In short: The benchmark for China's catch-up: 18 months after DeepSeek, five or six Chinese labs match US models while the chip gap has widened (no Blackwell, let alone Rubin) — "the cat is out of the bag." Chinese AI has 25% US share, 50% in six months.
DeepSeek is the Chinese AI lab that shocked markets about 18 months ago with a cheap, capable model. Woo uses it to measure how fast China has caught up. US export bans keep China from buying Nvidia's newest chips, so China's hardware gap is actually bigger now. Even so, half of the six top-ranked AI models are Chinese, and five or six Chinese companies are racing. His conclusion is that the US can't afford to slow down, and that is why Washington may simply ban Chinese AI.
33:48They haven't even gotten the Blackwell chips. They're still using the H400 chips. So from that point of view, if you think about where we are now, the state of Chinese models versus let's say 18 months ago when Deep Seek was first released. When Deep Seek was first released, they were probably using the H400 chips and then OpenAI was using the A400 chips.
In short: Paulo — evidence, not a stance: a new DeepSeek release in the last 12 hours that nobody noticed because of the bond sell-off, "literally right behind Astra. And 1.4% of the cost. And it's ahead of Fable" — "these open models are almost there and they cost a fraction." A way for the AI-IPO trade to "get snaked" before the bankers get the deal away.
In short: Evidence for model convergence: "there is very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek… in practical composite terms." He runs blind Pepsi-Coke tests behind a harness and "inevitably no one can tell the difference."
DeepSeek is a Chinese AI lab whose models are cited here purely as a yardstick.
Kedrosky's claim is that the gap between the best model in the world and a much cheaper one has essentially closed: on composite measures, an Anthropic frontier model, Qwen and DeepSeek land in the same place. He runs blind side-by-side tests — a Pepsi-Coke challenge behind a harness — and "inevitably no one can tell the difference."
If that's right, the commercial value of being the frontier lab is much smaller than the spending implies, because customers can swap to a cheaper supplier without noticing.
36:33And that variance has collapsed. There is very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek or somewhere else, in practical composite terms. People delude themselves and say there's a huge difference. But I do some testing every once in a while just to piss people off where I'll literally do like a Pepsi-Coke test and say, "Okay, you think you can tell the difference? I'll put them in front of you and see if you can.
In short: The most surprising line item in CATL's 150+ strategic stakes, raised by Stig: "one of them is DeepSeek… it's China's AI darling… this is the lab that shook everyone in Silicon Valley in early 2025. They're building this frontier model on the cheap. And in June, they took outside money for the first time — roughly $7 billion at a valuation above $50 billion USD. CATL itself put in $700 million. I should say for the record that these numbers come from press reports" — private deals with no filed terms. Stig's read: "the rationale is straightforward. AI centers are power hungry. CATL are investing in a future customer," and it is not a uniquely Chinese habit — "you have Nvidia committed more than $40 billion to equity stakes in 2026," though "Chinese companies hold roughly 10% of their market cap [in strategic stakes], whereas Americans, it's only 4%." Manish declines to value it: "the CATL investment portfolio is more strategic than financial… their investment in DeepSeek I believe is part of their AI data center integration play, although investing in a foundational AI model lab is going to the extreme end of vertical integration."
DeepSeek is the Chinese artificial-intelligence lab that startled Silicon Valley in early 2025 by building a frontier-quality model far more cheaply than anyone expected. It is private, so there is nothing to buy — it comes up because a battery company owns a piece of it. CATL holds stakes in more than 150 companies, and when DeepSeek took outside money for the first time in June — about $7 billion at a valuation above $50 billion, according to press reports rather than filings — CATL put in $700 million.
The logic is less strange than it first sounds: AI labs consume enormous amounts of electricity, and electricity at that scale needs storage. CATL is buying a future customer. Manish's caution is that this is vertical integration taken to its limit, and that none of these 150-plus stakes should be valued as an investment portfolio. He simply uses the roughly $11 billion carried on the balance sheet as long-term equity investments — about one year of operating profit — which, if credited at book value, would lower the valuation of the core business from about 18 times operating profit to about 17.
1:04:55It was this is the lab that shook everyone in Silicon Valley here in early 2025. They're building this frontier model on the cheap. And in June, they took outside money for the first time. So it was roughly 7 billion at a valuation above 50 billion USD. CHL itself put in 700 million. I should say for the record that these numbers come from press reports.
In short: The named face of Eisman's "cheaper open-source Chinese models" — the substitution he says enterprises are already making "in order to control costs." Host adds the price gap: DeepSeek V4 Flash at 14¢/28¢ per million input/output tokens, ~3¢ per test.
3:02Then, in this example, it's the LLM providers that are the true gold miners. They could strike gold, but the business is much more risky. And this is Steve's first potential domino, the fact that the LLM providers don't really have moats. In fact, right now, they're in fierce competition with one another, and they're now in competition with drastically cheaper Chinese providers like DeepSeek.
In short: Agreed with the host's point that cheap Chinese open models are real competition — Microsoft wanting government permission to use DeepSeek because it's "so much more affordable than Open AI and Anthropic" as token costs soar: "No, that's true."
44:10models. Sure. We even hear companies like Microsoft is basically ask the government's permission to say, "Look, we really want to start using DeepSk and some of these Chinese options because they're so much more affordable than Open AI and Anthropic, right? Token costs have gone through the roof.
In short: The Chinese open-source model whose feared impact is now real: gets 90% of tasks done at ~1.5% of the cost — the pricing wedge forcing US frontier labs to cut prices; Microsoft is even considering using it.
DeepSeek is a Chinese open-source AI model. A year ago people feared it; now the fear is reality: it reportedly does 90% of what you need at about 1.5% of the cost of the expensive US "frontier" models. Because it's so much cheaper, companies are switching to it in droves — which forces OpenAI and Anthropic to slash their own prices. Even Microsoft is reportedly considering using it. It's the wedge collapsing the pricing assumptions the whole AI buildout was justified on.
14:08So instead of using the frontier models, the OpenAI and Anthropics which are more expensive, although even then they're being forced to slash their pricing, they're using these Chinese models. And there's a whole bunch of them from DeepSeek, which if you remember a little over a year ago, or March of 2025, — there was a great concern about DeepSeek and what the impact would be.
In short: Strongest-positioned in foundation models with Zhipu. V4 official launch set for mid-July with new peak/off-peak API pricing (peak at 2x); the DSpark speculative-decoding framework made V4 serving 57–85% faster with no quality loss; 1.6T-parameter V4 Pro runs at a fraction of US SOTA size and cost.
DeepSeek remains the reference name for cheap frontier AI. Its V4 Pro model is 1.6 trillion parameters — a tenth or less the size of leading US models — and its engineering keeps compounding: a June framework called DSpark made its models respond 57–85% faster without changing their output at all. Goldman scores it, together with Zhipu, as the strongest-positioned foundation-model player in China.
Two telling details in the note: DeepSeek is about to launch the official V4 in mid-July, and it's introducing peak-hour pricing — charging double during Chinese business hours — because demand for its compute now exceeds supply. A price war at the bottom of the market and surge pricing at the top is exactly what "critical point of adoption" looks like. As a private company there's nothing to buy directly; it matters here as the pace-setter that keeps compressing what the world is willing to pay for AI.
Full passage: premium transcript (PDF).
In short: (Private, China.) The cheap-pricing threat the desk keeps returning to: it can do it "so much cheaper," can't be stopped from US use, and a Chinese company/government "not concerned with profitability — they just want the lead." Weiss: "pricing always calls the end of something." The latest copilot pivot destination.
In short: The disruptor: Microsoft adopting a heavily-modified, Western-safety-overridden, Azure-self-hosted DeepSeek v4 (~1/3 the input / ~1/7 the output cost of Claude) — stripping OpenAI/Anthropic of pricing leverage. A geopolitical paradox: a Chinese-born model inside one of America's biggest tech institutions because frontier-model economics are unsustainable.
Full passage: premium transcript (PDF).
In short: Last Feb's "DeepSeek moment"; another such shock — China actually passing the US in AGI — could hit the Mag 7 hard.
DeepSeek is the Chinese AI model that shocked markets when it appeared cheaply capable. He warns of "another DeepSeek moment" — if China is shown to have actually pulled ahead of the U.S. in advanced AI, he thinks the big tech leaders ("Mag 7") could drop as much as 50%.
42:57We saw with Deep Seek uh last February there was a big kind of moment where everybody said jeez these guys are making some they're stealing the code and now this the theft um and we're talking like 20,000 to 25,000 fake accounts. Um and where these these these these three platforms in China are stealing data from Claude stealing code uh putting into this Kimmy regime and then also um and and then and throwing it out into open source and where really the piranhas of open source can just rip it apart and so that's just like stealing
Nothing matches this filter.
Verbatim excerpts from the public transcripts (auto-pulled at each mention's timestamp, lightly cleaned). Timestamps link into the video; "source page" opens that commentator's full analysis at this row. Click a mention's header line to expand it (one open at a time). For personal study — not investment advice.