DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23 · primary
Themes: Open Source as Strategy · US-China Framing · Economic Argument · Spectrum of Motivations
Chinese labs release weights because openness is their competitive strategy against closed frontier labs.
Then on the matter of open source, we thought it through very clearly from the start. First, the first reason is vision; second, we believe that to succeed commercially with AI, open source has benefits.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Compute, not algorithms or talent, decides who can compete at the frontier.
So the difference between us and the US, I believe, is a difference in resources. We might believe that all the differences we see—including the talent difference, the model capability difference, the application difference—can all be considered as due to the difference in compute resources.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Talent isn’t the bottleneck—resources are the biggest bottleneck. Resources first affect talent cultivation, because with little compute, our opportunities to run experiments are relatively few, so our talent overall has a gap with the US. The talent gap is essentially also because of the compute gap.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Compute resources—on one hand, cards simply can’t be bought domestically; on the other hand, our capital investment is less than the US. We’re much less at the capital investment level, and talent-based salaries account for a very low proportion within this. You see them offering salaries of a hundred million dollars, but calculating it, talent salaries still account for very little—the bulk is still compute.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Architectural and training efficiency is what lets a latecomer reach the frontier, so scarce compute is not disqualifying.
Another reason cost goes lower is: the lower the cost, the more I can train larger models, the more I can afford larger models. On the same compute, when compute is limited, if my computational efficiency is higher, I can afford larger models.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
This narrative is lagging one to two years, but using only one-twentieth of its compute. So in the future we want to rewrite this narrative—that is, we use one-nth of its compute, but shorten this time even more, shorten it to 6 months, 3 months—I think this is a goal.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
So the gap between us and the US might be lagging the US by 12 months, lagging the US maybe 12 to 18 months, or 6 to 12 months. Anyway, simply put, lagging the US by two years, then using only one-twentieth of the US’s compute to accomplish this thing.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Cheaper and more available models increase total consumption of intelligence rather than displacing it, and the volume creates economies of scale.
I just said why we don’t keep cutting prices—it’s because it’s inelastic. That is, if I cut prices further, demand won’t increase more, or if I cut prices further, demand increases very little. Because at this price everyone can already afford it, everyone feels this price is satisfactory, and won’t stop using it because the price is expensive.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
So cutting prices—first, the company won’t have more revenue; for society there’s no more value either, because everyone is already satisfied with this price. Even if you lower the price further, it won’t increase society’s sense of happiness much. Right, OK, but on this issue just now, on this matter of pricing, we’re definitely not...
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
This is our standard. It’s actually not profit-maximizing—if it were profit-maximizing, we should set prices higher. Because in this price range, user demand is inelastic, meaning if I cut the price in half, or if I raise the price by double, the token consumption doesn’t differ much.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Open models keep AI continually available, rather than dependent on any one provider's decisions.
But AI is big enough—ultimately it might occupy, say, ten percent of human society’s GDP. That’s actually a very large number. One person monopolizing this thing—you can’t monopolize this thing, you must share with others, otherwise you definitely can’t survive.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
This is different from previously open-sourcing a piece of software, because that software’s market wasn’t that big. But AI is simply too big. If we want to monopolize this benefit, then we’re bound to be abandoned by history. I think the main point is that this is an objective law, this is a view of history.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Open weights commoditise the model layer, moving durable value to distribution, data, and the infrastructure underneath.
So on this matter of open source, actually my judgment is that it’s sustainable. There’s no conflict between open source and commercial paid services—the premise is that under sixfold profit there’s no conflict.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
We provide as much help as possible to the open source community, assisting everyone to deploy our models. I’m not worried they’ll steal this business from me, because this market is big enough. I’m only worried they can’t deploy it—that they got some details wrong, made the effect relatively poor, or that their cost is relatively high.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
This also shows there’s actually no conflict. All of last year, on the C-end I basically open-sourced, and then the C-end service saw no conflict, really saw no conflict. So, this is the open source part.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Open weights prevent AI from being controlled by a few actors, because anyone can run them.
It’s not that if I don’t open-source, I can monopolize this market—that theoretically doesn’t conform to objective reality. You’ll definitely encounter many obstacles, there will definitely be other methods to stop you from achieving this goal.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
A coordinated slowdown is worth pursuing even at the cost of advantage, because uncoordinated speed is the greater risk.
First, this concession—for our company internally, we’re very happy, everyone is happy, employees feel a sense of accomplishment, and we’ll have cohesion because of it. And this concession is good for society, society is also happy, other peers or ordinary people will be happy. So this restraint, my understanding is, this restraint from a long-term perspective can increase our probability of achieving AGI.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Open models damage the business case of closed models by putting downward pressure on prices.
But if you wanted to earn a hundredfold profit, then open source would indeed affect your earning a hundredfold profit, because third parties would deploy—they might be at twentyfold cost, which would be lower than yours.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
The open source part I can say more about, because many previous questions were about open source. First, I think we will open-source, and then even our strongest model will probably be open-sourced too. Because I don’t see the benefits of closed source, I don’t see the inevitable benefits. ByteDance’s model is closed source—what benefit does it have? I don’t see any benefit.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
So open source, I think has no impact whatsoever on our business model. The premise is that we only earn sixfold profit—ten months to recover cost roughly corresponds to sixfold profit. Under the condition that we only earn sixfold profit, open source won’t have any impact.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Then I’m also not worried about others deploying our models and then competing with us—not worried at all. We even hope they can deploy it successfully.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Open models move the binding scarcity to hardware — compute, chips and the networking between them.
Under this situation, domestically we can consider it as having just started in the last half year, so in time, I think more time is needed. This has little relation to capital investment, because even without more capital investment, the original capital is enough for it to expand at the fastest speed.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
But this speed has a ceiling—the bottleneck isn’t that I can immediately have more people, it’s not stuck on money, nor stuck on cards. But it’s indeed in a rapid-expansion process.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Building national AI capacity on sovereign infrastructure is sound policy for smaller countries, not merely a way to sell deployments.
Will we build large clusters ourselves in the future? I think building large clusters ourselves is definitely necessary—we’ve always been doing this ourselves, all our clusters are self-built. But whether to develop our own chips in the future, I think depends on how big the returns here are, depends on how big the returns are.
DeepSeek's Liang Wenfeng Breaks His Silence · 2026-07-23
Positions are read from the claim graph, and every passage is fetched from the archive when this page is built. A passage the archive cannot supply fails the build.