In the realm of artificial intelligence (AI) and semiconductors, the dynamics are shifting, raising fundamental questions about the future of technology investmentsRecently, stocks related to AI semiconductors, especially those of Nvidia, faced increased scrutinyInvestors began to wonder whether the significant capital poured into AI innovation correlated with tangible returnsThis skepticism comes at a time when China’s DeepSeek technology is rapidly gaining traction, allowing the Hang Seng Index to enjoy a 15% uptick over the past month, making it the best-performing index during this period globallyIndeed, the excitement surrounding DeepSeek has led to a paradigm shift in investor sentiment.

In response to a surge of inquiries from global clients, major international investment banks like Goldman Sachs, Morgan Stanley, and UBS have released multiple research papers focusing on DeepSeekThese institutions caution against jumping to conclusions about the impacts of DeepSeek’s innovations on the broader AI ecosystem, particularly concerning cost and demand dynamics for leading models such as OpenAI’s GPT-4, Anthropic's Claude 3.5 Sonnet, and Meta's Llama 3.2. Nevertheless, the consensus among analysts indicates that applications and platform layers stand to benefit from heightened competition among machine learning models and a decline in computing costs.

Cass, a former global market application head at OpenAI who specializes in AI and business strategy, remarks, “As AI capital expenditures rise and models expand, U.S. investors grow increasingly enthusiasticThey have historically supported such narratives.” The impetus behind this enthusiasm was exemplified by a recent announcement from the U.S. concerning a colossal $500 billion AI infrastructure plan dubbed 'StarGate.' Cass asserts that while DeepSeek might not be a groundbreaking innovation, it undoubtedly compels a reevaluation of existing industry narratives.

DeepSeek has rapidly lowered the cost structure associated with AI models

Advertisements

Goldman Sachs observed that, to date, the market has rewarded companies making hefty investments in AI—such as Amazon, Microsoft, and Google—as well as those providing essential tools and infrastructure like Nvidia and BroadcomYet, the low-cost model offered by DeepSeek raises critical questions about the sustainability of previous investment norms within the AI ecosystemInvestors are left deliberating whether such vast expenditures remain necessary moving forward.

The MoE (Mixture of Experts) architecture utilized by DeepSeek-V3 exemplifies a clever design borrowing from the concept of “divide and conquer.” By leveraging numerous specialized routing experts alongside a common shared expert, the model achieves significant expansion in capacity without incurring extensive resource costs.

Goldman Sachs had flagged excessive spending on AI in several reports, including one titled "Generative AI: Too Much Spending, Too Little Return?" However, Wall Street appeared unfazed, continuing to embrace the narrative that heavier capital investments in AI should correlate with increasing stock pricesData reveals a steep climb in capital expenditures among tech giants like Google, Meta, Amazon, Microsoft, Apple, and Oracle, culminating in an estimated total of $160 billion for 2023 and a projected $200 billion for the following yearThis trend exhausts a substantial portion of these companies' incremental free cash flow.

UBS shares insights indicating that DeepSeek might be disrupting the AI investment landscapeThe launch of DeepSeek R1 has prompted investors to question the supercycle anticipated in AI computing investmentsThis unease has caused a slide in stocks connected to AI semiconductors and related equipmentThe key concern lies in the reports suggesting that DeepSeek achieved extraordinary performance with minimal computational resources, successfully training a highly competitive base model while maintaining deduction costs far lower than those of competitors.

For illustrative comparison, DeepSeek charges just 1.4 cents per one million tokens generated—approximately equivalent to 700,000 words

Advertisements

In stark contrast, Meta charges $2.80 for similar outputs on its most robust model, placing DeepSeek’s costs at 1/200 compared to American equivalents and merely 1/50 of OpenAI's costsFurthermore, China's leading AI chatbot "Doubao" is reported to run at costs 85% lower than the industry average, showcasing significant advantages in cost management within the Chinese AI sector.

Nonetheless, there remains a sentiment of caution among institutional investorsUBS notes that DeepSeek's reliance on advanced techniques like “multi-head attention” and “mixture of experts,” while effective in reducing computational resource needs, may not translate seamlessly for all larger-scale frontier large language models (LLMs). Additionally, while DeepSeek adopts an open-source model, the challenge of its integration into existing AI ecosystems remains uncertainHowever, in the long run, diverse training methodologies may ultimately foster greater inferencing demands, suggesting that investments in AI computing will continue to grow.

Cass references Jevons’ Paradox, emphasizing that if a resource becomes more efficient, it tends to see increased usageHe posits that AI could potentially become as ubiquitous and affordable as the internetEven if expenditures on AI computing rise, the costs per unit may decline, leading to a more accessible AI landscape. “If AI models improve yet their prices increase, they risk being monopolized by elite sectors, exacerbating inequality, whereas DeepSeek's approach promotes affordability,” Cass elaboratesReflecting on his time at OpenAI, he recalls how the initial model cost $60 per one million tokens but has now dropped to just $2 for GPT-4—a trend he expects to persist.

As DeepSeek reshapes perceptions, liquidity managers on Wall Street are feeling the pressureLarge-cap fund managers are heavily invested in tech stocks, with the seven major U.S. technology firms contributing to 41% of the S&P 500's total return of 25% in 2024. Despite a recent market rebound, the anxiety regarding DeepSeek's implications looms large.

UBS assesses that the impact of DeepSeek on internet companies is multifaceted

Advertisements

Both Amazon and Google not only utilize AI models but also provide AI model services to external clients via Amazon's Bedrock and Google's Vertex AIShould the observed trends of enhanced efficiency and reduced capital expenditure persist, operational and capital expenses for these giants may decrease.

While assessing risks related to projected AI-generated revenues, Meta appears to be the least impacted, followed by Amazon and GoogleMeta has yet to derive substantial revenue from its open-source large model, Llama, while Amazon relies on multiple external AI model providers like Anthropic, and Google chiefly depends on its proprietary Gemini model alongside external ones.

Notably, current financial forecasts for Amazon Web Services (AWS) and Google Cloud Platform (GCP) do not account for massive growth in cloud computing and AI segments, which means a decrease in operational and capital expenditures may positively affect free cash flow significantly.

Goldman Sachs opines that within tech giants, Google and Meta are particularly well-positioned due to their advancements in AI at the application level.

Meanwhile, opportunities are burgeoning for larger numbers of small to medium businessesThe applications layer is expected to see more use cases develop—similar to the way 5G technology proliferated from upstream to downstream mobile applicationsMore tech companies are anticipated to harness generative AI technologies to enhance their product or service value, including firms like Canva, Adobe, and GitLab, which have yet to go public but hold significant monetization potential.

In the semiconductor sector, investment firms retain a strategy of buying on dipsNvidia, Broadcom, and others have faced the brunt of recent sell-offsUBS points out that computation remains the central driver behind AI performanceDespite heightened efficiency through new algorithms, demand for AI training will persist as the leading growth factor for AI computation over the next two to three years

Ultimately, AI computation is still in its nascent stages.

The overarching sentiment in the market has been invigorated by developments in the Chinese AI industryThe Shanghai Composite Index has climbed above the 3300-point mark, while the Hang Seng Index seems on the verge of a technical bull market this yearThe Hang Seng Tech Index has skyrocketed 23% since its low in January, serving as a barometer of overseas investor confidence.

Goldman Sachs posits that companies like Tencent, Alibaba, CenturyLink, and Global Data are all basking in the AI fervorTencent stands out due to its stronghold in the consumer-side AI agent application realm, bolstered by the WeChat super app ecosystem that offers integrated social, payment, and transaction capabilities, granting it a competitive edge in actualizing AI applicationsAlibaba, as China's largest public cloud computing enterprise, stands to benefit from the surge in AI applications.

Additionally, CenturyLink and Global Data, as prime representatives of the data center theme, are poised to be significant beneficiaries of the long-term growth in AI computation demand, propelling investments in public cloud and AI computing infrastructure.

Significantly, the consensus among analysts emphasizes that the application side will attract more attention across various sectors, including internet applications and manufacturing (robotics, intelligent driving). A private equity manager noted that companies once viewed as having weak product capabilities, like Kingsoft and Yonyou, are now being reexamined amid emerging opportunities.

However, Morgan Stanley's Asia Tech team has offered warnings of macroeconomic risksThey highlight that groundbreaking advancements from DeepSeek might trigger concerns regarding the overvaluation of Asian AI supply chain tech stocksIn contrast, conventional non-AI tech firms are presently struggling due to geopolitical uncertainties and weak global demand, potentially attracting renewed investor focus.

The uncertainty surrounding tariffs remains an underlying concern for the market

Advertisements

Advertisements