Vol.156 Industry Watch 27 | Seven Core Questions About DeepSeek and a Breakdown of DeepSeek’s Key Technologies
Summary
- 刘鹏琦 breaks DeepSeek’s viral rise down into “20% technical innovation and 80% open-source influence plus the China factor,” making the first investment distinction the gap between a capability leap and an emotional re-rating. V3 approached leading pretraining models at dramatically lower cost, while R1 was the first to publicly demonstrate a path to reproducing o1-level reasoning through reinforcement learning; the bigger shock came from a latecomer using open source to break the closed-source oligopoly, prompting a re-rating of expectations around whether “a Chinese team can produce original innovation.” “The barefoot have nothing to lose.”
- V3 rewrote the cost curve first, with its moat coming from the system-level coordination of algorithms, infra and hardware optimization rather than any isolated concept. 颜黔杭 cites roughly 2.78M H800 GPU-hours and training costs of more than $5M, versus about $63M for GPT-4 and $58M for Llama 3.1; its 671B total parameters activate only 37B per pass, while MLA, MoE, FP8, DualPipe and PTX jointly reduce compute, memory and communication overhead. 刘鹏琦’s characterization is more restrained: “not much conceptual innovation,” but “an extremely high engineering barrier.”
- R1’s significance was not inventing reinforcement learning or test-time compute, but turning the closed-source era’s reasoning paradigm into open-source engineering that could be reproduced and distilled. R1-Zero applied RL directly to V3 using ground-truth tasks such as math and coding, showing that long-chain reasoning could emerge without SFT; R1 then added cold-start data, multistage SFT and RL to fix readability and language mixing. The key constraint is that the base model must be strong enough—“it can find an answer once in 100 attempts”—or the same framework may not train at all.
- Scaling Law has not failed; it is shifting from pretraining toward post-training and inference. Both guests believe pretraining is constrained by the depletion of high-quality data, with the marginal return on adding parameters and compute declining; post-training currently consumes less than 1% of pretraining compute, while R1 may already be near 10%, leaving room to scale, although rewards, expert QA and CoT data are becoming new bottlenecks. At inference, longer outputs and self-correction can be exchanged for accuracy: “the model’s output process is itself a computational process.”
- In the competitive landscape, DeepSeek may have a larger impact on domestic To C model makers, while its long-term To B impact still awaits validation through real deployments. Kimi and MiniMax are not technically weak, and Kimi 1.5 has received positive reviews, but both must answer whether closed source can sustain frontier-level capabilities and whether open source could upend existing business models; overseas majors still lead in multimodality, generalization, tooling and developer ecosystems. DeepSeek must also decide how to handle its “windfall from heaven”: raise capital and expand, or preserve a small team and research focus.
- DeepSeek is not unambiguously negative for Nvidia: it expands total AI demand while weakening some scarcity-based pricing. Market education is driving more companies to buy compute, and guests have heard that H100 and H200 prices actually rose; constantly evolving model architectures also favor general-purpose GPUs. On the other hand, MoE reduces interconnect requirements, while PTX-level optimization shows that high-level CUDA libraries can potentially be bypassed, opening a window for Chinese chips and new architectures; the long-term contest will still turn on compilers and software ecosystems, not just single-chip performance.
- The truly tradable application-side variable is the ROI expansion created by falling token costs, but domestic expectations are clearly running ahead of actual performance in the short term. DeepSeek educated governments, corporate executives and ordinary users within weeks, while distilled models could bring reasoning capabilities to PCs, phones and AI hardware; however, tests by Peakview-backed companies found that DeepSeek did not lead on every industry task, and some distilled Qwen models may even underperform the originals on many tasks. The longer-term question is whether RL can transfer from math and code to Agents, embodied intelligence and the physical world, but for now the right conclusion remains: “let the bullets fly a little longer.”
Deep dive
1. DeepSeek’s Technical Breakthrough Explains Only a Small Part of Its Viral Rise
刘鹏琦’s core judgment is “20% technical innovation, 80% open-source ecosystem and the China factor”: low-cost V3 and open-source R1 are both strong pieces of work, but technology alone cannot explain the intensity of the global market reaction.
A Chinese team approaching frontier models under resource constraints exposed an information gap overseas and triggered skepticism. 刘鹏琦 believes the short-term impact is mainly emotional, while the long-term value lies in rebuilding market confidence that “China can produce genuine technological innovation.”
Open source gave the latecomer an asymmetric advantage. Compared with overseas majors, DeepSeek had fewer legacy constraints and could challenge the closed-source dominance of OpenAI and a series of models including o1—the “barefoot have nothing to lose.”
2. V2 Already Broke Through API Pricing; V3 Cut Training Costs to One-Tenth
颜黔杭 traces the story back to May 6, 2024, when DeepSeek-V2 offered near-GPT-4 capability while cutting input and output prices to RMB1 and RMB2 per 1M tokens, respectively. The GPT-4 model he cited on June 13 priced input and output at RMB217 and RMB434, triggering the industry’s first price war.
Released at the end of December 2024, V3 extended the shock from APIs to training: roughly 2.78M H800 GPU-hours translated into training costs of more than $5M, versus about $63M for GPT-4 and $58M for Llama 3.1, according to 颜黔杭.
Early media reports misread the “$5M training cost” figure. 刘鹏琦 urged people to revisit the claim after the relevant information was corrected, but the conclusion remained unchanged: V3 achieved performance close to GPT-4o and Claude 3.5 at one order of magnitude lower compute cost, while outperforming Qwen 2.5, Llama 3.1 and other open-source models available at the time.
3. MLA Compressed the Most Expensive KV Cache by More Than 90%
颜黔杭 begins with Attention’s Q, K and V: when a new token is generated, the K and V of historical tokens are repeatedly called and therefore must be cached. With long contexts, KV Cache quickly makes memory capacity and bandwidth bottlenecks before raw compute becomes the constraint.
MLA’s approach is to apply low-rank compression to the full KV matrix. 颜黔杭 compares it to preserving the shadow of a three-dimensional object on a plane: the full structure need not be stored; a low-dimensional matrix is enough.
The V2 paper reported that MLA reduced KV Cache by more than 93% versus earlier DeepSeek models—“leaving only one-tenth of the original.” 刘鹏琦 summarizes the trade-off as “using time to save space,” thereby reducing memory and memory-bandwidth requirements.
4. MoE’s Key Is Not Just More Experts, but Making Every Expert Actually Work
V2 had 236B total parameters and activated roughly 21B during inference; V3 expanded to 671B total parameters while activating only 37B. 刘鹏琦 describes a routing structure with one shared expert and 256 independent experts, calling the shared expert plus only 8 of the others each time—“using space to save time.”
颜黔杭 explains that DeepSeek expanded the typical roughly 32 experts to 256, then added a shared expert carrying general knowledge. “Three mediocre experts can match one Zhuge Liang”: more numerous and narrowly specialized experts can reduce the knowledge bias of any single expert and produce sparser computation.
The challenge is a Matthew effect during training: strong experts keep receiving data while weak experts are called less and less. DeepSeek uses Loss-Free Load Balancing to dynamically correct for each step’s load and prevent “the strong from staying strong and the weak from staying weak.” 颜黔杭’s summary is: “Many hands make light work, and everyone has to contribute.”
The two guests differ slightly on the standard for originality. 颜黔杭 calls MLA an innovation that went “from zero to one, then from one to ten and 100,” while 刘鹏琦 believes most of V3 stands on the shoulders of predecessors and that the truly hard-to-copy advantage is the engineering barrier created jointly by algorithms, frameworks and hardware.
刘鹏琦 also notes that Dense Models retain advantages in specialized To B applications and small edge models, so MoE cannot be declared the model architecture that “rules them all.” He raises another open question: were V3’s innovations deliberate choices or forced adaptations to compute constraints? If more compute had been available, would these innovations still have emerged or been necessary?
5. FP8, DualPipe and PTX Turned Low Cost into a Systems Engineering Exercise
颜黔杭 says DeepSeek built a high-performance, cost-efficient AIPC cluster with several tens of thousands of Nvidia GPUs, alongside extensive engineering optimization in node-to-node communication, pipeline parallelism and FP8 mixed-precision training. On the algorithm side, it also uses MTP, or multi-token prediction, to improve inference efficiency.
V3 uses FP8 mixed precision to reduce numerical storage requirements while preserving training accuracy as much as possible. DualPipe fills idle time in the GPU pipeline, minimizing device wait time as tokens traverse nodes.
刘鹏琦 places particular emphasis on PTX: the team did not rely solely on high-level CUDA libraries, instead writing lower-level code directly to compress compute and communication overhead and extract the maximum from the hardware. This capability must move in lockstep with algorithm design; otherwise, new bottlenecks created separately by MLA and MoE can cancel out their theoretical gains.
V3’s cost is therefore not the result of a single trick. 颜黔杭 believes DeepSeek’s algorithm-infra coordination ranks in “the industry’s first tier, perhaps even the best tier,” which explains why the same ideas from the same paper can produce completely different efficiencies in different teams.
6. R1-Zero Proved That Pure RL Can Elicit Long-Chain Reasoning
R1-Zero used V3 directly as its base model, without first adding high-quality human dialogue or reasoning data, and applied RL to tasks with ground truth such as math and code. The result showed that the model could elicit long-chain reasoning on its own.
The trade-off was equally clear: without language constraints, R1-Zero was difficult to read and mixed languages. 颜黔杭 explains that the reward function mainly pushes answers toward correctness; it does not automatically require the reasoning process to conform to human communication habits. The language mixing did not affect the final result.
刘鹏琦’s cooler assessment is that GRPO and objective-answer rewards are not particularly complex, and OpenAI had already pointed to the direction of RL and test-time compute when it released o1. DeepSeek’s contribution was moving faster, proving out a viable path first and making it public.
7. R1 Used Multistage Training to Turn an Experimental Model into a Usable Product
R1 did not stick to R1-Zero’s pure route. It first used cold-start data for SFT, then applied RL to activate reasoning capabilities; the intermediate model continued generating new cold-start data, with RL repeated iteratively to improve language expression and overall performance.
颜黔杭 calls it the industry’s first reproduction of a GPT-o1-level reasoning model, as well as “the Llama of the reasoning-model era”: closed-source capabilities entered the community for the first time in the form of a low-cost open-source base model, training path and distilled models.
Test-time compute is the external manifestation of this capability. The “thinking aloud” users see is not decoration; by expanding output time and token count, the model uses the generation process itself for search, verification and correction.
8. A Strong Base Model Determines Whether RL Can Converge at All
刘鹏琦 identifies search hit rate as the key mechanism: V3 may be able to find an answer “once in 100 attempts,” providing enough reward signal for rapid iteration. A weaker model that only hits the answer by chance once in 10,000 attempts may not train at all under the same RL framework.
DeepSeek also used R1 as a teacher to distill Qwen and Llama, releasing 6 small models ranging from 1.5B to 70B. 颜黔杭 says the 32B model was broadly consistent with GPT-o1, and all of these distilled models were stronger than the results of applying RL post-training directly to the base models.
His analogy is that a good teacher can materially raise a student’s starting level, while only a “particularly smart child” might surpass someone who has been taught through self-study alone. Distillation therefore benefits edge models, while pure RL remains highly dependent on a strong base model.
9. DeepSeek’s Quant DNA Transferred Nanosecond-Level Optimization to GPUs
刘鹏琦 believes the industry knowledge gap is narrowing: academic and industrial methods are gradually converging, and as pure conceptual innovation slows, the engineering execution at which Chinese teams excel is becoming decisive.
幻方量化, which incubated DeepSeek, supplied an unusual genetic advantage. In quantitative trading, even a nanosecond-level execution edge can determine profit or loss, so the team was already accustomed to customizing FPGAs and embedding assembly in C before transferring that low-level optimization capability to GPUs, PTX and training communications.
Organizational form matters as well. 颜黔杭 says DeepSeek “doesn’t look much like a commercial company,” but more like a highly collaborative research team. 刘鹏琦 emphasizes that its modest revenue and commercialization pressure, open-source culture and long-term orientation allow the team to assess a paper’s value quickly and engineer it immediately, rather than first subordinating it to short-term commercial goals.
10. Scaling Law Is Moving from Pretraining to Post-Training and Inference
The shared bottleneck in pretraining is a shortage of new data: continuing to add parameters can lead to overfitting, and 刘鹏琦 sees GPT-5’s “long gestation” as one early sign. Compute is not useless, but the marginal return on simply adding data, parameters and GPUs is approaching its ceiling.
Post-training still has substantial room to scale. It currently uses less than 1% of the compute used for pretraining, while 刘鹏琦 cites analysis suggesting that R1 may already be close to 10%. Expanding RL further could improve capabilities, but defining rewards and obtaining expert QA and high-quality CoT data have become new scarce resources.
Inference expands computation through longer thinking time: the model produces a draft, finds errors, reflects and adjusts. 刘鹏琦’s conclusion is not that Scaling Law has failed, but that “the angle from which it manifests has changed.”
11. Grok 3’s 200,000 GPUs Did Not Prove That Brute-Force Scaling Is Still Worthwhile
Grok 3, released on February 18, uses 200,000 GPUs. Both guests acknowledge gains on math, science and coding benchmarks, but believe it used one order of magnitude more compute than OpenAI and two orders of magnitude more than DeepSeek for only several percentage points of incremental improvement.
颜黔杭 says Grok 3’s cost input was more than 100x DeepSeek’s compute cost and may in reality have reached several hundred times. He also questions the decision to set chart axes at something like 40 to 100 rather than 0 to 100, which visually exaggerates the gap. Given that Grok 2 also posted strong standard scores while requiring its real-world performance to be discounted, his position is to “let the wind blow for a day or two.”
刘鹏琦 goes further: current gains come mainly from reasoning capabilities shaped through post-training rather than from pretraining scale itself. Grok 3 therefore reinforces the case that the old pretraining Scaling Law is weakening.
12. Model Competition Will Force Domestic Vendors to Rechoose Their Open-Source and Commercialization Paths
DeepSeek is still primarily a language model and lacks multimodal capabilities, so comparing it directly with full multimodal products is unfair. 刘鹏琦 cautions that OpenAI, Anthropic and Google still hold advantages in multimodality, generalization, tooling, developer ecosystems, technology reserves and compute: “There is still a long way to go before genuine superiority.”
The pressure on domestic To C players is more immediate. Kimi and MiniMax are not technically behind, and the engineering work on Kimi 1.5 has received academic recognition. The challenge is whether closed source can sustain frontier-level performance and whether open source could in turn overturn existing pricing models.
To B customers will eventually return to rational vendor selection, but in the short term emotion has driven concentrated trials of DeepSeek. It has educated the market, but has not automatically removed the barriers of delivery, adaptation and performance evaluation.
For DeepSeek itself, traffic is a “windfall from heaven.” Whether it raises capital or remains relatively small and research-focused will determine whether it can preserve its organizational advantage.
13. Nvidia Benefits in the Short Term, but Its Hardware Moat Faces a Re-Rating
刘鹏琦 believes Nvidia’s core moat is not just single-chip design, but chip interconnects—including Infinity Fabric and NVLink—combined with the ecosystem barrier created by the CUDA software stack.
DeepSeek has educated the market, lifted confidence in AI applications and may prompt more companies to buy hardware. 刘鹏琦 heard through several channels that H100 and H200 prices actually rose after DeepSeek released its application; constantly changing model architectures also increase demand for general-purpose GPUs and may not favor specialized chips.
On the other hand, MoE reduces the requirements for chip-to-chip interconnects. 刘鹏琦 calls PTX a programming technology below CUDA, but it is also Nvidia’s underlying technology. DeepSeek’s direct use of lower-level code demonstrates the possibility of reducing reliance on high-level CUDA libraries. If more models adopt relatively fixed architecture paths, other hardware could also support them.
In the short term, DeepSeek will drive the use of lower-performance cards, including Chinese domestic cards. Longer-term opportunities may emerge in 3D packaging, high-speed interconnects, high-memory designs and the broader compiler ecosystem. 刘鹏琦 stresses that single-chip or interconnect capability alone is insufficient; a complete software ecosystem is the key to long-term success.
14. Application ROI Has Improved, but Industry Adoption Still Has to Clear the Real-World Testing Hurdle
Lower token costs directly expand application profit margins. 颜黔杭 believes AI startups previously had to squeeze profit between model-inference costs and user payments; the price war has finally given developers cheap, flexible room to experiment. OpenAI’s release of o3 and price cut for o1 after R1 reflect the same competitive dynamic.
Within weeks, DeepSeek got government and corporate executives, ordinary consumers and even parents’ generation to start using it, materially increasing willingness to adopt on both the To C and To B sides. But 刘鹏琦 is explicit that the domestic market is overestimating it in the short term, particularly as real enterprise performance may fall short of expectations.
Field tests by Peakview-backed companies provide an important counterexample: customer demand for deployment is high, but DeepSeek does not lead on every industry task, and some Qwen models distilled from it may even underperform the original models on many tasks. Distillation “may” benefit PCs, phones and AI hardware, but the industry still needs to validate the claim.
The longer-term question is whether RL, already validated in objective domains such as math and code, can transfer to real-world reasoning, AI Agents and embodied intelligence. 刘鹏琦 remains hopeful, but also argues that we should “let the bullets fly a little longer.”
15. Open Source Is Both a Distribution Mechanism and the Infrastructure for the Next R&D Cycle
颜黔杭 cites an entrepreneur: “DAU is not a valuable factor for model R&D, but open source is.” After DeepSeek open-sourced its products on January 27 and launched its application in app stores, DAU quickly surpassed ByteDance’s Doubao and Kimi’s chat assistant, generating heat and traffic at almost zero cost.
Its deeper value resembles early Android: developers can deploy, study and iterate, gradually forming a platform consensus. The party that establishes that consensus gains talent, attention and downstream commercialization advantages, while also forcing open-source models such as Llama to keep innovating and creating pressure for closed-source vendors.
R1’s objective rewards currently apply mainly to math and code. 颜黔杭 invokes the shift from AlphaGo to AlphaZero to ask whether reducing human supervision and increasing self-reinforcement could allow models to break through the ceiling imposed by human data.
The longer-term vision remains explicitly hypothetical: “Perhaps reinforcement learning is the only way to achieve AGI that truly surpasses humans.” Until reward design and real-world transfer are solved, 刘鹏琦’s closing position is still to respect the breakthrough while continuing to watch: “Let the bullets fly a little longer.”