(Preview) Xi Jinping and China's Tech Companies, The Long-Run Implications of the Chip Ban, and a Pessimistic Outlook for Taiwan
- DeepSeek changed China’s AI psychology by showing that a chip-constrained company could build an internationally competitive open-source model: “Yeah, we can do this.” Bill Bishop called it a turning point that briefly “crashed the US stock markets” and NVIDIA while lifting Hong Kong-listed AI stocks; Ben Thompson noted that markets subsequently recovered almost completely.
- Open sourcing gave China a rapid global distribution channel while disrupting domestic model pricing. Anyone can run DeepSeek instead of paying Anthropic for Claude or OpenAI, while Baidu, Tencent, and other Chinese providers were pushed toward free offerings—leaving the business model “not at all clear.” Ben cautioned that DeepSeek may have made the open-source decision itself rather than following a central-government directive.
- The supposed $6 million miracle was a misreading of DeepSeek’s V3 paper, not a claim that the full R&D effort cost $6 million. Ben stressed that the figure covered one training run and explicitly excluded experimentation and R&D. DeepSeek’s breakthroughs remain genuine, but the mythology outran the disclosed economics.
- R1’s viral breakout combined broad access, visible reasoning, China anxiety, and fear that billions in Western AI capex might be wasted. For many users it was their first reasoning model, and watching “my little AI friend” work through an answer made it feel better than paywalled alternatives; Bill questioned whether the surge was fully organic, then retained the view that it was “some mix” after hearing Ben’s “perfect storm” case.
- DeepSeek prompted a competitive response from U.S. AI companies Bill called “fat and happy.” Ben cited falling AI pricing, more aggressive releases, a GPT-4o update with less of an “HR voice,” and Google models that were even cheaper and arguably comparable or better.
- DeepSeek and xAI pursued very different routes to strong models. DeepSeek redesigned mixture-of-experts training around bandwidth limits and, Ben believed, H800 rather than H100 hardware; xAI raised “$16 billion or $12 billion,” bought NVIDIA hardware, used NVIDIA’s networking capabilities, and produced the apparently state-of-the-art Grok 3 only 19 months after founding. Ben noted that o3 may be better but is a distinct reasoning model unlikely to be released directly.
- The unresolved investor question is whether ingenuity can keep compensating for restricted hardware over the next one or two years. Bill asked whether a real separation emerges if DeepSeek effectively can buy only Huawei Ascend chips while xAI and OpenAI keep buying better NVIDIA systems; the preview ends just as Ben begins laying out his concerns.
1. DeepSeek turned technological scarcity into national confidence
Bill’s central claim was psychological: despite “this chip blockade,” DeepSeek found “very creative ways” to build an internationally competitive model. Making it open source transformed that achievement into “Yeah, we can do this.”
The market response dramatized the shift: Bill said DeepSeek “crashed the US stock markets, crashed NVIDIA” and sparked a Hong Kong AI-stock “melt up,” although Ben emphasized that prices later returned almost to their prior levels.
DeepSeek had originally been a hedge fund and quant fund that bought substantial NVIDIA hardware. Its CEO, Li Wenfeng, met Xi Jinping on Monday; Ben joked, “This is my quant.”
2. Open source globalized DeepSeek and disrupted domestic pricing
Bill’s distribution thesis was that anyone worldwide could download a strong Chinese model instead of paying Anthropic for Claude or OpenAI. Baidu and, Bill thought, Tencent integrated it, making open source “an incredibly powerful thing for China.”
Asked why DeepSeek open-sourced the model, Bill called it “a bait.” Ben cautioned that people overestimate how much the central government knew or cared about the decision and suggested DeepSeek may have made its own choices all along.
DeepSeek also disrupted China’s AI market: companies that had charged for models now had to go free. Bill saw no clear business model; Ben replied that the U.S. has the same problem, despite OpenAI and Anthropic generating subscription revenue that still does not cover costs.
Ben’s consumer thesis was sharper: ChatGPT had achieved tech’s hardest asset, a consumer brand with meaningful market share. Advertising was the “inevitable end state,” and OpenAI needed to reach it quickly so free users could receive its best possible models.
3. The viral moment was a perfect storm, not a $6 million trick
Bill wondered whether DeepSeek’s sudden rise on X and in the App Store was partly inorganic. Ben judged it “pretty authentic”; after hearing the case, Bill said he thought it was “some mix” and deferred to Ben’s more compelling case.
V3 arrived over Christmas after years of published papers. Its roughly $6 million figure covered only the specified training run—not experimentation or R&D—and the paper disclosed those exclusions plainly. Ben said people were twisting the figure to serve their own agendas, even though the paper was clear about what it did not include.
R1 then gave many users their first access to a reasoning model and exposed its thinking. Comparable reasoning models were paywalled, while OpenAI withheld its process for competitive reasons. Add China anxiety, billions in AI investment, and a stock market resting on that spending, and DeepSeek became “a current thing for a weekend.”
4. DeepSeek and xAI optimized for opposite constraints
Ben believed DeepSeek’s technical story because its architecture matched its claimed bottleneck. It reworked mixture-of-experts training to scale around bandwidth limits and, Ben believed, used H800s rather than H100s; DeepSeek’s CEO and employees had also identified chip access as the largest constraint.
Grok 3 offered the inverse case: xAI, founded 19 months earlier, produced an apparently state-of-the-art model by raising “$16 billion or $12 billion,” buying NVIDIA chips, and using NVIDIA’s networking capabilities to wire them together at scale. Ben qualified the comparison by saying o3 might be better, but that o3 was a distinct thinking model unlikely to be released directly; its Deep Research version was impressive but had clear flaws.
Ben’s comparative-advantage analogy was that startups must capture “lightning in a bottle,” while cash-rich incumbents can de-risk by buying the people behind a breakthrough. Facebook abandoned its Poke response and put Stories into Instagram. DeepSeek optimized scarcity; xAI purchased abundance. Both approaches fit their circumstances.
5. Competition improved quickly, but the chip-access verdict remains open
Ben saw a competitive response: AI pricing had already come down, OpenAI had become more aggressive in releasing products, and the GPT-4o update appeared less scolding, with less of an “HR voice.” Google offered models that were cheaper and arguably as good or better.
Ben stressed that DeepSeek’s engineering included genuine breakthroughs that would be, or already had been, adopted globally. The episode’s reaction reflected a mix of AI mythology and a reality that was itself “fairly spectacular” but not fully appreciated.
The long-run question remained unanswered: if DeepSeek continues to lack the best NVIDIA chips and can effectively buy only Huawei Ascend chips, while xAI and OpenAI keep buying better NVIDIA hardware, Bill asked whether “a real separation” emerges within one or two years. The preview ends as Ben says, “So let’s buckle up and get into it.”