In a significant development that may reshape the artificial intelligence landscape, Bank of America has recently spotlighted the innovations brought forth by DeepSeek, specifically the launch of their advanced AI models, Deepseek-R1 and DeepSeek-V3. These models leverage the energy-efficient NVIDIA H800 chips, delivering competitive pricing for both training costs and API usage, thus challenging existing large language models (LLMs) like those offered by OpenAIThe implications for various sectors of the tech industry and the broader supply chain are profound.

DeepSeek's open-source approach empowers nearly every software company to intrinsic AI capabilities into their offeringsThis appears particularly beneficial for SaaS firms, like Kingsoft and Kingdee, who may prioritize embedding this technology into their products, enhancing user experience and productivity.

While the immediate repercussions for data centers are limited, there is a belief that as AI applications proliferate, the burgeoning demands will catalyze an uptick in data center utilizationA notable shift may also occur as computational resources transition from model training towards inference tasks, which is critical for AI functionality.

The public cloud segment stands to gain significantly, as rising requirements for AI inference point towards a burgeoning opportunity in the marketThe concept of "model as a service" (MaaS) may solidify as a promising new revenue stream as demands for such services grow.

In the semiconductor sector, the proliferation of edge AI devices, including AI glasses and toys, could see a surge, given the increased competitive edge of domestic chips in inference tasks

Advertisements

However, it is noteworthy that the overall supply is still restrained by U.S. technology export limits.

DeepSeek's innovation represents a significant reduction in AI application costs, heralding prospects for a thriving AI ecosystem in China and fortifying local companies' competitiveness in software, cloud computing, and edge computing domains.

However, the ramifications of DeepSeek on the storage market remain ambiguous: investors have expressed concerns that advancements from DeepSeek may lead to diminished demand for high-bandwidth memory (HBM). Yet the current market data indicates a continuous shortfall for HBM3e, with strong orders for the 12-layer variant, while SK Hynix anticipates finalizing its HBM4 supply agreement with NVIDIA in the upcoming months.

On the part of Samsung Electronics, their cautious forecasting for HBM sales in early 2025 is attributed primarily to the 12-layer HBM3e still undergoing sampling and not yet entering mass production, contrasted with a decline in 8-layer HBM3e demandIn contrast, SK Hynix is witnessing persistent growth in HBM shipments, boosted especially by orders from clients like NVIDIA for their 12-layer HBM3e.

The storage market outlook suggests that the demand for HBM continues to be robust, with significant capital expenditures projected from American tech companies in 2025-2026. For example, Meta plans to invest $62 billion and $68 billion during 2025/26, starkly up from anticipated expenditures of $27 billion and $37 billion in 2023/24.

In the DRAM and NAND storage markets, as both Samsung and Hynix reduce capacity for traditional memory types (DDR5 and NAND), prices are anticipated to witness a rebound in the first half of 2025.

Current predictions for DRAM spot prices indicate stability in January 2025, but there may be slight increases in February to March, particularly due to stocking impacts from data centers

Advertisements

DeepSeek's influence hasn’t yet shifted the storage industry dynamics significantly; HBM market demand sustains its vigorSamsung faces formidable challenges in this sector, while SK Hynix is poised to exceed Samsung’s HBM revenues by more than 100% in 2025, capitalizing on technological advantages and production capacity.

Goldman Sachs also reflects on how DeepSeek might herald a transformative phase, likening it to a potential "GPT moment" for Asian stock markets.

Goldman Sachs strategists Tim Moe and Alvin So raised these considerations in a market insights report, which delves into the implications of the advancements from China’s AI sphere and their potentially significant impact on regional equities.

They assert that while the ongoing development of DeepSeek technology could mimic the disruptive effects seen previously with generative pre-trained transformers (GPT), there’s a dual opportunity alongside corresponding transaction strategiesThey have outlined a framework to help investors create portfolios that hedge against potential risks while capitalizing on DeepSeek advancements.

The report identifies several firms that are poised to optimize user experiences and fuel productivity through LLMs, with many originating from China’s cloud computing, software, applications, and IoT device sectors.

The report differentiates between two investment baskets: one focuses on companies listed in Hong Kong (GSCBHAIT) and the other caters to firms listed on the Chinese A-shares (GSCBCAIT), each boasting over $700 million in daily trading volumes.

The Hong Kong basket comprises major players such as Tencent, Alibaba, Netease, JD.com, and Baidu, with a weight of 7% attributed to each entity.

Meanwhile, the Chinese A-share basket includes Beijing Kingsoft, Shenzhen Transsion Holdings, and 360 Security Technology, weighted slightly lower at 6.5%.

Additionally, the report outlines derivatives trading strategies: (1) swaps suggesting long positions in GSCBHAIT at a rate of SOFR + 70bps and GSCBCAIT at SOFR + 50bps; (2) options incorporating a zero-cost structure that finances calls by selling puts.

Goldman Sachs expresses cautious optimism regarding the developments around DeepSeek, allowing for the possibility of disruption within the global AI chip demand and upstream supply chains

Advertisements

Investors are encouraged to consider structured options to hedge against short-term volatility.

Lastly, investment concerns regarding DeepSeek have resonated throughout the semiconductor sector, prompting a sharp sell-off of all AI-related stocks.

As China embarks on developing an open-source large language model that rivals GPT-4o yet operates with far less computational power, market reactions have reverberated broadly across the AI semiconductor domainDeepSeek’s architecture incorporates mixture of experts (MoE) and multi-head potential attention techniques alongside high-quality parameter processing, prompting a renewed focus on return on investments (ROI) within the AI industry.

Despite the remarkable efficiency of DeepSeek’s models, no notable commercial AI applications have yet emerged from this innovationThe potential revision of computational power needs could predict a decline in AI capital expenditures by 2026 or curtail growth.

The industry’s natural apprehensions toward a perceived slowdown in computational demand are compounded by ongoing scrutiny of ROI in AI venturesFor instance, despite Nvidia anticipating $200 billion in GPU revenues for 2024, stakeholders do not see substantive returns on such vast investmentsWhile advancements in models demonstrate substantial cost, the absence of specific commercial successes to validate such expenses raises further concern.

We postulate that DeepSeek's success could catalyze two primary industry strategies: one where firms perpetuate the drive for enhanced computational power to expedite model advancements, and another focusing on efficiency and ROI, pointing to a decrease in computational demands post-2026.

With ample capital in the market, overseas AI firms have relentlessly pursued refinements in modeling

However, DeepSeek's outcomes may compel investors to reassess the viability of such investments in computational capacitiesHence, American AI companies will likely face increased pressure to substantiate the rationale behind growing AI capital expenditures as 2026 approaches.

The ramifications of these developments on the AI supply chain—including GPUs, original design manufacturers of servers, printed circuit boards, and liquid cooling systems—could be detrimentalIn contrast, specific integrated circuits (ASIC), high-bandwidth memory (HBM), power supplies, and data centers may exhibit greater resilience.

When considering the smartphone impact, smaller models performing satisfactorily could indeed present a positive scenarioHowever, consumer recognition of AI-enabled smartphones remains tepidRunning larger models necessitates advanced hardware upgrades (such as advanced packaging combined with fast dynamic random-access memory), thereby inflating costs.

For instance, Apple's models, largely based on MoE technology, still possess merely 3 billion parameters, insufficient to deliver valuable services to consumersHence, while DeepSeek signifies a sliver of hope in this domain, its success does little to shift the short-term outlook for AI smartphones.

Lastly, amid the complexities of the US-China tech competition, the focus on chip restrictions has made China the sole market aggressively pursuing improvements in large language model efficienciesAs a consequence, there is a growing realization that further chip restrictions may inadvertently accelerate Chinese innovationsA shift toward relaxing AI diffusion policy may soon be seen, as U.S. policymakers could recognize such risks and opt for more lenient approaches.

In light of these unfolding developments, Jefferies plans to host two expert conference calls centered on the topic of DeepSeek, with dates potentially aligning with the Chinese New Year around February 10. Stakeholders are encouraged to stay attuned for further details regarding these insights.

Advertisements

Advertisements