DeepSeek has done it again. Less than 14 months after its R1 model rattled global AI markets by delivering frontier-level reasoning at a fraction of conventional compute costs, the Chinese AI lab quietly dropped another technical bombshell — and HSBC Qianhai Securities is already telling clients to pay attention.

In a research note published March 2, 2026, HSBC Qianhai analysts Yiran Liu and Heng Zhang argue that DeepSeek's newly unveiled DualPath inference system represents a structural inflection point for China's software industry, one that could meaningfully accelerate AI monetization across enterprise software, infrastructure, and edge computing. The bank maintains Buy ratings on six Chinese software names: Sangfor Technologies, Yonyou Network Technology, Glodon, Kingdee International Software, Kingsoft Office, and ThunderSoft.

What DualPath Actually Does — And Why It Matters

On February 25, 2026, DeepSeek published a research paper introducing DualPath, a dual-path key-value (KV) cache loading mechanism designed to eliminate a fundamental bottleneck in large language model (LLM) inference. In the prevailing disaggregated architecture, loading large-scale KV caches from external storage creates a lopsided bandwidth problem: storage network interface cards (NICs) on prefill engines become saturated while those on decode engines sit largely idle. The result is a system-wide throughput ceiling that constrains complex, multi-turn AI workloads.

DualPath breaks this constraint by introducing a second data path — storage-to-decode — that transfers KV caches into the decode engine and routes them to the prefill engine via RDMA compute networks, bypassing congested links and avoiding interference with latency-sensitive model execution. A global scheduler dynamically balances load across both engines.

The performance numbers are not trivial. According to DeepSeek's own benchmarks across three production-grade agentic workloads, DualPath improves offline inference throughput by up to 1.87x and online serving throughput by an average of 1.96x without violating service-level objectives.

"GPU utilization improvements will directly lower per-inference latency and compute consumption, driving a significant reduction in the marginal cost of LLM inference," the HSBC analysts write. The implication: agentic AI applications — those requiring long-context processing, multi-step reasoning, and iterative interaction — just got substantially cheaper to deploy at scale.

Infrastructure First, Then Applications

HSBC's investment thesis flows through two channels. The first is infrastructure. As agentic AI systems grow more complex — featuring multi-agent clusters, distributed execution, and stringent data privacy requirements — public cloud alone cannot satisfy enterprise demand. That creates a structural growth opportunity for on-premise hyper-converged infrastructure (HCI) and software-defined storage (SDS).

The bank's top infrastructure pick is Sangfor Technologies (300454 CH), currently trading at RMB 131.10 with a target price of RMB 191.00, implying 46% upside. HSBC forecasts the company's cloud business gross margin will expand to 46% by 2027 (from 44% in 2024), supported by scale effects and AI-driven average selling price increases. Three catalysts underpin the thesis: deeper agentic AI adoption, ongoing CPU and memory price hikes boosting demand for cost-effective HCI/SDS alternatives, and accelerating VMware replacement.

The second channel is enterprise software. Domestic vendors can now integrate DualPath-optimized inference — either natively or via cloud-based LLMs — to deliver better response latency, smoother multi-turn interactions, and more complex task handling at lower compute cost. HSBC argues this lowers the deployment threshold for enterprise AI adoption broadly, which should accelerate AI order growth across the sector.

Stock-by-Stock: Where the AI Orders Are Building

Yonyou Network Technology (600588 CH, Buy, TP RMB 21.00, 48% upside) disclosed that its AI order volume surpassed RMB 1 billion (approximately US$138 million) in 2025, exceeding 10% of annual revenue. HSBC projects AI orders could reach RMB 2.4 billion by 2027, representing roughly 20% of projected revenue — a trajectory the current share price does not yet reflect.

Kingdee International Software (268 HK, Buy, TP HKD 17.70, 76% upside) is the most deeply discounted name in the basket. HSBC projects AI product orders of HKD 698 million by 2027, implying a 50% CAGR from 2025 levels. The company is expected to return to profitability in 2025, with net profit reaching RMB 665 million by 2027.

Glodon (002410 CH, Buy, TP RMB 17.00, 22% upside) has upgraded its AI agent product to read and interpret engineering diagrams and calculate workloads autonomously. The company booked over RMB 100 million in AI-related orders in 2025. HSBC projects the AI agent market for construction cost estimation could reach RMB 10 billion by 2030.

Kingsoft Office (688111 CH, Buy, TP RMB 353.00, 17% upside) is the largest name in the group by market cap at roughly US$19.9 billion. WPS AI 3.0, launched in July 2025, has driven strong user engagement. The analysts see WPS 365 revenue growth as the primary earnings catalyst, supported by domestic software substitution tailwinds and improving enterprise AI adoption.

ThunderSoft (300496 CH, Buy, TP RMB 82.00, 12% upside) is positioned as the pick-and-shovel play on AI hardware integration. As AI glasses, smart vehicles, and IoT devices proliferate — ByteDance is reportedly targeting a Q1 2026 launch for a display-free AI glasses product — embedding AI capabilities into edge hardware requires intensive software engineering work. ThunderSoft's expertise in on-device model deployment and OS/chip adaptation makes it a direct beneficiary. HSBC forecasts a 49% CAGR in the company's smart IoT revenue from 2024 to 2027.

The Bottom Line

HSBC is not changing earnings estimates or price targets on any of the six names — this is a conviction-reinforcing note, not a revision. But the analytical message is pointed: DualPath is not an incremental tweak. It is another demonstration that DeepSeek can compress the economics of AI inference through software engineering alone, without requiring next-generation hardware. For China's enterprise software sector, every incremental reduction in inference cost is a demand multiplier — and the latest data point suggests that multiplier is still accelerating.