Summary

DeepSeek’s core cost advantage is its MoE architecture plus infrastructure designed for MoE. Its 671B “full” model activates only 37B parameters for each generated token; the number of parameters read “may even be less than that of a 70B model.” Combined with DeepSeek’s joint optimization of the inference framework around the expert dimension, R1’s API cost is close to one-tenth that of GPT o3 mini. Ji Yu believes the cost curve has not yet been fully mined: deep optimization for MoE “can itself drive costs down further.”
A Chinese company built it because chip bans created “choice pressure.” Silicon Valley can buy H100s without limit and expects to receive GB200s, charging ahead along Scaling Law inertia with “no choice pressure.” Chinese companies may not be able to obtain H100s, and what they do get is generally a cut-down version, forcing a choice between “small dense models or ultra-large MoE models.” DeepSeek’s redesign around extreme MoE therefore created a major competitive advantage—it “was forced out by the environmental changes created by the ban.”
The moat of compute resources may be “unable to withstand a single blow.” Ji Yu believes the cloud vendors’ path of building around OpenAI’s algorithms and NV GPUs, with 10,000-GPU clusters as the threshold, “artificially created a technical barrier” that DeepSeek has broken. “Resource moats are easy to break with technology,” and building barriers before AI reaches large-scale economics is “essentially forcing the industry to grow prematurely.” True moats are often built on business models, ecosystems, and organizational culture.
The investment battleground is “AI for all” versus “mainframe-ization.” The NV-plus-Scaling-Law system is replaying the mainframe era, when millions or tens of millions of machines were needed to run the best models. Ji Yu is betting on an x86-style disruption: “A blue giant like IBM was disrupted by x86 chips that looked weak and low-end.” If chips costing thousands to tens of thousands can run full DeepSeek, and clusters built from them can deliver more aggregate service capacity than expensive mainframes, DeepSeek is only “the first shot.”
The real opportunity for China’s domestic chips is not “being compatible with DeepSeek.” Everyone announcing that domestic chips can run DeepSeek is following the same logic and “will find it hard to make a genuine impact.” The breakthrough is to build a system “more attractive and more competitive” than the NV ecosystem. The open gap lies in the middle: edge devices and 10,000-GPU training are both covered, but laptops, mini PCs, and home labs “still do not meet the demand.” The bottleneck is the ecosystem—these devices must run Windows or macOS—not the technology.
The Scaling Law pretraining path has indeed hit obstacles; the next paradigm is exploration at inference time. Ji Yu believes the gains from GPT-3 and GPT-4 mean “you cannot expect this path to keep delivering continuous improvements.” O1/R1 rely on mathematics and coding, where rewards can be verified quickly; generalizing to open-ended domains without rewards “may eventually run into another wall.” The larger opportunity lies in language’s own capacity for self-reflection, rather than “continually finding more rewards.”
Organization and culture are important moats. Ji Yu strongly agrees with 梁文锋 that “technical advantages are temporary.” DeepSeek’s success came from joint innovation across infrastructure, algorithms, and models; 梁文锋 knows “exactly what every thread is doing” inside a CUDA operator. The team relies heavily on domestic graduate students rather than big-tech pedigree. China’s past 40 years were powered by efficient empiricism, but innovation requires moving beyond it and “building up a new way of organizing.”

Deep dive

1. Xingyun Turned Itself into a Customer: Private-Deploying DeepSeek for AI Coding, “From Disdain to Surprise”

Ji Yu’s team began privately deploying models after AI coding took off around August or September last year: cloud solutions carried high subscription fees and information-security concerns. They tested Qwen, DeepSeek, and other models internally, and found that “after private deployment, performance was extremely high—much higher than cloud APIs.” They ultimately chose DeepSeek 2.5 because its results were clearly the best.
The key discovery was that small models are fine for code completion, but once users move to agent-style tools such as Continue and Cline, “only models at this particularly large scale can work with reasonable stability.” Small models lose instruction-following ability and cause plugins to crash. Verilog training data is scarce for a chip company, so engineers initially “tried every possible way to test it,” but “everyone went from complete disdain, to finding it pretty good—甚至有点惊喜—after using it, and then to intensive use.”
Li Gang added a consumer-side observation: Online models, apart from OpenAI, “are all extremely stingy.” Products marketed as having long context windows often actually read through search, and cannot translate tens of thousands of Chinese characters in full. Running a 32B quantized model locally on a 3090 “gets slower and slower, but it can complete your task.”

2. The Breakout Formula: Open Source and Free + R1 Needs No Prompt Engineering + the Spring Festival

Ji Yu was not surprised that DeepSeek would become popular—its engineering and base-model capabilities had always been strong, and it was open source and free. But “becoming this popular was genuinely beyond expectations,” with the model elevated to the level of “national destiny” and “AI equalization with Chinese characteristics.” One underestimated factor was going home for the Spring Festival and personally teaching parents how to use it, which helped drive mass adoption.
The usability jump was technical as much as cultural. Previously, “you had to prompt your model well to use it,” while parents asking simple questions often got terrible answers. R1 thinks for itself, and DeepSeek’s technical report even recommends against over-prompting—“let the model understand you on its own.” As a result, ordinary users can ask questions in all kinds of bizarre ways and still get good answers.
A useful comparison: Kimi K1.5’s language was “highly dialectal, with many northern colloquialisms,” while DeepSeek was “more proper” and more literary. Open-source models with genuine deep-reasoning capability remain scarce because “deep reasoning depends heavily on the model itself being good enough.” After screening the models on the market, DeepSeek may currently be the only large model with reasoning that clears the bar; Llama’s largest version “may not represent a particularly large improvement” over 70B.

3. The Mechanics of One-Tenth the Cost: Sparse MoE Activation + Infrastructure Rewritten for the Expert Dimension

The core math is straightforward: 671B parameters at 8-bit precision require 671GB of storage, but MoE is sparse, so only 37B parameters are activated for each generated token. A traditional 70B dense model may need to read the full 70G for every token. So although R1 looks like a 600B–700B model, “the number of parameters that must be read for each token may even be less than that of a 70B model”—the first-principles explanation for why R1’s inference cost is close to one-tenth that of o3 mini.
The second layer is DeepSeek’s joint innovation across infrastructure and algorithms. The industry’s dense-model inertia led it toward distributed approaches such as TP, but DeepSeek “completely shifted” distributed training and inference toward the MoE expert dimension. Its entire inference framework was customized for that purpose, and “there is still more room to mine.” Ji Yu attributes this to the team’s stronger cross-domain coordination and originality.
Li Gang’s characterization: MoE brings the capabilities of much larger models down to relatively low-cost hardware, while the longer outputs enabled by thinking and reasoning improve answer quality. Together, they produce a “Chinese-characteristics, high-value-for-money, high-quality” model—a convergence of timing, conditions, and people.

4. Why a Chinese Company Fired the Shot: The Choice Pressure Created by the Ban

Ji Yu’s framework is worth preserving in full: US companies “can buy large quantities of H100s at any time, and expect to receive GB200s that are more powerful than H100s in the future.” They “have no choice pressure; every path keeps moving forward.” In Silicon Valley, Meta will train ultra-large dense models, while Silicon Valley overall is roughly split between MoE and dense. Under the ban, Chinese companies may not be able to obtain H100s, and what they get is “generally a cut-down version,” leaving them with essentially 2 long-term routes: small-scale dense or ultra-large MoE. “That is what choice pressure produces.”
Existing technology stacks were developed along the Scaling Law path from “small dense to large dense,” so DeepSeek’s targeted optimization of MoE created a major competitive advantage. Ji Yu’s conclusion is direct: “Why was it a Chinese company? I think it was instead forced out by the environmental changes created by the ban.”

5. Scaling Law Has Hit Obstacles; the New Paradigm Is an Extension, Not a Replacement

On rumors that OpenAI’s training has stalled and GPT-5/4.5 keep slipping, Ji Yu’s view is that “there have in fact been some obstacles.” GPT-3 and GPT-4 made major gains over the previous generation of deep learning, “but you cannot expect this path to keep delivering continuous improvements.” Market chatter about Scaling Law hitting obstacles had already emerged last September—an event relatively independent of DeepSeek breaking through.
Ji Yu had long believed inference-time scaling was inevitable. Pretraining gives models a platform of natural-language capabilities, and “they naturally have to move toward the next paradigm to create the next leap in performance.” Paradigms are “not mutually exclusive; they are successive layers”—just as post-training builds on a sufficiently strong V3 base model to produce R1.
The current route has clear limits. O1 and R1 are built around mathematics and coding, where they can “quickly obtain a large amount of reliable reward.” Strong language generalization lets reasoning spread into domains not tuned by reinforcement learning, but “if you rely only on these 2 domains to drive it, generalization across more domains will grow weaker and may eventually run into another wall.” Conversely, “it is precisely because resources are constrained that people look for new solutions.”

6. The Next Step Needs Reward, Not Data—but Language Itself Is the Foundation

Asked whether data is running out and AI will have to generate its own training data, Ji Yu shifted the frame: reasoning models such as R1 “may not need data; what they really need is reward.” The reinforcement-learning methodology is to generate a batch at random, use external signals to rank quality, and adjust the model. Data is simply “an extremely strong supervisory signal”; a broader signal can be weakened to “this one is a little better, that one is slightly worse.”
He stressed that this is a personal view and a key point of disagreement: “People may think reinforcement learning is the most important thing; I actually think natural language itself remains the most important.” Language is the underlying foundation that provides the capacity for thought. Humans can “think quietly for an entire afternoon, carrying out logical deductions in their own minds and arriving at deeper conclusions.” AlphaGo-style self-play is “not yet a sufficiently defined methodology that can fit into the existing toolbox,” leaving substantial room for exploration.
The 2 paths are not in conflict. Reinforcement learning across more domains can “activate different aspects of the model’s capability through differences in data characteristics across domains.” ChatGPT’s breakthrough came from “training code and natural language in the same model”; training them separately had not worked particularly well.

7. The Moat Test: Resource Barriers Force Premature Growth; Commercial Ecosystems Count

Ji Yu’s core thesis is that “resource moats are easy to break with technology.” Barriers built on 10,000 cards, data volume, or GPU counts “may gradually be broken”; true moats are often business models and ecosystems. DeepSeek’s biggest impact on NV is that it broke the race among cloud vendors and giants to build around OpenAI’s algorithms and NV GPUs, with 10,000-GPU clusters as an artificially created technical threshold.
His further judgment is that AI “has not yet entered the economic system and generated enormous commercial value.” Building barriers at this stage “is essentially forcing the industry to grow prematurely.” If companies such as DeepSeek first allow the industry to develop, “once you have an economic cycle, you will naturally get commercial structures, business models, and corresponding new moats.” Before technology becomes broadly economic, these so-called barriers “may be unable to withstand a single blow.”
On calls to completely ban graphics cards from use in China or a handful of developed countries, his response was blunt: “It clearly shows that they are getting anxious; when challenged, their first reaction is to become even more closed-minded.”

8. AI for All vs. Mainframes: the x86 Disruption of IBM Replayed

This is the core investment narrative for both Ji Yu and Xingyun: the NV-plus-Scaling-Law system “is essentially pushing computing toward a mainframe architecture,” increasingly resembling the 1980s, when “millions or tens of millions of machines were needed to run the best model available today.” On the other side of history, “many people may not know what IBM’s position in the industry was before the 1980s, but that blue giant was disrupted by seemingly insignificant, weak, low-end x86 chips.” Both the PC and internet revolutions were built on “a more economical computing foundation.”
Applied to AI, if chips costing thousands or tens of thousands can run full DeepSeek, “a cluster assembled from chips costing tens of thousands could have greater aggregate service capacity than a single extremely expensive, high-value machine.” The mainframe architecture would “increasingly lose its value.” DeepSeek “is certainly the first shot,” but digging up the moat “may only be the beginning.” The next stage is a contest between an AI-for-all industrial structure and the small circle seeking a monopoly through hardware resources, with chips inevitably playing a central role.
For Xingyun, private deployment shifted from “where is the demand, exactly; perhaps this market is zero today” to a market consensus. The current market is split between “full models costing several million” and “low-precision distilled models.” Xingyun wants to turn full, unquantized models into the standard everyone can afford—“there is really no need to settle.”

9. “Domestic Chips Can Run DeepSeek” Is Not a Breakthrough; the “Kuda” Ecosystem Won’t Be Built by Slogans

Ji Yu poured cold water on the wave of announcements claiming to “connect to DeepSeek” or that “domestic chips are compatible with DeepSeek”: “If the logic is simply ‘the US has OpenAI, so I have a domestic model; the US has Nvidia, so I have a domestic chip,’ that is not the real reason DeepSeek broke out today.” DeepSeek broke out because it found a better path than the existing one; extreme MoE achieved “better economics” in the overall software-hardware tradeoff. If hardware merely says “Nvidia can run DeepSeek, and our chips can too,” it will “find it hard to make a genuine impact”—that is still the old logic of domestic models following the US path. A real breakthrough requires building a system “more attractive and more competitive” than today’s NV-centered ecosystem.
Open-source model standardization does give domestic chips an opening: “No matter what chip you use, as long as you can run the model well, you can connect it to an API and offer it as a service.” But once the excitement fades, the market “will certainly return to whose solution is best and most economically coherent.”
Li Gang identified an unfilled tier: edge inference is covered, with 8B models reaching 200 token/s; server-side private deployment and 10,000-GPU training also have players. “The middle seems not to meet the demand yet.” Li Gang sees significant opportunity in laptops and mini PCs, but the core issue is not the technology—it is where to enter the ecosystem. Outsiders cannot break into mobile SoCs; with laptops, “you cannot simply make your own chip and run Linux on it—you have to run Windows or macOS.” The demand for a home lab—a mini PC dedicated to large-model capability—“may not yet have been truly activated.”

10. 梁文锋’s Example: After Rejecting Empiricism, Build a New Way of Organizing

On 梁文锋’s view that “technical advantages are temporary, while the true moat is culture and organization,” Ji Yu expresses “complete agreement.” When OpenAI first emerged, nobody could match its technology, but “there are so many outstanding people in the world that this can certainly be replicated; it is difficult to block others through technology.” DeepSeek’s success was joint innovation across infrastructure, algorithms, and models, demanding exceptional organizational capability; many core members are domestic graduate students or newly minted PhDs. Ji Yu heard from a junior colleague at DeepSeek that “their boss knows exactly what every thread is doing” inside the library or operator the colleague wrote—whereas many infrastructure engineers today do not necessarily understand DeepSeek’s optimization in depth.
The methodology of innovation is not simply to gather the smartest people from different fields and expect value to emerge. Pure brainstorming and trial-and-error “is actually very inefficient,” and innovation itself requires efficiency. The safeguard is comprehensive first-principles thinking; organizers must be able to go deep into the details, understand what every layer is doing, and then break through each layer’s boundary conditions.
His diagnosis of China’s existing reflexes is that investment and entrepreneurship alike rely on empiricism: “trust industry experts, trust big names, trust labels.” There is nothing inherently wrong with that approach—China’s economic takeoff over the past 40 years was built on it, and its success rate is genuinely high. But “anti-empiricism is not simply throwing all of this away; destroying something is easy.” It must also “build up a new organizational model.” 梁文锋’s definition of talent is different as well—he may look more for students still in school rather than people with big-tech pedigrees. By entering the arena himself, he became the first example and a starting point for the reversal. A shift cannot be confirmed by a single event: “Only when many similar cases appear can we say this has genuinely changed.”
Li Gang’s framework for the past 30 years: 2000–2010 was Chinese manufacturing—mass production at low cost; 2010–2020 was high quality at low prices, Xiaomi-style, perhaps delivering 80% of overseas performance or capability at 20% of the price; the past 5 years have been me better—“in some industries, achieving capabilities far beyond overseas peers while keeping prices relatively low.” In investment terms, that means backing AI-native founders.

11. The Application Landscape 3 Years Out: Private Productivity, Cloud-Based Consumer Entertainment, Foolproof Interaction

Ji Yu’s first step is to make full models affordable to everyone; “low-precision compression can be left for later.” The analogy is computer memory today: “you do not care that your browser uses 1G of memory, because memory is already sufficiently large and cheap.” Once AI is widespread, it will penetrate every industry as mobile and PC ecosystems did. The entry point should be high-value scenarios where users are willing to pay, such as AI coding, until “people no longer need to think about whether it is worth spending money, because the cost is basically negligible.”
The end state will not be entirely private deployment: “the cloud still has enormous opportunity.” The internet is built around a cycle of experience, traffic, and commercial returns. DeepSeek’s servers are “busy every day”; under a flood of traffic, servers would need to expand many times over while commercial returns declined. The way out is like “Google once accumulating a pile of x86 machines to build search infrastructure”: assemble large-scale clusters from extremely cheap machines, making it possible to serve tens of thousands of applications. The principle is: “Do not divide commercial structures according to hardware resources; use whichever commercial structure is most suitable.” Productivity tools are suited to private deployment; consumer entertainment is suited to the internet.
Li Gang’s vision: OpenAI will retain a “signpost effect” in the short term, and personal assistants in the style of Deep Research will spread. Endpoints may not even need screens; a speaker connected to a cloud model could deliver powerful functionality. Interaction will evolve by generation: “People of our generation are good at typing, those born in the 2000s all use touch controls, and the era of those born in the 2010s will be pure voice.” But he retains a clear-eyed view of the present: dazzling applications still run on OpenAI’s o1 with usage limits, and context windows are not yet long enough. “Today, we have not really seen people make good use of large-model capabilities; perhaps we still need to wait for infrastructure to change.”

Transform Your Life with AI Health Breakthroughs | EP #149 The ARC Prize 2024 Winning Algorithm [Daniel Franzen and Jan Disselhoff]