Hu YanpingShanghaiHai FinanceUniversity Distinguished Professor Author of the Intelligent Economy Series Report

Some time ago, "People Miss DeepSeek" hit the screen. The article mentioned one point - DeepSeek has promoted the cost reduction of global large models, allowing users and industries to enjoy cheaper Tokens.

The key issue is that the crazy "Token burning" of intelligent applications such as "Crayfish" has once again pushed up the user's cost of use. In this case, the important task of promoting cost reduction and efficiency improvement in the entire industry has fallen on the shoulders of DeepSeek.

Calculating the time, it has been more than a year since the release of DeepSeek V3 and R1. The outside world originally expected that DeepSeek V4 would explode during the Spring Festival this year, but in the end their hopes came to nothing. However, judging from a series of developments such as recent downtime and the launch of expert mode, it feels like DeepSeek V4 may be getting closer to us.

So, this may also be the last time to "urge" DeepSeek.

In this reminder letter, I want to talk to those friends who miss DeepSeek about the narrative of China's AI, the wave of technological evolution, ecological competition, and Token economics.

China’s AI narrative has changed

During the Spring Festival of 2025, DeepSeek R1 will be unveiled with low cost + high performance + open source, and its release will be its peak. It not only dominates the field of domestic large-scale models, but also becomes popular around the world. Internet platforms, IT giants, and all walks of life have joined in and embraced open source. All kinds of DeepSeek all-in-one machines are trying to steal the show.

During that time, when it came to Chinese AI, DeepSeek was definitely the name. It is no exaggeration to say that grandparents on the street were probably talking about or even using this domestic AI assistant.

However, in the past year, the artificial intelligence industry and China's AI narrative are no longer the same: the "China Group", "China Chain" and "China Ring" surrounding artificial intelligence - three narratives are intertwined and formed. The AI ​​China narrative supported by DeepSeek alone has lost its color.

So, looking at large models and artificial intelligence from this perspective, what we lack is not only computing power and electricity, but also time windows.

About "Chinese Group", I summarize it as "(3+1)+6+N", among which "3+1" refers to the four major manufacturers, corresponding to Byte, Alibaba, Tencent and Baidu. The latter three are the three giants in the Internet era, known as BAT. The number "6" corresponds to the "Six Little Tigers" of the big model era - Kimi, Zhipu, MiniMax, Step Star, Baichuan, and Wall-Facing Intelligence. The main ones have completed the launch or are in the process of being launched while DeepSeek is immersed in self-research.

"Six Little Tigers" originally included Kai-Fu Lee's Zero One and All Things, but in the first Hundred Model War, Zero One and All Things began to fall behind, so here we put the wall-facing intelligence in, but in fact, Baichuan's voice has gradually weakened in the past year or so.

"N" is actually not just one company, it corresponds to other vertical models and AI companies in the professional market.

A total of 10/category companies form the leading position in China's large model industry. They are no longer scattered soldiers, but an industrial legion with cluster competitiveness. They are also the opponents that DeepSeek must surpass on its way to becoming a god again.

Growing simultaneously with the "China Group" also includes the "China Chain" - from chip computing power, clusters/clouds, data corpus, algorithms/models, agents, and AI application development ecology, it has completed the entire chain and has become the only country in the world with a full industrial chain of smart technology. It is expected to provide an additional option for global smart infrastructure, and it is also hoped to provide new public goods for global smart inclusiveness through the output of capability economy.

Don't doubt this. DeepSeek R1 did establish a Chinese model overseas as a brand back then, but now manufacturers like MiniMax are also doing great business overseas.

As for the "China Loop", it covers three aspects: industry, application and investment - the industrial closed loop from AI to AI4S to modern industrial groups, the closed loop of market applications from AI technology to thousands of industries and hundreds of millions of households, and the capital closed loop from early investment to listing and exit. The initial formation of the closed loop not only means that artificial intelligence is running through China, but also means that the large and small cycles at different levels of the smart economy are connected.

From groups, chains to rings, China’s AI narrative has changed.

Regardless of the company's free marketing strategy, since the beginning of 2026, Liu Xiaohu's model has successively led the share of Token consumption on international platforms such as OpenRouter, and has an overall share of more than half, mainly overseas users.

In summary, China's open source power has changed the global artificial intelligence development pattern in 2025. By 2026, China's artificial intelligence development will enter the stage of exporting capabilities.

From the perspective of the global large-scale model and artificial intelligence industry, the diversification of technological paths has enhanced the vitality of talent flow and is conducive to the resilience of the supply chain. For downstream application developers, the existence of multiple optional suppliers means stronger bargaining power and lower lock-in risk.

In China's AI narrative, another good phenomenon is that the market has not been monopolized by a few oligopolies. This is a good thing for competitive innovation and talent ecological construction, and is also conducive to the formation of cluster advantages in the Sino-US AI competition.

The four waves have passed

Chinese classical mythology always says that "a day in heaven is a year on earth." In this year of DeepSeek's "interruption", artificial intelligence has passed four waves - programming, multi-modal, agent, OpenClaw (crayfish).

When AI programming tools such as GitHub Copilot, Cursor, and Claude Code swept the developer community, in the story of Vibe Coding, it was difficult for people to remember the existence of DeepSeek, although it was also used in programming scenarios.

Programming, the underlying driver of artificial intelligence sweeping all industries and the scene most in demand by developers, is now firmly occupied by Anthropic and others abroad, and has become a battlefield for Kimi and others in China.

In the multi-modal wave, Gemini 3 Pro and others have performed well in the fields of visual understanding and image generation. What everyone can remember is Nano Banana, and in the field of video generation, it is Byte's Seedance 2.0.

DeepSeek seems to be a slow player. It was not until V3.2 that it started grayscale testing of millions of Token contexts, and multi-modal capabilities have not yet arrived.

Some people say that in the field of large models, once the technical route of a generation of products is wrong, an era will be missed? It's hard to say whether DeepSeek is stuck here.

The third wave is Agent-multi-Agent-swarm intelligence. Compared with the understanding and dialogue capabilities of AI assistants, Agents have evolved to the execution level, shifting from "answering questions" to "problem solving" - in the past it was "passive response", but now it is "active execution". In this wave, the emergence of products such as Manus marks that AI Agent is moving from concept to implementation, and Kimi Agent Swarm has pushed this wave to a climax.

DeepSeek is used more as a model in this wave, not as a builder of the Agent ecosystem. The model itself has limited support for Agents, tools, and codes.

When the time comes to 2026, the wave of mobile intelligence represented by OpenClaw and other types of Claw, Claude Code, Claude Cowork, etc. has begun to appear. Their capabilities have actually surpassed the Agent level and become a takeover application operating system - application AI OS.

However, products such as OpenClaw are also nicknamed "Token black holes", and their token consumption for a single task is dozens or even hundreds of times that of traditional conversational AI. This high-input, low-output model faces sustainability problems in industrial-scale applications. The product itself is rough, unstable, and has multiple destructive version iterations, just like a rough house.

So it is not surprising that some people are shouting "People miss DeepSeek". After all, it has disappeared in several waves. After all, people need it to promote cost reduction and efficiency improvement of China's large models.

But what must be said is that OpenClaw confirmed that the logic of applying AI OS and universal mobile intelligence is established, and the time has come. It tells everyone that AI is no longer just a tool, but can be an all-powerful takeover agent.

So during the "National Shrimp Farming" trend in March, you see how quickly everyone copied the work. In order to promote local products, everyone began to send "cyber eggs", because OpenClaw allowed major manufacturers including Anthropic to instantly understand that the all-in-one application OS and mobile intelligence are right in front of you. If you have a brain and can perform tasks, it is not easy to grow limbs and become a general-purpose intelligent agent!

It is also for this reason that Anthropic has the fastest reaction and counterattack, and it also has the greatest impact on Claw. Claude Code is outflanking OpenClaw, and other major manufacturers are quickly copying the work of Claude Code and OpenClaw. This is what is happening right now.

The reason why military strategists must compete is because the entrance status, huge value and future ecological dominance of this matter are no less than the model and no less than the previous three waves.

If large models are accumulating power, multimodality is broadening scenarios, and Agents represent sowing seeds, then large-scale harvesting of the ecology relies on the application of AI OS and universal mobile agents. Now it has more or less the meaning of the end and the shadow of the ultimate form. When it reaches the stage of EI endogenous intelligence and II autonomous intelligence, it may be a different matter.

However, in terms of today’s input-output ratio of OpenClaw, it may not be the one that can occupy the ecological niche of AI OS and general mobile intelligence.

So in this last reminder letter to DeepSeek, we also want to ask a question: DeepSeek, which did not jump into these four rivers immediately, is choosing to accumulate strength, hoping to "make a big one" through V4 and subsequent base models?

However, the market never waits. Users' attention, developers' enthusiasm, and capital flows are all diverted in wave after wave. In the fields where these four waves are located, the competition threshold has been sharply raised, and the cost of ecology has also increased significantly.

DeepSeek’s story can only stop at the Spring Festival in 2025?

Full ecological competition is coming

My previous view was that leading companies have reached the stage of full ecological competition. At this stage, full-stack AI capabilities are the basis for the next giant competition. The best example is Google.

The reason why Google received great attention in the Gemini 3 Pro wave is because their accumulated "thickness" advantages gradually emerged in four aspects: model principle, force, lasting evolution(Evolutionary Index), data depth(Data Index), full-chain ecological breadth(Ecological Index), intelligent connectivity(Connectivity Index).

Google CEO Pichai has been in office for almost 10 years. In the interview just now, he recalled the unforgettable past when Transformer was preempted by ChatGPT. However, he does not think that losing the first-mover advantage will result in loss. He summarized Google's advantage as full-stack vertical integration.

So with Gemini 3 Pro, based on this full-stack integration, Google has made a beautiful turnaround.

We can make a bold prediction. In 2026, the competition at the head of the US artificial intelligence industry will probably take the lead. Anthropic will take the lead first, followed by Google, and OpenAI, which is ahead of the competition, will face a situation of double-teaming. In the end, the top four will become the top three, and the one that falls behind will be Grok, which has further widened the gap.

In the early warm-up phase of GTC in 2026, Huang Renxun rarely wrote an article and proposed the "five-layer cake theory": energy → chip → AI infrastructure → model → application.

But if you want to break it down in more detail, the competition in artificial intelligence is also reflected in chip computing power, data corpus, model base, development tools and developers, agents and tool skills, and application services. The loss of each link may lead to a decline in overall competitiveness, and the threshold for competition and investment has become a heavy asset game with tens of billions or hundreds of billions of dollars.

Innovation is no longer limited to "overtaking on corners", but also lies in system competition and system confrontation. In particular, factors such as capital, computing power, algorithms, and data that large models rely on have become key decisive factors. Taking a powerful pill or a bowl of sea cucumbers will not solve many problems.

DeepSeek, in the landscape of full-scale ecological competition, is based on the principle of generating force - bottom-level breakthrough - although it still has advantages, its shortcomings are also obvious: it lacks the support of the industrial ecological chain of IT giants, product application functions are relatively thin, and multi-modal and Agent ecological construction needs to be strengthened.

The Rise of Token Economy

The Token economy is gaining momentum in the new year. The Token economy is a value closed loop of the smart economy as a capability economy. This is my view in the interview with CCTV.

In the past, in the industrial era, the energy unit was kilowatt hours, in the digital era the traffic unit was GB, and the supply unit of capability products in the intelligent era was token. Token makes AI's "capabilities" become measurable, priceable, and tradable commodities.

You can understand it this way: Token has become the "settlement unit" that connects technology and business, thus forming a business closed loop of capability economy.

Token consumption is expanding at a geometric growth rate - China's average daily Token calls have jumped from 100 billion in early 2024 to 140 trillion in March 2026, an increase of more than a thousand times in two years. The more consumed, it represents the vigorous development of the capability economy.

For an enterprise, achieving an increase in gross profit margin through price leverage means that part of its profit model has been fully operationalized.

However, Token is a unit of measurement, not a unit of mass. The industry cannot just look at the number of Tokens, but also pays attention to the “quality of capabilities” behind them. Therefore, I think the differentiation of the Token economy will be obvious in the future - Tokens with high capabilities make money, while Tokens with low capabilities lose money, and the latter may even be eliminated.

So, when Xiaomi's Luo Fuli "brought goods" to the MiMo large model package, she said: "The current global computing power supply can no longer keep up with the demand for tokens created by Agents. The real way out is not cheaper tokens, but co-evolution - the collaboration of a more token-saving Agent framework and a more powerful and efficient model."

There is a very typical trend this year. Users are shouting that Tokens are expensive while at the same time paying for Tokens. In essence, part of the Token consumption is converted into productivity. When Token payment becomes a trend, companies can gain revenue before investing in the research and development of higher-level models. This is to create blood for the smart economy.

The two most direct ways to commercialize models and agent companies are: either relying on paid subscriptions to generate revenue, or using APIs to generate revenue through Token rate packages. There are too many uncertainties in OpenAI's approach of associating ads with AI assistant conversations, and no other company in the industry has followed suit.

I think that in the era of reasoning-driven Token economy, I think there are three types of scenarios that are the first to succeed: high-value and high-density scenarios(such as financial risk control, medical diagnosis, customers are willing to pay a premium for "no mistakes"); High-frequency and high-demand scenarios (such as intelligent customer service, code generation, cost dilution based on scale); and scenarios where Agent agents are widely used.

In the future, Token will become a basic service like water and electricity, with low profit, inclusiveness and ubiquity. The unit Token cost will continue to decline, but the Token economy will be stratified: Tokens with conventional capabilities will tend to make small profits and win by volume; Tokens with high capabilities and high value may continue to maintain a premium.

More specifically, companies that can build a scenario + data + platform + model closed loop and provide high-value agent services will receive a premium.

DeepSeek, who has a background in quantitative investment, is not short of money, but from the perspective of sustainable development, it also needs to embrace the Token economy.

The open source ecosystem is waiting for the third turning point

Over the past year or so, the open source ecological landscape has changed.

In early 2025, DeepSeek completed its first detonation of the open source ecosystem. At the beginning of this year, OpenClaw completed its second great assist to the open source ecosystem. The first detonation caused some closed source models to move closer to open source. Domestic giants such as Baidu joined the open source camp, and overseas such as OpenAI, Google, etc. are also increasing open source efforts.

According to the OpenRouter platform's analysis of 100 trillion Token call data, the market share of the open source model has climbed to 33%. The sudden rise of China's open source model is particularly eye-catching. At one time, five of the top six companies in the OpenRouter platform were Chinese open source models.

The rise of the open source model is the result of technology iteration, user needs and economic factors. The core motivations for enterprises to choose open source models have become very real: closed source API costs are strongly related to call scale, and marginal costs are uncontrollable; self-hosted open source models have significantly reduced unit costs in high concurrency, long context, and agent scenarios.

To put it bluntly, as long as the capability is online, the more open source models are used in privatized deployment scenarios, the cheaper they will be. As a disruptor in the open source model ecosystem, DeepSeek will most likely boost the open source industry again in 2026.

This expected promotion covers the industrial impact of computing power costs, the detonating effect of the user market, the stimulating effect of the open source ecosystem, and the boosting effect on market confidence, etc., and may appear again.

This is the underlying logic why people miss DeepSeek, and the price is just an appearance.

Although open source is good, there is still a long way to go to build it.

For DeepSeek, it also needs to form a developer ecosystem as soon as possible, support the Agent development ecosystem, and establish Apps and skill encapsulation and distribution channels similar to Skills to improve openness and flexibility while attracting more developers to participate.

Looking forward to DeepSeek becoming a key player in the open source ecosystem again.

Expectations for V4 are not the standards of the past

The suspense on the other side of the ocean is how far the next generation models of OpenAI and Anthropic can reach, whether Super App can become an application OS and universal mobile intelligence like the ecological Claude Code, and who is the fastest in Coding, the ecological bottom knife. These three things will affect the direction of this year's trend.

Judging from the current situation, Anthropic's fire has almost reached OpenAI's base camp. This can be seen from the financial data of the two companies disclosed by the Wall Street Journal. Anthropic may turn losses into profits before OpenAI.

In this context, what can we expect from DeepSeek?

To summarize the previous points, it should include V4, R2 to achieve intergenerational leap, a context window of 1 million Tokens (just started grayscale testing), native multi-modal capabilities, and a trillion-parameter level basic model should be the most basic starting point.

However, these are past standards and should not be the upper limit of the capabilities of V4 and R2. In today's time period, what DeepSeek needs is to achieve breakthroughs in multi-agent capabilities, tool usage, computer operation, and the super strong coding capabilities behind it.

There is no need to be overly anxious. Although AI Agent is very popular, it is still at the stage of integrating capabilities and is still far from a truly autonomous agent.

In the future, AI Agent may take four paths: cloud virtual machine integration, device-side hybrid mode of local and cloud collaboration, intelligent interconnection through protocols, or reconstruction of all high-frequency application portals in the form of a "super OS". No matter which path is taken, it will eventually become the hub of personal intelligent services and the strategic commanding heights of future competition.

The old standard no longer matches DeepSeek V4, so in this update letter, I expect it to be not only a more powerful language model, but also an intelligent agent base that can autonomously perform complex tasks, integrate multiple tools, and interact efficiently with the external environment.

As mentioned before, we expect it to be "big", and DeepSeek's actual exploration of model principles and product technology progress seem to confirm this "big" rhythm.

Since October last year, DeepSeek has continued to accelerate its paper releases and partial product updates in the field of large models, forming an intensive innovation rhythm.

From the release of DeepSeek-V3.2 in December 2025, to the release of three core architecture papers such as MHC, Engram, and DualPath in January 2026, and significantly updated and expanded the previously released R1 technical report, the overall research and development has shown a three-dimensional advancement trend covering architectural innovation, reasoning efficiency, multi-modality, and agent capabilities. This series of work is generally regarded as the technical prelude to the next generation flagship model DeepSeek-V4.

DeepSeek has not officially confirmed how these innovations will be integrated into V4's final architecture, but the paper's author signatures (including founder Liang Wenfeng), code leaks, and visible changes to the platform all point in this direction.

The DeepSeek-OCR series in October 2025 explores the possibility of compressing text information through visual representation, subverting the traditional assumption that "text tokens are more efficient than visual tokens". The visual causal flow mechanism of OCR 2 further allows the model to "understand" documents based on layout logic like a human being, rather than mechanical scanning. This provides a new way of thinking for multi-modal models to understand and process extremely lengthy documents (such as entire books, entire financial reports). It is expected to expand the context window of large models to tens of millions of tokens without having to bear the square-level increase in computational complexity.

mHC technology targets the fundamental problem in trillion-parameter-level model training: signal explosion, breaks through the bottleneck of large-scale development of "deep network stability", and paves the way for the training of trillion-parameter-level open source models. It also helps to achieve deep model expansion through architectural innovation without relying on advanced process chips.

Engram is oriented to engineering solutions for long context and continuous learning. Its conditional memory mechanism theoretically supports cross-session persistent memory, breaking through the limitations of "stateless" reasoning in current large models, and reasoning efficiency is no longer hindered by knowledge density. It challenges the traditional Transformer design paradigm of "exchanging calculation for memory". This method stores static knowledge in an external sparse table, allowing the model feed-forward network to focus on dynamic reasoning. This "neuro-symbolic" hybrid architecture allows the model to significantly reduce inference costs while maintaining million-token-level context.

The V3.2 version in December 2025 has initially demonstrated the ability of "cross-tool memory retention", solving the problem of traditional AI Agent losing the reasoning chain when calling multiple tools, and using the sparse attention mechanism to reduce the cost of 128K long context reasoning several times and reduce memory usage by 70%.

In addition, DeepSeek, together with Peking University and Tsinghua University, released a new paper to launch the agent reasoning framework DualPath. Aiming at the storage bandwidth bottleneck of agent large model reasoning, it innovates a dual-path KV-Cache loading mechanism, allowing data reading and GPU calculations to be paralleled, completely solving the problem of idling computing power in traditional architectures. The measured offline reasoning throughput is increased by up to 1.87 times, and the online Agent operating efficiency is increased by 1.96 times. Pure software optimization is used to double the performance. It can be called a disruptive breakthrough in AI infrastructure. The style of improving cost efficiency is very DeepSeek.

There are various signs that the upcoming new generation flagship model DeepSeek-V4 will most likely integrate text, image, and video generation capabilities, and use native multi-modal pre-training instead of post-stitching. The model parameters exceed one trillion, and it has strong memory, tools, code, learning capabilities, and good support for agents.

The combination of domestic models and domestic computing power

In addition to the model, another expectation for DeepSeek V4 is that it hopes to work together with domestic computing power after running in and exploring.

There have been many reports discussing that before releasing V4, DeepSeek did not provide a preview to American chip manufacturers such as NVIDIA and AMD. Instead, it chose to open access to Chinese chip suppliers including Huawei several weeks in advance to ensure that the model could be fully adapted and optimized on the domestic computing platform.

This is also a key reason why the outside world believes that DeepSeek V4 is delayed.

Adapting to domestic computing power is a very difficult road for domestic models, but it has to be done in the longer term. Something that has to be done must have a starting point, and perhaps DeepSeek V4 is that starting point.

When models extend an olive branch, pressure is placed on domestic computing power, which requires efficiency, production capacity, and effective supply to be able to further keep up and form an ecological synergy with the development of models and agents.

If DeepSeek V4 and R2 are proven to be able to perform at world-class levels on domestic chips from training to inference, and at a lower cost, there is hope that they can significantly get rid of their dependence on overseas computing power and break the "King of Tokens" label that Jen-Hsun Huang put on himself through SemiAnalysis.

If you still remember, the night DeepSeek R1 came out, NVIDIA plummeted by nearly 17%, and its largest single-day market value evaporated by US$589 billion.

Nvidia's plunge is not a good thing for technology stock investors, but if it is driven by DeepSeek, then we would like to see this happen again.

Treatment of sugar water intelligence and force intelligence

At the end of this letter, if there is one more expectation to mention - DeepSeek can make a breakthrough in another Scaling Law.

This breakthrough is not in the traditional sense of "the bigger the model, the stronger the capability", but the ability of small-volume models to continuously scale large-volume models.

Based on the two technical routes of "principle-algorithm-training-evolution of thinking and reasoning capabilities" and "intelligent compression-distillation-internalization", the small-volume model in each stage continues to reach the capability level of the large-volume model in the previous stage, and even continues to approach and reach the daily high availability level, and then on this basis, capabilities-applications-scenarios-value are gradually layered.

Small model, regular intelligence serving simple and basic daily life, good at quantity, open, end-side, edge deployment and cost efficiency are better - this is "sugar water intelligence", the soup part of the Token economy.

Large models and super intelligence serve business-productivity-professional technology-heavy tasks in the enterprise industry, and high standards generate high premiums - this is "force intelligence" and the meat-eating part of the Token economy.

As for the evolution of small model capabilities, Google Gemma 4 is a good reference system. It includes four versions: 2B, 4B, 26B and 31B, covering all scenarios from mobile phones to workstations. Among them, the 31B Dense model ranks third in the Arena AI open source rankings, and the 26B A4B MoE model ranks sixth. All four models support image and video input, support more than 140 languages, and have built-in switchable thinking modes. This is not simple parameter compression, but intelligent distillation and internalization - through more efficient knowledge transfer, more precise quantitative pruning, and more advanced distillation technology, small models have great wisdom.

We hope that DeepSeek can surpass Gemma-4 in the three high-quality models of 30B-70B-120B. Enterprise-level deployment can surpass the level reached by the "Six Little Tigers" with an open source model with trillions of parameters and create a new pattern.

On the device side, DeepSeek is also expected to achieve the same breakthrough in the lightweight models of 1B-8B. When client-side models can run smoothly on consumer-grade graphics cards and even mobile phones, when there are hundreds of millions of client-side models in personal mobile phones and computers, and when every ordinary user can have strong AI capabilities, it will be an inclusive form of smart equality and smart economy.

Write at the end

2026 is the year of "jump development" for a new generation of cutting-edge models and takeover agents. Each AI company will play its own trump card, triggering a new round of industry reshuffle.

"China Group" needs the return of DeepSeek, the open source ecosystem needs the promotion of DeepSeek, the Token economy needs DeepSeek's deep force intelligence, and domestic computing power needs the verification of DeepSeek.

Now, there is almost no gap between the model capabilities of China and the United States in the conventional intelligence part of daily question and answer, but there is still a gap in the deep intelligence of long-range complex tasks. This gap makes everyone look forward to DeepSeek.

This is the last reminder and the final call. V4 and R2 carry expectations not only for the iteration of models, but also for the development and advancement of an era. From model war to full-ecological war, from single-point breakthrough to full-stack AI competition, from imitation to independent innovation - can DeepSeek's next step define the next step for China's artificial intelligence?

I hope that DeepSeek's "silence" for more than a year is for a better outbreak.