Why Nvidia Wants to Sell Chips to China, Answering Intel Objections, KPop Demon Hunters Conquers the World

Summary

• Nvidia’s $46.7 billion July-quarter sales and $54 billion Q3 guide do not establish an AI-demand plateau while GPU supply remains the binding constraint. Ben Thompson calls whisper-number parsing “a bunch of nonsense”: Nvidia also allocates scarce chips strategically, favoring CUDA-centric customers such as CoreWeave while not giving Amazon nearly as many GPUs as it wants as Amazon builds Trainium. Until supply can satisfy demand, the demand curve is unknowable, while percentage growth must decelerate off a much larger base.
• Nvidia’s push to sell H20s—and potentially modified Blackwells—into China is ultimately about preventing a Chinese CUDA replacement from becoming global. H20 bandwidth limits huge training systems, but its memory makes it useful for inference and superior to available domestic options. If Chinese developers must switch, “life will find a way”: open alternatives could “seep out” to Western buyers, eroding Nvidia’s pricing power and creating a long-run opening for AMD.
• Thompson sees the April H20 sales ban and licensing regime as a self-inflicted US mistake that raised the issue’s prominence and gave Beijing’s self-sufficiency advocates ammunition. Howard Lutnick later said America wanted Chinese developers “addicted to the American technology stack” — strategically correct, Thompson argues, but reckless to announce on television. His preferred regime is permissive on Nvidia chips and TSMC fabrication but absolute on semiconductor equipment, with the EUV restriction potentially outweighing the broader controls’ mistakes.
• The US government’s 10% Intel stake is a costly attempt to preserve irreplaceable foundry knowledge, not merely factories or equipment. The deal covers up to 433.3 million shares and repurposes much of $11.1 billion in federal awards; dilution and governance risks are real. But if Taiwanese and South Korean expertise were lost in a conflict, Intel would remain “by far the best option” for rebuilding advanced US production.
• Intel needed government ownership because it could not credibly promise that its foundry would still exist when an external customer’s designs were ready. Lip-Bu Tan said Intel would not pursue 14A without an external customer, yet no responsible customer would design for a foundry that might disappear within 24–36 months. Thompson grants that the stake creates an “unfortunately far” slippery slope, but argues the relevant horizon is 2035–2037, when America cannot suddenly recreate decades of tacit manufacturing knowledge.
• Rare earths expose the same collision between economically rational globalization and national security. Domestic mining and refining are expensive, environmentally fraught, and vulnerable to China flooding the market, so “there is no market solution” to independence. Yet Thompson’s default remains that war is unlikely precisely because America and China are “so freaking enmeshed” — a deterrent that governments cannot responsibly treat as guaranteed.
• KPop Demon Hunters shows how dominant distribution can make original content less risky while turning capped studio upside into platform-scale franchise value. The film reached 236 million Netflix views on an unusually linear, word-of-mouth curve; Netflix funded the reported $100 million production while Sony earned a guaranteed return Thompson described as $20 million in this case. Sony rationally limited both downside and upside, while Netflix gained potential sequels, merchandise, games, and a theatrical event built after audiences already knew the songs.

Deep dive

1. Nvidia’s guide cannot reveal demand while supply sets the ceiling

• Andrew Sharp framed the earnings anxiety with $46.7 billion in July-quarter sales and a $54 billion Q3 forecast, slightly above consensus but below the market’s whisper number. Thompson’s response: “It’s just a bunch of nonsense,” especially for anyone day-trading Nvidia on tiny deviations from expectations.

• Thompson compared Nvidia’s apparent predictability with Apple’s early iPhone years. Carrier purchase guarantees let Apple repeatedly nail forecasts through the iPhone 6 and its China Mobile expansion; once distribution stopped governing sales and consumer demand became decisive, Apple suddenly discovered that it “actually had no idea” how to forecast the market.

• Nvidia faces the inverse problem: hyperscalers are expanding CapEx while saying they will buy every GPU available. Production capacity, advanced packaging, and Nvidia’s own allocation decisions govern shipments; it favors Nvidia-centric operators such as CoreWeave while not giving Amazon nearly as many GPUs as it wants, as Amazon develops Trainium as an alternative.

• The arithmetic alone forces percentage growth lower. Adding 100 units to a 100-unit business doubles capacity; adding 100 to a 1,000-unit business adds only 10%, even if the absolute gain matches or exceeds prior quarters. Thompson’s bottom line: “We won’t truly know what AI demand is until there’s sufficient supply to meet it.”

2. Chinese customers still want H20 despite its deliberate limitations

• Nvidia’s Q3 forecast assumed no H20 shipments into China, leaving potential licensed sales as upside rather than committed revenue. CFO Colette Kress said geopolitical questions, purchase decisions, and licensing remained unresolved, while Nvidia could build more H20s if approvals and interest arrived.

• Thompson rejected the premise that Chinese companies do not want the chip. H20’s chip-to-chip bandwidth constrains large training systems, but it remains capable, carries substantial onboard memory, and is particularly useful for inference — “better than anything that these companies can get in China.”

• H20 was designed to fit Biden administration export specifications. Sharp argued that once Washington chose technocratic performance thresholds instead of a blanket ban, Nvidia would have been “irresponsible not to build a chip” around them; Thompson agreed that officials unhappy with the outcome should have written better rules.

3. CUDA makes China a future-of-the-company problem for Nvidia

• Thompson’s framing: Nvidia’s moat includes networking and leading performance, especially as US power constraints make tokens per watt increasingly decisive. But its deepest lock-in is CUDA, the “clunky” yet groundbreaking programming layer that created a common way to use GPUs as supercomputers.

• Developers know CUDA, employers hire for it, and every major cloud offers Nvidia hardware. That low-level software familiarity resembles the historical staying power of x86: customers may dislike the stack, but “no one wants to do that work all over again.”

• Excluding Nvidia from a market as large as China forces developers to create alternatives. Huawei’s proprietary approach has helped Nvidia, in Thompson’s view; a fully open-source CUDA replacement with genuine Chinese adoption would be far more dangerous because its software and developer support could spread globally.

• That makes the restrictions potentially favorable to AMD over the long run. AMD loses Chinese sales too, but a hardware-neutral software layer would weaken Nvidia’s ecosystem moat and let Western customers substitute AMD silicon. “It’s not just a China problem,” Thompson said. “It’s a future of the company problem.”

4. Washington advertised its strategy and strengthened Beijing’s hardliners

• Thompson believes the Trump administration reached the right destination — allowing Nvidia sales — only after imposing the April H20 sales ban and licensing regime. Officials’ insistence that they had not enacted a ban was, to him, empty pedantry: “Sorry, you screwed up. You don’t get to undo it.”

• The interruption raised the prominence of the dependency issue. When the controls were relaxed in July, Howard Lutnick publicly explained that America wanted to sell enough chips for Chinese developers to become “addicted to the American technology stack.” Thompson called the logic correct but added, “You probably shouldn’t talk about it on TV.”

• Sharp’s pushback — worth keeping: Chinese policy cannot be reduced to reactions against Washington, and Lutnick’s perceived insult alone probably did not determine Beijing’s position. Still, both agreed his remarks gave Chinese officials advocating technological self-sufficiency “real ammunition.”

• Beijing now faces an authentic trade-off: take superior US chips and capture AI benefits immediately, or accept potentially years or decades of disadvantage while building a fully indigenous stack. A modified Blackwell approval may be the tell; Sharp also floated the possibility that China is squeezing Jensen Huang so he will lobby Washington for broader access.

5. The defensible control point is semiconductor equipment, not chips

• Thompson’s preferred package is deliberately stark: let China buy even Nvidia’s best chips, let Chinese firms fabricate at TSMC, and impose a total ban on semiconductor-manufacturing equipment. His objective is to preserve Chinese dependence on Taiwan while keeping CUDA’s network effects global.

• He called the denial of EUV machines “the one single most effective chip control.” Because EUV equipment is enormous and difficult to move covertly, Sharp noted, it is inherently easier to restrict than individual chips circulating through global trade; Thompson conceded that this one measure might justify the broader policy package despite his objections to most other controls.

• Nvidia’s proposed 15% payment to the US government remained legally and operationally murky. Kress said no money could leave without a genuine regulatory document under company controls and SOX procedures; Thompson noted the Constitution’s prohibition on export tariffs, while acknowledging the administration might argue that foreign-made chips never enter US territory.

6. Intel’s strategic asset is knowledge accumulated inside the foundry

• The government’s 10% Intel deal covers up to 433.3 million shares and dilutes existing holders. Its funding largely repurposes CHIPS Act money: Intel had received $2.2 billion, was due another $5.7 billion, and had a separate $3.2 billion award, for $11.1 billion altogether.

• Thompson had repeatedly predicted that Intel was “barreling toward nationalization.” The surprise in public debate reflected, in his view, a misunderstanding of what America is preserving: not x86, buildings, or equipment, but the foundry’s role integrating equipment, materials, process design, and partners.

• The ASML discussion illustrated the stack’s depth. Thompson first misstated and then corrected the laser supplier’s name to TRUMPF; that San Diego technology contributes US leverage over ASML, but no single supplier is “the most important company.” Foundries coordinate the pieces and develop tacit pattern recognition that “cannot be replaced” overnight.

• Nationalizing plants after a conflict would therefore solve little. Thompson invoked countries seizing oil infrastructure only to see output collapse when expertise fled: machinery might operate briefly, but without the people who know how to adapt and improve it, the capability decays.

7. Overseas fabs cannot replace expertise located near China

• Sharp’s listener suggested relying on TSMC and Samsung’s US plants, then nationalizing them if Taiwan or South Korea were attacked. Thompson answered that many essential engineers, R&D centers, and sources of knowledge remain “within, like, a five-minute missile flight of China.”

• A wartime evacuation cannot be assumed. If the people carrying process knowledge were killed or trapped, a nationalized Arizona fab would be less useful than Intel’s imperfect domestic organization; despite Intel’s failures, it would be “by far the best option to fill in the gap.”

• Thompson credited both Biden and Trump for pressuring a reluctant TSMC and Taiwan government to expand in Arizona. Those facilities and their growing economies of scale are valuable, but TSMC remains Taiwanese and Samsung South Korean; allies are not ultimately equivalent to a durable US-controlled knowledge base.

• TSMC’s dominance is not evidence of market failure. As leading-edge fabrication grew more expensive, AMD spun out GlobalFoundries and competitors such as IBM fell away; whoever got there first and was in the best financial position survived. The problem is precisely that this “market success” conflicts with US security requirements extending 10, 20, or even 100 years.

8. Intel ownership breaks a circular commitment problem around 14A

• Thompson once preferred purchase guarantees: promise demand for Intel’s 18A output and attach incentives. But Lip-Bu Tan’s statement that Intel would not develop 14A without an external customer changed the available options by signaling that leading-edge foundry investment might end.

• Sharp spelled out the loop: no responsible company will devote resources to a foundry that might stop producing in 24–36 months, yet Intel will not continue without such a customer. Thompson did not necessarily accuse Tan of manipulation, because the economics may be truthful, but the statement effectively “guarantee[d] that Intel does not get an external customer.”

• Government ownership supplies the credible promise Intel cannot make alone: the foundry will still exist when a customer’s design arrives. Political pressure on companies to use an inferior Intel service may be objectionable, Thompson conceded, but absent voluntary customers, “that’s kind of what needs to happen.”

• His cleaner design would separate the foundry from Intel’s product business. x86 remains profitable but is shrinking, AMD is “kicking Intel’s rear end,” Arm is growing, and spending is shifting toward GPUs; Washington’s strategic interest is manufacturing, while potential foundry customers also fear that integrated Intel could appropriate their designs.

9. The Intel stake is a bad option chosen over a worse one

• A listener compared Thompson’s Intel position with his former support for blocking Facebook’s Instagram acquisition. Thompson accepted the analogy and every institutional objection: the precedent creates a slippery slope, “national interest” can stretch uncomfortably far, and he trusts neither this nor future administrations with the power.

• His antitrust thinking changed partly because TikTok demonstrated that Facebook’s dominance was not permanent. Government could not reliably distinguish Instagram from benign acquisitions, while the competitive stakes proved low enough for markets to respond; the danger of clumsy intervention exceeded the likely benefit.

• Advanced fabrication differs because the loss may be irreversible on the relevant timeline. “The problem isn’t in 2025 or 2026 or 2027,” Thompson argued; it is 2035, 2036, or 2037, when America cannot decide to reconstruct a foundry and instantly recover decades of accumulated expertise.

• Thompson’s honest position is not that nationalization becomes good, but that policymakers face “an array of horrific options.” He would have intervened differently and earlier, yet the present choice is between Trump’s flawed stake and allowing Intel Foundry’s leading edge to wither; “this is the better choice… even though it stinks up and down.”

10. Rare earths reveal why strategic independence needs state support

• Sharp and Thompson treated rare earths as the closest analogue. Domestic mining and especially refining have weak economics: projects are capital-intensive, environmentally damaging, exposed to lawsuits, and vulnerable to China flooding the market before private investors earn returns.

• China consequently possesses real leverage, explaining why people complain that Trump is going easier on China than on the EU. Rebuilding capacity cannot emerge from ordinary market incentives; Thompson’s categorical conclusion was, “There is no market solution. It has to be government action.”

• The same logic complicates complaints about manufacturing outsourced to China. Those decisions were economically rational at the time, just as abandoning Intel Foundry is rational now; reversing globalization requires interventions that free-market advocates will find intrinsically uncomfortable.

• Yet dependency runs both ways. America resents reliance on Chinese rare earths, while China resents reliance on US technology and Taiwanese fabrication. Thompson’s default is that war probably does not occur because the countries are “so freaking enmeshed”; Sharp’s answer was that avoiding war remains too important to plan solely around that assumption.

11. Sony rationally capped both its KPop downside and upside

• KPop Demon Hunters had reached 236 million Netflix views and led the US box office during a limited theatrical release. Sharp initially read an inaccurate ChatGPT synopsis calling its heroes a fallen group; Thompson immediately objected that they were already “the top of the game,” prompting Sharp to admit the source and call it “horrible.”

• Thompson described the Sony arrangement as a guaranteed 25% margin or $20 million, whichever was higher, then said Netflix bore the film’s $100 million production cost and Sony made $20 million. Sony earned a return without building a streaming service, carrying subscriber-acquisition costs, or risking billions in churn-driven losses.

• The opportunity cost is enormous in hindsight. Netflix gained the core franchise, potential sequels, likely merchandising leverage, and an obvious gaming premise — “K-pop demon hunters… fight demons. Sounds like a good game” — while Sony cannot capture the full value of a phenomenon potentially worth billions.

• Thompson nevertheless defended the original decision. Netflix’s COVID-era purchases helped Sony avoid layoffs and preserve an animation studio he considers the industry’s best; Sony systematically chose to make content rather than imitate Disney or Netflix. “Whenever you limit downside risk, you limit upside risk.”

12. Abundant distribution turned movies from defaults into destinations

• Thompson used Warren Buffett’s distinction between franchises and businesses. Local newspapers were once unavoidable delivery systems for news, sports, comics, classifieds, and advertisers; after the internet let everyone reach everyone, success depended on customer acquisition and management, with The New York Times becoming a destination rather than a default.

• Movie theaters once had the same default status. In the 1980s and 1990s, audiences went on Friday night and selected whatever was showing; once a small group of Hollywood gatekeepers approved an original film and placed it in theaters, being made largely solved distribution.

• YouTube, TikTok, podcasts, Sharp Tech, and countless other choices removed that guarantee. A film must now persuade someone to leave home, making sequels and familiar franchises valuable as “embedded free marketing”; Thompson said that if a studio spends $200 billion, it may need to spend $200 billion on marketing to make money.

• IMAX survives by intensifying the destination: a giant screen, enveloping sound, and something unavailable at home. Sharp called it closer to a concert than an ordinary movie, and Thompson agreed that this exclusivity explains why the format remains economically coherent.

13. Netflix’s low-friction funnel created organic demand and a theatrical event

• KPop Demon Hunters did not launch like a star-driven Netflix film, with an immediate spike followed by a plateau. Its view curve rose gradually in an almost straight line and had not flattened — a pattern Thompson recognized as “pure word of mouth,” with each social circle handing the film to another.

• Netflix made experimentation cheap because subscribers had already paid. The downside of trying the movie was merely time on a night when someone might watch something anyway, not tickets, travel, parking, popcorn, and the commitment required by a theatrical destination.

• Thompson’s through-line: studios believed owning content would let them defeat Netflix, but theatrical economics pushed them toward recycled properties while Netflix’s distribution allowed originality to surface. Its catalog contains plenty of failures because it “throw[s] a bunch of crap against the wall,” but “the distribution gives the space for greatness” and preserves the upside when something sticks.

• The later theatrical success did not disprove the streaming thesis; viewers went because they already knew the film and songs. Like a concert, the screening converted familiarity into communal participation, making the movie what Sharp and Thompson called a “killer example” of Netflix as a surviving monoculture.

14. Thompson cut into his desk to make vertical monitors ergonomic

• Thompson prefers stacked displays because four visible quadrants support writing, browsing, references, and ChatGPT without constant head-turning. Side monitors remain useful for status information such as messages or recording levels, but he dislikes performing primary work outside his central field of view.

• Moving from 21-inch LG displays to 24-inch screens forced the lower monitor into a steep angle, so he cut a rectangle into his standing desk and sank it below the surface. His eyes sit high enough to see the full panel and bottom dock; the eventual plan is two stacked Apple Studio Displays, potentially after cutting the opening deeper around the desk supports.