DeepSeek's Five-Step Ladder to AGI: A Restrained Roadmap Built on 20,000 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 or discuss KPIs. Instead, his opening statement centered on a single concept—restraint. This restraint is reflected not only in product pricing but throughout his overall roadmap toward AGI (Artificial General Intelligence).

AGI (Artificial General Intelligence) refers to an AI system capable of performing intellectual tasks across any domain, much like a human, in contrast to the currently dominant "narrow AI" (ANI, Artificial Narrow Intelligence). Narrow AI excels only at specific tasks, such as image recognition or language translation, whereas AGI implies the ability to reason, learn, and adapt autonomously across domains. While academic definitions of AGI remain contested, 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, but their paths differ. Notably, AGI is not the same as "superintelligence" (ASI, Artificial Superintelligence), which refers to systems that comprehensively surpass human capabilities; AGI instead refers to intelligence roughly "on par with human capabilities."

This article is based on an analysis of this conversation by the Bilibili channel "Value Investing Partners," laying out how DeepSeek plans its path to AGI through a self-consistent "five-step ladder" logic, under the constraint of having only 20,000 GPUs of equivalent computing power. It should be noted that the following content is compiled from public information and stated remarks, and does not constitute investment advice.

The Restrained Pricing Logic: Good Enough Is Good Enough

DeepSeek's API pricing follows an extremely simple standard: recover equipment costs within ten months. In other words, however much money is spent buying GPUs should be earned back from API revenue within ten months—not a penny more. By this calculation, this corresponds to roughly a sixfold profit.

Intriguingly, Liang Wenfeng stated that "sixfold profit is too high," and he hopes to reduce it to threefold or fourfold in the future. He even admitted that within this price range, demand is almost completely inelastic—doubling the price would barely reduce the number of tokens users consume, and total revenue could simply double. Having money on the table but declining to take it while saying "it's enough"—is this vision or naivety?

He shared a detail: the company previously had a model priced too high, and the team wasn't happy about it. He later cut the price directly to one-quarter, and the company chat erupted in cheers. This pricing philosophy produced unexpected results—one round of casually reduced C-end traffic left users who "couldn't be driven away," while B-end API revenue is reportedly on track to reach hundreds of millions of dollars.

Liang Wenfeng calls this a dimensional-reduction strike: when you stand on the technological high ground of AGI, serving the C-end and B-end markets requires very little effort. This restrained strategy generated positive effects at the organizational level—C-end users can't be driven away, and B-end business grows naturally.

DeepSeek's Five-Step Ladder Roadmap to AGI

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

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

The first step is chain-of-thought reasoning. At this stage, AI's ability to solve olympiad math problems and write programs already surpasses the most elite humans. This rung is considered already 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 how humans "work through a problem on scratch paper." This technique dramatically improved large language models' accuracy 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 help models spontaneously develop long-chain reasoning capabilities, already surpassing top human performance on the AIME math competition and coding benchmarks. CoT's limitation is that it remains "single-conversation" reasoning—the model cannot accumulate experience across sessions. This is precisely why Liang Wenfeng defines it as the "completed" first rung: it is the foundation of capability, but not the endpoint.

Step Two: Agents—The Current Main Focus

The second step is agents, which is DeepSeek's current primary focus and also the arena where the whole world is competing fiercely. Agents allow AI to chain multiple steps together to solve more complex problems.

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

But Liang Wenfeng candidly admitted that even after completing both CoT and agents, 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 be careful about when asking him for something." But for AI, you'd have to feed it all of this—which is simply impossible in reality.

Step Three: Continual Learning—The True Global Bottleneck

The third step, "continual learning," is the real 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 AI. Its central difficulty is "catastrophic forgetting"—neural networks tend to substantially degrade performance on old tasks when learning new ones. The human brain achieves integration of new and old knowledge through the hippocampus's memory consolidation mechanism, but existing deep learning architectures lack a similar biological mechanism. Current academic approaches include Elastic Weight Consolidation (EWC), Progressive Neural Networks, and Experience Replay, but none have achieved breakthroughs on large-scale language models. Liang Wenfeng's characterization of it as a bottleneck for which "no one in the world has an answer" is no exaggeration—it is precisely the hardest technical chasm to cross between being an "excellent tool" and a "true colleague."

Liang Wenfeng's attitude toward this is distinctive: rather than concentrating large amounts of computing power to tackle it, he lets everyone in the company try. He calls it "drawing lots"—the barrier is very low, anyone can try, and whoever hits it first gets the next ticket aboard. 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 a concentrated assault.

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

He calls the fourth step the "starting-point model": AI begins to iterate on itself, developing its next version independently, no longer needing humans to help it improve. The fifth step is embodied intelligence, where AI enters the physical world to do housework and provide elder care.

Embodied Intelligence is a frontier direction that deeply integrates AI with robotics technology. Its core proposition is that 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—that the formation of human intelligence fundamentally depends on continuous interaction between the body and the environment. On the implementation level, embodied intelligence must solve the real-time perception-decision-execution closed loop, involving multiple subfields such as computer vision, tactile perception, motion planning, and force control. Companies like Tesla Optimus and Figure AI are competing fiercely in this arena. Liang Wenfeng placing embodied intelligence as the fifth step rather than an earlier stage reflects a pragmatic judgment on technical priorities: before language intelligence breaks through the continual learning bottleneck, investing resources in embodied intelligence prematurely is inefficient.

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

The Compute Gap: 20,000 GPUs and an Order of Magnitude

Although the roadmap is clear, every step requires computing power to support it. The reality is that DeepSeek currently has only about 20,000 H-series-equivalent cards, "more than an order of magnitude" behind the United States.

Understanding the deeper context of this gap requires knowledge of the Scaling Law logical framework. The Scaling Law was systematically articulated by OpenAI in its 2020 paper "Neural Scaling Laws." Its core conclusion is that model performance follows a power-law relationship with model parameter count, training data volume, and computing investment—as long as scale continues to expand, performance continues 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. When Liang Wenfeng says "China hasn't even touched the wall," he implies DeepSeek is still far from the point of diminishing Scaling returns—under compute constraints, their model scale hasn't yet reached the range where marginal returns clearly decline. This actually becomes a unique strategic advantage: every additional card can bring significant performance gains.

To train a model of the same scale as the US, one would need 50,000 top-tier NVIDIA cards or 200,000 Huawei 950s. DeepSeek currently has only trained models at the level of a few billion active parameters—not because he thinks it's enough, but because "you train as big as your resources allow." He explicitly stated that China is still far from the ceiling of Scaling, and "hasn't even touched the wall."

Can Huawei fill the gap? Liang Wenfeng mentioned that Huawei provided 16,000 950 cards, but converted, this only equals about 4,000 NVIDIA B-series cards—he himself admits this isn't enough to train the next-generation model. However, he is actually more optimistic about domestic computing power than the market. Huawei's 950 supernodes can substitute for GB200 and GB300 in both performance and price, at the cost of four Huawei cards replacing 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 V3 was trained on NVIDIA cards, it barely relies on the CUDA ecosystem anymore. Instead, they developed their own advanced compiler, TileLang, and completed all their 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 AI training. Almost all mainstream AI frameworks (PyTorch, TensorFlow, JAX) deeply depend on CUDA, constituting 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 competitors like Huawei and AMD struggle to shake NVIDIA's market position. TileLang's technical approach resembles 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 bypasses 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 the entire Chinese AI industry with a viable path to break free from dependence on the NVIDIA ecosystem.

Now all that's needed is to re-run this workflow on Huawei cards, and the ecosystem problem of domestic cards is essentially solved. His exact words were: "This will be validated within a year," and what remains is merely a production capacity issue. This provides a clear timeline and viable path for domestic computing power substitution.

The Fallback Logic: AGI on the Upside, Selling APIs on the Downside

Can this company sustain itself? Liang Wenfeng did the math: if inference demand continues to expand, it's entirely possible for API annual recurring revenue (ARR) to reach hundreds of millions of dollars, which is already enough to cover all R&D expenses and turn cash flow positive.

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

He even laid out his ace: in the worst case, with technology frozen right here, going all-in on selling APIs is enough to support a publicly listed company. This forms DeepSeek's strategic structure—upward, the imaginative space of AGI; downward, the cash flow safety net of the API.

The Biggest Risk Isn't Technology, but Team Stability

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

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

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

Conclusion: Five Key Signals Worth Watching

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

Going forward, five signals deserve close attention:

1. When B-end ARR breaks $500 million—determines whether the company can achieve self-sustaining growth;
2. When the complete deployment of TileLang on Huawei cards is officially announced—validates the feasibility of domestic substitution;
3. Whether the next-generation model demonstrates an early form of continual learning;
4. Whether the core team remains stable and whether any executives depart;
5. When Huawei 950's production capacity no longer constrains scale expansion—determines the speed at which the compute gap narrows.

Under the constraint of 20,000 cards, DeepSeek has chosen a restrained yet self-consistent path. Whether this path can truly lead to AGI—the answer may lie hidden within these five signals.

Key Takeaways