Ben Thompson

Thompson reads Apple’s mid-cycle price increases as a bearish signal because it missed the memory shock twice: first in procurement, then in March pricing. Jensen Huang was visibly securing supply while “we didn’t see Tim Cook drinking soju nine months ago”; Apple then launched new Macs without increases, only to raise prices by hundreds of dollars out of band. That breach of Apple’s stable-pricing promise suggests management “just doesn’t know what’s going on” with AI.

Apple may still deliver a good Siri in 2026, but Thompson argues it will embody what was possible in 2024. AI was already good enough then to support “tons of products,” so Apple can fulfill its old promise while remaining “nowhere close” to what AI is capable of in 2026 and beyond. Sharp’s framing is also harsh: Apple has become “a bunch of normies out there in Cupertino.”

The memory shortage could turn China into a dominant player by letting it climb the learning curve profitably. Thompson says Chinese producers cannot yet make current HBM, though they can make HBM1 or HBM2 while the frontier is HBM4. DRAM is fungible, and as the big three suppliers focus on HBM, scarcity lets China sell older DRAM at profitable prices, make a lot of it, and improve.

Export controls and attempts to capture scarce supply may accelerate the competitor they are meant to constrain. If Nvidia is not allowed to take Chinese memory, Huawei and others will take it. Thompson’s categorical conclusion is that “the die is cast” and China will become dominant in memory, while Sharp sees possible security-driven US investment over the next 5 or 10 years despite weak economics.

Vibe coding collapses the distance between an idiosyncratic need and purpose-built software. Thompson wanted a lookup system for household objects—not inventory or task-management software—designed around one assistant doing the cataloging and one owner who does not want to do the administrative work. “I don’t need permission to do it. I don’t need to hire a developer.”

The resulting app combines a physical-location hierarchy, AI vision, and QR codes into a narrowly optimized workflow. An assistant chooses where everything belongs; a photo gets roughly 95% of objects right; scanning a shelf’s QR code assigns a later arrival to its location. The specificity is the point: conventional software accumulates features to serve millions, while AI enables software tuned to one user.

The lighter mailbag answers expose contrasting personal operating systems rather than investment calls. Sharp consumes a Venti iced coffee, a Grande Americano, and on three of five workdays a 20-ounce yerba-mate drink, despite hitting “a point of diminishing returns.” Thompson usually drinks one morning pot and avoids caffeine past noon. His broader productivity maxim is sharper: “You don’t fix your weaknesses; you make more money so you can pay for someone else to fix them.”

The Commerce Department’s directive suspended Fable 5 and Mythos access for foreign nationals, including those inside Anthropic, while Claude was pulled from the market entirely. Thompson thinks the administration may be demanding a technically impossible guarantee from probabilistic software, but Anthropic’s own conduct and rhetoric created the mistrust now driving policy. “Who gets final say?” is the real dispute.

Blocking frontier models may buy “weeks or months, maybe a year,” but it also delays the defensive work needed before comparable capabilities proliferate. Finding a vulnerability and helping patch it can be “the exact same request,” while defenders benefit from possessing their own code. Thompson’s preferred response is “the Manhattan Project we need”: inspect and repair web software and open-source projects at scale.

Anthropic’s safety narrative has created expectations of control that the technology cannot satisfy. After years of comparing AI with nuclear weapons and initially restricting Mythos through Project Glasswing, the company opened access roughly two months later. Sharp’s diagnosis: Anthropic promises something close to perfect security, then policymakers panic when jailbreaking proves unavoidable.

The clash is as much about institutional authority as cybersecurity. Sharp cited reports that Anthropic was slow to disclose 111 organizations with advanced Mythos access plus roughly 50 additional entities, including a South Korean company; Thompson pointed to the Department of War dispute, model “nerfing,” and Anthropic’s red line that “we get to decide how this is used.” The government may be technically misguided—and the directive might not survive court review—but Anthropic cannot credibly position itself above the state.

An AI bubble correction may punish companies that merely “plaster on AI to existing workflows.” Thompson expects the durable value to come from rebuilding products and organizations around probabilistic computing, comparable to replacing ads beside articles with algorithmic feeds. That transition takes years because users must first reset their assumptions about what software can do.

AI will probably be deeply disruptive, but confident forecasts about the exact outcome deserve skepticism. Thompson rejects both “nothing is going to happen” complacency and Anthropic-style prescriptiveness: the internet reshaped society over 25 years in ways no one in 1993 could have mapped, and AI might be larger. The missing ingredient is “epistemic… humility.”

World Cup hydration breaks illustrate how a transparent money grab can also improve the product. Thompson argued that the extra commercial inventory lets exhausted players reset, citing Croatia’s comeback before England ultimately won 4–2; Sharp agreed that “fewer exhausted players is a good thing” but said the pauses should simply be called commercial breaks. The broader media call is Thompson’s enthusiasm for the “eventization of sports”—premium tournaments over routine weekly inventory.

Sora’s shutdown looks like a clean product failure but a consequential warning about AI economics. The Sora team’s farewell promised timelines for the app, API, and preserving users’ work. Sora briefly reached the top of the App Store before proving to be a “novelty,” leaving Ben to call its closure a “very bloodless killing.” OpenAI’s product instincts were defensible; the standalone social network appears not to have reached critical mass.

Copyright controls may have removed the feature that gave Sora its best chance of surviving. Ben argued that early interest may have come from manipulating recognizable figures and characters; without that shared cultural material, interest may have fallen sharply. Andrew saw a plausible recurring meme tool—not an Instagram rival—and Ben called the unrealized version “the greatest meme maker of all time.”

Cloud video generation reverses the zero-marginal-cost conditions that helped earlier social products compound. Instagram began with filters computed locally and distribution handled by the App Store, then added infrastructure as its value grew. Sora incurred heavy compute costs from day one—Andrew cited an unverified “$15 million a day” estimate—creating an “all-or-nothing scenario” before OpenAI had figured out how to make money from it.

OpenAI’s enterprise turn may be economically necessary even as it leaves the consumer-platform thesis untested. Ben once envisioned a roughly billion-user consumer base plus advertising creating a revenue-and-R&D flywheel, but OpenAI was late to start monetizing and “we don’t know if that would’ve worked.” Enterprise customers will pay directly for productivity, Codex is compelling, and servers must be funded—though competition with Anthropic, Microsoft, and others will be “tooth and nail.”

Private-company revenue disclosures obscure the OpenAI–Anthropic race more than they resolve it. Ben said reported Anthropic ARR may resemble GMV by including total customer payments, while OpenAI’s numbers reflect the amount above what goes to cloud providers. Even so, “trajectories are real”: Anthropic’s appears more vertical, while OpenAI’s absolute lead might still be larger. Andrew said both companies could go public within the next couple quarters; Ben viewed that as public-company financing working as it should.

ChatGPT Image 1.5 narrows the image-generation gap that had been giving users a reason to open Gemini, but Ben Thompson’s more important call is that OpenAI keeps winning at product. Gemini produced his single favorite image, yet ChatGPT won across the set and wrapped existing capabilities in an interface that helps ordinary users discover what AI can do: “You have to lead them down that path.” The same week’s GPT-5.2 update used the same base model with more post-training; Ben called it a good response ahead of a supposedly next-generation model in the new year.

ChatGPT’s contextual continuity turned image generation from a tool awaiting instructions into a product that stimulated more demand. Reusing identical prompts, Ben found that ChatGPT automatically preserved characters and visual style across images, creating a comic-book-like thread that “was pulling ideas out of me”; with Gemini and Nano Banana, he generally started only after already having an idea.

OpenAI’s popularity is simultaneously its moat and a compute-intensive catch-22. A ChatGPT feature may reach 900 million people, and AI retains a marginal cost, so deployment resembles Apple manufacturing Face ID at tens-of-millions scale: invention is insufficient without reproducibility. Ben revived his monetization critique—“I should’ve been even harsher on ads”—arguing that a 2023 launch would have produced a mature ad product and habituated users by now.

Disney licensing characters to Sora may not revive the app, but it reinforces the rising value of universally recognized IP. Ben called the agreement bullish for Disney “even if Sora doesn’t amount to anything”: when AI makes content highly customized and individualized, common cultural “totems” become more valuable because audiences need shared reference points.

The AI competitive map increasingly looks like vertically integrated Google against a coalition, although Google’s economics remain formidable. Disney’s Sora agreement and near-simultaneous cease-and-desist to Google place OpenAI, NVIDIA, and Disney on one side, while Google owns its chips and models and dislikes deals because “deals don’t scale.” Yet Gemini 3 Flash is “extremely cheap” and strong enough that its API economics may matter more than flagship-model rankings.

Ben’s strategic case for Netflix’s Warner Brothers pursuit is built around attention, where YouTube is the central threat, while narrower markets remain relevant antitrust questions. Netflix is working to pull at least 18 podcasts’ video episodes from YouTube through iHeartMedia and Barstool deals beginning in early 2026—another step toward becoming a stand-in for all television and entertainment.

SpaceX’s nearly $20 billion of spectrum purchases put scarce wireless capacity at the center of this preview. EchoStar agreed to sell another $2.6 billion after September’s roughly $17 billion deal, alongside a separate $23 billion sale to AT&T brokered under pressure from federal regulators. Ben Thompson rejects the idea that EchoStar was simply “forced” out: it declined to fund the towers needed to become a genuine fourth carrier.

The carrier business is ultimately a capital-spending contest with weak differentiation. Thompson’s blunt formulation is that “one cellphone tower is like another cellphone tower,” leaving profitability tied to how aggressively operators build coverage. Verizon’s network may still be the best for getting at least 4G nearly anywhere, he hedges, but maintaining that advantage means continually spending.

Integration creates leverage when it enables an experience that was previously impossible, but it becomes a tax once the broader ecosystem catches up. NTT DoCoMo’s tightly integrated Japanese mobile experience let users do things that U.S. phones at the time generally did not; AOL similarly made getting online easy. Once the open internet matured and Google made it navigable, AOL “fell down under its own weight.”

Apple’s durable integration of hardware, software, operating system, apps, and an existing iTunes user base let the iPhone reverse the carrier power structure. Verizon rejected Apple’s terms because Apple would not permit carrier branding or software control; Cingular accepted, rebranded as AT&T, and watched Verizon customers switch networks “strictly for the phone”—behavior even Motorola’s Razr never induced.

The iPhone commoditized carriers while making them finance the infrastructure that increased Apple’s value. Heavy iPhone data consumption forced AT&T and rivals into an “astronomical” network build-out, fixing early congestion without Apple “spending a dime on it.” Thompson’s conclusion: “The whole world did that” for Apple, while Apple preserved its premium-margin hardware model.

Apple converted carrier dependence into contractually guaranteed iPhone sales, making earnings unusually predictable until it exhausted new distribution partners. Operators bleeding subscribers came “begging for the iPhone” and accepted minimum sales guarantees. After NTT DoCoMo and China Mobile joined around the iPhone 6 era, Apple ran out of major carriers to add; around the iPhone 6S, results began missing forecasts both upward and downward.

Substrate is a semiconductor moonshot that founder James Proud has described as having 1% odds, while its potential payoff has become large enough to attract financing. The company claims particle-accelerator-powered X-ray lithography could halve manufacturing costs, eliminate some multi-patterning, and produce high-resolution layers comparable to those of leading foundries. Ben Thompson remains explicitly uncertain—“It’d be nice to see more evidence”—but highlights Proud’s posture of operating with “100% conviction” despite the long odds.

Trying to disrupt ASML and TSMC simultaneously sounds absurd, yet the industry’s integration makes that strategy internally coherent. Thompson argues that lithography is the central step around which the entire fabrication process is optimized, so a radically different tool probably cannot simply be dropped into an incumbent fab. Substrate can buy conventional etching, coating, and doping equipment, but probably must rebuild how those pieces work together “from first principles.”

ASML and TSMC’s co-evolution created formidable incumbent lock-in. Lithography now represents perhaps 30%-35% of fabrication cost; Thompson roughly estimates current machines near $250 million and the next generation closer to $500 million, while stressing that he is “pulling these numbers out of my rear end.” A roughly $30 billion fab that is not filled and running leaves its owner in serious trouble, making TSMC “almost the brake on the AI bubble” because it bears enormous upfront equipment risk.

The strongest defense of today’s AI bubble is not durable GPUs but coordinated invention across many improbable projects. Big Tech spending flows through NVIDIA to TSMC and ASML, while simultaneously making outsiders such as Substrate financeable: “A lot of people working on a lot of hard problems all at the same time.” Thompson expects excess and failures, yet argues that the future will depend on what “emerges from the rubble” and could not otherwise have been attempted.

A thermodynamic chip promising theoretically 10,000-times-more-efficient image generation is the episode’s best specimen of bubble-enabled experimentation. Its proposed mechanism excites a chip according to a probability distribution, then uses the system’s physical decay to compute probabilistic outputs. It sounds “totally insane,” but Thompson sees a conceptual fit between probabilistic AI and probabilistic hardware that could better align computation with probabilistic outputs than today’s binary, deterministic chips.

Semiconductor controls, ideological ambition, and AI mania are all forcing functions against industrial stasis. Thompson says a new US foundry would otherwise go bankrupt “100 times over” while descending the learning curve; chip controls instead motivate China to absorb uneconomic development costs, while without them TSMC could “rule the roost forever.” Breaking that equilibrium may require “religious fervor”: Substrate is ideological about the US winning, while AI companies act as if they are “inventing God.”

Sora 2 plausibly clears the more defensible “GPT-3.5 moment” bar: not a new ChatGPT-scale event, but the point when generated video becomes a coherent, consumable finished thought. Ben Thompson initially “poo-pooed” OpenAI’s comparison, then distinguished GPT-3.5 from ChatGPT itself: short clips now hold together well enough to astonish mainstream users, even if experts have already seen related capabilities from Google’s Veo models. “It’s sufficiently finished to be something meaningful like GPT-3.5 was.”

Andrew Sharp sees ChatGPT-like virality but not ChatGPT-like utility. Sora 2’s photorealistic output, intuitive interface, and unsolicited excitement from people outside the early-adopter tech crowd feel unusually consequential, yet Sharp wants to leave after roughly 90 seconds and cannot identify a daily-use case beyond entertainment. His pushback: ChatGPT was immediately useful, while Sora currently looks “useless”—a condition he explicitly says might change.

OpenAI’s repeatable advantage may be productization, not merely model quality. ChatGPT, the Studio Ghibli image-generation moment, and now Sora represent three viral launches; Thompson’s investor-relevant conclusion is that “three times is worth keeping track of.” Sora reached No. 3 in the App Store while Meta AI was roughly No. 125; Gemini was No. 1, albeit invite-gated.

AI video could expand the creator class enough to produce an order-of-magnitude content shock. Thompson’s 1%-create, 9%-distribute, 90%-consume model may break when generating video stops being a “big pain in the ass”: moving creators from 1% to 2% already doubles supply, while a tripling or even 10× increase could surface ideas previously blocked by production constraints. The structural change is “the breaking apart of the idea and the substantiation of the idea.”

Meta’s attention franchise faces a potentially disruptive format it cannot simply clone inside Instagram. Thompson says Meta’s power and advertising business are predicated on attention. TikTok took attention, but Meta could answer with Reels; Sora presents two harder problems if momentum persists—Meta’s AI may not be good enough, and users may not want AI video shoved into Instagram. His call is conditional but blunt: “Meta might be in big trouble here.”

The hostile reaction to Meta Vibes exposed narrative—not visual novelty—as the missing ingredient for most users. Listeners criticized Vibes as neither funny nor sexy and as lacking narrative; Sharp added that it also offered no apparent information or instruction. Thompson’s own screensaver comparison reflects his atypical taste, and he conceded that treating ambient, narrative-free imagery as broadly compelling was “a massive blind spot.”

Sharp treats today’s AI-video apps as half-baked prototypes whose destination is still unclear. He compares them with computers in 1993 or 1994: an interesting glimpse of where things might be in four or five years, even though he largely agrees with the Vibes backlash. Thompson revisits his optimism about AI-generated worlds, acknowledging that fantastical environments may attract a committed niche without becoming broadly compelling.

A recurring outside critique is that frontier resources are being directed toward attention decay. Andrew read Ashkay Kothari’s question about why people keep dedicating “our brightest minds, billions of dollars, and the most powerful GPUs on Earth” to another app optimized for attention decay, noting that versions of this criticism had recurred about both Vibes and Sora.

Ben’s central point is that being first matters more than Apple’s mythology admits. ChatGPT became the defining consumer AI product and gained compounding advantages in awareness and user preference; even if OpenAI went bankrupt tomorrow, “the bidding war to acquire ChatGPT would be astronomical.” Google invented the transformer but was caught flat-footed, and the result was disruption.

Google may have missed the first consumer-AI product while remaining strongly positioned for the broader market. Ben describes it as a combination of an irrepressible Loch Ness Monster and zombie; Andrew recasts it as a giant zombie jellyfish. Its competitive model, cheap infrastructure, and Cloud prospects could let it envelop markets despite an incoherent strategy. Ben’s self-critique is that Google’s frustrating lack of a legible top-down strategy may have biased his analysis against it.

Apple and AWS remain formidable, but their incumbent strengths may not transfer to AI’s new competitive plane. Andrew notes that Tim Cook and Andy Jassy made nearly identical cases: AI is early, real use cases will matter, their chips are best, and they have valuable data. Those claims may be true while missing the new strategic question; Ben is not predicting Apple’s or Amazon’s doom.

Apple’s moat may become less decisive if the phone becomes primarily a conduit for an AI companion. Touch interfaces, the App Store, and ecosystem control matter less if competition shifts to “how good the AI that is accompanying you is.” Ben’s blunt test for Apple’s insistence on control is: “What if you suck at it?”

Nokia is Ben’s cautionary analogy for choosing differentiation over a viable platform. Nokia could have adopted Android and probably remained around and fine, though less profitable or dominant and more under Google’s influence; instead it clung to Symbian and then chose Windows Phone because it “had such a need to be different.” Apple may likewise be better served letting Anthropic, OpenAI, or another partner provide Siri.

Jason’s Apple contrarianism is compelling if Apple partners for AI rather than building everything itself. He compares Apple’s cited $4 billion in next-year AI spending with Sony sitting out the streaming wars while distributors bid for its content. Ben accepts the business logic but says it implicitly requires a partner; Andrew worries Apple is falling behind in AI as its core competencies also erode.

Intel’s 14A disclosure turns a long-running strategic problem into an explicit capital constraint: internal products cannot generate acceptable returns without a meaningful external customer. Ben Thompson traces that back to missing mobile volume, but warns that industrial foresight is not the same as a trade: “If I had shorted Intel, I would’ve been both right and bankrupt.” The cleaner long-duration winner was TSMC.

Intel’s obstacle is less process technology than a vertically integrated culture that never learned to serve outside customers. Its historic obsession with performance tolerated poor yields, inefficient manufacturing and extensive product binning; Thompson now believes “Intel’s culture is fundamentally incapable of becoming a foundry.” A viable US foundry may therefore require a split and direct government support rather than another corporate turnaround.

Backside power captures Intel’s predicament: 18A and 14A are technologically ambitious but insufficiently matched to what customers actually want. Apple’s roughly 7–8-watt phone chips do not need the feature, while Nvidia can pay TSMC’s higher-cost, bonded-chip approach for processors drawing 700–800 watts and selling for $30,000–$60,000. Intel remains “more technologically interesting and creative,” but “the problem is it’s not meeting anyone’s needs.”

Samsung’s $16.5 billion Tesla deal could establish the second US leading-edge supplier before Intel finds its anchor customer. Tesla’s AI6 chip is returning to Samsung at two nanometers and anchoring its Texas fab; Apple and AMD appear locked into TSMC, Nvidia historically distrusts Intel, and Qualcomm has kicked the tires most seriously but has said no. If Samsung becomes viable, Thompson asks whether there is room for three suppliers at all.

US semiconductor resilience looks materially better even as Intel deteriorates. TSMC now refers to Arizona as a “GigaFab,” with its own R&D and a figure Thompson recalled as roughly 30–40% of its advanced chips eventually being made in the US; Samsung’s Texas investment could acquire similar gravity. Thompson’s defining line: “TSMC is becoming Intel, and they’re becoming Intel much faster than Intel was able to become TSMC.”

AI increasingly looks sustaining for Big Tech, whose distribution, infrastructure and free cash flow can overwhelm venture-funded challengers. Google Cloud is exploiting differentiated AI and workload gravity as Google’s CapEx rises from $55 billion in 2024 to $75 billion and then $85 billion, even though success lowers the parent company’s margin profile: “They are investing heavily to get worse margins, which is exactly what they need to be doing.”

OpenAI remains the most credible cross-platform disruptor, but Meta can weaken it simply by making the race prohibitively expensive. ChatGPT has won substantial consumer mindshare, yet OpenAI cannot stop model R&D without falling behind; Zuckerberg’s reported $100 million talent offers either recruit its people or inflate its costs, “like a boxer doing body blows,” financed from Meta’s free cash flow. Google Search and Apple’s position remain genuinely unresolved fronts.

The episode’s broader business-model lesson is that durable financing often matters more than the celebrated product. Grantland vanished when ESPN stopped funding it, while The Ringer’s monetizable podcasts let Spotify and Bill Simmons preserve lower-return writing; text acquires audiences, podcasts monetize them. Thompson’s epitaph is memorable: Grantland is “the Kurt Cobain of websites,” permanently great because it disappeared before it could fade.

Cloudflare’s current pay-per-crawl structure will not revive incumbent publishers. Ben agrees with Finn that AI companies can index or cache pages instead of repeatedly paying to crawl them; he would not enable the system and sees no “meaningful revenue stream.” He says the proposal is directionally pointing at a real need for content markets, but making one work would require someone with enough power to impose it. At most, it may facilitate brand-oriented licensing deals for large publishers such as The New York Times—“the folks that kinda need the least help.”

Publishers face a structural leverage problem: AI needs content, but it does not need any particular publisher’s content. Ben compares the imbalance to Google needing the web while individual sites need Google, a dynamic also undermining copyright plaintiffs. Andrew adds that search referrals are falling, the advertising model has deteriorated for 15 years, LLMs are becoming the portal, and ambitious media cost structures already fail; expecting licensing to “make it 1997 again” is a “fantasy.”

If AI does create sustained demand for more material, the winners will probably be new suppliers built specifically for that market. Ben’s model is “ghost kitchens for content”: just as delivery-only restaurants optimize every workflow for food consumed 30 minutes later, AI-native producers would optimize for machine ingestion. The market comes first and summons its suppliers, as Apple’s App Store, Meta advertising, and Amazon’s seller ads did.

The resulting labor market would be a barbell, not a broad revival of professional media. At one end, prestige creators and institutions retain direct-paying audiences; at the other are outsourced “salt mines of content” producing fungible inputs for models. “The LLMs, they don’t even want content. They want tokens.”

The New York Times is the Taylor Swift of publishing—proof that an exceptional model works, not a template for the median outlet. Its advantage came from aligning editorial choices with subscription value, bundling early, and making work readers would “feel good about paying” for; its dominance can leave everyone else competing to be the second subscription. Ben’s harsher diagnosis is that AI is “accelerating the death” already underway and might amount to a “mercy killing.”

A Spotify-like pool could eventually price machine-consumed material, but it would price a commodity rather than preserve prose economics. Payments might be allocated by measured share of usage, avoiding uncapped per-crawl liabilities, though analytics would be difficult. Andrew notes that songs remain distinct while facts—such as who won a Wizards game—are available everywhere; Ben’s conclusion is that the market would be “mining tokens.”

Windsurf’s employees received a last-minute rescue, but the deal sequence shows how value can now be extracted from a startup without buying it. OpenAI’s proposed $3 billion acquisition struggled to close; Ben estimated the trouble reached day 61 after a roughly 60-day exclusivity window. Google then struck a deal worth roughly $2.4 billion to hire CEO Varun Mohan, co-founder Douglas Chen, and select staff. Cognition subsequently agreed to acquire the “remain co” and accelerate employee vesting, creating a narrow “happy ending” after Windsurf had been “completely lobotomized.”

Ben’s core call is that regulators have killed Silicon Valley’s acquisition compact by making selective hiring cheaper and cleaner for Big Tech. Microsoft–Inflection, Google–Character.AI, Meta–Scale AI, and now Google–Windsurf demonstrate the workaround: take the differentiated people, avoid redundant staff and layoffs, close immediately, and skip merger review. “Everything about this is better for big tech,” Ben argued, concluding, “I think regulators killed the Silicon Valley model, and I think it’s dead.”

The endangered asset is career insurance for rank-and-file startup employees, not founders’ willingness to gamble. Silicon Valley paired plentiful $50,000 seed checks and $100,000–$200,000 cloud credits with an implicit downside floor: even a failed startup with useful technology might land its team four-year jobs at Google or Facebook. Remove that floor, especially in a tight hiring market, and designers, HR staff, and other necessary employees face a materially worse bargain after supplying the 80- or 90-hour weeks that founders need.

Both hosts now reject indiscriminate antitrust rules, though they retain the case for blocking genuinely foreclosure-heavy deals. Andrew has moved away from his earlier sympathy for something close to a per se ban, while preserving Adobe–Figma as a positive intervention that left Figma headed toward an IPO; Ben still regards Facebook–Instagram as a serious competition problem. His political conclusion is harsher: if regulators cannot execute case-by-case judgment and the practical choice is blanket permission or blanket prohibition for small acquisitions, “I would rather blanket allow than blanket ban.”

AI’s current talent concentration may make some of these restructurings economically rational even if the broader precedent is damaging. Character.AI no longer needed to train its own models once the industry appeared to be consolidating around a few foundation-model providers, so Ben thought transformer inventor Noam Shazeer was better deployed back at Google. Cognition’s Devin and Windsurf’s IDE may likewise be complementary: Windsurf supplies customers and a gradual, data-rich path toward agents, while Devin represents the more aggressive endpoint.

OpenAI’s abnormal nonprofit structure may have been the recruiting wedge that made its breakthrough possible. Vikesh’s counterfactual—that Ilya, Dario, and other “true believers” might not have joined a conventional company paying less than Google—struck Ben as plausible. He now separates Sam Altman’s likely talent-recruitment shortcut from Elon Musk’s incentive to create AI capabilities benefiting Tesla and his other companies, while maintaining that powerful technology is safer when profit supplies “a limiting factor” than when governance rests on promises about mission and safety.

Cloudflare’s crawler blockade is a credible attempt to force a market for AI-era content, but its default settings may deepen Google’s moat. Cloudflare sits in front of roughly 20% of websites and can enforce blocks that robots.txt merely requests, giving it collective leverage to demand pay-per-crawl compensation. Yet publishers cannot afford to block Googlebot, and Google classifies AI Overviews and AI Mode as search rather than Gemini training, so blocking Google-Extended does not stop the most economically threatening AI use.

Copyright doctrine targets copying when the publisher’s real problem is transformation without traffic. Traditional search displayed exact snippets but sent users through; AI summaries can be legally transformative precisely because they eliminate the click, the ad impression, and the publisher’s revenue. Andrew doubts fractions of a cent per crawl can rebuild the lost ad-supported web, while Ben argues that bilateral licenses cannot scale and that some marketplace must reward differentiated, current content—even if getting there requires Cloudflare to “crack a few eggs.”

The Anthropic ruling gives model developers a major fair-use win for training while allowing claims over allegedly pirated inputs to proceed. Ben Thompson backed Judge William Alsup’s distinction between transformative use and unlawful acquisition: digitizing used books Anthropic purchased for training was fair use, but “just because your usage of it is fair use doesn’t mean your acquisition is now suddenly legal.” Andrew Sharp accepted the legal result while warning that courts, rather than Congress, keep deciding technology’s largest policy questions.

Copyright liability should turn on infringing output, not merely on a model having absorbed copyrighted work. Ben argued that no individual book is material to a corpus of roughly seven million books and that an LLM’s output is “unquestionably generating original work”; Andrew’s pushback was that machine-scale ingestion differs meaningfully from human reading. Ben’s answer: “If you think that’s a problem, then make a law about it. Don’t try to retrofit copyright, which is about copying.”

Making model vendors responsible for every possible reproduction would structurally favor closed incumbents over open models. Claude’s aggressive filters helped Anthropic because the plaintiffs did not allege that its product reproduced their books, but Ben said those filters should not be a legal prerequisite. Hold a user liable for deliberately copying an article, he argued, while recognizing that a general-purpose LLM has “far, far, far more” lawful uses—the Betamax principle, not Napster.

Congress could turn lawful training-data access from an incumbent moat into a U.S. competitive advantage. Ben proposed directing the Library of Congress to digitize its holdings and provide a lawful training corpus to U.S. AI companies; otherwise, startups must scour the world for used books and build scanning operations while foreign developers or rule-breakers proceed more cheaply. Andrew was partly persuaded that “you have to prioritize AI,” but still wanted nationally consistent rules.

Elite AI researchers now operate in the closest thing tech has seen to an NBA labor market. Their specialized work is unusually portable across Google, OpenAI, Anthropic, and Meta, the pool of proven talent is small, and outputs are somewhat measurable—so “there’s a bank in every city.” Headline figures such as $6 billion around Jony Ive, $14 billion around Alexandr Wang, and perhaps $1 billion around Nat Friedman and Daniel Gross dwarf athlete pay nominally, yet remain smaller relative to the value of their companies.

The clearest near-term LLM return may come from bespoke software for obscure workflows that conventional software companies never discovered. A disability day program replaced manually counted Medicaid TXT-file spacing with a spreadsheet workflow built first through ChatGPT and then Gemini; checking a box now produces the billing output. It still required “a few hours of back and forth” and personal agency, but the endpoint is an ambient assistant that notices repetitive work and builds the automation—“Clippy’s final redemption.”

ChatGPT could become a child’s primary computing environment before social media or even a smartphone. One seven-year-old already uses ChatGPT, Roblox, and Minecraft, while her parent expects to withhold social accounts until age 14 or 15 in 2032; Ben saw that as a real platform risk for Meta because parents may prefer AI interactions to social feeds. The unresolved strategic question is whether AI products should avoid social features so they do not inherit social media’s parental distrust.

Starting an independent media business in 2025 likely means owning the product, accepting a smaller audience, and charging dramatically more. Ben would avoid dependence on TikTok or YouTube and potentially charge a few hundred or thousand committed readers $1,000 or even $10,000 a month, echoing research analysts who charge $100,000 annually; Andrew suspected he would have tried media briefly, then gone to law school. AI can help build a unique product, but using it as the content generator makes the result “completely non-differentiated.”

Apple’s most consequential WWDC signal was retreat: after promising AI beyond what OpenAI, Anthropic, or Google could deliver, it deferred unfinished features and demonstrated only what it could ship. Ben took the L for granting Apple too much benefit of the doubt, but called the reset “sensible” and “getting your house in order” — a necessary return to execution over aura.

Opening Apple’s on-device models gives developers cheap experimentation and gives Apple asymmetric platform upside, even though the models remain “vastly inferior” to cloud AI. A solo developer such as Overcast’s Marco Arment might add podcast transcription without cloud infrastructure: if it fails, the app absorbs the complaints; if it succeeds, high-end iPhones gain differentiated software. Ben’s caveat was blunt: “These models do still stink,” so this is the right move that may not move the needle.

Deeper ChatGPT integration could create a genuinely compelling AI reason to choose an iPhone, despite Apple not owning the underlying intelligence. iOS 26’s Visual Intelligence can send on-screen material directly to ChatGPT, while signed-in users can continue the conversation on a computer. Ben wants Apple to accept being “a ChatGPT vessel”: Apple supplies the tortilla chip, and OpenAI supplies “some tasty salsa.”

Apple’s distribution retreat may have arrived too late to preserve control, particularly after the appeals court declined to stay Judge Yvonne Gonzalez Rogers’ anti-steering injunction. Ben still favors locking down games, where consoles and Steam provide precedent and spending revolves around “whales,” gems, dances, and digital clothes, while freeing productivity apps and Kindle. The new Games app could support that bifurcation, but Andrew’s verdict was that Apple “lost so hard that you’ve now lost control.”

Liquid Glass is now a test of whether Apple still possesses its historical product advantage, not merely a cosmetic redesign. Ben expects roughness and says judgment may require one or two years, as iOS 7 did not settle until iOS 9 or 10; John Gruber’s favorable early view came with the possibility that Apple “shipped it a year early.” If the interface proves Settings-app bad rather than temporarily unfinished, the strategic question becomes whether Apple can still do what it traditionally does well.

Meta’s reported $14.8 billion investment for a 49% stake in Scale AI looks less like a capabilities acquisition than one of the world’s most expensive acqui-hires. Against Meta’s roughly $65 billion-$70 billion of annual CapEx and $42 billion of Q1 2025 revenue, the price is manageable; the stranger point is that Meta ownership may drive OpenAI and Google away from Scale’s data business. Installing Alexander Wang therefore signals “a real reset” after talent losses, a disappointing Llama 4, and an AI organization that appears messier than outsiders realized.

Meta is behaving from fear even though today’s LLMs could strengthen almost every layer of its existing business without achieving AGI. ChatGPT mobile usage reaching up to 20 minutes per user per day makes attention competition real. Andrew called losing time to OpenAI “classic losing,” and Ben agreed. Ben’s sharper organizational call was that Meta cannot progress until it removes Yann LeCun from the initiative—not because LeCun’s AGI skepticism is wrong, but because Ben sees his stance as undermining recruiting and being a poor fit for Meta’s immediate opportunities, such as advertising against “every single pixel.”

OpenAI’s nearly $6.5 billion all-stock purchase of Jony Ive’s io is its clearest bid to own the consumer AI interface rather than remain an app on someone else’s platform. Ben says an acquisition makes more sense than a loose partnership and fits OpenAI’s broader effort to become “the interface for AI.” His strategic framing: with Google certain to compete through models and Android, OpenAI sees an opening to become “the Apple of this space.”

The deal forces OpenAI toward a consequential choice between Apple-like integration and Google-like service ubiquity. Owning hardware could capture the full experience and economics if AI retains meaningful marginal costs, but it will also turn potential distribution partners into competitors and bias OpenAI’s services toward its devices. Ben’s warning from Google’s Android history: “The tail started wagging the dog.”

The $6.5 billion headline buys a concentrated Apple hardware organization, but not Apple’s supply-chain leverage. The operation has recruited former Apple leaders including Evans Hankey and Tong Tong, along with specialists spanning industrial design, manufacturing design, and operations; Ben says they have “plucked the best people.” Yet those veterans are accustomed to Apple dictating supplier terms, while OpenAI must build a smaller organization and perhaps a family of products amid supply-chain decoupling: “Hardware is really hard.”

The product opportunity is real, but neither host believes a new interface automatically displaces phones or chat. Altman called today’s interaction model the “terminal phase” preceding AI’s graphical interface; Ben counters that asynchronous text already mirrors how humans prefer to communicate and keeps users connected to the physical world. Glasses offer an outward camera, with audio potentially handled in the ears—though Ben notes that wearing devices in the ears for hours is problematic—and Andrew doubts any product could persuade non-wearers to adopt them daily.

Ive’s record supplies credibility and a warning that world-class design still needs a ruthless editor. His teams produced the colorful iMac, iPod, MacBooks, iPhone, and iPad, but Ben faults late-era Apple for elevating form over function through products such as the butterfly-keyboard MacBook Pro and $17,000 gold Apple Watch. Steve Jobs was “the world’s greatest editor in chief”; the nearly ten-minute, self-congratulatory launch film suggested that io still needs somebody willing to cut.

Google I/O showed industry-leading AI capability and a persistent inability to turn that capability into standalone consumer products outside Search. Ben praised Google’s models, image generation, and Veo 3, yet struggled to identify how ordinary users would discover the demos; even Google lacks a Gemini app for Mac, while its Flow interface felt rough. His concise diagnosis: Google remains “a one-product company, and that product is search.”

Google’s defensible plan is to use AI Mode as a proving ground, then “graduate” winning features into Search and Android at roughly six-billion-user scale. Real-time data connections, generated visualizations, AI Overviews, Lens, defaults, and existing habits give Google distribution that OpenAI must persuade users to seek out. That scale is simultaneously the moat and constraint: advanced features cannot reach main Search until they are reliable and cheap enough, while the $250 plan primarily attracts highly switchable enthusiasts.

Veo 3’s coherent eight-second clips with audio are already astonishing, and current cost and model capability are the immediate constraints on longer mass-market video. Ben described generation costs as “astronomical,” with each clip consuming substantial credits; as a Gemini Ultra customer, he now has about $550 per month in free APIs. As costs fall, his optimistic response to synthetic-media chaos is that people may finally assume “everything online is fake,” trust physical reality, and “go and touch grass”—a future Andrew hopes for but does not confidently expect.

Apple’s iPhone user base—not developer goodwill—is the App Store moat, leaving government intervention as the only credible constraint on its economics. Ben Thompson rejects the usual chicken-and-egg framing: “The reality is users come first,” and valuable users make developers tolerate almost any obstacle. His categorical call: developers will not flee, but Apple’s conduct can invite regulation or an adverse judicial decision.

Thompson says the App Store’s theoretical market-clearing rate is either nearly everything or nothing, making 30% an arbitrary exercise of platform power. Apple could demand “99.9” because developers cannot abandon iPhone users, while a prohibition on charging would still leave Apple supplying APIs because apps are indispensable to the device. He nevertheless defends commissions for Apple’s integrated in-app purchase flow; the abuse is preventing developers from telling users about outside commerce, not merely requiring them to use their own web checkout.

Apple has optimized itself into a brittle dependence on Google’s roughly $20 billion to $25 billion in annual search payments. Thompson describes the money as pure profit and frames it as perhaps a quarter of Apple’s profit, making Eddy Cue’s warnings about AI search strategically conflicted: Apple needs to demonstrate competition without losing the cash. His broader Cook critique is that Apple has “squeezed the juice out of every aspect of their business” across Google, China, and the App Store at the expense of resilience.

The first reported decline in Safari search volume in two decades is an unusually concrete signal that generative AI is taking activity from Google, not merely adding another app category. Thompson calls AI usage a “one-way ratchet” closer to the personal computer or phone than to Instagram: serious users tend to use it more, not burn out and leave. Google still controls high-value commercial queries, but a shift in Safari options or defaults could accelerate the change if Apple adds ChatGPT, Perplexity, or Anthropic options.

Google Search faces real product substitution, but its advertising machinery may endure far longer than the bearish narrative implies. Thompson cites commercial-query dominance, rising CPMs, difficult-to-replicate ad infrastructure, and Google’s claim that AI Overviews monetize as well as conventional search. “These things hold on longer than you think,” while new products often layer on top rather than fully replace incumbents.

ChatGPT is exposing higher education’s weak underlying economics more than it is destroying a previously healthy institution. Thompson’s blunt thesis is that “AI is not ruining college. College was ruined,” with degrees functioning mainly as filters for conscientiousness and basic intelligence after employers outsourced credentialing to universities. The wage premium remains large enough that he still considers sending children to college obvious, even as credentials become costlier and less connected to learning.

AI may divide the workforce between high-agency people whose output is dramatically amplified and a much larger population that delegates effort while consuming endless entertainment. Andrew Sharp insists that struggling through writing builds critical thought; Thompson concedes the concern but asks whether many students ever sought learning rather than the credential. The darkest scenario is economically productive in aggregate yet socially hollow: a few people gain organization-scale leverage while many become “vaguely dissatisfied” but comfortably distracted.

Ben Thompson’s explicit conclusion is that Apple needs new leadership—not because Tim Cook failed, but because Apple now needs to reverse parts of the system he successfully built. Cook remains especially valuable in balancing Apple’s relationship with China amid tariff uncertainty, and “you can’t argue with the results.” Yet privacy absolutism, strained developer relations, and supply-chain concentration all point toward a phase where “Apple is in a different phase, and the world is in a different phase.”

Apple’s privacy advantage has hardened from a business-model benefit into an “unshakable religion” that may now constrain product quality. Thompson argues that computing always requires trusting someone; customers chose Apple because selling excellent hardware aligned it with them, not because Apple eliminated trust. Private Cloud Compute will, by design, struggle to match the scale and capability of shared infrastructure from ChatGPT or Gemini—and users’ revealed preference is performance.

The strongest AI move would be to turn Siri into an API endpoint and let users select ChatGPT, Claude, Gemini, DeepSeek, or another system-wide assistant. That would make the iPhone “the ultimate platform play,” eliminate pressure for Apple to win the model arms race, and let it benefit from competitors spending “billions and billions and billions of dollars.” Apple could offer Siri as the private default while letting users decide whom else to trust.

Opening the iPhone to outside AI could weaken app lock-in or help another company build the device that eventually replaces it, but Thompson thinks that erosion is coming anyway. The choice is between leaning into Apple’s platform strength or prolonging the old app layer until “the eventual loss of your position” becomes more abrupt. His Windows analogy: Microsoft defended its legacy application model instead of designing the best internet-native phone, and Apple risks repeating that error with Siri and app intents.

AI should be the bridge to Apple’s next device, not an unnatural extension of the iPhone’s existing software hierarchy. Thompson’s rule is that “every affordance you make to support the old paradigm is going to compromise the experience for the new paradigm”; even Vision Pro’s support for iPad apps may be a handicap, though he concedes that case is less clear if the product should really be a Mac. Apple’s historic strength was breaking compatibility and moving forward: “Stop trying to be backwards compatible.”

China is the second long-cycle risk: Cook’s supply-chain system helped save Apple, but the resulting concentration may take painful years to unwind. Apple financed and trained Chinese component suppliers partly to escape Western and Taiwanese supplier margins, helping create an ecosystem that later served drones and rival smartphones. Thompson expects “a really painful few years” of diversification and doubts the architect of that system is the right person to reverse it.

Intel is the warning that strategic failure can remain invisible in earnings and the stock price until the next platform arrives. As fabs rose from roughly $500 million to $25 billion, Intel needed outside volume and ecosystem expertise, yet its integrated culture resisted becoming a customer-oriented foundry; once it lost the process lead, it was also difficult to work with. Thompson stresses this is not a short call on Apple: “These are the timelines that actually matter in technology”—five to 10 years between decisions and consequences.

Apple may require visible pain before any successor has the authority to change a company that still looks highly successful. Its executives are generally in their upper 50s and its board in its upper 60s, leaving Thompson doubtful that the senior ranks adequately represent 2035. Separately, Apple TV+ reportedly loses more than $1 billion annually despite about 45 million subscriptions and less than 1% of connected-TV viewing, although Sharp sees its increasingly reliable “solid B+” programming as a new-age HBO and defensible marketing spend.

Meta’s history divides into three eras, making the FTC’s case look temporally mismatched. The agency draws on emails from Facebook’s early social-network era, invokes dominance from its roughly five-year monopoly era, and seeks a breakup after TikTok forced Meta into a broad attention market. Ben Thompson’s verdict: the case treats a dynamic “4D image” as a static 2D one.

Facebook’s 2006 feed created the internet-native advertising machine that later powered Meta. Infinite content plus personalization turned disconnected profile pages into a continuously replenished attention product—“that is the internet”—even as users protested and usage reportedly rose 100X. Mobile then forced Facebook to abandon platform ambitions and become a focused, full-screen aggregator.

The Instagram acquisition was consequential but far less obviously decisive in 2012 than hindsight implies. Facebook went into its IPO around $38, fell to $19 within three months, and reached that low the same month the FTC finally approved Instagram; Thompson challenges today’s critics to explain why they did not buy the stock if the outcome was self-evident. He still believes the acquisition should not have happened, but now prefers preserving Silicon Valley’s acquisition system to imposing the blanket ban that political “blunt instruments” would likely produce.

Instagram Stories marked Meta’s peak ability to turn its existing social graph into a competitive weapon. Facebook copied Snapchat’s Stories feature, engineered it better, and gave users immediate access to friends already on Instagram—what looked “unseemly” to Andrew Sharp was also straightforward product competition. Stories initially monetized poorly and contributed to a historic stock decline, then created roughly two years of “astronomical” advertiser ROAS before measurement improved and large buyers erased the arbitrage.

Meta demonstrated monopoly power when it could knowingly suppress the video content users wanted, but TikTok ended that luxury. Zuckerberg prioritized friends-and-family connections over “video slop,” reducing usage in a move Thompson calls economically worse even if normatively defensible. TikTok proved that algorithmically selected entertainment from eight billion potential creators beats a limited social graph, forcing Meta into the same “absolute garbage that is incredibly addicting” as every other attention platform.

The ad-pricing dispute turns the usual monopoly critique on its head: Meta’s auction may feel extractive because it creates near-perfect competition among sellers. Sharp argues small businesses must surrender too much margin to online gatekeepers; Thompson counters that Meta does not set auction prices and that falling Shein and Temu demand should mechanically lower everyone else’s costs. “Perfect competition sucks” for suppliers because every merchant can reach the world—and must compete against everyone in it.

Both hosts end up treating antitrust as the wrong remedy for the internet’s race to the bottom, while disagreeing about regulation. Sharp entertains percentage caps and predicts future guardrails will make today’s unregulated feeds look “psychotic”; Thompson warns that price controls create scarcity and that regulating addictive technology can exchange social dysfunction for surveillance and red tape. His darker formulation: technology enables bad behavior, then supplies the tools citizens demand to suppress it—one route by which “technology leads to totalitarianism.”

Treasury stress appears to have drawn Trump’s attention amid the equity selloff. Trump paused the broad reciprocal tariffs for 90 days at 10% while raising China to 125% after seeing investors get “a little queasy”; Ben said disorderly 10-year moves can trigger margin calls, basis-trade unwinds and eventually a credit freeze. “It’s always the best way at the end of the day to get anyone’s attention.”

Ben’s core call is that the old trade system is unsustainable, but blanket tariffs destroy the American demand that could reshape it. The US is strategically dependent on a peer, yet indiscriminately taxing imported components raises costs for domestic manufacturers and weakens their ability to diversify. His prescription: identify critical inputs, publish credible tariff schedules and use guaranteed purchases to build US or trusted-allied suppliers — “leverage your demand.”

Apple’s China dependence is about flexibility and capability as well as wages. Moving manufacturing risk off its balance sheet helped Apple escape its late-1990s crisis, rationalize inventory and scale an iPod from a February idea to shipment eight months later. The uncomfortable conclusion is that “there’s a good argument that Apple does not exist without China.”

China became the TSMC of manufacturing: a modular platform whose aggregate volume compounds expertise and R&D. Whereas older US factories were optimized for one integrated product, Chinese suppliers became a horizontally scaled client-service system capable of serving startups within weeks and accommodating last-minute Apple redesigns. That ecosystem cannot simply be “lifted” into America, because its advantage is the accumulated demand of many customers.

Apple’s rational supplier strategy has produced strategic costs for the US and its allies. It helped Chinese competitors challenge Largan Precision and funded Yangtze Memory’s development as an alternative to SK Hynix and Micron; Ben said the resulting capabilities now extend toward HBM and Huawei. Chinese companies moved from a minority to a majority of Apple’s supplier list because “demand controls supply.”

Full reshoring and reducing dependence on China are separate goals requiring distinct tariff designs. Blanket tariffs can force broad onshoring only by making everything “crazy expensive,” while also erasing the incentive to move component production out of China; targeted measures could instead build a trusted manufacturing sphere. Ben’s categorical judgment: “The full onshoring is not viable. It’s not possible.”

The Taiwan-chip dependency makes decoupling potentially more dangerous, not automatically safer. China buying Taiwanese chips gives Beijing something to lose from conflict; exclusion from the global economy removes that restraint while its weak consumer market and excess industrial capacity already resemble “a wartime economy.” The longer-term threat extends to AI-driven robotics, where China dominates “almost everything that goes into” the physical supply chain.

DeepSeek is better understood as Moneyball for AI than as proof that compute no longer matters. Its $6 million figure covered one final V3 training run—not all the experiments, data collection, or R&D—and the disclosed optimizations are “a tremendous gift to the US.” Thompson’s analogy: the Oakland A’s found the tactics, but American labs can become the Red Sox by pairing them with vastly larger budgets.

DeepSeek achieved genuine efficiency breakthroughs without overtaking the US model frontier. Distillation is nearly impossible to prevent—rate limits merely turn it into a piracy-style cat-and-mouse game—and OpenAI itself used distillation to make ChatGPT economically viable. Producing GPT-4o-, Sonnet-, or o1-level performance months later means “by definition you’re not ahead”: these were “efficiency breakthroughs,” not capability breakthroughs.

Constraints made DeepSeek exercise an optimization muscle that abundant compute had allowed US software culture to neglect. Its mixture-of-experts training, expert-balancing method, and hand-programmed communication work showed that “the nerfing worked.” Thompson conceded his own prior bias against optimization and argued for dual tracks: one team pursuing capability without constraint, another thinking about efficiency exclusively.

The apparent shortage of transformative AI products looks more like a normal adoption lag than proof the technology has failed. ChatGPT launched only in November 2022; the web needed roughly 13 years before Facebook’s 2006 feed found a native commercial format. Chat clients already deliver “massively untapped efficiency gains,” but adoption remains gated by user imagination—and Apple Intelligence, which Sharp expected to make AI “idiot-proof,” has been “an absolute disaster.”

Cheaper models could increase rather than destroy aggregate compute demand because reasoning models can manufacture training data. A reasoning model generates chains of thought, those outputs train a cheaper base LLM, the improved base produces a better reasoning model, and the cycle repeats: “the AI’s making the AI better.” Thompson’s conclusion was categorical about the direction, though not the timing—this era could demand “a basically infinite amount of compute,” including inference performed specifically for training.

Microsoft now holds attractive AI optionality while OpenAI and SoftBank pursue a venture-scale frontier bet. Microsoft retains OpenAI API access and first refusal on compute while stepping back from funding models that may become commodities within months. Thompson put the alternative at perhaps a 5% chance with astronomical upside—reasonable for equity, not debt—and said the evidence currently tilts toward commoditization.

DeepSeek exposes the geopolitical wager embedded in US chip controls: a durable AI lead must compensate for America’s industrial weakness. Thompson remains “in a pile of mud” because EUV is an enforceable chokepoint, yet cutting China off may accelerate domestic substitution, weaken TSMC’s deterrent value, and encourage America to defend past innovation instead of advancing. The policy ultimately shares OpenAI’s takeoff premise: if models diffuse and commoditize, the supposedly compounding technological advantage may not endure.