YT YouTube2Text

YouTube2Text → Transcripts → Why Big Corporations are Secretly Shifting to Chinese AI Models

Why Big Corporations are Secretly Shifting to Chinese AI Models — Transcript

by Nick Saraev · 2,463 words · 368 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Big corporations are shifting to Chinese
0:01models. Uh Coinbase now runs on Chinese
0:04AI. They split their ad bill by
0:05approximately half by doing this.
0:07Airbnb's customer service runs on
0:09Chinese models. More than half of all
0:11open router tokens are currently on
0:13Chinese models. In this video, I'm going
0:15to explain why. So, the first thing you
0:16got to understand is the scale. Most
0:17people think it's just a couple of
0:18startups here and there using these, but
0:20it's only a few people that have
0:21actually talked about it. Big chunk of
0:23American companies are now using Chinese
0:25AI models for a variety of reasons. Um
0:27one of them being privacy, another being
0:28security. We're going to talk that in a
0:30second. You know, OpenRouter, for those
0:32of you guys that don't know, is a
0:33marketplace that people use to call a
0:36bunch of different AI models through one
0:38API. They basically simplify the process
0:41of dealing with AI models by giving your
0:43agents a simple replaceable API call
0:45format that allows you to swap it out
0:47for whatever model you think might
0:49necessitate the task. About a year ago,
0:51American models were about 70% of all of
0:53the traffic on OpenRouter. Today, it's
0:56about 30%. In July, Chinese models
0:59actually took all five spots of token
1:02volume. And they also carried
1:03approximately 58% of all tokens that US
1:07companies sent through it. That means
1:08that American companies were
1:09preferentially using Chinese models
1:11versus their own. If you compare that to
1:13Hugging Face, where people download and
1:15reshare open models, um Quen had over 2
1:19billion downloads this year. That's 2
1:21billion. That's about a quarter of the
1:22total number of human beings in
1:24existence. Meta, on the other hand, had
1:26only about 230 million. Because of this
1:28divide, a massive amount of
1:30infrastructure has actually been built
1:31on top of Quen. We have over 150,000
1:35models that have been post-trained,
1:36fine-tuned, and ultimately modified from
1:38the initial Quen branch. Coinbase,
1:40Airbnb, DoorDash, Cursor, um the search
1:43engine Ecosia, all of these use Chinese
1:45models right now. Amazon, Google, and
1:47Microsoft are also providing Chinese
1:49models at scale to customers. So, why is
1:51this happening? Three main reasons. The
1:53first is price. The Claude family of
1:56models currently offered by the flagship
1:58LLM provider Anthropic currently charge
2:00anywhere from two to $10 per million
2:04input tokens. OpenAI's flagship models
2:07cost anywhere from 10 to, assuming you
2:09use fast mode or ultra-fast mode, $60
2:13per million input tokens. We're not
2:14accounting for output tokens here, which
2:16can go as high as $300 per million. But
2:18this, for example, is Kimi K3.1 Pro,
2:21which is Moonshot's newest flagship
2:23model. Moonshot is releasing the open
2:25weights on places like Hugging Face and
2:28a couple of other providers as well. But
2:30if you want an out-of-the-box
2:31experience, you can also just pay API
2:32pricing directly to the Kimi team, who
2:35host the model, optimize it, and then
2:37also run inference on it for you. And
2:39right now, they only charge $2 per
2:43million input tokens. If you go a little
2:45bit back in time, you could see prices
2:47even lower at 95 cents per million
2:49tokens for K2.7 code. I know what most
2:51people are probably thinking. As of the
2:52time of this recording, frontier models
2:54from Anthropic and OpenAI are very close
2:57to on par with Kimi K3.1 Pro on a
3:00variety of benchmarks. But here's the
3:02thing I don't think you fully
3:03understand. We're already well past the
3:05point of good enough. Most coding
3:08applications do not actually need the
3:10most frontier model of frontier models.
3:12What they need is they need something
3:14that's fast, and they need something
3:15that is reasonably performant to act as
3:18the fingers of a broader hand that
3:19orchestrates things.
3:21In this way, Kimi K3.1 Pro and other
3:23models like it are a much more
3:25cost-effective way to do the same thing.
3:27And you can even fine-tune these models.
3:29The base model intelligence you get out
3:31of the box on Kimi K3.1 Pro might look
3:33just a tad lower on benchmarks than,
3:35let's say, Opus. And you can download
3:37these. I'm currently on Hugging Face
3:39here, which is a website where you can
3:41download a variety of open models.
3:43Moonshot says that the full K3.1 Pro
3:45weights will land here on exactly
3:48November the 4th. Kimi K3, their
3:50previous model, is up right now if you
3:52guys wanted to download that. But on
3:54November the 4th, you'll also be able to
3:55download K3.1 Pro, assuming you have
3:58enough hardware, and then even fine-tune
3:59it to the specific tasks that your
4:01company needs it to. Now, historically,
4:03this would cost a lot of money, but big
4:04corporations have both the funds and the
4:07wherewithal to take one of these models,
4:09spin up their own GPU clusters, or be
4:11more economical in who they're sending
4:13their hard-earned money and time to, and
4:15then fine-tune it and train it on tasks
4:17that they specifically need. Things
4:19like, you know, Airbnb's customer
4:20service, Lindy, which is an AI
4:22automation tool that lets you automate a
4:24bunch of business processes, recently
4:26moved their whole stack off of
4:28previously Western models to
4:30significantly more performance Chinese
4:31models. Bunch of people asked why, the
4:33founder Flo explicitly said, "Well, you
4:35don't need God to write your email, do
4:37you?" And that's more or less what we're
4:39getting at. You don't need the greatest
4:41and best generation of models in order
4:43to do most business tasks. And people
4:45are slowly waking up to the fact that
4:46the capability has moved far beyond what
4:49people are using it for, especially in
4:51most consumer applications. The next big
4:53reason is control. One thing that a lot
4:55of big companies are worried about is
4:57large language model companies in the
4:59West like Anthropic, OpenAI, Google
5:02DeepMind, and so on and so forth,
5:04technically have access to every query
5:06that any developer has ever sent to
5:08them. If you're working on something
5:10that is extraordinary data-intensive,
5:12let's say something that is super
5:14private and something that you don't
5:15necessarily want to share. It's your own
5:17intellectual property, it's your own SOP
5:19or process, and it's very unique to your
5:21company, it's how it makes money. Well,
5:23you might just be in a situation where,
5:26by working on what you want to work on
5:27with an American model, you're actually
5:29giving it all of your sensitive IP and
5:32information. It's not rare. This has
5:34happened a couple of times now where,
5:36you know, a set of researchers or a set
5:38of theoreticians were working on answers
5:40to very difficult problems, like a
5:41recent Millennium math prize problem,
5:43and then a couple days ago a large
5:45frontier lab posts the solution, almost
5:47like it's a magical coincidence. In that
5:49way, large language model companies are
5:51actually actively soaking up all of the
5:52moat, and then using that moat to train
5:54the next generation of models. Well, the
5:56reality is big companies have a lot of
5:58financial incentives, and they don't
6:00necessarily want to share those
6:01incentives or upside with AI companies
6:04like Anthropic or OpenAI or Google
6:06DeepMind. What they would prefer to do
6:08is keep their intellectual property,
6:10keep their SOPs, and sort of private
6:12workflows and ways of doing things, and
6:14instead just download open weights
6:15models like Kimiko 3 and have them run
6:17for them. Because they're open weights,
6:19because they're on their own devices,
6:21because they control the whole
6:22fine-tuning stack, they don't need to
6:23worry about somebody, I don't know,
6:24leading the legal department at one of
6:27these large language model companies
6:28creating some new legal product based
6:30off of their own legal workflows.
6:32Basically, despite the fact that
6:33everybody, myself included, calls them
6:34Chinese AI models, you don't actually
6:36have to send your data to China in order
6:38to use the AI model. Contrast that with
6:40Anthropic, OpenAI, Gemini, and so on and
6:43so forth, and you kind of have to send
6:45your data at least outside of the bounds
6:46of your business, maybe not the country,
6:48but certainly outside of the bounds of
6:50your company in order to use it. Now, if
6:51you don't believe me, you can actually
6:53go and use these Chinese AI models right
6:56now, and you can test it for yourself.
6:58I'm currently on the Kimi homepage. You
7:00can get to that at kimi.com or kimi.ai.
7:03And just like you see ChatGPT,
7:06Anthropic's Claude, and so on and so
7:07forth, you actually have an interface
7:09that looks just like this. So, hey, how
7:11is it going? And you can see they have a
7:12cute little thing that's kind of
7:14reminiscent to X's Grok bot. I'm going
7:16to ask it to build me a simple landing
7:19page for a hypothetical coffee mug. You
7:22can actually click the thinking tab and
7:24see what is going on under the hood
7:26reasoning-wise.
7:28And yeah, there's really nothing special
7:29going on here. I'm currently running
7:31this using the Kimi chat app, which
7:33means that the data that I am working on
7:35is technically leaving the confines of
7:37the country. It is currently being sent
7:38over to, presumably, China. Depends on
7:41where they have their GPUs and their
7:42inferences being done. Um, but this is
7:45something that you guys could very
7:46quickly internalize in your company,
7:48assuming that you are a big corporation
7:49with a few hundred thousand dollars to
7:51spend on localizing and making inference
7:53run within the confines of your
7:55business. Now, you can see it's actually
7:56going through and writing all of the
7:58code. It's even designing the SVGs in a
8:01super clean, extraordinarily succinct
8:03way. None of this stuff is coming to me,
8:05by the way. This is all just a reasoning
8:07tokens. One of the reasons I like using
8:09Kimiko 3.1 Pro specifically is because
8:11the reasoning is a lot clearer to me
8:13than a lot of what's going on with, you
8:16know, Open AI's models, let's say, or
8:17Anthropic's models. If you guys have
8:19ever looked at like GPT-OSS-120B
8:21and expand the reasoning, it is freaking
8:23bonkers. I have no idea what this thing
8:25is saying. But this is still very much
8:26like in English. Okay, so it's actually
8:28going through and it's building me the
8:29SVG of the landing page. As you see,
8:31it's pretty cute there. We got a little
8:32coffee mug. And this is actually being
8:34built directly in the browser
8:35experience. I mean, you know, obviously
8:38I'd prefer an actual image of some
8:39hypothetical mug. I'm not going to lie,
8:41this is pretty damn clean. If your job
8:44is just designing simple and
8:45straightforward interfaces like this for
8:47a company internally, do you really
8:48need, you know, the next latest and
8:50greatest Gemini model? Or is something
8:53like Kimiko 3.1 Pro totally fine? So not
8:56only is it cheaper and faster, you also
8:58get it totally 100% localized. Anyway, I
9:00know most people are going to accuse me
9:01of glazing at this point, but yeah,
9:03Kimiko's awesome deep seek pretty cool.
9:06Obviously, they have some quirks and
9:07there are some questions as to how these
9:08were trained initially, but models like
9:10this are fueling the next generation of
9:13corporations. And because we're
9:14fine-tuning them, you actually achieve
9:16far better results on specific tasks,
9:18usually tasks that are highly wrote, of
9:20which a big corporation will have a lot
9:21of data and examples of of of being
9:23done. So what should you actually do?
9:25Okay, here's how I would decide if I
9:27were running AI inside of a mid-sized
9:30maybe big big company today. Um first, I
9:32would build infrastructure inside the
9:35company that allows me to swap models at
9:37will. So I'd actually build some layer,
9:39just like Open Router, within the
9:40business that at any point in time I
9:42could press one button and then all
9:43requests could be sent to a different
9:45model provider, presumably with the same
9:47request structure, reasoning tokens,
9:49efforts, and so on and so forth. The
9:51idea is you want to be able to choose
9:53the best and highest performing model
9:55for your needs. Second, I would start
9:57all of my requests with the most
9:59intelligent models possible. What I mean
10:01by that is just look at whatever new
10:03benchmark or new model drop just
10:05occurred and try routing a request to
10:07that smartest model. Then I would
10:09evaluate and see, can that smartest
10:10model do my task? If the answer to that
10:12question is yes, then what I do is I'd
10:13roll it back one generation. And then
10:15see, okay, now we're back one
10:17generation, things are probably cheaper
10:18and probably a little less performant.
10:20Can the model still do my task? If the
10:22answer to that question is yes, then
10:23roll it back another generation.
10:25Continue doing this process until you
10:27get to a point where the model can no
10:29longer do your task at the level of
10:31quality you within your organization
10:33believe it needs to. And then if the
10:34level of model that is doing that,
10:36presumably a Western model, is at or
10:38below the level of the best Chinese
10:41model on the market that is also open
10:42source, then make it open source. It
10:44really doesn't have to be more
10:45complicated than that. Go download, you
10:48know, the open weights of Kimmy.
10:49Obviously verify you have the
10:50infrastructure or the means to acquire
10:52the infrastructure in the first place,
10:53but go go get it, actually get it up to
10:56speed on some infra, optimize it if you
10:58need for speed or performance, and then
11:00even fine-tune it if you want for better
11:02performance on a variety of tasks. Not
11:04only are you not going to be beholden to
11:06a major large language model company in
11:07the West, you won't be beholden to
11:09anybody. Essentially, you will have your
11:11own brain operating in your own server
11:13acts. And that is worth more than
11:14virtually any intelligence upgrade
11:16because anybody could pull the plug on
11:18you at any point in time. Better to rely
11:21on your own hardware or software than
11:23somebody else's. Hopefully, you enjoyed
11:25this video and market opportunities are
11:26increasing in both frequency and
11:28magnitude. They're also occurring a lot
11:30faster and for shorter periods of time.
11:32If you want to learn how to take
11:33advantage of these market opportunities
11:34and build a business doing so,
11:36definitely check out Maker School down
11:38below. It's my 90-day accountability
11:39program where I'll guarantee you your
11:41first paying customer for an AI
11:42automation service in that time period
11:44or I'll give you your money back. Aside
11:45from that, please like, subscribe, and
11:47leave a comment down below with what you
11:48want me to talk about next. See you
11:50later.

About this transcript

This page contains the full transcript of Why Big Corporations are Secretly Shifting to Chinese AI Models by Nick Saraev, generated from the public captions YouTube serves with the video. The transcript has 2,463 words across 368 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.