Ben Thompson argues that AI shopping agents matter most when they create purchases that would never happen otherwise, not when they copy Amazon checkout. On this episode of Sharp Tech, recorded the week after Jensen Huang resigned from Nvidia, Thompson and Andrew Sharp also explain why Huang is not coming back, test Meta’s Muse agent on real purchases, defend United Airlines for blocking booking agents, and land on a surprising verdict about Apple’s foldable iPhone Duo.
TLDW
Thompson says the agentic commerce debate is anchored on replicating the old way of shopping, when the real prize is new demand: an agent helps you design a project, builds the cart from five suppliers, and you buy things you never would have clicked through yourself. He thinks small specialty retailers that block agents are making a self-defeating mistake, while United is right to block them because agents turn an airline into a commodity supplier. On Nvidia, the hosts argue Huang’s real offense was opposing the chip ban rather than breaking it, and that a total ban will build the Chinese chip industry it is meant to stop. The episode closes on the iPhone Duo, which Thompson expected to dislike and now loves, and on what expertise is worth when Claude can design the parts.
Thoughts
The best argument in this episode is that the agentic commerce question has been framed backwards. Most of the debate asks whether an agent can buy the thing you already wanted more conveniently than Amazon can, and Eric Seufert’s answer, which Thompson largely accepts, is no: going to Amazon and looking at the right item is better than texting a chatbot and hoping it picks the right one of forty listings. Thompson’s turn is to point at his own basement. When Claude helped him design a hardware project and then listed 250 bolts in individual quantities, plus electronics from five different suppliers, that was spending that simply would not exist without AI. He compares it to the feed, which did not make newspaper ads cheaper online but invented a new kind of advertising. His sharpest refinement is that the underrated step is neither discovery nor checkout but “cart construction,” the moment where the agent hands you a complete basket after a conversation about what you are trying to build. That is a more concrete claim than “agents will shop for us,” and it is testable.
The second idea that deserves attention is Thompson’s split between who should block agents and who should not. United, he says, is right to cut them off. An airline is a commodity business that earns its margin from loyalty, credit card fees and customers who go straight to its app, and an agent that compares fares is just Kayak or Google Flights again, turning the airline into a supplier whose profit gets competed away. But a specialty retailer has no ad business and no loyalty moat to protect, and the agent is the one thing that can find it in the long tail. His example is B&H Photo, which grew huge on Google and now blocks agents, against the small fastener shop that let Claude assemble a strange order and made a lot of money from him. “The agents are the aggregator’s enemy,” he says, and the retailers copying Amazon’s defensive posture are protecting a position they do not have. Sharp’s worry is the right counterweight: if three quarters of sites block agents, many of them through a service provider’s default setting, the new behavior could be smothered before anyone notices it working.
The closing exchange about a listener who had Claude design a 27-body electromechanical assembly is the most honest part of the show about AI and skill. Thompson does not claim the tools make expertise obsolete. He says you get out of them what you put in, that someone who knew what they were doing would have spotted the flaw in his light-box measuring rig before he built it, and that a language model will happily agree to bake three bad ideas into a system because they are theoretically possible. The skill that survives is judgment about what not to build, and defining what “works” so the machine can build to that standard. His point that human-written code was never bug-free is fair, but the question of where the feedback loop comes from for physical parts with little gears is left open, and that is the right place to leave it.
Thompson’s iPhone Duo review fits the same theme as his commerce argument, which is probably why it lands. His complaint is that Apple fell in love with the unfolding animation and designed the phone as one consistent experience across two screens, when in practice it is two different products in one piece of hardware: a small screen for a quick look on the go, and a big one you never close at home. The developer beta that locks the phone the moment you shut it, like an old flip phone, fixed his biggest irritation. It is the same mistake he sees in agentic shopping, carrying the old paradigm into a new form and getting a pale imitation of it.
On China, the strongest point is not that the Chinese will eventually build EUV machines, which Thompson treats as near certain because knowing something is possible makes it achievable. It is the mechanism. China wins categories like TVs, drones, fiber optics and 3D printers by being cheaper, which brings volume, which brings the customer feedback loops that made ASML, TSMC and Apple so hard to catch. A total ban forces Huawei, SMIC and their toolmakers into exactly that loop, while Sharp adds the risk that rarely gets discussed in Washington: China supplying its own industry while American tech stays almost entirely dependent on Taiwan.
Key Takeaways
- Thompson believes Jensen Huang will not return to Nvidia; on the earnings call, the chairman opened by saying he is not coming back and announced a national security and export controls committee that will include outside members.
- He argues most of the Megaspeed reporting described legal activity: Chinese firms renting chips outside China was allowed, Oracle and Microsoft did it too, and the Biden administration explicitly wanted it.
- In Thompson’s reading, Huang’s “ultimate crime” was opposing the chip ban, not violating it. Chips became national security items, tech companies were expected to act like banks, and Huang did not adjust.
- Responding to a listener, Thompson says China does not need to invent reusable rockets or EUV first. Copying what it knows is possible, then winning on price and volume, is enough to cut out the bottom of the market.
- Anthropic is the biggest winner of Huang’s exit in Thompson’s view: least dependent on Nvidia, opposed to open models, and about to IPO as institutional money comes free.
- Thompson agrees with Eric Seufert that replicating Amazon shopping through an agent is a worse experience, and that Amazon will fight it because of its logistics and advertising businesses.
- Where he disagrees is new demand: projects designed with AI generate purchases, like 250 individual bolts or parts from five suppliers, that would not happen without an agent.
- He calls “cart construction,” the step between discovery and checkout, one of the best opportunities for agents.
- Small retailers that block agents, including B&H Photo, are hurting themselves; he sees agents as the long tail’s best chance in ages and criticizes Cloudflare for blocking agents by default.
- Meta’s Muse found a sold-out Walmart onn 4K Pro streaming box within a minute, though it first tried to buy one for pickup in a town 36 miles away.
- United blocking booking agents is “totally expected” to Thompson, because agents reduce an airline to a commodity supplier under an aggregator. Sharp, who wants to book flights with Muse, disagrees as a customer.
- Sharp says Muse handles about a third of the tasks he tries, but that third is excellent: a tax spreadsheet that would have taken hours took six minutes.
- Thompson bought GTA VI for $99.99 to salute “the last great human-made” blockbuster game and doubts he will ever play it.
- After expecting to dislike it, Thompson loves the iPhone Duo and says its multitasking beats anything the iPad offered, though the iPhone Air remains his all-time favorite phone.
Chapters
3:16 Megaspeed, Nvidia and Why Jensen Huang Is Not Coming Back
A listener asks whether the legal gray area around Megaspeed renting compute to Chinese firms could clear the way for Huang’s return. Thompson says no, and that the legality is exactly the point: almost everything people were upset about was legal, including introducing Alice Huang’s company to Alibaba in early 2022, before the chip controls. He reads the affairs as the means to remove Huang, not the reason, and expects there was an ultimatum to go the easy way or the hard way. Sharp adds that Huang was the anti-Tim Cook in Washington and left plenty of ammunition lying around for his opponents.
16:47 Can China Build Its Own Chip Industry?
A listener argues the Great Firewall analogy fails because China’s internet boom came from a capitalist window that closed around 2012, and that the CCP can only copy what it has seen. Thompson pushes back on both halves. Solar, EVs and autos all boomed after 2012 under brutally competitive conditions, and copying a known destination is enough. Cheaper products bring volume and feedback loops, as Apple’s role as TSMC’s first customer showed, and the ban forces those loops into existence for Huawei and SMIC.
32:09 Anthropic’s Gain and How Much Blame Huang Deserves
Thompson names Anthropic the biggest winner, citing its independence from Nvidia, its opposition to open models and the timing of its IPO, and he describes its ideology as top-down control of the technology. A listener then argues the hosts could have used the same facts to blame Huang for damaging America’s position. Thompson concedes the parallel with Anthropic’s fight with the Department of War, and says doing the right thing badly or ham-fistedly poisons the ground for the cause.
47:01 Agentic Commerce and the Feed Analogy
Prompted by a listener on Booking.com’s struggle to optimize for LLM traffic, Thompson revisits his interview with Eric Seufert, who has been right for two years that agentic shopping would play out poorly. Thompson agrees that an agent replicating Amazon is a pale imitation, just as online newspaper ads were, and argues real economic impact arrives when something new is invented, as the feed did for advertising. Sharp compares agentic commerce to e-commerce in 1994 or mobile in 2009.
53:43 The Purchases Ben Would Only Make With an Agent
Using toilet paper and Amazon’s old Dash buttons as the tedious-purchase case, Thompson shifts to what he is actually doing: designing a one-button camera device with AI, which picks the parts and fills the baskets. That process generates demand that did not exist, and since the whole decision came out of an agentic experience, he wants his own agent to finish the purchase rather than restart with Alexa. Sharp agrees that people will each have one agent they work with.
1:02:03 Why Specialty Retailers Should Not Copy Amazon
Thompson says Amazon blocking agents is reasonable because it protects a massive profit center, but most retailers have no such business. B&H Photo blocks agents; Albany County Fasteners, which sells individual quantities, let Claude build his order and got his money. He criticizes Cloudflare for blocking agents by default across a large share of the web, and recalls how John Gruber turned Daring Fireball’s feed into native advertising when Google Reader broke his subscription model.
1:11:20 Ben Sends Muse Shopping
Thompson wanted more Walmart onn 4K Pro boxes, which run Android TV and feed his Channels setup, but they are sold out everywhere. Muse found one within a minute for pickup only in Watertown, so he stopped the purchase and narrowed the rules to Madison and nearby towns. It later found stock and pinged him to buy it himself. His verdict: thumbs up, and cart construction is the underrated middle step.
1:16:27 Why United Blocks Booking Agents, and Is Muse Worth Paying For?
A 1K flyer is surprised that United banned agentic purchases. Thompson is not: airlines earn their margin from loyalty and direct relationships, and agents work like Kayak, putting them under an aggregator. Sharp objects that he wants to book his 2027 trips with Muse. Sharp also describes a six-minute tax spreadsheet, and Thompson urges him to pay for the better Wonder model.
1:26:36 GTA VI on Principle and How the iPhone Duo Won Ben Over
After buying GTA VI on principle, Thompson explains his phone saga: he cracked his iPhone Air, found the iPhone 18 Pro too heavy, and picked up a Duo through a friend’s stock alert. Folded it is a chunk; open it feels like a big Air. He argues Apple wrongly designed for consistency between screens, praises the beta that locks it on closing, and finds himself using voice dictation far more.
1:40:35 What Expertise Is Worth When AI Can Design the Parts
A listener had Claude produce a 27-body CAD model in minutes and asks whether we are heading toward physical objects nobody fully understands. Thompson says the tools supercharge people who know what they are doing, that humans never wrote bug-free code either, and that the future work is building feedback mechanisms and evals that define what works. Expertise now means knowing what is a bad idea, because the model will not say no.
Notable Quotes
“The world changed. Chips are going to be treated like national security items. Tech companies need to act like banks. And he didn’t adjust.”
Ben Thompson, on why Jensen Huang lost his job
“They’re cheaper. And by the way, when you’re cheaper, you get way more volume.”
Ben Thompson, on how China wins industries
“This is new shopping that’s made possible by AI.”
Ben Thompson, on buying 250 bolts through an agent
“This is the biggest opportunity for the little guy in ages, particularly in retail. The long tail, the AI can find the long tail.”
Ben Thompson, on retailers that block agents
“The agents are the aggregator’s enemy.”
Ben Thompson, on why United and small retailers face different choices
“This is two different experiences in one device.”
Ben Thompson, on what Apple got wrong with the iPhone Duo
“You get out of it what you put into it. And if you know what you’re doing, this supercharges you.”
Ben Thompson, on AI-designed CAD files and expertise
Watch the full conversation here.
Related Reading
- An Interview With Eric Seufert About Agentic Commerce (Stratechery) the interview behind the episode’s shopping debate.
- My iPhone Duo Review (Stratechery) Thompson’s written review of Apple’s foldable.
- Jensen Huang Was Always Going to Lose (Sharp Text) Andrew Sharp’s essay on Huang and the Anthropic parallel.
- United Airlines Moves to Block Personal AI Agents From Booking Flights (Skift) the news behind the United question.
- Daring Fireball (Wikipedia) background on John Gruber’s site and its feed-based advertising.