DeepSeek's Five-Step Ladder to AGI: A Restrained Roadmap Built on Twenty Thousand GPUs

Introduction: The Philosophy of Restraint Behind a Four-Hour Conversation

DeepSeek founder Liang Wenfeng once held a nearly four-hour, in-depth conversation with investors. He didn't list growth figures, nor did he talk about KPIs. He opened with just one word—restraint. This restraint is reflected not only in product pricing but runs throughout his overall roadmap for AGI (Artificial General Intelligence).

AGI (Artificial General Intelligence) refers to AI systems capable of performing intellectual tasks across any domain the way humans do, in contrast to today's mainstream "narrow AI" (ANI, Artificial Narrow Intelligence). Narrow AI can only excel at specific tasks, such as image recognition or language translation, while AGI implies autonomous reasoning, learning, and adaptation across domains. The academic definition of AGI remains contested, but the mainstream consensus includes core capabilities such as autonomous goal-setting, continual learning, common-sense reasoning, and cross-domain transfer. Top labs like OpenAI, DeepMind, and Anthropic all treat AGI as their ultimate goal, though their paths differ. Notably, AGI is not the same as "super AI" (ASI, Artificial Superintelligence)—the latter refers to systems that comprehensively surpass humans, while AGI refers more to a level of intelligence "on par with human capabilities."

This article is based on the analysis and breakdown of this conversation by "Value Investing Partners" on Bilibili, examining how DeepSeek plans its path to AGI with a self-consistent "five-step ladder" logic, constrained by an equivalent of only twenty thousand GPUs. It should be noted that the following content is compiled from public information and statements made during the exchange, and does not constitute investment advice.

The Logic of Restrained Pricing: Enough Is Enough

DeepSeek's API pricing follows a minimalist standard: recoup equipment costs in ten months. In other words, whatever is spent buying GPUs should be earned back from API revenue within ten months—not a penny more. By this calculation, it corresponds to roughly a sixfold profit.

Intriguingly, Liang Wenfeng stated that "a sixfold profit is too high," and he hopes to lower it to threefold or fourfold in the future. He even admitted that at this price range, demand is almost completely inelastic—doubling the price would barely reduce the volume of tokens users consume, so total revenue could directly double. To leave money on the table and still say "it's enough"—is this vision or naïveté?

He shared a detail: the company previously had a model priced too high, and the team wasn't happy about it. Later he cut the price directly to a quarter, and the company chat erupted in cheers. This pricing philosophy produced an unexpected result—a certain casual reduction in C-end traffic left users who "won't leave even if you try to drive them away," while B-end API revenue is reportedly poised to reach hundreds of millions of dollars.

Liang Wenfeng calls this a dimensional strike: when you stand on the technological high ground of AGI, doing C-end and B-end business simply doesn't require much effort. The strategy of restraint produced positive effects at the organizational level—C-end users won't leave, and the B-end grows naturally.

DeepSeek's Five-Step AGI Ladder Roadmap

The AGI roadmap Liang Wenfeng described has five clear steps, each demanding progressively more computing power.

Step One: Chain-of-Thought (CoT)—Already Completed

The first step is chain-of-thought reasoning. At this stage, AI's ability to solve math olympiad problems and write programs already exceeds that of the most elite humans. This step is considered to have been crossed.

Chain-of-Thought (CoT) reasoning is a prompt engineering technique formally introduced by Google's research team in 2022. Its core idea is to have the model display intermediate reasoning steps before giving the final answer, similar to the "scratch paper" behavior of humans solving problems. This technique has dramatically improved the accuracy of large language models on mathematical reasoning, logical inference, and multi-step problems. Models like DeepSeek-R1 upgraded CoT from a prompting trick to a training objective, using reinforcement learning to have models spontaneously develop long-chain reasoning abilities, surpassing top human levels on the AIME math competition and programming benchmarks. The limitation of CoT is that it remains "single-conversation" reasoning—the model cannot accumulate experience across sessions. This is precisely why Liang Wenfeng defines it as the "already completed" first ladder—it is the foundation of capability, but not the endpoint.

Step Two: Agent—The Current Main Line

The second step is the agent, which is DeepSeek's current focus and also a track the entire world is fiercely competing in. Agents enable AI to chain multiple steps together to solve more complex problems.

AI Agent is one of the most intensely competitive tracks in the current AI industry. Its core is to upgrade large language models from "answering questions" to "executing tasks." Through three key modules—Tool Use, Planning, and Memory—agents decompose complex tasks into multi-step action chains, autonomously completing real workflows such as code debugging, web operations, and file management. Since 2024, OpenAI launched Operator, Anthropic launched Computer Use, and Google launched Project Mariner—all important moves in the agent track. The technical difficulty of agents lies in the cumulative amplification of "hallucination" across multiple steps—even if the single-step error rate is only 5%, after ten steps the success rate drops below 60%.

But Liang Wenfeng candidly admitted that even after completing both CoT and Agent, AI still cannot replace your employees. The reason is: you cannot write all the context in your head into a Prompt. When you hire a new employee, they spend two months getting familiar with the environment and then know "who Xiao Wang is, how to find him, and what to watch out for when dealing with him." But for AI, you'd have to feed it all of that—which is simply impossible to do in reality.

Step Three: Continual Learning—The Real Global Bottleneck

The third step, "continual learning," is the true bottleneck: can a model, like a human, spend two months getting familiar with an environment on its own and then start working? No one in the world has found the answer to this question yet.

Continual Learning (also called Lifelong Learning) is recognized as one of the core challenges in the AI field. Its central challenge is "catastrophic forgetting"—neural networks tend to drastically degrade performance on old tasks when learning new ones. The human brain integrates new and old knowledge through the memory consolidation mechanism of the hippocampus, but existing deep learning architectures lack similar biological mechanisms. Current major research directions in academia include Elastic Weight Consolidation (EWC), Progressive Neural Networks, and Experience Replay, but none have achieved breakthroughs on large-scale language models. Liang Wenfeng's positioning of it as a bottleneck where "no one in the world has answers yet" is not an exaggeration—this is precisely the hardest technical gap to cross between "an excellent tool" and "a true colleague."

Liang Wenfeng's attitude toward this is unique: he doesn't concentrate large amounts of computing power to tackle it, but instead lets everyone in the company try. He calls it "scratching lottery tickets"—the threshold is very low, anyone can scratch, and whoever scratches first gets the next boarding pass. This strategy reflects the inherent unpredictability of the continual learning problem: such breakthroughs often come from unexpected directions, and casting a wide net is more likely to touch the answer first than concentrated efforts.

Steps Four and Five: The Starting-Point Model and Embodied Intelligence

The fourth step he calls the "starting-point model": AI begins to be able to self-iterate, developing its next version on its own, no longer needing humans to help it improve. The fifth step is embodied intelligence—AI enters the physical world to do housework and take on elderly care services.

Embodied Intelligence is a frontier direction of deep integration between AI and robotics technology. Its core proposition is: true intelligence must possess the ability to interact with the physical world, rather than existing only in digital space. This concept originates from the "Embodied Cognition" theory in cognitive science—the formation of human intelligence essentially depends on continuous interaction between the body and the environment. At the technical implementation level, embodied intelligence needs to solve the real-time closed loop of perception-decision-execution, involving multiple subfields such as computer vision, tactile perception, motion planning, and force control. Companies like Tesla's Optimus and Figure AI are competing fiercely in this track. Liang Wenfeng listed embodied intelligence as the fifth step rather than an earlier stage, reflecting a pragmatic judgment of technical priorities: before language intelligence breaks through the continual learning bottleneck, investing resources in embodied intelligence too early is inefficient.

His logic is clear: first get continual learning and the starting-point model right, and then let the model develop itself—the further along you go, the easier it becomes.

The Computing Power Gap: Twenty Thousand Cards and One Order of Magnitude

The roadmap may be clear, but every step requires computing power support. The reality is that DeepSeek currently has only about twenty thousand H-series equivalent cards, which is "more than an order of magnitude behind" compared to the United States.

Understanding the deeper background of this gap requires understanding the logical framework of the Scaling Law. Scaling Law was systematically expounded by OpenAI in the 2020 paper "Neural Scaling Laws." Its core conclusion is: model performance follows a power-law relationship with model parameter count, training data volume, and computing power investment—as long as scale continues to expand, performance will continue to improve, and this improvement is predictable. This discovery sparked a global AI arms race, driving the birth of super-large models like GPT-4 and Gemini Ultra. Liang Wenfeng saying "China can't even touch the wall" implies that DeepSeek is still outside the tipping point of diminishing Scaling returns—under computing power constraints, their model scale has not yet reached the zone where marginal returns clearly decline. This actually becomes a unique strategic advantage: every additional card can bring significant performance improvements.

To train a model of the same scale as the U.S., you would need fifty thousand top-tier NVIDIA cards or two hundred thousand Huawei 950s. DeepSeek has currently only trained models at the level of a few billion activated parameters—not because he thinks it's enough, but because "you train as big as your resources allow." He clearly stated that China is still far from Scaling's ceiling, "can't even touch the wall."

Can Huawei fill the gap? Liang Wenfeng mentioned that Huawei provided sixteen thousand 950 cards, but when converted, this only amounts to four thousand NVIDIA B-series cards—he himself admitted that these few cards aren't enough to train the next-generation model. However, he is actually more optimistic about domestic computing power than the market. Huawei's 950 supernode can substitute for GB200 and GB300 in both performance and price, at the cost of four Huawei cards equaling one NVIDIA card, while lagging two years behind.

The Key Trump Card: The TileLang Compiler

Facing the ecosystem gap, DeepSeek revealed a key trump card—TileLang. Liang Wenfeng disclosed that although NVIDIA cards were used when training V3, they hardly rely on the CUDA ecosystem anymore. Instead, they developed their own advanced compiler, TileLang, and completed all work based on it.

Understanding the weight of this trump card requires recognizing the essence of CUDA ecosystem lock-in. CUDA (Compute Unified Device Architecture) is a parallel computing platform launched by NVIDIA in 2006. After nearly two decades of ecosystem accumulation, it has become the de facto industry standard in the AI training field. Almost all mainstream AI frameworks (PyTorch, TensorFlow, JAX) deeply depend on CUDA, which constitutes NVIDIA's deepest moat—the so-called "CUDA tax." Once developers are deeply bound to the CUDA ecosystem, the cost of migrating to other platforms is extremely high. This is the fundamental reason why competitors like Huawei and AMD find it difficult to shake NVIDIA's market position. TileLang's technical approach is similar to LLVM in the compiler field—by building a hardware-agnostic intermediate representation (IR) layer, the same code can be compiled to run on different hardware architectures. This means DeepSeek has bypassed CUDA lock-in at the architectural level, achieving true hardware portability. Its strategic value far exceeds its surface meaning: it not only solves the ecosystem compatibility problem of domestic computing power, but also provides a viable path for the entire Chinese AI industry to break free from dependence on the NVIDIA ecosystem.

Now it's just a matter of re-running this workflow on Huawei cards, and the ecosystem problem of domestic cards is basically solved. His exact words were: "This matter will be validated within a year," and the rest is just a capacity issue. This provides a clear timeline and viable path for domestic computing power substitution.

The Safety-Net Logic: AGI on the Upside, Selling API on the Downside

Can this company support itself? Liang Wenfeng did the math: if inference demand continues to expand, it's very likely that the API's Annual Recurring Revenue (ARR) can reach hundreds of millions of dollars, which is already enough to cover all R&D expenses and achieve positive cash flow.

ARR (Annual Recurring Revenue) is the most core financial health metric in SaaS and API business models, reflecting the enterprise's predictable, sustainable revenue scale, as distinct from one-time revenue. For AI infrastructure companies, the ARR of API revenue represents the degree of continuous dependence of downstream developers and enterprises on the platform, and is a direct manifestation of "technological moat" converting into commercial value. DeepSeek's strategy of trading extremely low pricing for high usage essentially builds user stickiness through "token economics": when developers' codebases and business processes deeply depend on DeepSeek's API interfaces, even if there are price adjustments in the future, migration costs will become a powerful retention mechanism. Currently, Anthropic's Claude API and OpenAI's API business have both entered competition at the billion-dollar ARR level. DeepSeek, entering at a lower price, is competing for the foundational base of this incremental market.

He even revealed his trump card: in the worst-case scenario, if technology freezes here, going all in on selling API is still enough to support a publicly listed company. This forms DeepSeek's strategic structure—imaginative upside space with AGI, and a cash-flow safety net with API on the downside.

The Biggest Risk Is Not Technology, but Team Stability

When asked where this logic is most likely to fail, Liang Wenfeng gave an unexpected answer: not computing power, not competitors, and not technical bottlenecks, but team stability.

He said: "As long as the team doesn't leave, I will definitely achieve AGI." The biggest risk is others poaching people, insufficient options, or everyone losing confidence in the vision. This is also the most core motivation behind that funding round—to give the team enough options. From historical data, DeepSeek's talent turnover has always been lower than peers, and this is the one core interest he is unwilling to negotiate on.

Behind this judgment lies deep industry logic: top AI researchers are the scarcest factor of production today. The irreplaceability of a core engineer who can work at the Scaling frontier far exceeds that of any single technical patent. Amid the high-salary competition from giants like OpenAI, Google, and Anthropic, maintaining a team that "believes in the vision more than the paycheck" is itself an extremely difficult-to-replicate competitive barrier. Liang Wenfeng listing this as the number one risk is a clear-eyed recognition of placing human capital above technical capital.

Conclusion: Five Key Signals Worth Watching

Combining public information, what can be confirmed is: Liang Wenfeng's five-step AGI logic is self-consistent, the restrained pricing strategy has produced a dimensional-strike effect at the organizational level, and the domestic substitution path has a clear timeline. What cannot be confirmed is: no one in the world has solved the third step of continual learning, whether DeepSeek can break through first remains unknown, and the "four times plus two years" computing power gap is a hard constraint running throughout the entire roadmap.

Going forward, five signals are worth focusing on:

1. When B-end ARR breaks through $500 million—determining whether the company can achieve self-sustaining growth;
2. When the full deployment of TileLang on Huawei cards is officially announced—validating the feasibility of domestic substitution;
3. Whether the next-generation model demonstrates a prototype of continual learning;
4. Whether the core team is stable and whether any executives leave;
5. When Huawei 950's production capacity no longer constrains scale expansion—determining the speed at which the computing power gap narrows.

Under the constraint of twenty thousand cards, DeepSeek chose a restrained yet self-consistent path. Whether this path can truly lead to AGI—the answer may lie in these five signals.

Key Takeaways