Regular readers will know that ever since Deepseek came out, with token costs of three to five percent of US models, and open source which you can run on your own servers, I’ve been saying that China would win the AI war.
I Went From $3,000/Month on Claude to $5/Week on DeepSeek
And honestly? 80% of my work is identical.
For the past two months, I was burning $3-5K monthly on Claude Code. Every idea from design to development to testing – full end-to-end automation, even simulating users to test my products and provide feedback. Extremely token-intensive.
But Claude’s caching sucked, making it insanely expensive. Then I discovered DeepSeek V4.
The numbers: • Claude: $5 input, $25 output per million tokens •
DeepSeek: $0.28 input, <$1 output (with their current discount) • DeepSeek cached: $0.0002 – literally less than a penny The caching optimization is game-changing.
Once DeepSeek has seen content, it basically stops charging tokens. My result: $5/week vs $1,000/week for the same workload.
The Chinese have been optimizing for efficiency. Frontier models are a little better, at each generation, but the Chinese aren’t sitting still on quality either. What will happen is convergence: where all the models are about equally good. Until then, the Chinese models may be 3 months to 6 months behind, but that’s all and when they have the advantage of being, not just 95% cheaper as before but over 99% cheaper, and when they’re open source, so the vendor can’t just increase prices or restrict usage without users having any ability to run the model themselves, well, only an absolute moron would plan on using Anthropic/Claude or Open-AI.
The West is building vast numbers of data centers, far more than the Chinese, because nothing is optimized for efficiency. Our models use far more electricity, far more GPUs, and far more water. Additionally the Chinese are working hard on real-world AI uses: robotic AI, in other words, so that their AI can be used for actual production, and pushing on humanoid robots so they can take care of their old people, do household work and so on: Chinese AI is optimized to do shit work so you can read and write and paint, while Western AI is optimized to do creative work so you can shovel manure, do your own laundry and clean toilets.
I simply cannot fathom; I literally spent 5 minutes repeating to myself, “how can they be so fucking stupid?” what American tech leaders and the people throwing money at American AI companies are thinking. They lose money with every single query. Starbucks just got rid of their inventory AI because it couldn’t tell the difference between oath milk and cow milk even after months of trying to make it work. The only large use-case for American AI is programming and even there quality questions are rife.
But even if it winds up being as great as they say, they’re still going to lose, because it’s not going to be enough better than Chinese AI (if it’s better at all in a few years, which I doubt) to make up a 95 to 99% cost difference plus the safety that open source provides.
We are ruled by morons. Our rich people, whom we celebrate as brilliant, are idiots and fools. The moment Deepseek came out, the response should have been to figure out how it was so much more efficient and add that to American AI. No one even considered it, they just threw hundreds of billions of dollars more money at doing the same thing they were doing before.
You can’t look at American AI, as anyone who uses is, and not realize it’s massively subsidized and that at some point they’re going to have to charge the actual, real cost. As with everything else, the plan is to destroy alternatives, then once they’re damaged beyond repair (like Taxis and “rise sharing”, raise prices.”
But this won’t work with AI because the Chinese aren’t playing along and it’s software that can be run on any server, if you have access, and, again, Chinese models are mostly Open Source and way, way cheaper.
I’m shaking my head as I write this. This is the greatest mis-allocation of resources I’ve seen in my entire life. It makes the housing bubble (out of which we at least got some homes) look brilliant and wise.
Morons. Our leaders are morons. And this isn’t even a case of “well they’ll do well out of it and they don’t care about anyone else”, it’s a case of “they’re going to lose hundreds of billions of dollars.”
One more bailout. Maybe. This is the last time.
And the West is DONE.
Because we aren’t even ruled competent evil people. We’re ruled by utter idiots.
(I wanted to write that they’re the dumbest money in world history, but, alas, that still goes to the Dutch tulip bubble. Though, again, at least they did get a lot of beautiful flowers out of it, which is more than we’re going to get.)
What I write here is for the benefit of everyone, but alas, I live in capitalism and I, and the site, take money to keep running. If you value the writing here and can, please subscribe or donate.
bruce wilder
The Dutch flower markets are impressive.
warvigilent
Not to disagree that the rich and techbros are fucking stupid but remember between covid and epstine the balance between stupid and evil is clearly on the dark side. They are not so stupid as to not see this possibility, there is no way they don’t have a legion of toadies to find out. Its that they don’t care about making money because that doesn’t matter anymore. This is about the total surveillance state, this is to run their fantasy tech utopia where everyone else is just a slave. The data centers are to run all the ai backed surveillance to find any dissenters / terrorists etc , Eg ai to watch all the cameras all the time. real time tracking of everyone. In addition they will use this to replace most managers as well, or turn them in to just managers of “teams” of ai agents. Instead of the chore or medical care robots we have been working on military robots and already on the streets police robots, Eg the “dog” robots seen around, will be armed . soon there will be police drones everywhere first to track and stun suspects but soon armed more and more lethally . Then the bots will be favoured over police who will be replaced a la robocop. the once favoured classes are being swept up in this generation of layoffs for the new corporate ai gods , manager and the enforcers will both be culled.
both sides do it
Not that the main point of the post is wrong, but the DeepSeek focus is misleading.
And “The moment Deepseek came out, the response should have been to figure out how it was so much more efficient and add that to American AI” is just flat-out wrong.
DeepSeek has piggy-backed off of the corporate US models to train its own models through a process called “distillation”.
That’s why the cost is so low: siphoned R&D value.
(US corps figured this out and now look for signs of distillation attempts to stop it.)
Corporate espionage like this, as Ian has pointed out, is a time-honored way for rising industrial powers to gain at the expense of declining industrial powers, a la US chemical companies helping themselves to German company secrets around WWI.
But it’s fundamentally a different dynamic than what’s framed in the post.
China is still going to integrate AI into its political economy better than the US for reasons given in this post and previous posts:
– the focus on adapting AI to physical industrial applications
– legal rules forcing open-source for AI consumer options while also being the strictest “do not invent the Torment Nexus” regulatory regime on the planet
– nascent dominance in university research
– chiiiiiiiips (resource bottlenecks and, eventually, quality)
But this post is a l’il off-kilter in the specifics.
The vitriol is spot-on, and seconded. If not encouraged. (Take care of yourself, Ian.)
GlassHammer
Ian,
1. AI is the logical conclusion of Silicon Valley’s “Unicorn Hype Machine.
2. Western AI could only make money as a monopoly and achieving that is impossible now.
3. The actual plan for Western AI is a federal bailout after massive IPOs leave every passive investor “holding the bag” when the losses move through the market.
Ian Welsh
No, distillation doesn’t lead to the actual running of the model being cheaper. It helps with the costs of each advancement, however. Though note that Deepseek is better at some things than the frontier models, it’s not just a copy.
bruce wilder
The urgent need for massive investment in data centers remains a mystery to me. DeepSeek in its latest model iterations can require enormous computing power — the equivalent of a truly impressive gaming computer with massive allocation of GPUs and RAM.
But, to do what, exactly? What are the tasks to be done?
I see speculation about “making money” by displacing the work of white collar drones, writing memos and coding websites and “apps”. Maybe doing diagnostic review of x-rays and c-scans.
I see more than speculation about devastating education by enabling students to cheat their way through youth, to emerge functionally illiterate. And, like warvigilent, I see the beckoning surveillance society. That last is the most plausible explanation to my non-technical mind for what appetite for data could require such vast capacity to digest.
If warvigilent is right, the data center boom doesn’t end with the collapse of an ill-conceived bubble, but with a reaping of soylent data.
Of course, it could be both: a financial collapse followed by a reaping by the total surveillance society.
ventzu
The ruling idiot class are enthralled by the money they can make: consulting firm ceos implementing AI; corporate ceos being at the frontier, and cutting jobs; and politicians jumping on any idea that looks like growth. Even the political strategists think western AI can dominate the world, and that they can both earn economic rents and ensure USD pricing (‘pax silica’). Having no other ideas, they somehow think they can beat the Chinese open source initiative.
It is a classic “emperor wears no clothes”.
And this ignores the most obvious fact that gen AI dumbs people down, and can only lead to the end of anyone actually generating new content.
The ruling class put themselves on a pedestal, when the reality is they are just as dumb as everyone else; they have just learned the art of sycophantic networking better than others.
Feral Finster
The short term bet is to get out before the bubble pops.
The longer term bet is that western governments will firewall DeepSeek, buy it out, buy it off, or otherwise put it out of existence, sort of like how TikTok was neutered.
Basically, it’s a bet on the continued political influence of the Anthropics of the world.
Fair competition is for fools. Smart people grift and cheat.
I don’t have a dog in the fight, so to speak.
Feral Finster
Or yes, a bailout. Stupid cat!
GM
It’s a window of opportunity issue.
And it has nothing to do with profitability.
The bet is that whoever gets to a certain level of AI and with the infrastructure to run it will lock in global dominance in perpetuity, by being able to do absolute surveillance and immediate suppression of all dissent and resistance. And whoever gets there first will be the master of the world forever, because there will be no way to break the chains.
This is also why they are seriously going for deep underground data centers and data centers in space. The only way to get rid of the data centers in space is with ASAT weapons, but ASAT weapons pretty much require a state actor to put together. But once Russia, China, Iran and North Korea have been destroyed and subjugated (which is part of the plan, obviously), there will be no such independent actor. No rag-tag rebels will be able to amass the resources to deorbit the data centers in space, which will lock in the perpetual dominance of whoever controls them. Putting them deep underground is also a level of protection above that provided by the current practices of having them on the surface. I am not quite sure how cooling will work in either case though, but the intent is clear.
Warlords in the 6th century didn’t worry about the stock market and the quarterly profit/loss statements. Their goal was physically destroying their enemies and placing as much land and whatever peasants survived on that land under their control so that they can be exploited. That has always been the fundamental logic, it has never changed, the stock market and capitalism were just a very complex front for what is really happening. We will just be going back to the more direct ways of doing things.
spud
Keen has a appropriate video out today. gold investors should have paid attention as to why its done so far, the opposite of what has always been its hook that’s its increasing in value, or will hold its value in a deflating economy. you gotta sell in a deflating economy to raise scarce cash, to service your debts. and most so called investors, never really understand this.
Keen points out that is was the 1990’s where wages really started their downward spiral.
so in reality, we have been in a long depression like scenario since, with sudden downward jerks from time to time.
when the decision in 1993 to get rid of actual production, and rely on asset and paper manipulation was going to bring america nirvana, and keep its dynamic and innovative economy, gave us bubbles, derivatives, bitcoin, AI and other assorted ponzi schemes instead.
whilst handing the socialists our real wealth and power.
https://www.youtube.com/watch?v=oSYPZrGEiok
Professor Steve: The 2026 Financial Crisis Is About to Begin, Let Me Explain!
————
the expected bailout will only be a bailout for a few. it will have little to no bang for the buck for the rest of the western world.
most of the western worlds population, can’t borrow, when they can’t even pay back what they owe now, let alone eat, keep a roof over their heads, and get a tank of gas.
as Ian has said, its all over but the shooting. the chosen few that get the bailout, will agitate for killing off what they consider to be the parasites that were actually the real creators of wealth.
you really can’t make this stuff up.
Jefferson Hamilton
Most people don’t want to read, write, or paint, they want to scroll tiktok or Instagram or at best play videogames. Even if they wanted to do those things the truth is well over 99% of people suck at them and their taste is shit.
Anyway, I have assumed for a while that warvigilent is more or less correct.
marku
I’d suppose that the insiders sell out before the crash with massive earnings intact. The government buys out the leftovers and data centers, and runs a massive security state.
And the CIA dems will be just as happy with this as the Trumpers. They are already calling anti AI thoughts “terrorism”
Feral Finster
GM gets it.
Purple Library Guy
@Feral Finster They can’t kill DeepSeek, just like they couldn’t kill Linux, because it’s open source. Even if they do huge barriers to it as a Chinese product, someone in the US or Iceland or somewhere will just grab the code and do a startup.
Brian M
I’m not sure that it’s pure stupidity, although I’m that there is some of that. I think a lot of this (and much else that we suffer from in the US/West) is the result of the sea of American Exceptionalism in which we swim from birth. Simply put, if you believe (as in take it as an article of faith) in American Exceptionalism, then there can never be a reason to even consider anything that is not American. By definition, American is better. Always. Everywhere. In every way.
I think this is what constrains much of a thinking. The result is what amounts to cognitive dissonance. Their world view requires that they are exceptional, their companies are exception, their products are exceptional, but the metrics and facts suggest that this just isn’t the case. At best, they are marginally better. This is a reality that cannot be digested. Therefore, it must be stomped down. How? By spending ever more money doing the same thing while expecting a different (exceptional) result.
As Einstein is famously misquoted as suggesting… insanity.
mago
I’ve been railing against AI for so long that I have nothing left to say except I’m still astounded that so many people I know who are otherwise seemingly aware are drinking the kool aid. I just keep my mouth shut these days.
I, too, have to agree with warvigilant’s assessment. When the AI movement goes tits up all those sprawling data centers will be used by the surveillance state.
(Also Jefferson Hamilton got in a good hit.)
I’m so tired of those tech creeps with their Dark Enlightenment agenda and dreams of world domination I can’t even begin to relay my disgust, not that it matters. I think it’s all doomed to failure, but there’s going to be a world of pain on the way down.
Purple Library Guy
GM, that does seem to be the attitude, but I don’t think they can pull it off. They underestimate the reality of the world. It’s true that they’ve got this huge infrastructure of surveillance and all that, but what have they really been able to do with it? Decide what ads people look at. That’s mostly it. And they’ve been able to keep doing that because people mostly don’t really care.
Lately they’ve started doing weird things with pricing, where they charge people surveillance-based prices that are higher if they think that person can be suckered into paying more. But that’s barely started and there’s already moves to make it illegal, because people CARE about that, they can see where it means they are personally having their money stolen.
Sure, the techbros have also started serving surveillance information to governments who then use it to choose the slain, but that just underlines the fact that the techbros can’t do anything serious with all this crap by themselves, just using all their tech, they need meatspace infrastructure involving people and money and weapons to actually do things. In the end, you can be the techiest techbro of all time with the monopoliest monopoly on information gathering, and if the government where you live sends people to your home with guns, you are still dead or in jail.
Feral Finster
@PLG: good point, but an open-source program cannot be firewalled, akin to The Great Firewall of China?
StewartM
I am thinking that the AI market ‘enthusiasm’ is the result of an economy where we paradoxically have too much money available for investment, while having too little money available for consumer demand. Ergo, money starts flowing into bubbles and likely useless enterprises because the useful ones (the ones that actually meet human needs) are starved for demand.
Markets only “work” (and imperfectly at that) to satisfy human needs if there is a high degree of income and wealth equality. The more unequal income and wealth inequality, the more money flows away from addressing actual human needs into either unimportant luxury items or just bubbles. It’s the old saw about how ‘socialism means no IPhones”—which may be true to a degree, because socialist societies actually do a better job at meeting basic human needs (housing, food, healthcare*) and thus can put less effort into “wants” like pricey phones for the well-off.
*– this is actually true; normalizing for the same level of development, socialist countries do a better job at the basics than the capitalist ones. It’s why a poor country like Cuba has maybe the best health care in Latin America and Cuba despite being poor (and being sanctioned, and now blockaded) has equal life expectancy as the US.
Carborundum
I use AI for coding and I have to say it’s pretty obvious why they’re building out remote compute. Apparently, the current full versions of DeepSeek are specced for 8 GPUs, each with 141GB of memory – I don’t know what home systems other folks are running, but I feel safe in saying that there aren’t many with that level of compute at home.
I’ve spent some time running the distilled versions of DeepSeek that are locally viable and they are much less capable than what I can get via remote compute. Multiple orders of magnitude slower, less complete code, etc. In my benchmarking, stuff that took me literally 4 seconds using ChatGPT or Claude took 10 minutes locally on DeepSeek (and that doesn’t include the problem of DeepSeek repeatedly not parsing entire files for reasons known only unto itself) – and this was just porting SAS script to R, not complexly reasoned stuff.
I don’t have a lot of insight into whether they’re over-investing in compute, but I know for sure isn’t as simple as one can run DeepSeek on one’s own local server and get as good a result.
Purple Library Guy
@Feral
Possibly. But who would do it, exactly? The US government? They could only firewall the US. Google? They could make it disappear from search, theirs at least, but that probably wouldn’t be enough to really stop it, and might catalyze European efforts to come up with a Europe-based search engine.
And I think it would be a lot of work. I have a feeling that it’s almost as hard to Great Firewall one product as it is to do the whole schmeer–so would it be worth rebuilding the whole internet just to put in a block against one readily changeable item of software? Would the US even have the government capacity to undertake that kind of project?
elkern
OT, but related, maybe; an interesting anecdatapoint that I want to share here, reminded by a couple comments above…
I’ve been listening to Red Sox on radio for 30 years, but my local station was recently sold (to NPR!?!), and I’m now stuck listening online. The broadcast is the same, but the ads are different; notably, I now get some Spanish language ads (their targeting sucks).
In the last week, I’ve started to hear ads for… TikTok !?! (and bizarrely, most of them use a bad fake British accent)
My instant reaction is that this means that TikTok is *dead*, at least in the USA. Silly Rabbit, TikTok is for kids, not geezers who listen to baseball on radio….
Ian Welsh
Carborundum,
the post is about businesses. No business of any real size can’t afford to run the full version, plus I expect hosting companies to offer hosted versions of open source models within a couple years.
Jorge
The thing to remember about “Deepseek v.s. The West” is: the US companies all started as a buncha geniuses setting out to build a Machine God and also maybe make some money, while Deepseek was a fintech company filled with people building internal products and the Big Boss decided to build a bigger product.
ProNewerDeal
as a consumer, do you recommend trying Deepseek? It is not on the free list of AI prompts at duck dot ai . It seems that chat dot deepseek dot ai allows to try it using a Google login.
Ian Welsh
Why not give it a try?
But the frontier models are a bit better IF you don’t use AI much. If you’ve got a big bill, Deepseek can save you a lot of money.
What I wouldn’t do is integrate with Claude or Chat-GPT, because prices are going to go up a lot eventually, plus they keep changing usage amounts.
Prasket
@ianwelsh.net
I've been very happy using Qwen3 (alibaba cloud) models in past few months locally for all kinds of things. I have yet to give any money to AI companies for my usage of LLMs and honestly don't see why I would.
The future is Open Source Local LLMs IMHO.
both sides do it
Distillation allows smaller models with fewer weights. This lowers computation costs at runtime, beyond the training costs.
Larger DeepSeek models deploy comparable numbers of weights as larger models at US companies but as a mixture of smaller models with specialized subdomains hooked together. Less than a tenth of the weights are activated in any given query -> less computation cost. (Though the chain-of-reasoning architecture re-runs results back through the weights.)
Agreed that DeepSeek has def made advancements, they’re not just making copies.
Ian Welsh
I stand corrected!
Feral Finster
@PLG:
I know little of the technical issues, but europeans are always looking for an excuse to censor and to curry favor with the Americans.
Firewalling off Chinese aI would give them an opportunity to do both.
Purple Library Guy
@Feral Finster
Given the way the EU has been operating these days, I would expect them to make rules against Chinese AI, but also to fork it and have an approved government-developed “European” AI to avoid dependence on the American companies. Which has the same effect on the American companies.
Carborundum
What one defines as “of any size” matters a lot here. The vast majority of businesses below the quite large (call it a couple hundred employees or more) are not running a lot of their own services / servers in house. They’re overwhelmingly likely to be buying from providers rather than rolling their own.
I can totally see people moving to running something like a full DeepSeek model on independent hardware that they control (so far as I know, none of the major western providers actually allow this outside of NATSEC applications) – my point is that I don’t see that being physically located on-premises or actually directly owned by the users in a very large majority of cases. The concept of remote compute as utility is far too ingrained.
Medium-term I suspect that what is going to end up happening is that we end up with a contest between a large number of substantially more specialized models that perform better at particular tasks running on local compute and large centralized models as utilities (kind of a rhyming repeat of the PC vs. mainframe debate).
Altandmain
Ian, the main goal of the Western elites is to try to get AGI, namely an AI that is capable of truly original thought and creation. From there, they will rent seek the way Über does today after bankrupting the taxis.
It doesn’t look like the US elites are going to be successful. They are a victim of their own greed.
The Chinese are of the opinion that the Western approach is fundamentally flawed. Another consideration is that the US is trying to get rich people even more rich. China is trying to get a system to make the benefits of AI more egalitarian. It isn’t perfect, but it is working better.
This is the do or die moment for Western hegemony, so they have thrown it all in a reckless gamble.