The history of domestic large models: The competition in large models enters the "post-brute-force computing era"
TechFlow Selected TechFlow Selected
The history of domestic large models: The competition in large models enters the "post-brute-force computing era"
The younger generation evolves fiercely, while the older generation forges ahead through adversity.
Navigating Web3 tides with focused insightsAuthor: KeyFrame
Image source: Generated by Wujie AI
The AI arena is quietly witnessing a profound shift in technological power.
The transformation triggered by DeepSeek shows no sign of slowing down. The large model competition has entered the "post-brute-force computing era," where efficiency has become paramount and AI power is being restructured, continuously challenging OpenAI's previously dominant position.
As new waves evolve aggressively and pioneers fight through challenges, the battle for supremacy remains undecided. The key to victory lies in leveraging open-source to gain ecosystem support while using closed-source strategies for commercialization.
01. Chinese AI Projects Surge with Policy Tailwinds
The development of domestic AI has been brewing quietly. 2023 is regarded by industry insiders as a watershed year for artificial intelligence advancement.
AI scientist Fei-Fei Li once said: "Historically, 2023 may be remembered for its profound technological shifts and public awakening."
Prior to that, technological exploration and innovation in artificial intelligence had already been extensive.
In 1956, John McCarthy first proposed the concept of "Artificial Intelligence" at the Dartmouth Conference, marking the formal birth of AI as a discipline.
However, by 1973, due to bottlenecks in AI research, funding for AI drastically declined, plunging development into a "winter period."
It wasn't until 1986, when "AI godfather" Geoffrey Hinton introduced the backpropagation algorithm, that the revival of neural networks brought renewed hope to AI. Then in 2017, Google introduced the self-attention mechanism, replacing RNN/LSTM and becoming the core architecture for subsequent large language models (LLMs)...
Looking back at China’s AI development journey, 2023 was also the inaugural year of the "domestic AI era."
According to Tianyancha, there were over 20 financing events directly related to large models in just the first half of 2023. More than 100 large models had been released domestically, and by July 2024, nearly 200 generative AI large models had completed registration and gone live.
To this day, only about a dozen players remain with a chance to reach the final stage. Consulting firm Frost & Sullivan指出指出指出指出that we now have fewer than 20 competitors in the general-purpose foundational large model field, primarily led by internet companies, cloud computing giants, and AI startups.
All are participants in this silent "war." Looking back from the beginning of 2025, perhaps after the rigorous filtering of the 2024 "hundred-model battle," DeepSeek was able to deliver a thunderclap across the global tech industry at the start of 2025, enabling domestic AI to achieve a "critical leap" and solidify its standing.
Enterprises with sustained innovation capabilities are gradually dominating the market, rapidly expanding AI applications from image-text generation to video creation and multilingual ad production.
Meanwhile, large models and agent technologies have entered an accelerated development phase. Whether optimizing user experience on the consumer side or delivering enterprise solutions on the business side, agents and large models are redefining how technology connects with society.
Currently, three major forces occupy the final arena: first, internet giants and cloud service providers such as Alibaba and ByteDance entering the large model space; second, national AI teams represented by iFlytek, adopting a G/B/C integrated approach to deliver both solutions and hardware products; third, AI startups like Zhipu and DeepSeek, a minority still committed to foundational model innovation.
Differentiation exists across upstream and downstream players in the industrial chain, with divergent development paths among model vendors. Even the "six small AI tigers" face strategic splits—Baichuan AI has shifted focus to vertical industry models like healthcare; 01.ai outsources super-large model training to Alibaba; Moonshot AI and MiniMax concentrate on consumer-facing applications and products.
Industry insiders generally believe that compared to upstream and downstream segments, midstream model vendors commonly face profitability challenges. In 2025, the number of finalists capable of innovating at the foundational large model level will further decrease.
02. From "Burn Money Faith" to "Efficiency Revolution"
If "cost, AI Agent, multimodality" are the three current keywords in the AI industry, representing the evolutionary direction of large models in 2024, they may also represent critical junctures toward industrial deployment.
First, cost is undoubtedly a decisive factor for corporate survival. Training and deploying large-scale AI models require massive computational resources, imposing high computing and operational costs on enterprises.
DeepSeek-R1 precisely addressed pain points around efficiency and cost control, achieving performance comparable to—or even surpassing—top-tier models despite relatively lower computational investment.
Traditional AI development often relied on a "bigger is better" logic, pursuing ultra-large models and massive compute clusters. DeepSeek R1’s lightweight model design and open-source strategy have lowered the barrier to AI adoption, promoting the proliferation of mid-tier computing infrastructure and distributed data centers.
Upstream chipmaker NVIDIA has begun facing pressure to adjust demand structure due to DeepSeek’s emergence.
ASIC chip manufacturers, meanwhile, are seeing new opportunities. Since ASICs can provide hardware acceleration tailored to specific AI applications, they offer clear advantages in energy efficiency and cost control, aligning well with the trend toward distributed computing.
On the computing service side, regional data centers are increasingly handling latency-sensitive applications such as intelligent quality inspection in manufacturing and financial risk control, thanks to their low latency and proximity to application scenarios.
Cloud computing giants like AWS and Alibaba Cloud are adjusting construction strategies for large data centers, increasing investments in edge computing and distributed computing infrastructure.
On the application side, reduced computing costs will benefit various sectors, accelerating AI penetration in manufacturing, finance, healthcare, and beyond.
On code-hosting platform GitHub, numerous integration cases based on the DeepSeek model (awesome deepseek integration) have emerged, creating a virtuous cycle of "demand driving supply" and enabling mutual empowerment between "computing power + industries."
AI technology will accelerate its infiltration into all sectors, becoming a crucial engine for industrial upgrading and economic growth.
Notably, DeepSeek R1’s technical breakthroughs, while lowering the threshold for AI adoption, may also trigger the "Jevons Paradox."
The Jevons Paradox, proposed by 19th-century economist William Stanley Jevons, observes that as coal efficiency improved, overall coal consumption actually increased. This paradox reveals a profound economic principle: improvements in efficiency do not necessarily reduce resource consumption; instead, lower costs and expanded applications can stimulate demand, ultimately increasing total resource use.
Microsoft CEO Satya Nadella’s reference to the Jevons Paradox to explain DeepSeek R1’s potential impact hits the nail on the head.
Nadella believes that more affordable and accessible AI technologies will lead to surging demand through faster adoption and broader application. As AI becomes more accessible, new application demands will emerge in areas previously constrained by cost—such as small and medium enterprises and edge computing scenarios—leading to exponential increases in computing call density.
The explosion of emerging application scenarios will further accelerate the fragmentation of computing demand. Cutting-edge fields like autonomous driving and embodied robotics require enormous real-time computing power, far exceeding the pace of optimization achieved by DeepSeek. Even if single-task efficiency improves several-fold, concurrent demands from millions of intelligent terminals will still create a massive black hole of computing consumption.
03. Synergy Between "Open-Source" and "Closed-Source"
With the explosive popularity of the open-source large model DeepSeek, keywords like "open-source" and "free" have become ubiquitous.
If before DeepSeek, Chinese large model companies were divided on whether to go "open-source" or "closed-source," now calls for "open-source," "open ecosystems," and expanding developer communities seem to dominate.
Under the disruptive influence of DeepSeek—the "mackerel in the pond"—domestic large model companies are displaying more "open" postures, aiming to accelerate the establishment of their own developer and application ecosystems.
Key differences between open-source and closed-source models can be observed across three dimensions: foundational conditions, technical aspects, and commercialization.
In terms of foundational conditions, open-source models rely on public datasets and community-contributed data, supported by distributed GPU clusters owned by developers, offering equal access opportunities for developers, researchers, and enterprises, thus promoting innovation and sharing.
Closed-source models are developed by companies or teams using proprietary data—such as user behavior logs, private databases, and cleaned public data—and users can only access these models via interfaces or platforms provided by the company.
From a revenue perspective, open-source models do not generate direct income but typically monetize through value-added services like cloud computing, technical support, training, and custom development. Companies build sustainable revenue streams by offering premium services built atop open-source foundations.
Closed-source models follow a more direct commercialization path, generating profits through licensing, subscriptions, and platform fees. These models offer high profitability, as customers pay for usage rights and services.
Open-source and closed-source are not mutually exclusive. In the future, we’re likely to see a symbiotic relationship where open-source accelerates AI dissemination and innovation, while closed-source ensures commercial viability and stability.
The future winners will be versatile players capable of mastering both approaches—leveraging open-source to gain ecosystem momentum while using closed-source to capture value.
As Nadella put it, "ultra-large-scale AI won’t result in winner-takes-all outcomes; open-source models will balance closed-source dominance."
Epilogue
DeepSeek will play a pivotal role in today’s AI era, much like Android did during the mobile internet revolution.
It will reshape industrial ecosystems, trigger chain reactions, and accelerate both upper-layer application development and lower-level system unification. This will mobilize cross-cutting ecosystem forces spanning hardware and software, upstream and downstream, prompting greater investment in co-optimization and vertical integration across "models-chips-systems," further eroding CUDA’s ecosystem advantage and creating opportunities for China’s AI industry.
Through technological innovation, DeepSeek has reduced reliance on high-end imported chips during AI model training, demonstrating a viable technical pathway for domestic firms and significantly boosting confidence in indigenous computing chip development.
The competition involves more than just a technical choice between open-source and closed-source—it's a struggle for话语权, market dominance, and control over computing resource allocation. This battle for AI power has already begun.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
