Nomura Global Markets Research has released a comprehensive analysis of DeepSeek's anticipated next-generation large language model, DeepSeek V4, suggesting the upcoming release will primarily benefit China's domestic AI ecosystem rather than triggering another shockwave through global AI infrastructure markets as its predecessor did.

According to a report published February 10, 2026, by Nomura analysts Bing Duan and Joel Ying, DeepSeek's forthcoming V4 model—expected to launch mid-February—will likely incorporate breakthrough innovations in coding capabilities, ultra-long code processing, and inference reliability. The Information first reported the imminent launch, noting the model may achieve significant performance improvements across multiple dimensions.

New Architectural Innovations: mHC and Engram

Nomura's analysis focuses on two recently published DeepSeek research papers that likely underpin V4's architecture: "mHC: Manifold-Constrained Hyper-Connections" and "Conditional Memory via Scalable Lookup," introducing the "Engram" module.

The mHC framework addresses a critical bottleneck in transformer-based LLM training. While standard "residual connections" worked adequately when neural networks had fewer layers, they've become increasingly problematic as modern LLMs employ hundreds or thousands of layers. DeepSeek's solution enables richer, more flexible communications between layers through multiple internal information streams while enforcing strict mathematical guardrails to prevent signal amplification or destruction—issues that plagued previous "Hyper-Connections" attempts.

"The mHC framework provides an improved mechanism to the standard 'residual connections' in the transformer-based LLM training process, which may help to improve the LLM's performance led by better information flow scheme in between different layers," the Nomura analysts wrote.

The Engram module tackles a different challenge: the inefficiency of using computational resources for knowledge retrieval tasks. Current LLMs rely heavily on conditional computing mechanisms like Mixture of Experts (MoE) but lack native knowledge search capabilities, forcing them to "simulate" retrieval through expensive computation. Engram functions as a sparse memory table stored in system DRAM, retrieving static knowledge—entities, fixed expressions—in O(1) time, freeing up expensive High Bandwidth Memory (HBM) for priority workloads.

Breaking Through China's "Chip Wall" and "Memory Wall"

Nomura argues these innovations could prove particularly valuable for China's AI development, which has been constrained by limited access to advanced chips and memory due to export restrictions. By reducing training and inference costs while maintaining high performance, DeepSeek V4 may help Chinese AI developers "narrow down their gaps with global peers" through algorithmic and engineering innovations rather than hardware advantages.

"The Engram module helps to reduce training & inference costs by freeing up expensive memory (i.e., HBM) resources for prioritized workloads, and by offloading less time-sensitive workloads to cheaper memory (i.e., DRAM)," the report states. "We think this may help China's AI development to accelerate, as the advanced AI 'chip wall' and 'memory wall' have been a drag on LLM development in China."

The firm identifies potential beneficiaries in China's AI hardware sector, including server and switch company Unisplendour Corporation and optical transceiver manufacturer Accelink Technologies.

DeepSeek's Market Position One Year Later

The analysis arrives approximately one year after DeepSeek's V3 and R1 models drew significant attention—and sparked concerns about curtailed demand for computing power globally. Nomura's data shows DeepSeek remains the largest open-source model contributor by volume, though its dominance has declined as competition intensified. DeepSeek's two models accounted for over half of all open-source token usage on OpenRouter in late 2024 but now face a more fragmented market, with Alibaba's Qwen and Meta's Llama gaining ground.

Total token consumption for DeepSeek models on OpenRouter reached 14.37 trillion between November 2024 and November 2025, well ahead of Alibaba Qwen's 5.59 trillion and Meta Llama's 3.96 trillion.

No Repeat of Infrastructure Shock Expected

Critically, Nomura does not anticipate V4 will replicate the market disruption caused by V3 and R1. "We do not expect that the potentially new LLM from DeepSeek would send another big shock wave through the global AI infra market, as the major cloud service providers have been striving to pursue AGI (Artificial General Intelligence) with advanced computing power," the analysts wrote.

Instead, they suggest the more pertinent question is whether DeepSeek V4 can help global LLM and AI application players accelerate monetization progress, "thereby relieving the pains of the increasingly heavy capex burdens."

AI Agents and Software Applications

On the application front, Nomura expects more powerful AI agents to emerge, driven by DeepSeek's innovations and competitive responses. The firm noted recent developments including Doubao AI phones (in collaboration with ZTE Corporation and Alibaba's Qwen APP, which can perform multi-step tasks more automatically, signaling a transformation from conversational tools to comprehensive "AI assistants."

Contrary to market concerns about "LLM killing software," Nomura argues DeepSeek's innovations are "value-accretive to leading software companies in China, which could leverage these AI technologies to build more powerful software suites." The firm's top software picks include Kingsoft Office Software Corporation and Kingdee International Software Group.

The report suggests that as multi-task AI agents interact more frequently with LLMs, token consumption and computing power demand should increase if widely adopted—potentially supporting rather than undermining infrastructure investment cycles.