At the forefront of China's AI industry, DeepSeek has been thrust into the center of a dual gamble. On one hand, it is a leading domestic large model company, yet it heavily relies on NVIDIA and Huawei chips for training and inference. In the current global environment of tight high-end computing power and overlapping geopolitical risks, this reliance could turn from an advantage to a limitation at any moment. About a year ago, DeepSeek quietly initiated a self-research chip project aimed at model inference scenarios, and in the past few months has also privately recruited chip design engineers, attempting to use a path of “self-mastery of the critical lifeline” with its inference chip route to hedge against the price and supply uncertainties of external suppliers. If this pathway proves viable, it will significantly enhance its controllability of costs and resources. At the same time, in the first half of 2026, Chinese regulatory authorities held several meetings with top tech companies to discuss whether to restrict overseas users from accessing China's most advanced AI models. Related discussions were prominently disclosed by the media on July 7, yet no formal policy text has been established so far, necessitating all leading companies to pre-emptively simulate scenarios of "restricted exports." The technological route and regulatory direction have yet to take shape, but they have already formed a pair of scissors: one blade points to the technological autonomy of computing power and chips, while the other blade points to global compliance regarding cross-border access to models. DeepSeek's choice is merely a microcosm of a larger issue: how the entire Chinese AI industry can find an operational trajectory between these two blades that avoids being locked down or isolated.
From NVIDIA to Self-Research: DeepSeek's Hardware Breakthrough
Before the regulatory scissors have truly fallen, DeepSeek has already felt the pressure of another invisible knife – the supply of computing power. As a leading large model enterprise in China, its training and inference currently depend almost entirely on high-end chips provided by NVIDIA and Huawei. According to multiple Chinese tech and crypto media outlets citing Reuters and informed sources, DeepSeek's model operation is highly dependent on these two suppliers. In the context of tight global high-end chip supply and heightened geopolitical game, this dependence means that a single policy, a supply chain failure, or even a price fluctuation could rapidly amplify into delays in model iteration, uncontrollable inference costs, and uncertainties in service. DeepSeek's core capabilities are effectively locked within others' hardware stacks.
It is precisely under this pressure that, about a year ago, DeepSeek quietly launched its self-research AI inference chip project at an undisclosed time. Also according to Reuters and informed sources, this technical line clearly aims at model inference scenarios rather than directly replacing general-purpose training GPUs. In other words, it has chosen to cut into the closest link to users by optimizing the inference cost structure through self-research inference chips, keeping the most sensitive supply security in its own hands. In recent months, DeepSeek began privately recruiting chip design engineers to advance this project, quietly establishing a new technical base in the talent market without high-profile announcements, indicating that self-research chips have moved from conceptual proposals to substantial execution. Training continues to rely on NVIDIA and Huawei, while inference attempts to build a foundation on its own, allowing DeepSeek to carve a viable gap between technological autonomy and realistic constraints through this method of “self-rescue at the front end, pressure at the back end.”
Supply Chain Anxiety and Technological Sovereignty: The Pressures Behind Self-Research Chips
In a global landscape where high-end AI chips are long in short supply and export controls and quotas are tightening, any company that heavily relies on external computing power essentially hands over its lifeline to upstream supply chains. Over the past year, DeepSeek has been highly reliant on NVIDIA and Huawei chips in its training and inference processes. This dependence is not only a technical choice but also a systemic risk: once global supply tightens further, the shipping rhythm of a particular chip is adjusted, or new restrictions arise in the policy environment, the existing blocking pressures on Chinese enterprises regarding advanced AI chips will quickly transmit into specific computing cost and delivery capability issues. For DeepSeek, which requires large-scale inference to support product iteration, such external uncertainties mean that the limits of computing power and timelines may be defined by others, leaving it to passively queue up in the supply curve set by others.
In such a pressure field, self-research inference chips are designed as a crucial buffering layer. If this self-research route ultimately proves successful, DeepSeek will no longer be fully constrained by the prices and supply rhythms of external high-end chips in large-scale inference scenarios, allowing it to exercise more initiative in cost structure, capacity planning, and compliance assessment. On one hand, its own inference chips provide the company with a relatively stable “internal domain” during global computing power tightness, preventing its business model from needing frequent recalculation due to upstream disturbances; on the other hand, hardware self-control is itself a long-term emphasis of China's technology industry policy, thus self-research chips fall within a larger narrative of “technological sovereignty.” For DeepSeek, this is not merely an engineering optimization; in the dual context of tightening external blockades and internal regulations, it builds a layer of self-adjustable safety buffer in advance for future model capabilities and business boundaries.
Tightening Regulations Ahead: Overseas Users Face New Barriers
While DeepSeek quietly lays down a “hardware buffer zone” internally, the external institutional boundaries are also subtly shifting. In the first half of 2026, Chinese regulatory authorities convened top tech companies multiple times; according to Deep Tide TechFlow citing three informed sources from Reuters, one of the core discussions of these closed-door meetings was whether to impose restrictions on overseas users accessing China's most advanced AI models. By July 7, this regulatory trend was disclosed for the first time, although it has yet to be accompanied by any formal policy texts or implementation guidelines, it has sufficiently made overseas teams that rely on Chinese models realize that the current open landscape is not a given, but a channel that may be tightened at any time.
The concerns of regulators are not hard to understand: at the intersection of the risks of technological outflow and data security red lines, cutting-edge large models naturally find themselves in a sensitive zone. On one hand, model capabilities are considered key technological assets, and the degree of openness directly relates to China's technological leverage in the global AI contest; on the other hand, cross-border calls and data returns may touch security and compliance bottom lines, forcing regulators to think ahead about how to install “valves” for model exports and remote access. If relevant restrictions are truly implemented in the future, overseas users wishing to continue using advanced models from companies like DeepSeek will likely face more review and compliance barriers, and the global landscape of AI services will be redrawn beneath these invisible institutional boundaries. This potential restructuring has already become a long-term variable that the Chinese AI industry must confront.
Technological Self-Reliance Meets Export Red Lines: The Internal and External Game of Chinese AI
About a year ago, DeepSeek pressed the start button for its self-research AI inference chip internally, and in recent months has quietly stepped up its recruitment of chip design engineers, attempting to gradually pull “computing power lifeline” back from the external supplies of NVIDIA and Huawei to its own controllable realm. Almost concurrently, Chinese regulatory authorities convened top tech companies multiple times in the first half of 2026 to discuss possible paths to limit overseas user access to the country’s most advanced models. On one side, companies are seeking technological self-reliance at the hardware level; on the other, regulators are brewing new red lines in the model export process, which together constitute the “dual breakthrough” that the Chinese AI industry is pushing forward: it aims to grasp the underlying computing power while also building tighter safety gates for cross-border flows.
For companies like DeepSeek, these two forces will not remain mere strategic slogans; they will directly shape specific product strategies, overseas business rhythms, and cooperation models: under the current reality where self-research chips have yet to materialize and training and inference still heavily rely on NVIDIA and Huawei, the combination of hardware autonomy planning and policy discussions around export tightening means it needs to design different model versions for domestic and overseas markets more cautiously, evaluate technology integration paths and compliance costs with overseas partners, and recalculate “which capabilities can be opened and in what manner.” From a global perspective, this internal and external game will accelerate the potential differentiation of the AI ecosystem—beyond potentially more diversified and regionalized model sources, the access thresholds, review processes, and compliance costs surrounding China's advanced models are also elevating expectations. In the future, who can continue to access these models and at what cost will become a new dividing line in the global AI competition landscape.
A New Normal of Limited Openness: Chinese Models Move Towards More Restrained Globalization
Along the two main lines of self-research chips and export regulations, the long-term state of China's AI is being quietly rewritten: on one end, enterprises represented by DeepSeek continue to rely on NVIDIA and Huawei chips for training and inference while secretly advancing self-research chip projects aimed at inference scenarios and accelerating recruitment of design engineers in recent months, hoping to break the external chokehold of high-end computing power in the next stage; on the other end, regulatory bodies have convened top tech companies multiple times in the first half of 2026 to discuss whether to limit overseas user access to China's most advanced models. Although as of July 7 there are no public policy texts or implementation timelines, the uncertainty of “whether and how models can be exported” has been incorporated into all participants' planning. The result is that Chinese models are likely to shift from previous “open access” to more nuanced “conditional access”: overseas developers and companies will still have the opportunity to use these models but will need to meet more security and compliance threshold screenings, viewing access qualifications as an adjustable policy resource rather than a default right. Moving forward, whether the self-research inference chips truly work and whether regulations ultimately impose formal restrictions on overseas access will determine the extent of China's AI openness in globalization—regardless of how the path evolves, a “new normal of limited openness” that is more technologically independent and controlled is already beginning to exert a continuous and profound reshaping force on the global AI landscape.
Join our community to discuss and grow stronger together!
Exclusive Hyperliquid benefits for AiCoin: https://app.hyperliquid.xyz/join/AICOIN88
Exclusive Aster benefits for AiCoin: https://www.asterdex.com/zh-CN/referral/9C50e2
On-chain Telegram community: https://t.me/AiCoinWhaleData
On-chain community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin on-chain Twitter: https://x.com/aicoinwhaledata
Disclaimer: This article represents only the personal views of the author and does not represent the position and views of this platform. This article is for information sharing only and does not constitute any investment advice to anyone. Any disputes between users and authors are unrelated to this platform. If the articles or images on the webpage involve infringement, please provide relevant proof of rights and identity documents and send an email to support@aicoin.com. The relevant staff of this platform will conduct an investigation.