The recent surge in interest surrounding a Chinese AI company, DeepSeek, has sent ripples through the global technology landscape, particularly in the realms of semiconductor stocks and AI investmentsWhile NVIDIA and several Asian semiconductor companies have faced significant pressure, concerns over whether substantial AI investments correlate with adequate returns have begun to rattle even the mightiest of American tech giantsMeanwhile, DeepSeek's stocks have soared, contributing to a remarkable 15% rise in Hong Kong's Hang Seng Index over the past month—far outpacing similar global indices.

As investment queries flood in from clients across the globe, prominent investment banks like Goldman Sachs, Morgan Stanley, and UBS are racing to publish analyses centered around DeepSeekA consensus appears to be forming: it is too early to assess how DeepSeek’s reported innovations in training will impact the larger ecosystem, especially concerning the costs and demands of cutting-edge models like GPT-4, Claude 3.5 Sonnet from Anthropic, and Meta’s Llama 3.2. Nonetheless, there is a collective agreement that applications and platforms are expected to benefit from increased model competition and decreasing computation costs.

According to Zack Kass, a former global market application lead at OpenAI and an expert in AI and business strategy, the trend toward larger investments in AI is favored by American investors, who have typically rallied around such narrativesHe remarked upon the recent announcement of a substantial $500 billion investment plan for AI infrastructure by President Joe Biden, noting it resonates with investors' affinity for ambitious storiesWhile DeepSeek may not represent a groundbreaking innovation from zero to one, its emergence prompts a reassessment of the sprawling AI narrative.

Delving deeper, Goldman Sachs has noted that the market is currently rewarding both well-established companies that are heavily investing in AI—such as Amazon, Microsoft, and Google—and those supplying the necessary tools and infrastructure, including NVIDIA and various semiconductor entities

Advertisements

However, the low-cost characteristics of DeepSeek’s models seem to shake investors' confidence in the overall AI ecosystem, raising questions over whether hefty investments will remain necessary going forward.

DeepSeek's V3 model features a MoE (Mixture of Experts) architecture, which encapsulates an innovative design inspired by a "divide and conquer" strategyBy onboarding various specialized routing experts alongside a universal shared expert, DeepSeek has achieved efficient expansions in model capacity.

Despite warnings from Goldman Sachs as early as Q3 of last year regarding the risks of excessive AI expenditure—illustrated through their report titled "GenAI: Too Much Spend, Too Little Benefit"—Wall Street has remained largely unfazed, eager to embrace the narrative that increased AI capital spending should subsequently elevate stock prices.

Current data showcases that total capital expenditures from major players like Google, Meta, Amazon, Microsoft, Apple, and Oracle have seen pronounced growth, with 2023 expenditures estimated at a staggering $160 billion, surging to $200 billion projected for 2024. This upward trend has notably consumed the majority of these companies' incremental free cash flow; for instance, Microsoft expects to allocate around $80 billion to capital expenditures this year, nearly matching its total cash flow.

UBS highlighted that DeepSeek's emergence might disrupt the AI trading market, casting doubts over what many considered a super-cycle of AI computing investmentsThis, in turn, pressured semiconductor and equipment stocks across the sectorA critical factor underscoring investor anxiety is the report that DeepSeek has trained a highly competitive foundational model with minimal computational resources, yielding inference costs substantially lower than those of comparable modelsDeepSeek's V3 model has reportedly introduced a deflationary impact on Chinese AI computing costsConsequently, investors may now view DeepSeek as undermining the prevailing paradigm dictating the need for incessant investments to train larger models.

For context, DeepSeek charges a mere 1.4 cents for generating one million tokens, translating to about 700,000 words

Advertisements

In stark contrast, Meta charges $2.8 for the equivalent outputs from its largest model—a cost ratio that positions DeepSeek's pricing at 1/200th of American large models, with its expenses roughly at 1/50th of OpenAI's costsFurthermore, China's most popular AI chatbot, Doubao, operates with costs 85% lower than the industry average, signaling a potential competitive edge in cost management within the Chinese AI industry.

However, uncertainties remain regarding the broader applicability of DeepSeek's approachUBS pointed out that DeepSeek heavily relies on technologies such as Multi-head Attention (MLA) and Mixture of Experts, which demand minimal computational resources to complete model trainingMOE has garnered widespread adoption in Chinese AI paradigms, effectively partitioning training tasks into smaller sub-models to reduce resource demandsThis methodology, however, may not universally translate to larger-scale frontier large language model (LLM) training.

Moreover, although DeepSeek employs an open-source model, its seamless integration into the existing AI ecosystem remains questionableEven so, there is a prospect that its methodologies may inspire adaptation within prevailing ecosystemsLong-term, multiple training methods may increase inference demands; thus, spending on AI computation is likely to continue its upward trajectory.

Kass emphasizes an interesting perspective rooted in the Jevons Paradox: as a resource becomes more efficient, its overall usage tends to increaseHe posits that AI services may evolve to be as affordable and ubiquitous as the internet, suggesting that while investment in AI computation may rise, the per-unit costs may decrease, leading to wider accessibility of AI technologiesHe conveyed that the initial model at OpenAI cost $60 per million tokens, whereas GPT-4 launched at $2—an extraordinary drop in two yearsHe anticipates that this trend toward affordability will persist.

On the international tech front, anxiety has emerged amongst Wall Street investment managers, particularly given that many large-cap fund managers have heavy stakes in tech stocks

Advertisements

The so-called "magnificent seven" tech companies have contributed immensely to the 25% total return of the S&P 500 for 2024. While the market has seen a rebound, concerns surrounding a potential tech correction linger amongst investors.

UBS argues that DeepSeek's implications for internet companies are multifacetedCompanies like Amazon and Google play dual roles as consumers of AI models while simultaneously offering AI model services through platforms like Amazon's Bedrock and Google's Vertex AIIf the current trends of enhanced efficiency and reduced capital investment hold, it could lead to reduced operational and capital expenditures for these stated companies.

When assessing risks associated with projected AI-related revenues, analysts concluded that Meta appears less affected, followed by Amazon and GoogleMeta has not derived substantial revenue from its open-source Llama model, while Amazon relies on multiple external AI model providers, such as Anthropic, and Google focuses mainly on its proprietary Gemini model alongside third-party models.

It is noteworthy that current financial projections concerning Amazon's AWS and Google's GCP do not factor in a sharp growth acceleration in cloud compute/AI business, meaning reduced operational and capital expenditures could present significant positive impacts on free cash flow.

Goldman Sachs, conversely, believes that Google and Meta are relatively well-positioned among tech titans, given their advanced strides within the "application layer" of AI.

For smaller firms, opportunities are also on the horizonThe application layer appears poised for an uptick in use cases, mirroring the dispersion of the 5G narrative from upstream base stations to downstream mobile applicationsThis transition could enable numerous tech companies to leverage generative AI technologies, enhancing their product or service valueNoteworthy examples on the application side include companies like Canva, Adobe, and GitLab, all poised for significant monetization potential, yet to hit the public market.

In the semiconductor sector, analysts maintain a buy-low approach

Stocks like NVIDIA, Broadcom, and Marvell Technology have experienced considerable impact from recent sell-offsUBS asserts that computational enhancements remain the core drivers for improving AI performance, suggesting that despite emerging algorithms, demand for AI computation in the next two to three years will continue as a principal growth engine for the sectorFrom a long-term view, AI computation is still in its infancy.

The overall market sentiment has also significantly benefitted from the advancements witnessed in the Chinese AI sectorThe Shanghai Composite Index has surpassed 3300 points, and the Hang Seng Index is approaching a technical bull market year-to-dateAs a bellwether for overseas investor confidence, the Hang Seng Tech Index has soared 23% since its January lows.

Goldman Sachs attests that companies such as Tencent, Alibaba, Century Link, and GDS Holdings are all benefiting from the AI renaissanceTencent is particularly well-positioned in relation to consumer-targeted AI applications thanks to its WeChat super-app ecosystem, which integrates social features with payment and transaction capabilities, making it a formidable player in deploying AI solutionsAs China’s largest public cloud computing entity, Alibaba stands to gain from ongoing growth in AI applications.

Furthermore, firms like Century Link and GDS Holdings exemplify the data center theme, as it is anticipated that long-term AI computational needs will drive investments in cloud services and AI infrastructure, making these data center companies pivotal beneficiaries of the AI wave.

The collective consensus underscores a heightened focus on application layers, with expectations that sectors ranging from internet applications to manufacturing (robotics, autonomous driving) will be symbiotically bolstered by AI advancementsA prominent private equity investment manager shared insights that even companies like Kingsoft and Ufida, previously regarded as lacking robust product capabilities, are undergoing a re-evaluation based on this emerging opportunity.

However, Morgan Stanley’s Asian tech team cautions against macroeconomic risks, suggesting that technological breakthroughs from DeepSeek could ignite concerns regarding the high valuations of Asia's AI supply chain stocks

Advertisements

Advertisements