The Undervalued New Narrative of Nvidia: "In-House Open Source Models" and "Supply Chain Lock-in"

Nvidia is quietly building a moat on two fronts that have rarely been fully priced in by the market.

According to a new research report released by HSBC on August 20, analyst Frank Lee believes that in addition to Nvidia's continued "outperformance," the new catalysts driving a re-pricing of its stock will come from two previously overlooked narratives.

First, a strong bet on open-source small language models (SLMs), which could expand the potential customer base from a handful of hyperscale cloud providers to millions of developers and sovereign nations; second, by locking in supply chain capacity years in advance through a series of long-term procurement agreements, an advantage that will become increasingly apparent as competitors find it harder to secure key manufacturing resources.

The market significance of these two new narratives lies in their common core proposition: Nvidia’s growth engine is evolving from a structure dependent on a few hyperscale clients to a broader, more resilient customer ecosystem.

If this transformation is recognized by the market, it will be a key variable driving valuation re-rating.

The HSBC report points out that after the hype around sovereign AI and neocloud providers faded, Nvidia had yet to establish a new narrative strong enough to drive a significant stock re-rating—one key reason why its performance has lagged the Philadelphia Semiconductor Index so far this year. However, its strategic bet on open-source AI is filling this gap.

Nvidia is currently making aggressive moves in the open-source model space, positioning itself as the world’s largest contributor to open-source AI.

According to Nvidia’s disclosures, open-source models have now become the second most popular category by token generation. HSBC believes the strategic significance of this move is that open-source small language models (SLMs) are quickly becoming the go-to inference engine for intelligent agents and on-device applications.

SLMs are favored for three reasons:

• Latency and throughput advantages—Autonomous agents need to run frequent cycles of parsing intent, calling APIs, and evaluating results; every micro-step executed in a large frontier language model (LLM) creates a severe latency bottleneck. SLMs offer significant advantages in inference cost and throughput;
• Task specialization—SLMs excel at constrained, deterministic tasks. Enterprises are embedding domain-specific SLMs into software platforms to solve complex business issues;
• Edge deployment capability—Models with fewer than 1 billion parameters can be easily loaded into local GPU memory or edge devices, supporting localized operations.

Nvidia’s open-source product suite is already quite comprehensive, covering multiple core scenarios:

• Nemotron series focuses on inference and language tasks;
• Cosmos targets "physical AI" for robotics and vision fields;
• GR00T N1 is positioned as the world’s first open, general-purpose foundational model for humanoid robots;
• Alpamayo is dedicated to autonomous driving scenarios;
• NVIDIA Agent Toolkit and NeMo are used for building and customizing enterprise-level AI agents and generative AI applications, respectively.

HSBC points out this free, highly optimized open-source ecosystem is, in essence, Nvidia’s strategic lever to guide developers to prioritize running their AI applications on its hardware.

HSBC's report notes that the demand for AI compute continues to outstrip supply due to capacity constraints. Nvidia has systematically locked in key supply chain capacity in advance packaging, memory, optical devices, and energy infrastructure for 2026 through a series of multi-year agreements.

HSBC judges that supply chain constraints across multiple segments may intensify further by 2027. Nvidia’s advance procurement strategy will generate far greater competitive value than peers and is likely to earn a higher market premium.

According to The Information, in advanced packaging and memory, Nvidia signed a multi-year contract worth $1.5 billion with Amkor Technology in July 2026 to support expansion of advanced semiconductor packaging and testing capacity in Arizona, USA;

In the same month, Nvidia reached a comprehensive cooperation agreement with SK Group worth as much as $50 billion, covering joint development with SK Hynix on next-generation AI memory (including HBM) and plans to build a 2 GW Vera Rubin AI factory in Korea.

In foundry capacity, Nvidia has pre-booked 63% (2026) and 52% (2027) of TSMC's CoWoS-L advanced packaging capacity. This means GPU and ASIC competitors must turn to other suppliers for alternatives, facing potential yield risks with non-primary suppliers.

In optical interconnects, as AI network infrastructure shifts from copper to optical connections, Nvidia has signed multi-year strategic agreements with Lumentum and Coherent, each investing $2 billion to support R&D and the construction of US domestic manufacturing capacity, and securing future access to advanced laser components;

Meanwhile, it has also signed a multi-year commercial and technical cooperation agreement with Corning to expand US-based manufacturing of advanced optical connection solutions.

In energy and land, Nvidia is locking in critical assets by directly taking stakes in infrastructure developers. According to Bloomberg, Nvidia has invested in Cloverleaf Infrastructure, Lancium, and SB Energy, securing power resources to ensure its chips have the "places to run" in the future, and thus embedding Nvidia’s full hardware and software stack into the early design stages of related facilities.

Among these, Nvidia announced, in partnership with SB Energy and OpenAI, it has locked in land, power, and construction capacity at the PORTS-Pike Technology Park in Ohio. The initial design supports 4.25 IT-GW of AI factory capacity, with Nvidia’s cumulative payment obligation capped at $105 billion, plus a $1.5 billion investment in SB Energy.

In addition, Nvidia plans to invest $1 billion in NAVER to expand the "GAK Sejong" AI plant from 55 megawatts to 200 megawatts by 2028, with a longer-term goal of reaching 1 GW of sovereign AI infrastructure.

From the competition perspective, these investments are essentially a bundling strategy. Cloud computing giants and AI labs typically mix and match chip, network equipment, and custom cable suppliers. However, by taking stakes in infrastructure developers, Nvidia has gained an important lever to ensure future facilities are designed around its complete technology stack.