The recent surge in the stock values of companies associated with artificial intelligence in Asia, particularly those connected to DeepSeek, has raised eyebrows within the investment community, drawing stark contrasts to the pressures faced by American tech giants like NVIDIAIn a world where the allure of substantial investment in AI technology is palpable, questions loom large about the returns these investments may yield over time, particularly as the DeepSeek phenomenon continues to escalateOver the past month, the Hang Seng Index has witnessed an impressive 15% increase, marking it as one of the top performers globally during a tumultuous period.

International investment banks like Goldman Sachs, Morgan Stanley, and UBS have launched various analyses centered on DeepSeek, responding to an unprecedented influx of inquiries from global clients intrigued by this latest player in the AI landscapeConsensus among these financial institutions suggests that the ramifications of DeepSeek's reported innovations in training models are still up for debateSpecifically, how these innovations will influence operational costs and demand for industry-leading models such as OpenAI's GPT-4 and Meta's Llama3.2 remains uncertainHowever, there is agreement that the emerging competition among models, coupled with the potential for reduced computational costs, will likely favor application and platform developments.

Zack Kass, a former global market application head at OpenAI and a business strategy expert, recently weighed in on the unfolding AI narrativeHe indicated that as capital expenditures in AI rise and model sizes expand, U.S. investors are increasingly drawn to these narrativesTheir willingness to invest heavily—upwards of $500 billion—illustrates a broader trend, as Kass pointed outHe posited that while DeepSeek may not represent a groundbreaking innovation in its entirety, it is prompting a reevaluation of existing market perspectives.

DeepSeek has positioned itself as a catalyst for cost efficiency and accessibility in AI models

Advertisements

Goldman Sachs previously observed that up until this point, the market has favored companies heavily investing in AI, such as Amazon and Google, as well as those providing essential tools and infrastructure, like NVIDIA and BroadcomThe competitive pricing structure of DeepSeek models has begun to shake investor confidence in long-standing patterns of investment in AI, as these models challenge the necessity of enduring significant expenditure for training future models.

The architecture of DeepSeek-V3 leverages a sophisticated design based on the principle of "divide and conquer." It introduces multiple specialized routing experts alongside a commonly shared expert to effectively enhance the model’s capacityThis innovative approach reflects a growing trend towards optimizing resources while achieving high-performance outcomes.

Noteworthy data reveals that capital expenditures among tech giants like Google, Amazon, Microsoft, Apple, and Oracle have dramatically increased, with total expenditures reportedly climbing to approximately $160 billion in 2023 and projected to reach $200 billion in 2024. Such significant financial commitments tend to consume a large chunk of these companies' incremental cash flowsFor instance, Microsoft has forecasted an expenditure of $80 billion this year alone, matching its annual cash flow, a scenario implying thinner profit margins due to heavy capital outlays.

UBS has drawn attention to how DeepSeek's emergence has disrupted the AI investment landscapeThe rollout of DeepSeekR1 has left investors pondering the sustainability of the so-called super cycle in AI infrastructure investmentMarket reports suggest that DeepSeek has managed to build a highly competitive base model while utilizing a fraction of the typical computational resources, driving down inference costs significantlyIntriguingly, as the V3 model has influenced cost structures in Chinese AI computing, investors now appear to be grappling with the radical idea that industry norms of relentless investment for the development of larger models may be shifting.

DeepSeek’s operational efficiency is showcased in its pricing model, which reportedly charges just 1.4 cents per million tokens—equivalent to around 700,000 words

Advertisements

In stark contrast, Meta charges $2.80 for the same output with its largest modelThis pricing strategy presents a ratio of about 1/200 compared to U.S. counterparts and only about 1/50 relative to OpenAIFurthermore, China’s leading AI chatbot, Doubao, exhibits an operational cost that is 85% lower than industry averages, underscoring a compelling edge in cost management among Chinese AI firms.

However, skepticism remains as UBS warns that DeepSeek's reliance on sophisticated methodologies like “multi-headed potential attention” (MLA) and “mixture of experts” (MoE) might limit the scalability and applicability of this training paradigm across larger modelsThese techniques often break down tasks into smaller, manageable elements, effectively minimizing computational demands but may not be universally applicable to the broader spectrum of leading large language model (LLM) training.

Moreover, questions linger regarding the adaptability of DeepSeek’s open-source model into existing AI ecosystemsDespite its potential, how seamlessly the broader AI community can integrate these new methodologies remains uncertainIn the grand scheme, while various models may pave the way for evolving methodologies, the demand for computational power is expected to rise continually, suggesting that AI infrastructure expenditures are likely to grow.

Kass, in his reflections, highlighted the implications of Jevons Paradox, which posits that as a resource becomes more efficient, its overall consumption might increase rather than decreaseKass drew parallels between the future of AI and the ubiquitous nature of the internet, projecting that while AI spending may continue on an upward trajectory, the costs associated with its operational use will drop dramatically, facilitating a more widespread incorporation into society.

The ongoing ripple effects from DeepSeek's advancements are triggering anxiety among Wall Street investment managers, particularly those holding substantial stakes in tech stocks

Advertisements

The collective contribution of America's tech giants accounted for a staggering 41% of the anticipated total returns for the S&P 500 in 2024. Although market sentiment recently indicated a slight rebound, lingering concerns persist among investors about potential volatility.

UBS suggested that the ramifications of DeepSeek's innovations might present a complex landscape for internet companiesGiants like Amazon and Google are not just consumers of AI models but also provide AI model services themselves via platforms such as Amazon's Bedrock and Google's Vertex AIShould the current trends of heightened efficiency and reduced capital investment continue, we could see operational and capital expenditures drop for these tech titans in the future.

In an evaluation of risks tied to AI-related revenue expectations, institutions have indicated that Meta appears to be less affected, followed by Amazon and GoogleWhile Meta has not significantly profited from its open-source model, Llama, Amazon relies heavily on a range of external AI model providersConversely, Google primarily integrates its proprietary Gemini model and various third-party models into its offering.

Significantly, current financial predictions for Amazon's AWS and Google's GCP have not accounted for accelerated growth in cloud computing and AI businesses; thus, potential reductions in operational and capital expenditures could strengthen free cash flow significantly moving forward.

Goldman Sachs pointed out that among major tech players, Google and Meta stand in particularly advantageous positions due to their advancements within the application layers of AI technology.

Smaller enterprises in tech also find new opportunities emerging as AI applications proliferate, resonating with parallels drawn from the spread of 5G technology into consumer marketsCompanies could harness generative AI to enhance their product offerings significantly, as foreseen with platforms like Canva, Adobe, and Gitlab, which represent vast monetization potentials, even before going public.

In the semiconductor realm, institutions maintain a buy-the-dip ethos, particularly for stocks heavily impacted by market sell-offs, including NVIDIA and Broadcom

UBS emphasized that computational power remains central to enhancing AI performance, positing that despite the emergence of more efficient algorithms, demand for AI computations will drive growth in this domain over the next few years, indicating that the AI computation market is still in its nascent stages.

Within the Chinese market, the sentiment remains high, buoyed notably by DeepSeek's recent developmentsThe Shanghai Composite Index has climbed above 3,300 points, and the Hang Seng Index has nearly reached a technical bull market since the beginning of the yearSince its low point in January, the Hang Seng Tech Index has skyrocketed by 23%, reflecting a robust resurgence in investor confidence.

Goldman Sachs noted that tech giants like Tencent, Alibaba, and Century Internet are thriving amid the AI heatwaveTencent, with its extensive WeChat ecosystem—which offers a rich blend of social, payment, and transactional capabilities—holds a prime position in consumer-targeted AI applicationsMeanwhile, Alibaba, as China’s largest public cloud computing entity, is also poised to benefit from the sustained growth of AI applications.

Additionally, companies like Century Internet and GDS Holdings are garnering attention as representatives within the data center themeAnalysts widely agree that the long-term growth trajectory of AI computation demands will invigorate investments in public cloud and AI infrastructure, with data center firms set to become noteworthy beneficiaries of this AI wave.

There is a shared belief among analysts that application development will increasingly capture attention across various sectors, including internet applications and manufacturing (such as robotics and intelligent driving). A prominent private equity manager even mentioned companies like Kingsoft and UFIDA, which were previously perceived as having weaker product offerings but are now receiving newfound recognition amidst this wave of opportunity.

Nonetheless, Morgan Stanley's Asian tech team has sounded a note of caution regarding macroeconomic risks

Advertisements

Advertisements