DeepSeek is a large language model developed by the Chinese company Hangzhou DeepSeek Artificial Intelligence Co., Ltd. While not directly state-controlled, it operates under censorship aligned with the Communist Party's ideological framework (Gorlla and Tuttle, 2025, p. 6–10) and emerged within a system where private firms—particularly in strategic sectors—are subject to intense political oversight (Li et al., 2020; Almén and Carlsson, 2025). Scholars have also highlighted that such state influence over commercial actors is typical of socialist regimes (Rivero Silva, 2022, p. 13–14). Thus, while DeepSeek cannot be classified as a state-owned AI, its development likely depended on government support, and its functions reflect political alignment with Party standards. As Feakin (2025) argues, it functions as an instrument of Beijing's soft power.

The Chinese state has shown minimal tolerance for technologies beyond its control—going so far as to attempt bans on cryptocurrency (Chen and Liu, 2022). In this context, DeepSeek's rise has occurred with clear ideological authorization. Accordingly, some authors describe it as a tool of “AI diplomacy” (Truby et al., 2025, p. 4), understood as the strategic use of AI to shape international relations and advance national agendas.

DeepSeek—particularly its widely downloaded version, DeepSeek-V3—relies on a specialized neural network named DeepSeekMoE, referencing its Mixture of Experts (MoE) architecture, whose main goal, is to increase model size without proportionally raising the computational cost per token. In simple terms, MoE is an architecture that allows the AI model to activate only a small subset of its components (or “experts”) for each input, instead of using the entire network every time. This makes it possible to increase the model's overall size while keeping energy and computation use relatively low. Thanks to this unique neural design, according to its developers, DeepSeek, may achieve fivefold lower computational overhead in training compared to other models of similar scale. DeepSeek MoE's architecture translates directly into markedly lower operational footprints. Its developers report that a 16-billion-parameter model performs only 39.6 % of the computations required by a comparable dense baseline; moreover, the same design delivers nearly 2.5 times the inference speed of a 7-billion-parameter dense model. These figures could imply proportional reductions in electricity draw, rack-level cooling, and embodied-carbon demand during inference. Taken together, these data suggest that DeepSeek MoE may curtail per-query computation by roughly 60 % and eliminate the need for multi-GPU clusters, whereas GPT-style dense models concentrate their largest environmental burden in resource-intensive training cycles and still incur higher per-generation energy costs. In life-cycle terms, the sparse-activation strategy therefore offers a materially more sustainable pathway for large-scale language-model deployment (Dai et al., 2024, p. 17–18).

The adoption of DeepSeekMoE, according to its developers, represent a significant competitive advantage for this family of AI models over ChatGPT's “Self-Attention” approach (Vaswani et al., 2017). While DeepSeek V3 handles a 671-billion-parameter language model, only ~37 billion parameters are activated per processed token. In other words, for any given query or input, the model would only use around 5% of its weight drastically which could significantly reduce the computational demands for each user request (DeepSeek-V3 Team, 2025, p. 3). Cai et al. (2025, p. 21) report persistent rumors that ChatGPT-4 may incorporate a variant of the MoE architecture, although it is unclear whether such an implementation would be partial or comprehensive, and the claim has not yet been independently verified. Consequently, the publicly available evidence still indicates that OpenAI's large-scale language models rely on transformer-based neural networks.

A further defining feature of DeepSeek is Multi-Token Prediction (MTP)—a decoding strategy that improves speed and efficiency during inference. In simple terms, rather than generating one word (or token) at a time, the model predicts several upcoming tokens in advance and then verifies which ones make the most sense in context. This approach significantly accelerates response time and reduces the computational load.

In practical terms, MTP can yield up to threefold faster response times in specific areas like coding (Gloeckle et al., 2024, p. 3) By its very nature, it entails lower computational demands and consequently saves energy. Rather than generating one token at a time, the model contemplates an entire range of possible sequences in advance and selects the most coherent answer. This approach was previously implemented, with moderate success, in other generalist models like Llama, specifically adapted for voice recognition (Raj et al., 2025). However, another school of thought urges caution about MTP's outcomes, noting that claiming to handle multiple, complex scenarios solely via straightforward autoregressive predictions might not be as effective as some advocates claim (Bachmann and Nagarajan, 2024, p. 9).

Either way, the data presented by DeepSeek's development team suggests that training DeepSeek V-3 required a total of 2,788 GPU hours on H800s. Assuming an approximate price of 2 USD per H800 GPU hour, the creation of this model cost around 5.6 million USD (Liu et al., 2024, p. 5). Building DeepSeek V-3 is DeepSeek R1 Zero, the next iteration in the DeepSeek family. This version uses unsupervised learning via an algorithm known as GRPO (Group Relative Policy Optimization). Operating within the broader scope of RL (Reinforcement Learning), GRPO allows for (a) training without human supervision and (b) no need for a “critical” model to assess whether a given response is “good or bad,” a method sometimes referred to as “semi-supervision.” Essentially, it establishes a system of rewards and penalties, sampling a range of answers to estimate the advantage of each one—whether that advantage relates to format (the style of the response) or precision (the substance) (Guo et al., 2025, p. 5).

DeepSeek's dataset would also have been meticulously refined for maximum efficiency. It features a corpus of 2 trillion tokens that, according to its development team, is periodically expanded. By and large, it comprises a C4 or Common Crawl (an automated crawl representing the “general knowledge” of the web) from which several operations are performed: (a) deduplication, to eliminate redundancy, (b) filtering, to remove toxic or trivial material, and (c) remixing, to ensure sufficient variety and balance in DeepSeek's knowledge base (Bi et al., 2024, p. 4–5). Moreover, the tokenization process—essential for converting dataset inputs into information the AI can process—employs the Byte-level Byte-Pair Encoding (BBPE) algorithm, a mathematical technique capable of creating consistent response patterns and therefore encapsulating more information in fewer tokens. In general, this speeds up the AI's data processing (Ghafari et al., 2024, p. 2).

Additionally, DeepSeek leverages another model-reduction technique known as knowledge transfer, which has demonstrated significant improvements in energy consumption and inference speed (Yuan et al., 2024). Specifically, DeepSeek employs a “teacher–student” inference approach. A large, comprehensive model is used to train specialized, streamlined sub-models, each containing only knowledge relevant to a particular function (Xu et al., 2025, p. 6). The goal is to produce hyper-focused models whose datasets exclude content that is not essential for their specific purpose. Accordingly, “mini” DeepSeek versions can be found for various use cases, such as math (Shao et al., 2024), medicine (Zhang, 2025), or coding (Zhu et al., 2024). This approach tackles both critical concerns in the AI race: (a) by offering numerous domain-specific models, Chinese users may eventually be more inclined to rely on DeepSeek for specialized queries, thus keeping data within national borders; and (b) because these are “lite” versions, as described next, they can run on personal computers and smartphones, thereby reducing reliance on large-scale data centers. Nevertheless, the proliferation of “mini” models can pose a clear disadvantage in multidisciplinary contexts, since their development would have to be specifically tailored to a single, well-defined task.

In short, DeepSeek according to its developers, could combine a range of cutting-edge techniques aimed at minimizing the size, consumption, and environmental impact of LLM-based AI models. Indeed, the distilled AI sub-models referenced earlier do not even require a data center to function—they can operate offline on ordinary consumer-grade computers, reducing electricity usage and e-waste generation compared to ChatGPT (Ball, 2025). In essence, DeepSeek promotes a decentralized AI that does not bear the heavy sustainability burden of data centers. For instance, Sapkota et al. (2025, p. 17), based on developers' work, estimate that DeepSeek R1 produces merely 3.3 percent of the carbon emissions attributed to ChatGPT-4 during the pre-training phase. Although no conclusive results are available regarding the online inference phase, preliminary evidence suggests that, as will be discussed below in section 3.1.2, the DeepSeek R1 online version may produce even higher emissions than ChatGPT 4.5.

Likewise, although no exact figure has been disclosed, scholars widely agree that the model achieves a marked reduction in computational demand—and therefore, electricity consumption—(Bogmans et al., 2025, p. 13) at least during the pre-training phase. Should one choose to run DeepSeek offline, some scholars argue that it is ideally suited to reusing compact systems, such as second-hand smartphones or laptops (Ngoy et al., 2025, p. 6), capable of handling lightweight AI models—something that, from one perspective, could contribute to a circular economy strategy by facilitating the reuse of electronic devices, thereby potentially reducing carbon emissions and e-waste. However, an alternative view raises concerns that the widespread offline use of DeepSeek on outdated or energy-inefficient hardware—especially in under-resourced settings, might ultimately result in unpredictable and suboptimal electricity consumption, thus offsetting any presumed environmental benefit. These contrasting interpretations point to an unresolved question: whether the offline deployment of AI on consumer-grade devices will in fact deliver meaningful ecological gains or merely shift the sustainability burden in new and less visible ways—a question that only longitudinal observation and further empirical validation may fully answer. Be that as it may, and irrespective of the inference phase involving offline user interaction, the scholarly consensus is that the production costs for DeepSeek (particularly the R1 version) have been substantially lower than those of its U.S. counterpart, ChatGPT, to the extent that while DeepSeek V3 was purportedly developed for under 5.6 million USD, ChatGPT 4 may have cost between 80 and 100 million USD (Krause, 2025, p. 3).

Researchers at the Massachusetts Institute of Technology maintain a skeptical view of the environmental advances claimed by the DeepSeek development team. First, Knight (2025) critically reflects on the fact that it is DeepSeek itself that has highlighted how inexpensive AI pre-training can be—characterizing this emphasis as a calculated move. This distinction is crucial: the development of AI models generally involves two main stages—(a) pre-training, where DeepSeek claims significant efficiency gains, and (b) inference (online or offline), which refers to user interaction or the execution of natural language processing (NLP) tasks. On this point, O'Donnell (2025) argues that the favorable environmental results associated with DeepSeek—as well as the genuine competitive advantage of the aforementioned MoE and MTP architectures, which is not disputed—lie specifically in the pre-training phase, not in inference. Indeed, it is worth underscoring that the results reported by Sapkota, Raza, and Karkee—indicating that DeepSeek produces just 3.3% of the emissions generated by ChatGPT-4—refer exclusively to the pre-training phase, not to inference. Similarly, the technical challenges involved in comparing carbon dioxide emissions across different AI models, such as DeepSeek and ChatGPT, should not be underestimated—especially given the lack of unified standards in the sector (Luccioni et al., 2024, p. 88).

During DeepSeek's inference phase, a different scenario emerges from that of pre-training. According to Jegham et al. (2025, p. 7), DeepSeek R1, during online inference, consumes ~23.8 Wh at maximum and 2.1 Wh at minimum to process an input of 100 words and generate an output of 300 words. For the same task, GPT-4.5 consumes a maximum of 6.7 Wh and a minimum of 1.2 Wh. In other words, DeepSeek R1, during online inference, is more energy-intensive than ChatGPT. However, when handling longer tasks—for instance, processing 7,000 input words and generating 1,000 output words—both models consume roughly similar amounts of energy: 33.634 Wh for DeepSeek and 30.459 Wh for GPT-4.5. This places them in a comparable range of computational demand. This observation is relatively significant, as it suggests two key points: (a) DeepSeek may be “cheap to train” in comparison to ChatGPT, but its R1 online version consumes even more energy and water during online inference than its direct North American competitor, ChatGPT-4.5. That said, available data suggests that DeepSeekV3 online—the version preceding R1—has an inference consumption profile relatively similar to that of GPT-4.5.

Regarding DeepSeek's offline energy consumption, the academic literature is limited. Nonetheless, some sources suggest that certain distilled versions exhibit especially low energy use and can run on systems with modest computational capacity (Chintalapudi et al., 2025, pp. 14–15), although no precise figures are provided. It is important to note that DeepSeekR1—the most recent version of the DeepSeek family—comes in several model sizes, ranging from a lightweight 1.5B (billion parameters) version to a 1.3T (trillion parameters) variant. The following table (Figure 1), relating to users who download the full DeepSeek model from the Hugging Face platform for offline use, shows a clear preference for the distilled 1.5B and 8B versions—respectively the first and third lightest models available. In relation to the above, Figure 2 below shows a comparison of the number of downloads attributed to each DeepSeek model in the Hugging Face platform.

This is not a trivial issue: enabling offline AI functionality on mid-range personal computers could reasonably support the reuse of older hardware. On this matter, some scholars (Gould et al., 2024, p. 1) point to increasingly shorter hardware obsolescence cycles, driven in part by so-called “planned obsolescence” and growing repair difficulty. While other authors advocate for “design for repairability” (Roskladka et al., 2025, p. 22) as a response to such trends, no clear balance has been reached to date—especially considering that ~70% of global e-waste ends up in developing countries. In this broader context, DeepSeek acquires significant geopolitical relevance, a topic to be explored later in this work. As product lifespans continue to shorten, countries such as Nigeria, Ghana, and India—with limited internet access—frequently receive discarded old or obsolete computers from Western nations (Lopes dos Santos, 2021, p. 58), often under the guise of donations that are, in practice, a form of e-waste export.

Consequently, a model like DeepSeek—extremely inexpensive to train, capable of running in distilled versions with as few as 1.5 billion parameters, and operable across nearly any mid range device—is well suited for countries that (a) have a surplus of outdated hardware and (b) lack widespread internet connectivity. In this sense, DeepSeek, In this regard, DeepSeek's offline development appears to have been a thoroughly strategic move, given that 25.6% of its population remains without stable internet access (Cui et al., 2024, p. 1). While it offers a meaningful sustainability dimension compared to ChatGPT, particularly in terms of pre-training efficiency and offline inference, it does not renounce online inference or the significant consumption of energy and resources in certain contexts. In other words, it facilitates a more sustainable form of AI suited to modest regions, outdated devices, or environments with limited internet access. However, it would be premature to characterize DeepSeek as a truly green AI. Importantly, this scenario partially complicates the initial sustainability narrative surrounding DeepSeek: the prospect of millions of outdated computers running distilled versions of DeepSeekR1—even offline—suggests a globally inefficient energy footprint, largely centered in the Global South, which lacks access to state-of-the-art hardware.

Indeed, the evidence suggests a clear bifurcation shaped by politically and commercially driven decisions, reflecting divergent strategic logics in how each state approaches the AI race. For instance, ChatGPT currently serves 300 million active weekly users and includes capabilities such as video generation and voice analysis—functionalities that, by design, entail a higher environmental cost. In contrast, DeepSeek, which deliberately avoids such high-intensity features, only reaches 38 million users. As Li et al. (2020, p. 3) note: “Generating videos consumes significantly more carbon than generating text: the average carbon emission for a single 240p video frame is equivalent to generating 78 text tokens with comparably sized video and text generation models.” Similarly, DeepSeek decision to offer distilled, offline-compatible AI suggests a reduction in user data collection. The absence of video generation—and other—features in DeepSeek, while indicative of a deliberate trade-off prioritizing environmental efficiency, may also reflect other factors, such as a potential lack of the American expertise required to develop such applications. As a counterpoint to the aforementioned, it should not be dismissed that such an environmental perspective may, to a greater or lesser extent, stem from China's strategic need to position itself competitively in the realm of sustainability, in light of its limited capacity to rival the advancements in artificial intelligence functionalities introduced by ChatGPT. Similarly, the development of DeepSeek could be understood as largely reliant on the foundational and pioneering work carried out by OpenAI through ChatGPT.

That said, these models, ChatGPT and DeepSeek are not monolithic representations of national AI policy. DeepSeek—or its environmentally conscious offline deployment strategy—should not be viewed as fully representative of China's overall approach to AI. Nor has the United States, as will be seen in the following section, entirely neglected the environmental implications of AI development. Rather, each model serves as an indicative example—valuable for analysis, but not exhaustive of national agendas. For instance, while DeepSeek promotes low-energy offline functionality, its R1 online version has been shown to consume even more energy than ChatGPT for certain inference tasks, further complicating any binary interpretation of sustainability leadership in AI. The use of offline models may exacerbate DeepSeek's existing security shortcomings, posing additional risks to both data security and the integrity of the model itself—for example, due to missing security updates that protect against cyberattacks or the fraudulent use of AI.

Nevertheless, it is worth noting that, despite its sustainability achievements associated with the pre-training phase and the possibility of running distilled versions offline, DeepSeek has not produced similarly promising outcomes in security. Using algorithmic jailbreaking, Kassianik and Karbasi (2025) launched category-based attacks—cyber-crime, disinformation, illegal activity, and general harm—against DeepSeek-R1 with the HarmBench dataset. DeepSeek-R1 failed every test, yielding a 100 % attack-success rate, while competing systems showed at least partial resilience. Crucially, these safety assessments were conducted only in English. When Zhang et al. (2025) switched the lens to Mandarin with CHiSafetyBench, they revealed additional weaknesses: DeepSeek-R1 achieved just 71.14 % accuracy in risk identification (vs. 91.13 % for the strongest baseline) and struggled to refuse sensitive prompts (RR-1/RR-2 = 67.60 %/67.17 %, compared with 77.71 %/77.27 %). DeepSeek-V3 offered modest gains in overall Mandarin safety (accuracy = 84.17 %) yet still underperformed at screening sensitive queries (accuracy = 66.96 %).

Likewise, the specialized DeepSeek version developed for biomedical applications and traditional Chinese medicine has suboptimal results, mainly because it cannot access real-time, specialized databases (McGee, 2025, p. 648). Finally, regarding “distilled” models, some authors (Lian et al., 2025, p. 13) contend that although these sub-models are powerful, they are not necessarily “superior” within their specialized domains. Likewise, it is worth noting that, although ChatGPT has introduced models that may be classified as “lite”—for example, the o4 mini—reliable information on the computational demand of most AI systems remains wholly opaque, thereby precluding any well-founded assessment of their purported advantages (Chen, 2025).

The political process by which China enters the competition in the AI race is certainly complex. On the one hand, we must acknowledge that, despite having a pronounced state ideology, the country is no longer that centralized and bureaucratized system once so closely mirrored by the Soviet Union (Molina, 2021, p. 15). On the contrary, China has evolved into a political system that has successfully adapted to globalization, viewing socialism as a means of political survival rather than an end in itself. Certain authors (Pieke, 2025) refer to this as “neo-socialism.” Others (Palmer and Winiger, 2019, p. 4; Li and Christophe, 2024, p. 10) propose that the application of expert knowledge to specific problems, rather than “assembly-style” decision-making, constitutes a form of governance characteristic of a “neo-socialist,” (and thus post-revolutionary), model exemplified by present-day China. This is a well-established doctrinal concept (Callahan, 2023, p. 15) that examines China's response to globalization as it seeks to secure its own international hegemony.

This neo-socialist perspective, coexisting with globalization, is largely what has led Communist China to embrace AI from a standpoint that, according to some authors (Zeng, 2021, p. 2–3), is directly tied to national security and set in an unmistakable context of confrontation with the United States. In this regard, influential figures within the Chinese sphere have pointed out that the AI race essentially encompasses two key factors: (a) computational and energy-related considerations, and (b) algorithms and data (Zeng, 2020, p. 1442). DeepSeek, by virtue of its computing attributes previously discussed, fulfills each aspect of this approach: it is national AI (requiring either no data center or relying on national data centers), sustainable (owing to very low energy and computing demands for pre-training phase and offline use), and affordable (having required only six million USD to develop), enabling China to compete with the United States in terms of efficiency rather than efficacy. In essence, it seems that China has recognized the clear environmental degradation associated with AI's technological development and its impact on public health (Anwar et al., 2018, p. 5)—particularly through e-waste and the contamination of cooling water used in data centers. Consequently, this reality has been considered in the national approach to AI within the broader context of its confrontation with the United States (Roberts et al., 2021, p. 47–49).

In line with the ethical framework proposed by Crawford (2021), it is evident that the seemingly positive environmental outlook surrounding DeepSeek requires critical scrutiny that significantly undermine its perceived sustainability. As previously discussed, the online version of DeepSeekR1, particularly for short-form responses, exhibits inference performance that is even more resource-intensive than that of ChatGPT 4.5. This suggests that the environmental ethic—closely linked to the efficiency of AI models—has not been fully internalized across the entire DeepSeek project. Rather, it appears to apply primarily to the pre-training phase and to offline inference. This combination ceases to be environmentally sound once any online variant of DeepSeek is introduced, given its reliance on data centers. In this light, while there is a discernible orientation toward sustainability, it would be inaccurate to characterize China's technological policy as inherently or consistently “green.” In fact, from a scholarly standpoint, the older or partially obsolete machines on which DeepSeek is expected to operate offline—whether in rural or impoverished regions, in countries with restricted internet access, or in authoritarian regimes such as North Korea, or in African states receiving e-waste—are themselves energy-inefficient. This dynamic, when assessed through the lens of Koomey's Law—recently revisited by Prieto et al. (2025, p. 11)—ultimately diminishes the environmental benefit that would otherwise be associated with distilled offline AI technologies.

In any case, China's approach to the development of DeepSeek cannot be understood without reference to the concept of “techno-nationalism,” originally introduced by Segal and Kang (2006) and later developed by scholars such as Kennedy (2013). At its core, this concept reflects a desire to develop domestically developed technologies, in order to reduce dependence on foreign intervention—particularly from the United States. In essence, it is about leveraging domestic innovation to meet internal needs, such as delivering AI services to public institutions, universities, or hospitals, even in regions where internet access is limited or where state-of-the-art computing technology is unavailable. Equally relevant here is the concept of ecological modernization, first articulated by Mol and Spaargaren and developed in later work (Mol et al., 2014, p. 2), where it is defined as “the social scientific interpretation of environmental reform processes at multiple scales in the contemporary world.” The framework suggests that globalization and the resulting interconnection of societies can act as a driver for environmental reform. In this light, the emergence of offline and distilled versions of DeepSeek should be seen not only as part of the global race for AI dominance, but also as a domestic effort to fulfill national priorities—such as data sovereignty and the ability to sustain AI services with limited resources. This reflection points back to a clear contemporary application of ecological modernization theory.

Another crucial point of the AI “consumption” from which China seeks to escape is economic in nature. Data centers are more or less sustainable depending on whether they use clean or “brown” energy, and that aspect directly affects their maintenance costs. According to the authors, Zhang et al. (2011) renewable energy drawn from the grid can be more expensive than brown energy. For example, industrial solar power can cost 16.14 cents per KWh in sunny conditions and 35.51 cents per KWh in cloudy conditions, whereas the wholesale price for brown energy can be around 6 cents per KWh. As for the carbon footprint associated with data centers, China also has a vested interest in this matter. Some scholars (Zhang and Liu, 2022, p. 12) estimate that CO₂ emissions related to this technology in China could be cut by 90% by the year 2060. This aligns closely with observations by other researchers (Li et al., 2023, p. 8–9), particularly regarding China's strategic approach to digital infrastructure. Notably, China launched a Three-Year Action Plan for the Development of New Data Centers (2021–2023) in July 2021, emphasizing the construction of green, low-carbon data centers and the accelerated adoption of advanced green and low-carbon technologies.

Consistent with this trend, Huawei—a pioneer in sustainable ICT practices in China—has introduced various strategies to reduce its carbon footprint. One notable example involves deploying its intelligent automatic cooling system, iCooling, in its data centers, which could well be available to DeepSeek. This technology has reduced total power usage at cooling stations by ~8%−10%, yielding energy savings equivalent to 3.85 million kilowatt-hours—comparable to planting 79,500 trees (Cao et al., 2022, p. 12). Additionally, Huawei is actively involved in transitioning toward renewable energy. The company prioritizes the use of renewable sources across its operations and is expanding its photovoltaic (PV) infrastructure on its campuses. In 2020 alone, these installations generated 12.6 million kilowatt-hours of electricity.

From a strategic perspective, it may be counter-productive for the People's Republic of China to enter a scale-driven contest over AI efficiency if doing so would exacerbate environmental degradation, impose public-health externalities, and amplify national energy demand. A more sustainable course would be to relax strict data-centrality in favor of offline or Small Language Models (Wang F. et al., 2024)—DeepSeek “lite models” serving as a leading example—whose markedly lower computational requirements, may mitigate both carbon emissions and operating costs. Such an approach would leave the rising energy intensity and attendant ecological footprint of large-scale, cloud-based systems such as ChatGPT to be borne primarily by competing U.S. platforms, thereby reallocating the sustainability burden without compromising China's long-term technological ambitions. Such energy consumption is, according to some recently published studies (Hosseini et al., 2025, p. 2), inexorably destined to occur. Notably, training GPT-3 reportedly consumes about 1,287 MWh of electricity and emits ~552 tons of CO₂, OpenAI's GPT-4 training is said to have used about 6% of the water consumed by West Des Moines, Iowa (population 75,000), and xAI's training lab (responsible for Grok) uses as much power as 80,000 households.

The previously mentioned efficiency-driven approach, embodied by DeepSeek, carries an environmental dimension that—albeit with many nuances—differs substantially from the paradigm of efficacy supremacy, which places particular emphasis on data accumulation. Returning to the internal power struggles as a defining feature of international races, the American arena appears tumultuous in this respect. While DeepSeek offline is championed in China as a cornerstone of future sustainability, in the United States multiple AIs vie for national leadership. Although some stand out—like Google's Gemini or Microsoft's Copilot—Grok, created by magnate Elon Musk, has garnered special attention, seemingly addressing an “anti-woke” discourse (de Carvalho Souza and Weigang, 2025, p. 2). In other words, there is an endeavor to amass as much data as possible: ChatGPT for professional applications (such as customer service chatbots) and Grok for social media interactions (DR and IS, 2025, p. 5). Indeed, certain studies suggest that Grok may be more accurate than ChatGPT in some responses, chiefly those related to social media-derived context (Yeasir Fahim, 2024).

Notwithstanding the above, while the Chinese approach has often been associated with an environmental sustainability agenda, it would be inaccurate to equate the so-called “American perspective” with a wholesale disregard for environmental concerns in the development of AI. At the governmental level, the recent passage of the Artificial Intelligence Environmental Impacts Act of 2024 (U. S. Congress, 2024) is particularly notable. This legislation acknowledges the environmental risks posed by AI and mandates a formal report on the matter by the Administrator of the Environmental Protection Agency. Likewise, attention should be drawn to the Executive Order on Advancing United States Leadership in Artificial Intelligence Infrastructure (Biden, 2025), in which former President Biden outlines that national AI development must be guided by three key priorities: (a) achieving supremacy over adversaries in the field of AI (with a veiled reference to China), (b) the economic dimension of AI, including the protection of domestic jobs, and (c) addressing environmental impacts.

Finally, the recent report issued by the U. S. Government Accountability Office (2025 p. 11) explicitly addresses the significant consumption of natural resources linked to AI systems. However, the report also notes that this issue remains “uncertain” at present due to persistent challenges in measurement and standardization. In short, the U.S. government has acknowledged the problem—although, to date, companies like OpenAI continue to roll out increasingly complex and resource-intensive functionalities, such as video generation, with limited legislative limitations. Paradoxically, environmental restraint in AI development appears more visibly embodied—through DeepSeek's offline model—in China, despite the fact that, according to some scholars (Zhang, 2025, p. 10), the country's regulatory framework, regarding environment protection, remains relatively weak or, at best, “relaxed” in certain respects.

Likewise, it is important to bear in mind that the consumption of natural resources in online AI systems depends not only on the algorithmic architecture of the model itself, but also on the data centers that host its inference capacity. In this regard, it is worth noting that, with respect to Azure servers, Microsoft is currently developing a server model known as “GreenSKU” (Wang J. et al., 2024, p. 2), which is purported to reduce carbon dioxide emissions by up to 8% across the entire data center infrastructure. Similarly, in the case of AWS servers, Amazon is reportedly experimenting with the integration of renewable energy sources, such as solar and wind power, to reduce the environmental impact of its data centers (Raza et al., 2024, p. 31). In addition, AWS has published a set of best practices under the title Cloud Sustainability Pillars (Amazon Web Services, 2025, pp. 2–5), which constitutes a clear statement of intent in this domain. That said, as of today, there is no independently verified evidence to confirm that these measures have been successfully and systematically implemented within Microsoft and Amazon's data centers. Therefore, it can be said that both the U.S. government and the major corporations responsible for operating the country's data centers have taken some initial steps—at the very least—to acknowledge the unsustainable consumption of natural resources associated with AI. However, this growing awareness has not prevented the nation's leading AI operator, OpenAI, from continuing to increase its weekly user base—reaching the previously mentioned 300 million weekly users—and offering increasingly complex generative AI capabilities, such as video creation through “Sora.” These developments significantly raise the energy and resource demands required for the routine operation of ChatGPT.

It is worth noting the link between the data accumulation interests of systems such as Grok and ChatGPT that it has led to the creation of “smart toys,” i.e., toys with integrated AI. A recent study (Pavliv et al., 2024, p. 172–175) demonstrated that these toys “listen” even when they are not interacting with children, raising serious concerns regarding the data they collect, and the likelihood of those data being used for commercial purposes Ultimately, the sustainability perspective of DeepSeek is most clearly manifested in environmental terms. ChatGPT and Grok do not allow their models to be used offline (Pham et al., 2024, p. 1–3). That is, they are permanently tethered to a data server, which has prompted some authors to describe the computational demand of such systems as “prohibitive,” particularly for small businesses (Hussein et al., 2025, p. 1; Joublin et al., 2023, p. 1). This is largely explained by its independence from large data centers—that is, through offline or “mini” offline models, plus the possibility of deploying online models on ordinary consumer devices like mobile phones or tablets, thereby eliminating reliance on so-called “mega data centers,” upon which the U.S. still depends. Each of these “mega data centers” hosts hundreds of thousands of servers and can draw tens to hundreds of megawatts of power at peak usage (Zhang et al., 2011, p. 2).

China's steps toward cheaper, offline versions—even at the cost of relinquishing some data control—suggest that the divide between the two models remains palpable. Although both nations remain interested in data control, China has recognized the untenable nature of accumulating data at any cost. It thus follows that this is a long-term race in which the strategic advantage will not necessarily belong to the AI boasting the most accurate answers or the largest user base. China has realized that genuine strategic leverage lies in attaining a level of sustainability that allows it to maintain a national AI presence, collect data as needed, yet avoid depleting its lakes or being burdened by an insatiable computing and energy demand. Put simply, it seeks to benefit from what AI has to offer while refusing to be overtaken by an international race that is, without question, already underway.

In relation to the above, the following Table 1 presents a SWOT analysis comparing two perspectives: one—albeit with nuances—centered on efficiency, as represented by DeepSeek, and the other—also nuanced—focused on effectiveness and the development of new functionalities, as exemplified by ChatGPT.