Summary

The capital-markets shock triggered by DeepSeek came from the narrative of “beating America’s lead at low cost,” not just from model performance. The episode cited these figures: it topped Apple’s App Store on January 26, reached 16 million downloads in 18 days—1.8x ChatGPT’s pace over the same period—and ranked first by downloads across 140 markets globally; Nvidia fell nearly 17% the next day, wiping $589B off its market cap. 陈羽北 believes the market mistook the low reported cost for total R&D spending and overlooked 幻方’s accumulated compute base: “The whole market is emotional.”
R1’s real starting point was the already powerful V3; reinforcement learning was neither the only secret nor the hardest part to replicate. 陈羽北 emphasized that when the base model generated 100 answers per question, its success rate was already around 10%, allowing rule-based rewards to turn sparse rewards into denser signals before methods such as GRPO reinforced correct answers; some online commentators believe PPO would also work, so “reinforcement learning is actually secondary here.” More important was V3’s efficient, solid base capability, achieved through MoE load balancing and KV cache optimization in the attention layers for a model exceeding 600B parameters.
DeepSeek’s technical value lies in a continuous combination of techniques, not one isolated breakthrough. V3 first improved architectural efficiency; R1-Zero then used search and verifiable rewards for self-improvement; finally, reasoning and planning from the large model were distilled into Qwen models ranging from 1.5B to the mid-30B range. The logic was: “First let the large model search its way to the answer, then let the large model teach the small model.” It also shows that academic approaches are converging: DeepSeek is highly capable, but 陈羽北’s preliminary view is that o1 generalizes better on ordinary tasks, and U.S. leadership cannot be declared over based on benchmarks alone.
For Nvidia, DeepSeek is a double-edged sword: positive for unit volume, negative for pricing power, with the final outcome hinging on whether applications take off in 2025. Lower costs could bring more startups into the market for training and inference compute, but they could also reduce customers’ need for the highest-end chips; John Yue summarized it as “lower premium, higher sales volume,” with the market-cap impact depending on the mix. If GPU prices were the main reason applications failed to gain traction, cutting costs to one-tenth or less would benefit Nvidia; if the bottleneck lies elsewhere, the incremental demand may not materialize.
DeepSeek proved that Nvidia’s wall can be climbed, but it did not make the CUDA and interconnect moats broadly lower. MoE reduces some interconnect pressure between Experts, while direct PTX optimization bypasses the high-level CUDA API, but John immediately qualified the point: it “did not completely bypass the CUDA ecosystem,” and the approach only serves specific models. AMD benefits in the short term and ASICs gain optionality, but the general-purpose software ecosystem remains the long-term barrier—the wall has not gotten lower; DeepSeek only proved “it’s possible to go over.”
Open source and low-cost APIs free application developers from the risks of relying on a single closed-source vendor. John said DeepSeek gave small teams “their own OpenAI,” removing the fear that OpenAI’s next release could swallow both their product functionality and pricing power; its API costs RMB1 per million cached input tokens, RMB4 for uncached input, and RMB16 for output, which the episode estimated made o1 roughly 26–27x more expensive across tiers. Closed-source models still have value if they remain ahead, much like Apple, but “if closed source is still worse than open source, it may have no meaning”; open source is more like Android, lowering the industry’s barriers to entry.
The guests rejected the aggressive view that AI will exceed most people in nearly every domain by 2026–2027; the core obstacle is a three-order-of-magnitude gap in learning efficiency. 陈羽北 compared the human brain’s 20 watts with roughly 20,000 watts for a 600B model, fewer than 10 billion tokens consumed by a person before age 20 with Llama 3’s 15 trillion, and a jumping spider’s millions of neurons with billion-parameter navigation models; John added that domain knowledge such as startup experience has not yet been digitized. A more meaningful AGI signal would be a model that can find its own learning signals, curate data, and “make every token count,” rather than simply burn through more data.

Deep dive

1. The Market Sold the Story of “Beating America’s Lead at Low Cost”

泓君’s reference points for the frenzy were these: DeepSeek topped Apple’s App Store on January 26, reached 16 million downloads in 18 days—1.8x ChatGPT’s pace over the same period—and became the most-downloaded app across 140 markets worldwide; on January 27, Nvidia fell nearly 17%, wiping $589B off its market cap in a single day.
陈羽北 sees three narratives stacking on top of one another: a previously low-profile, little-known team used “not the best compute” and very low reported costs to build a model comparable to o1, naturally prompting the media to ask whether “America’s leadership no longer exists.”
His caveat is equally important: DeepSeek is neck-and-neck with o1 on benchmarks, and may even be better, but his preliminary view is that o1 still generalizes better on ordinary tasks. The real concern is that technical approaches are converging—there are fewer innovations that make people say, “You articulated it, but I never thought of it”—rather than that one benchmark run has already transferred leadership.

2. V3 Was the Real Foundation for R1’s Leap

陈羽北’s contrarian view is that “reinforcement learning has been overemphasized this time.” When the base model generated 100 answers per question, its success rate before reinforcement was already around 10%; in his view, that was the prerequisite for subsequent self-improvement.
泓君 also noted that the paper used multi-head latent attention, but 陈羽北 focused on two core V3 efforts, both aimed at architectural efficiency: MoE improved load balancing among different Experts, while the attention layer reduced KV cache requirements. Together, they gave a 600B-plus, Megatron-scale model a respectable baseline.
This also explains why GRPO may not be the only answer. 陈羽北 considers its policy-gradient implementation “fairly rough,” and some online commentators have argued that PPO could work as well. If multiple RL methods are effective, the hardest parts to replicate are more likely to be the base model, data, and underlying architecture—not the algorithm’s name.

3. R1-Zero Turned Sparse Rewards into a Searchable Learning Signal

R1-Zero reduced the reward function to its essentials: the mathematical answer had to be correct, and the output format had to be correct. The model generated 100 answers at a time, then increased the weight of the correct ones, without separately training a complex process-reward model.
陈羽北 believes it bypassed reinforcement learning’s hardest problem: sparse rewards. If all 100—or even 10,000—outputs are wrong, the model receives no direction for improvement; once the starting success rate reaches roughly 10%, correct samples make the reward “relatively dense” rather than sparse.
The process is essentially an approximation of Search: the base model supplies candidate trajectories, rule-based rewards filter them, and learning amplifies the effective paths. 陈羽北 sees parallels with model predictive control, world models, and model-based RL, but says the approach starts with simple tasks whose answers can be verified.
Starting from a roughly 10% success rate matters for more than answering a few additional questions correctly. It shows that once “the model’s base capability is already pretty good,” the model may use data generated by itself to achieve bootstrap-style improvement.

4. Large-Model Search Can Be Distilled into Small Models

DeepSeek first had a 600B-plus model learn reasoning and planning through repeated generation and filtering, then distilled those capabilities into Qwen models ranging from 1.5B to the mid-30B range. 陈羽北’s summary: the small model may not be able to find the path itself, but it can directly learn the path already found by the large model.
The logic has a clear boundary: if the base model is not good enough, Search, model-based RL, and self-improvement will not work well. If correct trajectories already appear reliably among 100 candidates, one can first search them out directly and then decide whether further training is necessary.
That is why he views DeepSeek as a “combination punch”: V3 provides the foundation and efficiency, R1-Zero demonstrates self-directed search, and R1 carries the capability transfer forward. No single idea is necessarily unprecedented; the value lies in each step creating usable conditions for the next.

5. Nvidia Faces a Dual Repricing: Higher Volume and Lower Premiums

陈羽北 says reports describing “a few million dollars challenging OpenAI” are one-sided. DeepSeek is backed by 幻方, which has accumulated substantial compute, so total R&D spending is not low; he also qualifies his view with “if I’m not mistaken,” saying the team did use some of Nvidia’s best chips. There is therefore no basis for concluding that high-end GPUs are no longer needed.
John defines the impact as a double-edged sword: DeepSeek could reignite model and application startups, bringing more companies into the compute market, but the lower chip tier needed to train and deploy the same model could also weaken Nvidia’s pricing premium. “It could be that the premium declines while sales volume increases.”
The decisive variable is not the bill for one training run, but the application layer. If GPU prices were the main obstacle preventing applications from taking off, cutting costs to one-tenth or less could unlock demand; if the real constraints are product, data, or commercialization, chip volumes may not offset the pricing pressure.

6. DeepSeek Climbed the CUDA Wall but Did Not Tear It Down

John initially described Nvidia’s two major barriers as chip interconnect and CUDA. The episode first referred to the former as InfiniBand, while the host later used NVLink as the broader shorthand. The competitive focus is no longer single-card performance, but multi-card interconnect and an entire general-purpose software ecosystem.
MoE distributes different Experts across different cards and allows temporarily unused Experts to enter a dormant state. In John’s view, this reduces the importance of some interconnect traffic; optimization directly at the PTX layer also reduces dependence on high-level CUDA APIs.
泓君 asked whether this really meant bypassing CUDA, and John immediately corrected the wording: “It did not completely bypass the CUDA ecosystem.” PTX still sits within its lower-level stack, and model-by-model optimization is a major engineering undertaking that not every small company can reproduce.
He uses a wall as the analogy: people once assumed the wall could not be climbed, then DeepSeek jumped over it and said, “it’s possible to go over.” That does not mean everyone can immediately cross it—the wall has not gotten lower. What has weakened is Nvidia’s pricing premium, not the general-purpose ecosystem itself.

7. AMD and ASICs Get a Window, but Software Still Decides the Long Game

John expects AMD to benefit in the short term because it has already announced DeepSeek support; longer term, the outcome depends on whether DeepSeek becomes the gold standard. If the next generation of models changes architecture, AMD will have to adapt again, whereas CUDA’s advantage is that “one system works with whatever software comes over.”
For inference chips, he still expects Nvidia to remain ahead. Groq, Cerebras, TPU, and other solutions mainly differ in hardware through the mix of CUDA cores and Tensor cores; once Nvidia confirms that a particular ratio has a market, it can follow. The truly difficult part to replicate is software maintenance.
Groq’s expansion from chips into data centers and the cloud shows that insufficient software can force companies into a heavier vertical stack, but its financial resources are far weaker than Nvidia’s. John’s conclusion is that the more valuable chip startup opportunity is “building software better than CUDA,” not continuing to fine-tune DDR or the ratio of different cores.
Huawei “is possible,” but China faces constraints beyond CUDA: high-speed memory, assembly, packaging, TSMC manufacturing, and ASML lithography all sit within a restricted supply chain. A viable alternative requires closing the gaps in both hardware manufacturing and software ecosystems.

8. Continued Architectural Change Could Extend the Value of General-Purpose GPUs

泓君 points out that if the Transformer architecture had remained unchanged for 7 years and then frozen permanently, the market could design ASICs specifically for Attention, which would actually hurt Nvidia. CUDA’s generality is most valuable when model architectures continue to evolve.
She believes DeepSeek’s introduction of MoE and Multi-head Latent Attention once again changed the layers and the way computation is performed, which may not be bad news for Nvidia. The more frequently models change, the harder it becomes for specialized chips to pick the right architecture, and the more developers need a general-purpose platform that can support multiple architectures.
The shift could also pressure OpenAI. 泓君 relayed Sam Altman’s concern that if a better architecture than Transformer is found, the industry’s existing investment could be “in trouble.” DeepSeek-driven architectural change does not have the same implications for the chip leader and the model leader.

9. Open Source Lets Application Developers “Have Their Own OpenAI”

John recalled that application startups were in a slump in the second half of 2024, partly because developers feared that OpenAI’s next release could erase “80% or 90%” of their product moat. If a product is built on the OpenAI API, the supplier controls both functionality and pricing, leaving startups little room to compete.
After DeepSeek went open source, small teams gained “their own OpenAI”: a model that was capable enough, inexpensive, and offered continuity, without the daily fear that a new release such as o4 would overwhelm their product. John described that platform risk as “there was a dragon, and now it’s gone.”
Closed source has not therefore lost its purpose. John compares Apple with Android: centralized control is more likely to produce leading products, while open source drives adoption because everyone can use it cheaply. But “if closed source is still worse than open source, it may have no meaning”; incumbents that fall behind can only “fend for themselves.”
陈羽北 added from product experience that many open-source models cannot be deployed directly in production. But Llama, for example, provides a foundation that already handles roughly 90% of tasks. Stronger DeepSeek or Llama models will lower the barrier to productization, generate more real applications and feedback, and prevent the ecosystem from being dominated by a single model company charging excessive prices.

10. Low-Cost APIs Reflect Full-Stack Efficiency, Not Permanent Replacement of New Chips by Old Ones

泓君 cited R1’s pricing: RMB1 per million cached input tokens, RMB4 per million uncached input tokens, and RMB16 per million output tokens. By her estimate, the various o1 tiers cost roughly 26–27x more. John believes the gap comes from a full-stack optimization effort spanning PTX calls, MoE load balancing, and chip selection, not one isolated trick.
Models can expand beyond H100 and A100 to H800, H20, and other platforms, reducing the cost per token. But John also allows for another possibility: “OpenAI may have already brought the cost down internally,” without passing those savings through to retail pricing.
Relying on old cards such as V100 is not a long-term answer because Nvidia will discontinue legacy chips, while existing inventory will depreciate and become more expensive to maintain; gaming cards are also restricted from use in legitimate hosted deployments. The more sustainable path is to improve utilization on new chips through optimization and virtualization.

11. Small Models Push the Compute Economy from Renting Full Cards to On-Demand Slicing

John’s team has been betting for years that small models will proliferate. An A100 or H100 has 80G, while many distilled models need only around 10G; one 7B model they recently ran used about 20GB, and they carved roughly one-third of an H100 to run DeepSeek rather than paying for unused capacity on an entire card.
Inference workloads are inherently elastic: customer demand can rise or fall suddenly. If every physical card carries substantial unused capacity, the waste scales with every expansion. The value of GPU virtualization is “zero waste,” aligning deployment costs more closely with actual demand.
DeepSeek reinforces the trend: large models first generate reasoning trajectories, then distill them into more small models. Customers do not want to redo PTX optimization themselves; they want an out-of-the-box solution that is “simple, convenient, quick to deploy, and cheap.”

12. Edge-to-Cloud Systems Can Unlock Idle Compute and Data

陈羽北 has observed rapid progress in small-model capability over the past 6–8 months. On a single, clearly defined task, a sufficiently trained small model may differ little from one 10x or 100x larger. Microphone noise cancellation is one example: scaling the model further may not materially improve the final SNR, while a tiny neural network may be best suited for direct device integration.
A watch could handle queries and API calls locally, offloading complex tasks to the cloud; Qualcomm phone chips can deliver up to 50 TOPS and, in his comparison, are “not much worse than an A100,” enough to run small language models, VLMs, and basic Audio ASR.
His vision for AI infrastructure is a layered architecture spanning devices, the edge, and the cloud: sensors handle ordinary perception, edge devices execute more agentic functions, and only the most complex tasks go to the cloud. Since phones ship at billion-unit scale and sensors at an even higher order of magnitude, aggregate global device-and-edge compute could reach 100x the compute of HPC GPUs.
Data has a similar structure: roughly 90%–99% of data remains on devices and at the edge, while camera video is often “use it or lose it” and cannot all be uploaded. On-device AI can first identify valuable content and act as a data compressor for larger models; the selection criteria, however, remain “small, fast, and good enough,” and robustness cannot be sacrificed for cost.

13. Reasoning Matters, but a Three-Order-of-Magnitude Efficiency Gap Is the AGI Hurdle

陈羽北 defines reasoning as requiring intermediate steps to reach an answer. Checking the weather requires only a single API call and is not reasoning; finding the fastest route from Palo Alto to New York requires comparing airports, driving distances, flights, and multiple paths, making it a reasoning and planning task.
Rich Sutton’s 2019 essay “The Bitter Lesson” identifies two of the most general capabilities as learning and search, while warning: “do not try to be too smart.” Pretraining and fine-tuning lean toward learning; reasoning and planning lean toward search. DeepSeek combines the two by using self-directed search to generate candidates and rule-based rewards to drive learning.
Human data efficiency is still higher by several orders of magnitude. One estimate says a person receives fewer than 100 million tokens by age 13; another, based on 30 tokens per second for 12 hours a day, puts the total over 20 years at no more than 10 billion, versus 15 trillion used to pretrain Llama 3. 陈羽北 compares this with the steam engine’s efficiency rising from roughly 0.02% to around 20% and asks when we can improve “data-burning efficiency” by three orders of magnitude.
That is why 陈羽北 believes Dario Amodei’s 2026–2027 timeline remains challenging. Dario combines annual 4x or 10x efficiency gains from Scaling Law, algorithms, and chips with an increase in reinforcement-learning investment from $100K–$1M to $100M, predicting that AI will exceed most people in “almost” every industry by then. 陈羽北 says the question may be worth revisiting in 5 years, but from the standpoint of 2025, surpassing humans across “every industry” within the next 2 years remains extremely difficult.

14. World Models, Self-Improvement, and Undisclosed Data Will Decide the Next Round

陈羽北 uses three comparisons of scale to explain his caution: the human brain runs at roughly 20 watts, while a 600B model needs about 16 cards and consumes roughly 20,000 watts; humans encounter no more than 10 billion tokens before age 20, while Llama is trained on 15 trillion; and a jumping spider has only several million neurons yet can perform complex 3D navigation, while engineered models often contain billions of parameters.
John identifies another hurdle in data ownership: years of startup experience and other domain knowledge exist only in individuals, have not been fully digitized, and have not been handed to models. 陈羽北 sees the deeper issue as self-improvement—when will machines find their own learning signals, perform data curation, and “make every token count,” rather than continuing to depend on humans in the loop to prepare data?
陈羽北 views the world models promoted by Yann LeCun as a critical direction: given the current state and an action, predict the future. GPT can be seen as a world model without actions; DeepMind’s Genie 2, by contrast, can start from a single image and let an agent move forward, backward, and through the scene. Once the model is accurate enough, reality can be replaced by its copy, with reinforcement learning, model predictive control, or direct Search used to find trajectories.
The undisclosed parts of DeepSeek are precisely what he most wants to understand: its data composition, balance, curation process, and specific training details. John ultimately frames the industry impact conditionally: if DeepSeek continues to push the envelope and becomes the application gold standard, other chips will gain an opening and Nvidia’s premium will remain under pressure; if Llama 4 is far stronger, “everything goes back to square one.”

Breaking: Gemini 2.0 Flash Goes Live - Inside Google DeepMind's Latest Release with Logan Kilpatrick DeepSeek, Open Source, Tariffs, DOGE, Market Impact | BG2 w/ Bill Gurley & Brad Gerstner