflection of the technological shifts at play