The dynamic realm of Silicon Valley has been significantly stirred by a company called DeepSeek, particularly due to its cost-effective training technologies that promise to reshape the landscape of artificial intelligence (AI). This excitement has not gone unnoticed, as the financial titan Citibank has released an extensive report penned by analysts Atif Malik and Asiya Merchant, that meticulously explores the ramifications of DeepSeek’s innovations on the entire AI infrastructure ecosystemThis analysis sheds light on which segments of the industry stand to gain and which may encounter hurdles.

The report heralds DeepSeek's R1 model as a catalyst for widespread AI model adoption, particularly in both consumer and enterprise sectorsCitibank underscores that a decrease in computational costs will enable a noticeable uptick in the potential return on investment in AI technologyThis is a pivotal moment in the industry, as companies reevaluate the feasibility of incorporating AI into their operational backbone.

At the outset of the report, the analysts define “scaling laws,” acknowledging the contemporary perspective that categorizes these laws into three distinct phases: pre-training, post-training, and test-time scalingDeepSeek emerges as a prime example within the “test-time scaling” frameworkEach phase plays a crucial role in the development and optimization of AI models, creating a comprehensive spectrum of functionality and applicationThe initial phase, pre-training, involves the educator-like process of training machine learning models on large datasets to generate broad and adaptable featuresSubsequently, post-training utilizes techniques such as reinforcement learning and human feedback to fine-tune models after their initial training, enhancing their performance and reliability.

The final phase, test-time scaling, is where DeepSeek makes its mark, extending the model's reflection period through multi-step reasoning during the inference stage

Advertisements

This innovative approach can lead to more robust and smarter AI systems capable of making more complex decisions, thereby offering significant advancements to the industry as a whole.

The influence of DeepSeek's advancements on the AI infrastructure value chain is multifacetedWhile certain segments might face immediate challenges due to shifts in demand, the overarching narrative is one of continued robust growth in AI architectureCitibank foresees a surge in opportunities for components like GPUs, ASICs, and Data Center Interconnects (DCIs) due to the proliferation of inference stages in AI applicationsConversely, a decline in training demand could exert pressure on various segments, such as Retimers and optical modules, which traditionally thrived in training environments.

In this context, the importance of AI infrastructure cannot be overstatedThe growing capabilities of AI technologies position infrastructure development as one of the primary driving forces within the global technology sector, replete with new opportunities and challengesDeepSeek's advancements not only facilitate the wider acceptance of AI models but also stimulate each facet of the AI infrastructure, prompting a wave of adaptations across the ecosystem.

Impact on Specific Segments of AI Infrastructure

When scrutinizing the potential fallout across various segments, Citibank delineates clear expectations.

GPUs: Neutral

The essential role of GPUs in AI training and inference is evident, with market demand historically remaining strongAnalysts from Citibank argue that while DeepSeek’s technology may diminish the need for extensive training, the computational requirements for the inference phase will see an upswingThus, the GPU market is projected to maintain a neutral stance overall, balancing the shifts in training versus inference demands.

ASICs (Application-Specific Integrated Circuits): Neutral to Positive

ASICs exhibit particularly stellar performance during AI inference phases

Advertisements

Citibank anticipates that as inference demand rises, ASIC market shares will progressively expandWhile the tapering training demands might present challenges to the ASIC market, projections indicate that the long-term growth attributed to increased inference requirements will counterbalance these adverse trends, establishing a trajectory more aligned with inference tasks.

Retimers: Neutral to Negative

Retimers primarily facilitate high-speed data transfer and have traditionally been in demand during AI training incrementsHowever, Citibank suggests that as AI computation pivots from training to inference, demand for Retimers may diminish due to the lower computational intensity during inference phases, which in turn lessens the call for high-speed data transmission.

Optical Modules (Intra Server/DC): Neutral to Negative

Analogous to Retimers, the demand for optical modules has thrived during AI training stagesAs the industry transitions to more inference-centric models, particularly within data center connectivity, the market for optical modules could conceivably face a downturn, prompting a reevaluation of connection requirements.

DCI (Data Center Interconnect): Positive

The DCI market appears largely insulated from the fluctuations seen between training and inference stages, with its performance not contingent on the specific demands of AI models or workloadsCitibank posits that the anticipated growth in inference will unlock fresh opportunities for DCI, underscoring its robust relevance in the evolving AI landscape.

Switches: Neutral to Positive

Operating as the backbone of data center networks, switches are intricately tied to the spread of AI computationAlthough short-term fluctuations in training demand may apply downward pressure on the switches market, the burgeoning demand for enhanced network bandwidth during inference indicates a stabilizing force, suggesting a neutral to positive outlook for switch manufacturers as AI technologies continue to proliferate.

Connectors: Neutral to Positive

Similar to switches, connectors maintain a low correlation with training or inference directly, yet their utility aligns closely with the overall growth trajectory of AI infrastructure

Advertisements

Advertisements

Advertisements