Nvidia After Jensen Huang: AI, Chip Bans & Washington | Sharp Tech with Ben Thompson
- The hosts frame Huang’s reported removal as a governance shock whose costs extend beyond Nvidia shareholders. Andrew Sharp puts the stock down about 15% over five days, despite a Thursday rebound, while Ben Thompson describes a $4.5–5 trillion company engineered around one decision-maker. Nvidia was “an F1 car that’s built around Jensen Huang”; replacing the driver raises questions that a conventional succession plan alone would not address.
- Nvidia’s advantage rests on systems and supplier coordination as well as CUDA. Thompson credits the Mellanox acquisition with anticipating that AI would require interconnected chips, while Huang persuaded conservative Asian suppliers to invest ahead of demand. “It’s not a chip game anymore. It’s a systems game.” Losing the person who coordinated that system creates a different risk from losing a chip designer.
- Startups and Neoclouds may lose their most powerful patron just as financing becomes harder. Huang supported open models and alternative infrastructure providers because Nvidia needed customers beyond hyperscalers capable of building competing chips. He also became the “bank for the bubble,” helping bridge AI investment from free cash flow through credit and equity toward eventual operating revenue; Thompson questions whether an interim CFO or future CEO will sustain that role.
- Thompson argues that comprehensive chip restrictions manufacture the demand needed to displace America’s semiconductor ecosystem. Developers prefer Nvidia while it remains available; exclusion makes alternatives more necessary. His preferred mechanism is selling “N minus two, N minus three” generations: customers fund American R&D while retaining little incentive to build a competing ecosystem.
- The hosts distinguish Huang’s business incentives from the allegations used to explain them. They question treating Alice Huang’s supposed status as a trained Chinese spy as established. Sharp says the disclosure sequence “feels pretty coordinated,” while Thompson’s defense of Huang’s commercial logic stops at legal compliance: “Laws are laws.”
- The export-control disagreement ultimately turns on whether the long term still matters. Thompson sees a possible alignment between national-security hawks and people expecting decisive superintelligence within a few years. He rejects sacrificing durable industrial advantages to that forecast and questions how exclusive American superintelligence survives a disruption of Taiwan’s chip production: “What if you’re wrong?”
- Thompson’s home workshop supplies a concrete example of agents creating demand and shortening physical production workflows. Claude assembled hundreds of small hardware purchases, hit a retailer’s 250-item cart limit, and helped design original parts for a $1,500 Bambu Lab H2C. A screen-and-keyboard workflow for applying inventory labels went from idea to operation in “20 minutes tops.”
- Thompson revises his software thesis toward agents accessing systems of record rather than people opening bespoke apps. His inventory database remains useful even though he barely uses its interface: “I overstated the extent to which I’d want to create software.” The next constraint is execution access, illustrated by GPT-5.2 finally approving a Mac permission dialog through an external view of the machine.
1. The hosts see a shareholder loss inside a governance decision
- Sharp opens with reporting attributed to The New York Times, citing five Nvidia sources, about an alleged affair with an executive and company resources used to conceal it. The Wall Street Journal subsequently reported that the board was preparing to oust Huang. The hosts say his removal became official Thursday morning; Sharp says he is wearing NVIDIA green “supporting the new era with Colette,” and later refers to Colette Kress being in charge.
- Sharp’s puzzle: Nvidia shares were down about 15% over five days, despite a Thursday rebound, so why would shareholders benefit from removing Huang? Thompson applies the principal-agent problem to the board itself. Directors supposedly represent owners, but their incentives—including concern about personal liability—can diverge from what shareholders want. He believes most shareholders wanted Huang to remain.
- Thompson nevertheless accepts that misconduct can justify removal on long-term fiduciary grounds. His categorical description of sleeping with a subordinate is: “You always get fired.” Sharp adds that directors could know things still outside the public record. Their objection concerns the consequences of the decision; neither presents Huang’s importance as an exemption from company policies.
2. CUDA emerged from bets that nearly killed the company
- Thompson traces Nvidia’s history to the 1990s, an early failed graphics product, and a rescue from Sega in exchange for equity. The important technical progression was from graphics hardware executing predetermined drawing operations toward programmable shaders. Those small processors could eventually run programs themselves, opening possibilities beyond the original graphics workload.
- The underlying opportunity was an “embarrassingly parallel problem”: divide a screen into smaller regions and perform many relatively simple calculations simultaneously. Thompson connects that architecture to AI’s computational demands. CUDA supplied the programming capability, but the initial motivation came from graphics rather than a fully formed prediction of today’s language-model market.
- Adding programmability initially made Nvidia’s chips run hotter and less competitively. Thompson recalls ATI producing better graphics hardware while customers switched readily to whichever vendor performed best—an environment he compares with frontier-model switching today. CUDA was therefore not an immediately profitable moat; it was additional complexity that nearly sank the company before its broader usefulness became clear.
- AlexNet’s image-recognition results were the wake-up call, although Thompson does not know whether Huang envisioned the broader opportunity before or afterward. Subsequent GTC demonstrations threw CUDA applications against the wall until ChatGPT supplied an overwhelming winner. Sharp remembers their 2021 interview as Huang still searching for how the technology would become widely useful.
3. The moat expanded from programming tools to complete systems
- The AI boom also narrowed Nvidia’s story. Thompson remembers what he thinks was the 2024 GTC as “peak fame Jensen” and “least interesting Jensen”: product roadmaps displaced imaginative demonstrations. More consequentially, customers building on language models rather than directly on CUDA raised a question: how much protection would CUDA itself provide against alternative hardware?
- Thompson credits the Mellanox purchase around 2019–2020 with anticipating the next layer: many chips functioning together as a larger computational system. Networking became essential to scaling. “It’s not a chip game anymore. It’s a systems game.” A competitor could therefore win an individual-chip benchmark without matching what Nvidia delivered across an interconnected installation.
- His assessment of AMD is categorical on the distinction: strong hardware has repeatedly come with insufficient software and integration. He identifies TPUs as probably the largest credible training alternative at sufficient scale. He also says he believes Nvidia remains best on tokens per watt, which would favor it in an electricity-constrained world.
- Supply was another moat. Huang secured memory and components while persuading conservative Asian suppliers that AI justified investment before demand was certain. Thompson’s sharpest example is Huang’s five or six annual keynotes: presentations directed at suppliers coordinating the tens of thousands of pieces in a rack. Huang called them “the most efficient way for me to communicate to you en masse.”
4. Neoclouds protected Nvidia against its biggest customers
- Thompson’s value-chain framing: scarce chips let Nvidia capture exceptional margins, but abundance would strengthen buyers. Amazon, Google, Microsoft and frontier labs have the sophistication and resources to develop alternatives. A supplier dependent on a handful of such customers would eventually face pricing pressure, even if demand for computing continued growing.
- His comparison is AMD’s progress against Intel in hyperscalers versus its limited penetration, in his telling, among enterprises and government buyers. A large cloud provider will invest in making another supplier’s processors work because supplier competition is valuable. Smaller organizations stick with the established platform rather than undertake that compatibility work themselves.
- Nvidia therefore needed additional buyers committed to its ecosystem. Neoclouds, open models and startups expanded the market beyond customers with their own competing-chip ambitions. “NVIDIA did not want to become Intel.” Sharp emphasizes the resulting diversity; Thompson says startups losing the executive whose incentives supported that world is “a disaster.”
- That patronage also involved power. Sharp points to CoreWeave receiving chips early, while Thompson questions whether allocations always maximized immediate economic returns. He explicitly does not know whether Amazon received fewer chips because of competing initiatives. His broader claim is that Huang made long-term ecosystem bets, potentially choosing winners to protect Nvidia’s future bargaining position.
5. Nvidia’s organization was designed around one generalist
- Thompson’s personal encounters with Nvidia felt like dealing with a startup: improvised, fast-moving and charmingly ramshackle. Meta, by comparison, struck him as highly professionalized. Sharp observes that retaining startup characteristics at a roughly $4.5 trillion valuation sounds admirable. Thompson’s answer is that the same structure enabled Nvidia’s speed, annual release cadence and unconventional decisions.
- Drawing on Tae Kim’s The NVIDIA Way, Thompson calls Nvidia “an F1 car built around Jensen Huang.” Sixty direct reports mean specialists feed decisions to the center. Their competence is not in question; the organization assigns the integrating role to Huang. Underneath its many demonstrations and applications, Thompson argues, Nvidia essentially sells one product—a GPU whose possibilities software expands.
- Apple under Steve Jobs is his closest comparison. A functional organization hires specialists, then relies on its chief executive to integrate their work; a divisional organization distributes that responsibility among business units. Jobs’s famous reduction to four Mac products supported that structure. Thompson recalls criticizing Microsoft’s attempt to imitate it because Microsoft’s breadth required a different organizational design.
6. Succession requires an institution rather than another Huang
- Thompson suspects Apple’s experience between 2008 and 2011 was essential preparation: Jobs’s illness allowed Tim Cook to run the company temporarily, followed by a period with Jobs back. The lesson was not to identify another Jobs. It was to create an organization that could survive without “an exoskeleton that depends on the guy in the middle.”
- That comparison increases Thompson’s appreciation of Cook. Beyond supply-chain execution, he credits Cook with professionalizing Apple, including Apple University. He also recognizes the cost: writing down what makes a company distinctive begins a process of ossification. His analogy runs from walking through bicycles and cars to a bullet train—greater efficiency, progressively fewer directions available.
- Nvidia instead built “the fastest freaking car on the planet,” retaining flexibility while neglecting governance and succession. Sharp compares it with a Red Bull tailored to Max Verstappen that teammates struggle to drive. Huang had recently discussed working another 20 years; Sharp cites talk of running Nvidia until 120. Neither sees evidence of Apple-like preparation for a sudden departure.
7. The hosts separate the spy narrative from the commercial incentives
- The hosts describe a startling two-week collapse, but Thompson notes that Megaspeed had attracted New York Times and Bloomberg scrutiny a year earlier. Questions about a possible front and chips reaching China were longstanding. The allegation of a Chinese spy changed the story’s public reach; it did not introduce the underlying distribution issue.
- Sharp corrects the name to Alice Huang. He describes the framing that she was recruited young and trained in the “dark arts”; Thompson says sexual blackmail is a tactic used by spy services, and Sharp says he does not think that was the case here. Sharp argues that someone moving billions of dollars of important equipment could plausibly become known to or be in contact with Chinese state authorities without that establishing she was a spy. “Everyone is a Chinese spy, and no one is a Chinese spy.”
- Both reject the inference that the reported relationship explains Huang’s China policy. Sharp says Megaspeed represented about 1% of Nvidia revenue. Thompson says Huang’s longstanding commercial incentives to preserve Chinese customers and CUDA usage are the better explanation, and rejects explaining those incentives through Huang’s ethnicity.
- On military access, Thompson accepts the uncomfortable trade-off: chips entering China could reach the PLA. He criticizes assurances that military use was impossible and explicitly owns that consequence of opposing comprehensive restrictions. Sharp agrees that civilian-versus-military buyers offer little reassurance and questions whether even a ban can reliably prevent determined military buyers from obtaining high-end hardware.
8. The first compliant workaround began a lasting Washington feud
- Thompson traces the conflict to the initial Biden-era restrictions, which gated bandwidth rather than FLOPS. He disagreed with the policy but thought bandwidth was a sensible parameter for limiting large training systems. Nvidia produced a compliant, lower-bandwidth chip within a month. That chip was genuinely worse for scaling; Sharp’s suggestion that simply adding more chips restored equivalence gets rejected.
- Washington, in Thompson’s interpretation, saw compliance with the letter rather than the spirit. Huang saw a businessman legally adapting a product to serve customers. “He was circumventing what you thought you did.” Nvidia’s limited Washington presence compounded the clash: officials came to regard Huang as a bad actor, while he regarded them as ignorant of the technology.
- Huang subsequently gained access to Trump and helped reverse the H20 ban, the hosts say, while the earlier Biden restrictions remained. Sharp notes that China then would not buy H20s. Thompson sees continued appetite for Nvidia equipment through other channels as evidence that Chinese developers still preferred the hardware despite government efforts to move away.
- Sharp says the disclosures “feel pretty coordinated,” intended to delegitimize Huang and his policy agenda. Thompson interprets the sequence—spy allegations, an employee affair, another affair, suggestions of worse revelations—as an endgame. Their account attributes significance to the timing; it remains their interpretation of how longstanding policy enemies converted personal misconduct into Huang’s removal.
9. Export bans give rival ecosystems the customers they need
- Thompson’s objection to comprehensive restrictions centers on developer demand. A government can order an alternative to CUDA, fund it and announce its importance; it cannot manufacture an ecosystem by decree. “Ecosystems are pulled into existence by the developers who use them and want them and leverage them.” Those developers would generally prefer Nvidia if permitted to buy it.
- His best analogy is Windows Phone. Microsoft spent what he guesses were billions trying to establish a platform, but Android and iPhones already met users’ needs. Money could not create sufficient pull. Similarly, an immature Chinese chip ecosystem struggles to recruit developers while established Nvidia products remain accessible. Exclusion changes that calculation from patriotic aspiration to operational necessity.
- Thompson favors selling “N minus two, N minus three” generations. Older American hardware can still outperform a local startup’s offering, discouraging customers from investing time in that startup. Meanwhile, their purchases fund the supplier’s frontier R&D. He describes this as the longstanding technology-transfer approach: preserve a meaningful lead while getting disadvantaged customers to finance its continuation.
- His long-run forecast is categorical: China will eventually develop manufacturing and CUDA equivalents, then export that ecosystem globally. He sees comprehensive bans accelerating this outcome at the expense of American semiconductor and equipment businesses. Sharp adds a present-day observation: despite repeated reports of Chinese manufacturing breakthroughs, demand for Nvidia—especially for training—has remained persistent.
10. The superintelligence timetable changes which costs policymakers count
- Thompson suggests a coalition between national-security officials focused on China and people expecting imminent recursive self-improvement or superintelligence. Under the latter worldview, protecting semiconductor advantages over decades becomes irrelevant: “All that matters is the next two years.” He mocks that supposedly decisive two-year window for having persisted for roughly three years already.
- Sharp says this was the governing logic behind the initial restrictions in 2022. Thompson places Huang on the opposite side: planning Nvidia’s next 50 years, supporting open models and resisting chip restrictions. He suspects Huang genuinely considered continued Chinese dependence on American technology good for the United States, while acknowledging how conveniently that belief aligned with Nvidia’s interests.
- Thompson’s physical-world objection is Taiwan. A strategy assuming exclusive American superintelligence cannot ignore China’s ability to disrupt TSMC. His example is a reported issue involving high-speed-rail vibrations several kilometers from production facilities in Tainan. He says facilities that sensitive would not handle explosives well. “Where’s your ASI?” is his challenge to a strategy treating chip supply as secure.
- The question is “What if you’re wrong?” about a short timetable that excuses sacrificing the long term. Sharp identifies resentment toward Huang across administrations, mentioning Marco Rubio and Scott Bessent under Trump; Thompson stresses the bipartisan character of the security establishment. He expects stronger restrictions and computer tracking after Huang’s exit, and regards the long-run industrial consequences as disastrous.
11. The AI investment bridge may lose its banker
- Thompson places AI financing in a progression: free cash flow, then credit, then equity. “We’re well in the equity stage.” The necessary next step is sufficient revenue to repay obligations and restart investment from operating cash generation. That transition requires a bridge; infrastructure’s eventual usefulness does not itself pay today’s bills.
- Huang supplied part of that bridge by supporting and backstopping companies—“being the bank for the bubble.” Thompson questions whether an interim CEO, particularly a CFO thinking about the numbers, will make comparable commitments, or whether a permanent successor can justify them. This is a question about sustaining risky ecosystem financing, not merely continuing Nvidia’s chip roadmap.
- Neocloud credit conditions had already tightened, with concerns developing in the preceding months, Thompson says. Their weakness following the news therefore adds to an existing vulnerability. His railroad analogy matters: railroads were useful, yet financing structures could still collapse. Huang’s departure, in his assessment, significantly increases the probability of that kind of money crunch.
12. A sound commercial argument does not excuse ignoring the law
- Sharp says he began as a skeptic of Thompson’s export-control argument and was eventually persuaded. He distinguishes Thompson’s approach—which still restricts semiconductor-manufacturing equipment—from Huang’s position. But agreement on long-term strategy does not resolve the immediate issue: Nvidia is an American company subject to American laws, whatever Huang thought of their wisdom.
- Thompson explicitly marks his next claim as suspicion: Huang may have regarded the entire ban as illegitimate and treated apparent compliance as sufficient while chips continued reaching China. He suspects there is substantive wrongdoing behind the allegations. That differs from the earlier example of engineering an expressly compliant product; believing officials are wrong does not authorize disregarding their rules.
- Sharp is frustrated precisely because he likes Huang’s optimism and recognizes his systemic importance. Thompson’s explanation is hubris: a fighter accustomed to winning, surrounded by weak governance, repeatedly ignored off-ramps while supplying personal ammunition to enemies. His tragic framing preserves the contradiction: the willingness to make enormous bets created Nvidia and also contributed to Huang’s downfall.
13. Agentic shopping turned a workshop ambition into actual purchases
- Thompson’s practical AI example begins with a recurring frustration: driving about ten minutes to Ace Hardware and losing an hour obtaining a tiny quantity of bolts. He decided to build a hardware store in his basement. The $1,500 Bambu Lab H2C added another capability: “To vibe code an object” and then hold it was extraordinarily satisfying.
- Claude proposed the fasteners, nuts and washers he might need. Thompson reduced an initial list costing several thousand dollars, then let it populate shopping carts with hundreds of small bags, typically ten- or 25-packs. One retailer capped carts at 250 items, requiring two orders. Thompson made the final purchase himself; the agent handled the tedious selection and assembly.
- The distinction matters commercially: he says the agent “drove economic value” because he would not otherwise have undertaken all that shopping. It removed both the work of deciding what to stock and the friction of finding every variation online. Sharp calls the project excessive; Thompson accepts that it is also a hobby.
14. Original physical parts became another conversational workflow
- Thompson ordered roughly 12 shelves with eight bins each, then discovered that the bins sat loose rather than sliding securely. His solution was a printed rail system: four interlocking pieces for each long shelf. He measured dimensions and discussed constraints with Claude, including side-to-side movement, front-to-back movement, shelf columns and a rear crossbar.
- Sharp checks whether Claude merely retrieved an existing CAD model. Thompson says these were original designs, as were a keyboard tray and other objects Sharp had seen. Claude generated the model, handled slicing into printable layers and sent the work to the machine. “I don’t use 3D software. I don’t use Bambu Lab software.”
- The workflow also included observing the printer’s camera, sending progress pictures and detecting when Thompson removed a completed item so the next print could begin. For him, the important change is the distance from an idea to execution. Physical fabrication now joins coding as something he can direct through conversation without first mastering another specialized interface.
15. The inventory project becomes a system of record with disposable interfaces
- Because Claude had selected the hardware and planned its storage, it could populate Thompson’s inventory database before anything was labeled. The database records where objects belong rather than maintaining exact counts. Claude then obtained the bins’ label template and generated labels and QR codes from the existing records, reversing the earlier process of photographing and entering items individually.
- Applying around 1,000 labels created a new bottleneck. His assistant Josh initially consulted a Telegram bot on his phone to learn each destination. Thompson saw the friction and had Claude put a visual guide on a networked screen: the relevant shelf, bin and cup appeared alongside the current label.
- An old Bluetooth keyboard became a single “next” button: any key advanced the process. The printer produced a label, the screen showed its destination, Josh applied it, and a keypress requested the next one. Thompson puts the time from recognizing the problem to having that workflow running at “20 minutes tops.”
- This leads to a revision of his earlier thesis about personal software. He barely opens the inventory app; he asks an agent that accesses its database. “I overstated the extent to which I’d want to create software.” The lasting need may be a “system of record,” with interfaces generated or bypassed according to the immediate task.
16. Access and permissions determine how much an agent can finish
- Thompson still prefers ChatGPT for organized programming projects, including Gecko, while Claude better handles his overlapping conversations and household work. He praises “Opus 5.5, the GOAT” for speed and capacity, then describes GPT-5.2 successfully clicking a Mac permission dialog. His preferences are task-specific rather than a claim that one model wins every category.
- A KVM gave the agent an external view of the Mac Mini, including permission dialogs that ordinary programs on the machine could not see. Previously, it identified the obstruction and messaged Thompson to approve it manually. This week, it finally acted on his standing instruction that this was its own machine: “It’s my Mini, so I clicked okay.”
- The breakthrough mattered because agent-written software can otherwise stall on routine permissions. Thompson thinks it occurred while creating the keyboard remapping app. He also describes Claude refusing household-system access until Astra “broke in,” after which he handed Claude the key. His broader condition remains that agents need the right access and controls; capability alone does not complete the workflow.
- Local inference is consequently becoming something Thompson feels unable to live without. He describes “Apple Harbor” as an inference-server mode, assumes it would avoid operating-system overhead and permission dialogs, and considers backup Studios. Apple’s software stack still trails CUDA in his assessment. Dependence on available open models leaves a final appeal: “Meta, please do some more open source.”
Full transcript
Hello, and welcome back to another episode of Sharp Tech, presented to you by Stripe. Joining me in the studio, Ben Thompson. Ben, how are you doing?
I thought, Andrew, the hello was a little too enthusiastic.
Mm.
We are here on a day of mourning.
You know, it was supposed to be a day of celebration. This is our 300th episode of Sharp Tech.
Oh. Well, there you go.
But the celebration has been canceled.
It has been canceled.
I feel like I—
It’s been canceled harder than Jensen Huang.
You’re sitting here in all black. I missed a fashion memo.
I thought about going to buy a leather jacket, but I didn’t have time. So, yes, I had to wear black. I was looking for a black jacket. I had a Bucs jacket, which, number one, is great because it’s black, which is the effect I was going for.
Mm-hmm.
Number two, it’s a franchise that technically is not dead, but might as well be dead.
Not in great shape.
Yeah.
They did lose the MVP.
Yeah, and they’re 1–11 or something, so it feels like it actually matches.
Well, I’m going to say I’m wearing NVIDIA green here today, supporting the new era with Colette. In any event, that’s where we’ll begin and probably where we’ll live for much of this episode.
Wednesday evening on the East Coast, the news came through. The New York Times cited 5 sources at NVIDIA who said that Jensen Huang had an affair with an executive at the company and that company resources were used to cover it up. Shortly thereafter, The Wall Street Journal reported that NVIDIA’s board was preparing to oust Huang.
By the way, what were you thinking in terms of what we were going to do for Sharp Tech if he hadn’t been fired yet?
I was under no illusions that he wasn’t going to be fired in time. Eventually, he was going to go. That’s what I knew on Wednesday night.
Right. I was going to say, since we record on Thursday and publish on Friday morning, we would have to acknowledge it at the beginning and then not talk about it at all. The chances of going stale were high. Maybe video is good for us. I’ve complained about timing as far as tech events go.
Yep.
In this case, tech delivered.
Perfect timing.
Thursday morning.
Here we are. So, Thursday morning it became official. Let me ask a stupid question on behalf of the audience at the outset. NVIDIA stock is down about 15% in the past 5 days. It did rebound a little bit on Thursday, but if the stock is down double digits in the wake of the news that Jensen could be on the way out as CEO, why is it in shareholders’ best interest to make a change? Why did this happen?
It’s actually a very interesting question.
A little governance question at the top.
Well, it is one of the eternal challenges. One of the biggest challenges in business and management theory is alignment. There’s a specific term for it, usually in terms of a company’s management being aligned with shareholders, because their incentives aren’t always the same as the shareholders’ incentives.
Yeah.
For example, most executives want to keep their jobs. Maybe not Jensen Huang, but he does want to keep his job. They’ll act in ways that may not be best for the long-term company. This comes up in terms of—
Quarterly earnings—
—looking for quarterly earnings—
—versus longer horizons.
That’s right. It comes up in incentive questions. If an executive is compensated via stock, then, in theory, that’s in shareholders’ interest.
Mm-hmm.
But if they know they’re only going to be in the job for a year, or they’re worried about losing their job—
Let’s juice that share price.
That’s right.
Absolutely.
So, theoretically, the board should represent the shareholders and keep a rein on the executives to keep them aligned. But the board can also be misaligned with shareholders. Setting aside why he unquestionably was going to get fired, to your point, this is an alignment issue. The shareholders definitely want him to stay. The board does not, and they have to be concerned about liability for themselves. NVIDIA’s governance was not admirable, which, by the way, I’m going to defend slightly when we get to that point.
Mm-hmm.
It is an interesting observation that this is an outcome the vast majority of NVIDIA shareholders did not want, and they’re going through it—
And probably NVIDIA employees as well.
That’s right.
But board members can face personal liability if they’re aware of misconduct and fail to take action.
People will make the case, and there is a case to be made, that this is in NVIDIA’s long-term interest. There’s a fiduciary duty, and at the end of the day, there were, at a minimum, obvious violations of policies that exist in every company, like “don’t sleep with your subordinate.”
Yeah.
Sleeping with a subordinate is the kind of thing you always get fired for.
It’s a big red line. We’re going to be talking a lot about China today. China has red lines. That is a corporate red line in America in the current environment. There may be other stuff that the board was aware of that hasn’t yet become public, but in any event, we are where we are at this point. Where might NVIDIA go from here as we look ahead? It’s a loaded question. How dire is this?
It’s pretty dire, and it’s dire on multiple levels. Number one, I mentioned the corporate governance aspect. There is this issue where I think Jensen Huang has made this analogy. I first got it from Tae Kim, who wrote the book The NVIDIA Way.
Mm-hmm.
I don’t know if he got the analogy from Jensen, or Jensen has now adopted it from Tae Kim and is using it.
I wish I had read that book. Now, if I read it, I bet it’s actually going to be pretty sad to read through the 300 pages from Tae Kim.
Absolutely. It will, because Jensen is a singular figure in tech history. He built this company for 30 years, taking it to the point where it was the most valuable company in the world. But he did it multiple times in the face of very large bets, with basically no one believing in it.
Mm-hmm.
He drove it through with his own personal belief and conviction. When he started the company, no one saw the market for chips like that. There were other accelerator companies, like Voodoo from 3Dfx.
So this was in the 1990s?
This was the 1990s. Their first graphics card was a total failure because they designed it wrong. They got bailed out by Sega because they were making it for them, and Sega didn’t want to say it didn’t work. Sega actually forgave them in exchange for equity in the company, though unfortunately Sega then went out of business.
Coming out with the graphics card was one thing, but the brilliant piece was the concept of the shader, where the GPU—
Mm-hmm.
The first versions of GPUs—I had a TNT2, I think, the first NVIDIA card I had, back when I was in college. There were 2 ways to draw triangles on screen, basically: OpenGL and Microsoft’s Direct3D, I think. They were sort of dumb. They processed inputs and drew onto a screen, but everything was totally preprogrammed.
Mm-hmm.
It was designed around knowing what the OpenGL commands were. You could handle all those commands, generate triangles—because everything’s a triangle—and put them on the screen to create graphics. What NVIDIA did was come up with the concept that those little things shouldn’t just be dumb; they should be programmable.
Mm.
So you can adjust them and make changes. In the long run, they can actually run programs themselves.
That was the concept of the shader. That is the fundamental idea of the GPU.
Mm.
That’s where CUDA comes from. CUDA is the way you program those computers. All of this came out of being a gaming company, because the thing about being a gaming company is that you’re drawing on the screen. It’s an embarrassingly parallel problem.
Yep.
If you cut the screen into smaller and smaller squares, you can go faster and faster because there’s less work for any one of these processors to do. You can just scale up and do all these calculations in parallel at the same time. That’s what AI is. That’s where the whole GPU thing comes from: doing a bunch of relatively simple calculations in parallel, all at the same time.
When did it become clear that NVIDIA’s hardware was going to be essential for AI development? Was that around 2018 or 2019?
I think people point to AlexNet, this sort of image recognition that was built on NVIDIA chips using CUDA. That was the huge wake-up call. CUDA started out as a bet to do better graphics.
Yeah.
Everything about NVIDIA comes from graphics, and it was considered a fairly insane bet. It nearly sank the company because the added complexity in the first version that supported it meant the chip ran hot. It actually wasn’t that performant. It wasn’t AMD then. It was a company called—I think it was ATI, or something like that, that AMD bought. Their chips were just way better.
Mm-hmm.
In the graphics era, it was sort of like frontier models today, where everyone would just switch to whichever one was better.
Ah.
Because they were all relatively dumb, right?
Yeah.
They all used OpenGL. They all used Direct3D—DirectX, whatever it is. I’m going to come up with it by the end of the show.
Great.
Maybe Jim will type it on the screen and have mercy on me. NVIDIA went through this generation when they introduced this new capability, and their chips sucked because of it.
Mm.
They almost went out of business. It was really bad. But they pushed through, and then this capability to program the chips was used for AlexNet to do image recognition. That was the big wake-up call: this can be used for—
This is going to be transformational and unlock all kinds of upside that just wouldn’t be there—
That’s right.
—without this hardware.
And so, DirectX. Thank you, Jim. It used to be called, I think, Direct3D. DirectX wasn’t the original name. Meanwhile, I don’t know whether Jensen came up with the idea that this could be used for more than games before or after AlexNet. But that’s when CUDA development really took off. It wasn’t just for gaming; it could be used for anything. At the old GTCs, Jensen would show off 50 gazillion things. You’d think, “Oh, this is nuts.” But all he was doing was showing different variations of CUDA libraries.
Right.
“We could use it for this, use it for this, use it for that.” They were making all these demos, basically throwing CUDA demos against the wall to see what stuck. The big shift was ChatGPT, and that stuck, to say the least.
Right. From a normie perspective, my knowledge of Jensen is that his technology became economically essential basically overnight, and then he was tenacious about taking advantage of that and capitalizing on it across all sorts of different dimensions.
Right. What’s really interesting is that if you Google NVIDIA’s stock price and look at the full chart, because the AI thing is so big, it’s basically a totally flat line at the bottom.
Yeah. Everything’s relative compared to what we’ve seen.
There’s a little bump that I think came downstream of their sales really surging because of Bitcoin mining—or Ethereum in particular was done on NVIDIA chips.
Uh-huh.
That’s also an embarrassingly parallel problem. Then it kind of goes down again when the Bitcoin era crashed, and then it just shoots to the moon. It’s one of the wildest stock charts you’ll ever see. I think there’s an aspect of AI that was bad for Jensen and CUDA.
Hmm.
He was out there saying, “This is clearly a very useful technology to do this in parallel, this scalable capability.” He was talking about how everything needed to be rewritten for these systems, which I always pushed back on. No, it’s hard to program with, and no one’s going to go back and rewrite everything.
Mm-hmm.
Well, maybe now with AI they all will. But he was looking for something. This was the most fun time to watch NVIDIA, because CUDA was a thing and you could see them trying to figure out what they were going to do with it.
It’s a wonderful—
But then LLMs come along—
Well, I was just going to say, you can go back to 2021. You did an interview with him, and you could see he was sort of fumbling in the darkness for how exactly this would be unleashed and wildly useful.
Right.
Then a year later, obviously. But there was a period where it was like—
No, so—
—throwing spaghetti at the wall.
That’s what happened when the interview started. When we first did the interview, he really enjoyed it, and he emailed me right away. He said, “Let’s do it after every GTC.”
Yeah.
The reason was that he was out there campaigning for people to use accelerated computing. He saw it and thought, “Oh, you understood what we’re doing. You’re pretty technical,” and so on. Once ChatGPT came along, there was a reason the interviews stopped.
Everybody gets it.
They didn’t need me anymore. It was very funny, actually. I think I was flying to Taiwan that night after the first interview I did with Jensen Huang. I was trying to figure out how to do an interview, so I rented—or got—a hotel room near O’Hare. It was so noisy in the room that I think I got the mattress up against the door because it was super noisy.
Wow.
I was trying to construct something relatively soundproof where I wouldn’t look like a total moron on video. It was quite the experience. I was scrambling to make the room somewhat soundproof.
We’ll put links in the show notes if anybody out there wants to get emotional listening to old Jensen Huang interviews this weekend.
Yeah. It’s super interesting because there’s a little bit about LLMs, but not very much. It was super clear at a GTC in San Jose. I think it was 2023 or 2024—probably 2024—where they had a huge stadium. It was the first “Jensen Huang the rock star” sort of moment. And it was super boring.
Mm.
They talked about the roadmap. It was like, “What happened to all the cool demos?” Now this is the business. They brought that back a little bit in the last few. But that one in particular stands out because it was peak fame Jensen and suddenly the least interesting Jensen.
Less imagination. Okay.
At the same time, you suddenly had very real reasons for people to fight against CUDA, and very real threats to NVIDIA’s position. With all this growth, were those margins going to be sustainable? If people are building on top of LLMs and not on top of CUDA, is CUDA really a moat?
And that became the only question that NVIDIA was organized to get answers around.
1. Systems Become the Moat
NVIDIA—and Huang—was a great CEO. I can’t overstate that. One thing they did around 2019 or 2020 was buy a company called Mellanox. What Jensen saw was that they were going to scale not just on a chip, with lots of processors, but with lots of chips functioning as if they were one chip, even though they were multiple chips. It’s recursive, like chips upon chips upon chips. To do that, you needed networking.
Mellanox had this very high-speed chip interconnect capability that was later called NVLink, along with a couple of other things. He bought the company so they could take a systems approach. Their moat today isn’t just CUDA.
Another moat they have is that they sell systems, not chips.
Mm.
Like AMD, I’m not sure what the status is right now, but in a previous generation their chip was faster than NVIDIA’s fastest chip. It’s not a chip game anymore. It’s a systems game.
Okay.
And AMD isn’t building systems. So that’s moat number two. Number three was the supply chain. Pulling— I’m monologuing here.
No, it’s fine.
So please jump in if you want.
It’s NVIDIA history, you know?
So number three—
Before we get to the present—
Well, this really ties into what’s happening right now and what’s going to be a big issue going forward. Huang saw how big this was going to be. You saw him out there in 2024, too.
Mm-hmm.
The whole thing: he’s out there drinking in Seoul, partying in Taiwan with a girl next to him.
Right.
And what was he doing? He was securing memory. He was securing parts.
I was going to say—
He was securing—
The Japanese drinks—that was all about memory, right?
That’s right. Korean. But—
Sorry. Yes.
But yes. It’s okay. We all look the same to you.
The Korean memory makers.
I’ve got to tap into my Asian media.
That’s right.
So he was out there securing the supply. This is super important on a few levels. It’s not just that this gave NVIDIA a competitive advantage, and that they could get memory and other people couldn’t—
Mm-hmm.
—and the other components. One of the challenges, which we’ve discussed in the context of TSMC, with the supply chain moving to Asia, is a fundamental conservatism relative to Silicon Valley—
Wildcatters, yeah.
Yeah. They’re usually closely held businesses. They don’t want to overinvest and be left holding the bag. It takes a while for them to say, “No, this AI thing is real. Start building.”
Mm-hmm.
Jensen is the one who got them building. He got the whole supply chain geared up and realized this was going to be a big thing. If it weren’t for Jensen Huang, the number of chips in the world today—and all this other stuff—would be much, much less. Not just chips, but all the components, all the pieces.
Sure.
There are tens of thousands of pieces that go into an NVIDIA rack. He was out there selling this vision.
Mm-hmm.
And he talked about this. Why does NVIDIA have so many keynotes a year? He would do five or six. He actually made this very clear. The last time I interviewed him—or the second-to-last time—was in Taipei. It was interesting because he was very clear that he was speaking to suppliers.
Yeah.
To the ecosystem. He said this on stage: “The reason I do these keynotes is that it’s the most efficient way for me to communicate to you en masse.”
Mm-hmm.
“Look, this is where we’re going. Get ready. Prepare these things. Go forward.” For that reason, above and beyond NVIDIA chips enabling AI, he was the salesman in the supply chain who got Asia believing in AI so they could supply everything that’s going on today.
Wow. Yeah. And he got a lot of people on the American side funded. I mean, the creative financing he’s been overseeing over the last—
So—
—couple of years will also come into play.
2. The CUDA Preservation Trap
All these threats to their position—this is the key. This is the root of where it all went wrong for Jensen.
Okay.
At the end of the day, the thing that almost killed the company, and the most brilliant thing he did, was CUDA.
Mm-hmm.
He got so wrapped up in preserving CUDA as a moat that it led to all sorts of issues.
Lots of Chinese issues.
Well, that’s one of them. But to mention the Neoclouds—which, what did I call them? I called them something. I called them NeoLabs in my update this week. I don’t know what I was doing.
I read that and thought, “Are there NeoLabs in addition to Neoclouds?” Okay.
Yeah.
Good correction live on the show here.
The Neoclouds are—because the problem with Jensen is that he supplies the hyperscalers, and he supplies the frontier labs, sometimes through the hyperscalers. But there’s a big fight in the long-term AI value chain over who captures the margin.
Mm-hmm.
Right now there aren’t enough chips, so NVIDIA is capturing all the margin. In a world where there are plenty of chips, surely their margin will compress because I can buy chips from anyone, or I’ll make my own chips if I’m Amazon or Google.
Right.
In the long run, NVIDIA was going to face real pricing pressure if it only had a few buyers.
Mm.
If its only buyers were Amazon, Google, Microsoft, OpenAI, or Anthropic, those buyers would have more power over NVIDIA. When you have fewer buyers, they’re more capable of coming up with alternatives and doing something else.
Sure.
Why has AMD made so much more traction against Intel in the hyperscalers, while enterprises and the federal government all run on Intel?
Mm-hmm.
AMD hasn’t made any inroads there. Why? Because those buyers are relatively small and not sophisticated. They have some program that works on Intel processors, but not AMD processors. If you’re Amazon, you’ll put in the effort to make sure AMD processors work well—
Mm.
—and can run everything, because you want multiple suppliers.
Yeah.
That puts pressure on Intel, because AMD can be better and cheaper, whatever it might be. NVIDIA didn’t want to become Intel, or it would eventually become Intel. It needed other buyers who would use only NVIDIA.
That’s what I was saying on Sharp China earlier this week. The way I made it sound on Sharp China was that it was all done altruistically, because Jensen just wanted to see all these people succeed. But there was real—
No.
—self-interest in doing that. What I actually said on Sharp China was that there’s more diversity in the market as a result of Jensen at NVIDIA over the last several years, and what he’s incentivized to bring to bear in the market. Without him there, I’m curious how dire this becomes for the rest of the AI ecosystem. I’ll read this note from Brian Cho on Twitter. He wrote, “If we lost Sam, we have Dario. If we lost Dario, we have Demis. NVIDIA is singular. The gap between NVIDIA and AMD would be wider than the gap between OpenAI and Mistral, Europe’s top AI company, whose models don’t rank in the top 40.”
It’s arguably good.
Okay.
So this is the whole tension. The issue is that these need to be done in systems. There’s really still no—TPUs are probably the biggest competitor in terms of having large enough scale to do training.
Mm-hmm.
You need systems. You don’t just need chips. You’re not doing real training on AMD chips.
Yeah.
They’re not a system. They’re just a chip. This has been AMD forever. AMD has always been awesome at hardware and just bad at software, or not even paying attention to it. It’s like, “Here’s a super awesome chip. Go figure it out.” NVIDIA—ATI was the company AMD bought that became its graphics division—whereas NVIDIA had better software, better drivers, for example.
Mm-hmm.
Even if their cards weren’t as good, there was the idea of shaders, and then doing systems. NVIDIA was so much earlier to this than AMD. So NVIDIA is still definitely the best option and will continue to be the best option for most companies as far as training goes. Again, TPUs are probably the other alternative. The challenge is that, in a world like that, the systems approach arguably makes for a more efficient outcome.
Mm-hmm.
I believe NVIDIA chips are still the best in terms of tokens per watt today. In a world where we’re electrically constrained, that is arguably good for NVIDIA. At the same time, Jensen was ruthless in seeking to preserve those big, fat margins—
Right, and their primacy. Sure.
—so he pushed the line in a lot of ways. One example: we’re scarce, everyone needs chips. Who gets the chips? How many of those decisions were made based on the correct economic decision in the short run?
In terms of basically giving CoreWeave all kinds of chips way back when.
Right. And, say, Amazon—I don’t know if this is the case—you’re going to get fewer. We don’t—
Yeah.
—we don’t like what you’re doing to other stuff.
Or Megaspeed.
Right.
Who can say?
And picking and choosing winners and losers. I think the way to think about Jensen is that he is a very systematic thinker, and he really does think in the long run. He makes big bets for those long-run outcomes. That’s how CUDA gets there. That’s a big long-term bet that isn’t going to pay off for a very long time, until it pays off hugely. So, in the AI era, he can see all the ways the AI era is super great for NVIDIA’s financials right now and is probably bad for NVIDIA’s moat in the long run.
And he was extremely aggressive in trying to ameliorate that, and sort of fighting against it. So some of the share allocation, playing hardball with that, and going to bat for selling chips to China—
Mm-hmm. Over and over and over again. Well, we’ll get to that. NVIDIA without Jensen and with Colette Kress in charge—
Oh, wait, so this gets to the governance issue—
—always gets a little dicey, going from a founder and an engineer to—
It’s not just that he was a founder and engineer. My exposure to these companies on a personal level is usually through interviews.
Mm-hmm.
And it’s always kind of interesting. You get a sense of different companies by interacting with them, even a company like Apple, which doesn’t give me interviews. That also fits. Their PR team is very, very professional. But you don’t get—
No sit-downs with Tim Cook.
No, no.
Maybe the Turnip era will be a different story. But okay.
Hey, I’ve got takes, John. I know. I have—
Nobody doubts that.
I’ve always said the company that’s always impressed me is Meta.
Mm.
Meta is super professional. They get back to you super quickly and are very straightforward. It’s a pleasure to deal with their team in general.
Mm-hmm.
All the companies are good, but they’ve always struck me as a very professionalized operation. NVIDIA felt like you were talking to a startup.
Okay.
Kind of seat-of-the-pants, making it up as they go along, somewhat ramshackle in a very charming way. In a way—
I was going to say, that sounds like a compliment for a $4.5 trillion company.
Ten years ago, they were a gaming company.
Mm-hmm.
Five years ago—
Five years ago.
—they were a gaming and Bitcoin company.
Yeah.
It’s a company that moves so fast. This idea that they accelerated to a yearly cadence of releases to support this—they move so fast, have been so innovative, turn on a dime, and do things like the Mellanox acquisition. By the way, that was in the face of activist shareholder pressure at the time.
It’s because it’s a company that was—to go back to that F1 analogy from the Tae Kim book—sorry, I didn’t even finish my story. It’s an F1 car built around Jensen Huang.
Mm.
It’s engineered so that he is the decision-maker for everything. The only company that’s ever been comparable to this is Apple.
Yep.
3. A Company Built Around Jensen
One of my first articles on Stratechery was about Steve Ballmer wanting to reorganize Microsoft. The timing was beautiful because I had resigned on June 30, and I think on June 30 or July 1, he released a company-wide memo about a reorganization. They wanted to shift from being a divisional organization to a functional organization. This was peak Microsoft Apple envy.
Mm-hmm.
Windows 8 was coming out. We were going to have touch, and we were going to do it better than Apple. We were also going to reorganize the company to be like Apple, which is a functional organization. This was insane. A company like Microsoft is so much a divisional organization: you have different business units, and they’re almost like their own little companies within the company. Then they all roll up to the CEO. When you get large enough, that’s the only way to function.
Apple always prided itself on being a functional organization. We have engineering, core engineering, hardware engineering, design, and marketing. No one owns the whole thing. Everyone specializes. We hire specialists who are best at what they do, and we drive collaboration to make them work together to come up with products. They all flow up to the top, where the person at the top is the product manager for the company.
He keeps all the plates spinning, yes.
Steve Jobs could not be more on board with this. He talked about this endlessly. Steve Jobs was a student of organizational design, and he was 100% committed to a functional organization, which makes sense for someone like Steve Jobs.
I was going to say, if you’re Steve Jobs.
The whole company basically works for you.
Right.
It’s the ultimate manifestation of the CEO as ultimate decision-maker. He built the F1 car before Jensen did. When Jobs came back, Apple had gotten all divisional while he was gone. It wasn’t just that he came up with new products, working with Jony Ive and others—the iMac. He completely reorganized the company to go back to being a functional organization with him at the top.
Yeah.
This is very powerful and effective. It does not scale. You can only do a couple of products. Apple still says they’re a functional organization. They still are, to a much greater extent than anyone else.
Mm-hmm.
But you can see the seams more, and there are differences when things don’t work together right. No other company can really pull this off because it just doesn’t scale to multiple products. NVIDIA is like early Steve Jobs, early Steve Jobs 2. Everyone there is a specialist.
Mm-hmm.
They do what they do, and it all flows up to Jensen. When he talks about having 60 direct reports, that’s what it means. Everything just flows through him, and he makes all the decisions. That’s not to say people below him aren’t extremely competent and capable, or that they aren’t doing the actual work on the details.
Right.
But it’s designed so that everything comes up to him. It’s a company that is a manifestation of—how can they do that? At the end of the day, NVIDIA sells one product—
Mm-hmm.
—a GPU.
Yeah.
A GPU with—this is why the GTCs were so amazing. People say, “How do they do so many products?” I wrote about this ages ago, when I first wrote about CUDA. I said, “It’s not so many products. It’s one product.” This is the magic of software. Software makes your one product do lots of things. Even back at Apple, when Steve Jobs came back, all they had were Macs.
Yeah.
He reduced it to the famous two-by-two. We’re going to have a consumer Mac and a professional Mac, or a laptop and a desktop. We’ll have four products. He drew it, and that’s all we’re going to do. Steve Jobs’ reduction in products was part and parcel of reorganizing Apple to be a functional organization.
Mm-hmm.
NVIDIA is like that 100%. The two companies are so similar.
And is the issue just a lack of personal charisma in terms of managing that going forward for somebody—
In that—
—who’s not Jensen?
Well, in that sort of company, you don’t hire generalists, and Apple will tell you that. At Apple University, one of their precepts is, “We don’t hire generalists.”
Mm-hmm.
“We hire experts, specialists in their field.” The only generalist is the CEO.
Yeah.
In retrospect, I was just talking with John at Dithering about this when we brought up the Apple comparison. One of the most fortunate things that happened to Apple—it’s hard to say “fortunate” when someone died—is that Steve Jobs’ death wasn’t sudden. He had to take a sabbatical for six months or something because of his cancer. Tim Cook took over, and then Jobs came back for two, two and a half years before he died. I suspect that was a real wake-up call for Apple in terms of succession planning. Succession planning isn’t just identifying who the next Steve Jobs is.
Mm-hmm.
What it is, is professionalizing the organization so it is an entity in its own right. It’s not an exoskeleton that depends on the guy in the middle. So Apple, I think, in that time period, from 2008 to 2011, became a much more professionalized organization.
Resilient.
More structured—
Yeah.
More resilient. In anticipation, we’re not going to have the—
We can’t find another Michael Jordan.
—the master of puppets—
Right.
—that is going to be… Right. And so that’s how they could seamlessly have someone like Tim Cook come in—not Steve Jobs, the exact opposite of Steve Jobs. But it mirrors back to Tim Cook’s, in retrospect—
Mm-hmm.
—makes me actually appreciate Tim Cook more. It wasn’t just all the supply chain stuff and all the things that he did there. He’s the one who professionalized the organization. I go back to Apple University. Everyone talks about Apple University being a Steve Jobs thing. No, that was a Tim Cook thing. I don’t think it worked out as useful. It’s like a glorified HR training program now, but it was part of the overall question: How do we make Apple into a sustainable entity when Steve Jobs is gone?
Okay.
And I don’t think NVIDIA’s gone through any of this.
That’s one of the questions over the last week or so: How does—
Jensen was talking about working for 20 more years a month ago.
I saw an Information story where he was talking about working until he’s 120 years old and running NVIDIA until he’s 120 years old. It is pretty remarkable to step back and say, “Oh, boy, what a weak board and no succession plan whatsoever for a $4.5 trillion company.” But Jensen was dynamic enough to make it all work.
No, that’s arguably why NVIDIA is a $4.5 trillion to $5 trillion company.
Mm-hmm.
When I was at Apple University, part of the reason—it was an amazing, perfect internship. I’m very glad it was not my job. I learned so much about the company and had very high-level access. It was amazing. But part of my worry was, “This whole concept seems fishy to me.” If you’re trying to define what makes Apple Apple, when you write this stuff down—
It stops.
—it’s like ossification, right?
Right. Sure.
You’re starting the process. But that’s the case for any company. We talk about disruption. We talk about companies. We talk about Apple right now. When you’re building a company, when you start out, when you’re a startup, you—I might have used this analogy before. I’m going to pretend I haven’t and—
Okay.
—offer you a fresh one. You’re going for a walk.
Mm-hmm.
Your eyes are open. You can go anywhere. You might see a path. You take the path and see where it goes. Suddenly, you’re not very efficient, but you can do anything. Then maybe you get a bicycle. You’re much more efficient and can move faster, but you’re a little more constrained in where you can go. You can’t hike up some boulders or whatever.
You can’t bike up a flight of stairs, sure.
Well, you’re such a city, coastal elite. I’ve been using my head.
So I’m out mountain biking.
Like, the countryside.
Okay, great.
No, just like a prairie or something. Going up stairs.
Okay, continue through the corporate—
Unbelievable.
—progression. What’s after bicycle?
Maybe a motorcycle. Much more powerful, go much faster. Maybe not quite as flexible or able to weave in and out of traffic as you can on a bike.
Mm-hmm.
Then you’re in a car. You kind of need a road. You might go off-road a little bit, but not too much. Eventually, you’re in a truck. Eventually, you’re a train. You’re like a bullet train.
Mm-hmm.
You move so fast and you’re so efficient, but you can only go in—
Can you turn?
—one direction, right?
Right.
You see this with these aggregators, right? They seem like impenetrable business models, and then you ask them to do one thing slightly outside their business and they just utterly, completely fail.
Mm-hmm.
They’re so tuned to do what they do, and everyone in that company is a cog in the machine. Then they just can’t… This is what happens to companies. They get more efficient. This is where the analogy of the F1 car goes further. If NVIDIA never got to be a bullet train, they just built the fastest freaking car on the planet.
Right.
Which meant no good governance, no succession plan, none of this. But that’s also why they were such a dynamic company. You don’t want an F1 car to crash. So—
We’ve seen this in literal F1 cars, where the Red Bull is tailored to Max Verstappen and historically has been basically impossible for—
Anyone else to drive.
—any teammates to drive.
Yep.
Max can push the boundaries in all sorts of great ways and knows how to drive it. Jensen clearly knew how to drive NVIDIA. Going forward, you talk about his bets. I think about how much money they’ve sunk into open-source AI and stuff like that. I worry that there are going to be fewer of those bets going forward, and I worry what that means for the diversity in the ecosystem. But you said this could be a good thing.
Well, it depends on who you’re talking about. For startups, I think this is a disaster.
Okay.
There was one entity with tremendous financial resources whose incentives—
Just Verstappen people.
—were aligned.
Yeah.
Well, no, their incentives were aligned with open models, with—
More customers.
—also, if you started out and built on NVIDIA, you had a whole bunch of developers who could work for you. And there were lots of places to buy NVIDIA chips.
Mm-hmm.
Or NVIDIA servers, like the whole neocloud, Nebius, whatever it is. NVIDIA was very aligned with that world. And NVIDIA still is.
Right.
But they’ve lost the guy who got it, who was the patron of that world.
And he had the personal cachet within that company to rationalize any of those bets.
The company did what he wanted.
Yeah. With the board, with shareholders: “In Jensen we trust,” and everybody got rich following that plan.
Yes.
Now it will be interesting to see whether any of that trust can be established under a new regime. I’m not terribly worried about them in the immediate future, but it’s murky for everybody right now.
Well, the long-term threat—we should get to the China point.
Yeah.
Just to preview the discussion, because we’ve talked about it before: The reason he was so obsessed with China is, who’s going to create an alternative to CUDA? The only—
Somebody who can’t access it.
That’s right. And who has the resources to do that? China, broadly speaking.
A foreign government.
Well, no. The Chinese government is not going to create a CUDA alternative.
Mm-hmm.
People actually using AI, actual engineers, actual companies. This is where everyone in this debate from Washington, D.C., drives me bonkers.
Oh, boy. I am going to be playing the role of a deep-state D.C. guy on this show today.
You don’t create an ecosystem like CUDA with a top-down diktat.
Mm-hmm.
You can start the project. You can say, “Go do this.” Ecosystems are pulled into existence by the developers who use them and want them and leverage them.
Yeah.
Those developers, those people who would pull a CUDA alternative into existence, would all prefer to use CUDA.
Mm-hmm.
They’d all prefer to use NVIDIA chips. As long as there were NVIDIA chips in China, the government could want whatever it wanted. It doesn’t get to dictate the choices of individual developers using… What, are they going to look over your shoulder to see what programming language you use?
Sure.
There was never going to be the necessary pull, the necessary demand for an alternative as long as NVIDIA chips were available.
Mm-hmm.
That’s why he wanted to make sure NVIDIA chips were available.
Okay. Well, we will come back to that point. For now, today’s show, as I said earlier, is presented to you by Stripe. Today, we are focused on Stripe’s digital wallet, which is Link, and which can be used for agentic purchases across Muse, Intellect, and other consumer-facing agents. Agentic purchases with Link were up 38X in September, so clearly this is a use case that will continue to be relevant for all of us, Ben. I have seen people say that agentic commerce was solved with the introduction of Link. Curious what—
No, it was not solved.
Well—
It’s funny, there was a bit of a controversy about this last week. Someone made a purchase from Best Buy that was rejected because one part was a third party. They looked into the Link thing, so I did some research on this.
Okay. What needs to be solved?
Do you want to give your credit card to your agent, and have that agent give your credit card to—
A merchant.
Right.
Sure. Which I’m sort of fine with, for the record.
It’s getting a little dodgy there. Let me go back a few years. When I was in Taiwan, I was at the equivalent of a Home Depot there. I bought something and went to return it. In general, returns aren’t a thing in Taiwan, but the hardware store did accept returns. I had to give them my credit card number for the return, but I’d paid with Apple Pay. I didn’t realize at the time that when you pay with Apple Pay, the merchant doesn’t get your credit card number. Apple, in conjunction with the issuer, gets a token that’s a unique number used just for that purchase. If you go to your wallet and look at a purchase under a card, the info button will tell you the number, or last four digits, that was actually used for that purchase. That number is different from your card number. I was able to find it on the spot, and my return got processed.
Ah.
Credit card security is obviously a big deal. It used to go do-do. Do you remember that?
Sure, yeah.
Do you really remember that?
Of course. I’m not that much younger than you are. But for the first 15 years of digital life, everybody was terrified of putting their credit card online. Now we just take security for granted.
You also didn’t realize the extent to which companies were solving these problems and protecting you. Apple Pay is a good example. It’s a much safer transaction. In Taiwan, sometimes for really large purchases, you have to use your card, not Apple Pay.
Ah. Interesting.
That’s actually backwards. Apple Pay is much more secure than a card. Anyhow, when the agentic Link thing comes along, you just have to hack it together. Stripe is a credit card issuer. They issue a credit card on the spot for your purchase, called a virtual card. If you’re making a purchase at Best Buy for $150.28—
They will issue an anonymized number—
They generate a credit card that’s only useful for that purchase. You use that credit card to pay, and then you bought the credit card from Stripe. That was a hack to get this working. There are lots of problems. Best Buy was trying to charge multiple entities because it was a third-party marketplace, as well as Best Buy. Your credit card statement shows that you bought computer software, because you’re actually buying the card from Stripe, so you don’t get your rewards. Some cards give you 2X or 3X points if you buy from a restaurant. It’s a brute-force workaround, but it’s not ideal in the long run. What’s happening now is Visa, Mastercard, and American Express are doing the whole thing like Apple Pay. They generate the token, know what the card is, and can handle rewards and divide up the payment. The answer is to have these tokens handle the purchase, while your card is obfuscated from the other side. But getting this out there depends on the issuer. The bank has to sign up, just like with Apple Pay. It took a while for all the banks to support Apple Pay. Even now, because I just changed phones, we have to have another phone discussion. Tyler and Carter are such a pain to get onto Apple Pay. The long and short of it is that the person complaining about this needs to go to his bank and get on the agentic token. Probably his card isn’t yet supported, so he’s getting the hack, which is a virtual card.
But Link has most of the major cards already cooperating?
The whole point is this is really difficult—
Yeah.
—and hard.
So there are only going to be a few clients that can do it.
That’s the Stripe value proposition. This stuff is so messy. Everyone wants to stay outside of PCI scope, the rules that govern how payment information can be used. No one wants to mess with that. Meta doesn’t want to get into that business. You also want to scale to the whole world. Netflix needed to support gift cards and local payment methods, all sorts of weird ways people pay for stuff. Credit cards aren’t a thing everywhere. You want all of that abstracted away, so someone else handles it and you just have one interface.
Someone else’s problem.
Yes.
Sure.
So that is the idea of Link.
And it’s Stripe’s problem.
That’s right. Link is now becoming a thing for agents. From a user perspective, it’s safe. You can set rules for how much you want to spend.
It gives control and visibility to the human owner on what the agent is executing, and you can approve what it spends. It can come back to you if the final price changes, and eligible purchases can even come with protection if the agent makes a mistake. So there’s Link. That’s a longer ad than normal, but that was a pretty interesting conversation.
It’s funny: there was a bit of a controversy about this last week. Someone tweeted that a purchase from Best Buy was rejected because it was from a third party. They looked into the Link thing, so I did some research on it.
Okay.
So you have this issue: do you actually want to give your credit card to your agent, and have that agent give your credit card to—
A merchant.
Right.
Sure.
There’s—
Which I’m sort of fine with, for the record.
It’s getting a little dodgy there. Let me go back a few years. When I was in Taiwan, I was at the equivalent of a Home Depot there, called B&Q. There’s a B in it, and it’s orange.
Okay, great.
It’s very confusing. I bought something and went to return it. In general, returns aren’t a thing in Taiwan.
Yeah.
In this case, the hardware store did accept returns, but it wasn’t very easy.
Mm-hmm.
I had to give them my credit card number for the return, but I’d paid with Apple Pay. This was relatively early on with Apple Pay, and I didn’t realize at the time that when you pay with Apple Pay, the merchant doesn’t get your credit card number.
Okay.
Apple, in conjunction with the issuer, gets a token that’s a unique number used just for that purchase. If you go to your wallet and look at a purchase under a card, there’s an info button that will tell you the number, or the last four digits of the number, that was actually used for that purchase.
Ah.
That number is different from your card number.
Mm-hmm.
I was able to find it on the spot. I was trying to figure it out: “I paid with Apple Pay, but the number I gave them doesn’t match.” Then I thought, “Oh, I gave them the number. There it is. My return got processed.” In general, there’s this concept of credit card security, which is obviously a big deal. It’s kind of crazy. It used to go, do-do. Do you remember that?
Sure, yeah.
Do you really remember that?
Of course. I’m not that much younger than you are.
Okay. Just making sure.
But it is pretty funny, because for the first 15 years of digital life, everybody was terrified of putting their credit card online and what might happen. Now we just take security for granted.
You also didn’t realize the extent to which companies were solving these problems and protecting you.
Mm-hmm.
Apple Pay is a good example. I didn’t even realize that was what was happening. When I thought about it, I thought, “Oh yeah, of course that makes sense.” It’s a much safer transaction. It’s funny: in Taiwan, sometimes for really large purchases, you have to use your card, not Apple Pay.
Ah. Interesting.
That’s actually backwards. Apple Pay is much more secure than a card.
Okay.
Anyhow, you have this challenge: how do you avoid having these 16-digit numbers floating around? If someone has the numbers, why can’t they just do what they want with them?
Floating. Sure.
Right?
Mm-hmm.
When the agentic Link thing comes along, you just have to hack it together. Stripe is a credit card issuer.
Yep.
What they do is issue a credit card on the spot for your purchase, called a virtual card. If you’re making a purchase at Best Buy for $150.28—
They will issue an anonymized number—
Your number actually goes through. They generate a credit card that’s only useful for that $150.28 purchase.
Wow.
You use that credit card to pay, and then you bought the credit card from Stripe. That was a hack to get this working.
Mm-hmm.
There are lots of problems. Number one, Best Buy was trying to charge multiple entities because it was a third-party marketplace, as well as Best Buy. If you make this purchase, your credit card statement shows that you bought computer software, because you’re actually buying the card from Stripe.
Ah, okay.
So you don’t get your rewards.
Mm.
Some cards give you 2X or 3X points if you buy from a restaurant. It’s a brute-force workaround, but it’s not ideal in the long run. What’s happening now, and has been coming online over the last few months, is that Visa, Mastercard, and American Express—American Express just launched a little while ago—are doing the whole thing like Apple Pay. They generate the token, and then they know what the card is. They can handle the rewards.
Mm.
They can divide up the payment because they’re the ones holding the actual card. Stripe didn’t hold the actual card. They were in the middle of a transaction, but they didn’t have ownership on either side. The answer is to have these tokens go out and handle the purchase. Your card is completely obfuscated, or not seen by the other side, but then they can tell, “Oh, this was a restaurant purchase; give him points,” or split up the purchase, or whatever it might be. But getting this out there depends on the issuer. The bank has to sign up, just like with Apple Pay.
Mm-hmm.
Visa and Mastercard got Apple Pay ready to go. It took a while for all the banks to support it. Even now—because I just changed phones—we have to have another phone discussion.
Oh boy, I can’t wait.
Tyler and Carter are such a pain to get onto Apple Pay. It’s just a big problem. The long and short of it is that there was this controversial tweet last week from the guy complaining that this wasn’t going to work. He needs to go to his bank. He needs to get on the agentic token—
Ah, okay.
Because probably his card isn’t yet supported for this, so he’s getting the hack, which is a virtual card.
But Link has most of the major cards already cooperating?
The whole point is that this is really difficult—
Yeah.
—and hard.
So there are only going to be a few clients that can do it.
That’s the Stripe value proposition.
Yeah.
This stuff is so messy. There’s a whole new term I use: everyone wants to stay outside of PCI scope. PCI scope is a set of rules that govern how payment information can be used. No one wants to mess with that. That’s why the whole reason for the virtual card is that Meta doesn’t want to get into that business.
Sure.
It’s a messy business. And it’s not just that. You want to scale to the whole world. I remember talking to Netflix ages ago about this. They needed to support gift cards and local payment methods, all sorts of weird ways people pay for stuff. Credit cards aren’t a thing everywhere.
Mm-hmm.
You want all of that abstracted away.
Yeah.
You want someone to do all that, and you just have one interface.
Someone else’s problem.
Yes.
Sure.
So that is the idea of Link.
And it’s Stripe’s problem.
That’s right. It is.
Great.
That’s the value. Link is now becoming a thing for agents. From a user perspective, the idea is that it’s safe. You can set rules for how much you want to spend, and so on.
Well, no, because we were talking about Link this week, and I saw that tweet, and there was a lot of discussion about it.
And so I actually just called Stripe up. I’m like, “Oh, I’m gonna leverage this here.”
I’m curious. Great.
“Can you explain all this?”
Right.
Like, well, then what? This seems like a bad situation. It’s like, well, because this particular person, it’s like a hack. But in the long run, agentic tokens are actually a thing, so there you go.
Okay.
Ah, fair enough. All right. We talked about Jensen’s greatness. I’m gonna skip to this question we have from Samuel. He says, “What’s the Vegas line on the over-under for Jensen affairs? Personally, I find it naive to think that an affair with a Chinese spy whose company paid NVIDIA billions of dollars had no impact on Jensen’s decision-making. Sure, it’s pocket change to NVIDIA, but studies show that doctors prescribe significantly more of a drug when a pharmaceutical rep buys them a single lunch. Money plus sex equals influence.”
Setting aside that second sentiment there, the Vegas line on the over-under for Jensen affairs: while we’re recording this, Reuters came out with reports of another Jensen Huang affair with an NVIDIA employee. So, over-under, I’m gonna set it at five and a half in terms of how dark this could get, but I don’t know whether you have any thoughts.
It’s not really a conversation that I want to get into.
Fair.
I’m sure—I would guess it’s much higher than that. And there’s a—not to get all sociologist on you or whatever—why do great men do great things?
Why? For women?
Like, historically. That’s one of the traditional reasons. You’re strong, powerful. You get access to lots of women. I’m not excusing it. Obviously, in our moral context, that’s not something that’s allowed. It’s also not something that particularly shocks me, to say the least.
No, and I don’t think that the chance—
And I’m not condoning it, particularly the employee thing, to be clear.
We said it was a red line.
We’re moving into the “we’re not condoning” area, to be clear.
And I only read that question because I literally got a text message about an affair Reuters is reporting on as we’re recording this.
Right. Well, here’s the question. Samuel takes as a fact that Alice Huang is a Chinese spy.
That is a point of contention, at least in your coverage over the last couple days. What do you think? As we look at what happened this week, what’s your mental model for how we landed here? It is unbelievable. We shouldn’t take for granted how shocking it is that 2 weeks after the initial New York Times reporting, Jensen Huang is no longer the CEO of NVIDIA, the most important AI company in the entire world. Even Monday, that seemed inconceivable. So here we are. How did we get here?
Two weeks or several years? Megaspeed has been a topic. The New York Times wrote an article about them a year ago.
Sure.
Bloomberg wrote an article about them a year ago. These questions about whether this is a front, whether it’s being funneled to China—this is an issue that has come up. So, yes, what changed this week? “Oh, Chinese spy.” That changes things. Is ByteDance a Chinese state company?
I don’t know whether they are classified as such.
No. The answer is—right. But yet, why? What’s the objection to ByteDance owning TikTok?
Military-civil fusion in China.
Well, no. I’d say it’s more just that any company in China ultimately serves at the pleasure of the state. And to the point, that’s actually a law.
Basically, if the Chinese government tells you to jump, no matter what company you are, you have to answer, “How high?”
That’s right. Now, this episode we can ask to what extent that applies to the US as well. But, yes, the issue is, at the end of the day, everything is the Chinese government and nothing is the Chinese government. Everyone is a Chinese spy, and no one is a Chinese spy.
The way this was framed is that Alice Huang was recruited from a young age, trained in the dark arts—which, by the way—
That does happen.
That does happen, to be clear, right?
Sexual blackmail is definitely an MSS—
No, I believe it’s used by spy services everywhere.
Yeah.
It’s not exactly a new tactic. I don’t think that was the case here. Now, was she in contact with—by virtue of—her husband was originally in this game. I think he ended up in prison for something or other.
I don’t know.
But this whole world of smuggling, moving stuff around—this is a large world, okay? Whether it’s because of sanctions or just—there are people who are in this market. If you’re in this market and you’re suddenly moving very important equipment worth billions of dollars, is it likely that the Chinese state apparatus knows who you are and may be in touch?
Mm-hmm.
For sure. But is that a spy?
Who can say?
Who can say?
Right.
So—
I do think that’s a useful distinction, to the extent that we have normies listening to Sharp Tech. We have this note from Nira, who’s talking about his family talking about Alice Huang and how Jensen fell for this beautiful Chinese honeypot. I can confirm that this has crossed the mom-and-wife threshold, where I’ve been getting text messages all week about the honeypot scheme that Jensen fell for, and how could this ever happen? I’m not sure that’s really an accurate way to model what happened here.
No, but it’s an accurate way to—
To make people care about everything else.
It’s an accurate way to initiate the endgame. And I think that is—this week was the endgame. If Alice Huang was such a threat, why was this not addressed months ago, when she was identified as an issue, or when she was?
It’s Alice Huang, just for the record. Same name as my wife. You said Alicia, and it delights me to no end—
Oh, is that what I said the whole time? I apologize. Alice Huang.
Her name is Alice. Yes.
Alice Huang.
If Alice was a threat, I mean, that’s one of the number one questions when you look at the timeline here.
But “Chinese honeypot”—very salacious, gets a lot of attention, amps up the pressure.
Mm-hmm.
All eyes on. Next day, boom, employee affair.
Yeah.
Next day, boom, another affair. And the strong hint from everyone in DC: more is coming, and it’s actually way worse.
Mm-hmm.
What could be way worse?
I don’t know. You wanna guess? I have some guesses.
That the chips flowing through Alice Huang were not going to Alibaba, but were ultimately ending up with the Chinese military.
Yeah.
So, yes. Duh. There’s an aspect of the China conversation, which Jensen Huang is very much guilty of, that treats its listeners like they’re stupid. If GPUs are so important and valuable, if they’re going into China, the military’s gonna get their hands on them to the extent they want them. Of course they are.
I mean, that’s why Alibaba versus the PLA is a distinction without a difference, as far as I’m concerned.
And that applies to the US as well. Does the US have civil-military fusion? Of course we do, right?
Right? Is Microsoft a military company? Is Amazon a military company?
Yes.
Yes.
Yeah. Jensen would do this too: “Are these chips being used by the military? No, that’s impossible. No way.” You don’t know that.
Right. I mean, you probably do know that that’s just false.
Right, so you’re lying. Now, does he know—is he signing a contract with the Chinese military to sell them chips? No. But there’s this bit like, “Oh, it’s a big revelation that these chips might not have been for Alibaba.”
They were for the military. The chips were going into China. If the chips are going into China, there’s a good chance some of them are going to end up with the PLA.
Yeah.
Just duh. When I write that I’m opposed to the chip ban, or at least to the totality that the national security apparatus wants, I do so knowing—and owning—that this involves trade-offs. Of course this means the Chinese military is going to get some of the chips.
You’d be crazy to have any illusions that that wouldn’t happen. Also, you could argue that you’d be crazy to believe it’s possible to keep the Chinese military from getting high-end NVIDIA hardware—
Right.
—if they want it—
So this idea that—
—even with the chip ban, that is—
So overall, the issue here is that these chips are banned from China. Jensen is obviously lobbying to sell chips to China for the reasons we explained before. It gets—
And not because of this affair with Alice Huang.
Right.
So, like—
That’s so—
In terms of allocating hardware to Megaspeed, it was only about 1% of Nvidia’s overall revenue.
Now, which smuggling organization got more chips? Did Alice’s get more?
It could be the one that was sleeping with Jensen.
Right, right.
Sure.
But the idea and the insinuation really irritated me over the weekend. You’d have these national security types basically saying Jensen lobbies to sell chips to China because he’s sleeping with a Chinese spy. No.
You know—
That’s not the—
We did ultimately stamp those takes out, but for the first 3 days of this scandal, after the Financial Times reported that it was a romantic relationship, there was a lot of that. It was so stupid when you juxtapose it with Jensen Huang’s longstanding policy positions and the obvious financial incentives that have aligned with his agenda.
The other thing that comes up is the racial undertone: Jensen wants to sell to Chinese people because he’s Chinese. Jensen wants to sell to Chinese people because he’s a businessman. That is the way to understand Jensen. The reason he wanted to die in the Nvidia CEO chair is that Nvidia is his life.
Yeah.
This is why he did die this week. It is his life. It’s all he cares about. He feels like he’s carried this company from a Denny’s table through multiple near-death experiences, brought it all the way through, and is reaping the benefits now. He’s thinking long-term: “How do I preserve this? How do I keep what I have?” I think that’s why he wants to sell to China. It’s super important for his moat, above and beyond the fact that he wants to sell everywhere so he can make more money.
More customers also sort of preserve Nvidia’s pricing power. Question there—
Well, I think—
Okay.
A really illustrative episode: Nvidia had no presence in Washington, D.C. The China chip ban comes out under Biden, and it gates what you can sell based not on FLOPs, but on memory bandwidth. I actually thought Version 1 was the right way to construct it, though I disagreed with it, because what you’re constraining is the large systems you can use. Those are limited by bandwidth. So what did Jensen do? Within a month—
Engineered around it.
They had come out with a new chip that had less bandwidth. To be clear, that meant it was a worse chip. It wasn’t as useful for training because it had less bandwidth. They literally go into chips and burn off the connections, or weld them off.
And so you have to link more of the chips together to get comparable power, right?
It’s not going to be comparable. You’re fundamentally limited if you don’t have sufficient bandwidth. You’re not going to be able to scale as much or do as large of models, and so on. My interpretation at the time was that this immediately put Nvidia behind the eight ball with these national security types.
They’re like, “You’re complying with the rule—”
“We did not intend—”
“—but not the spirit.”
“—this to be circumvented.” Sure.
But that was Jensen saying, “Fine, I’ll follow the law, but I’m a businessman. I’m going to sell chips. I’ll serve my customers. I’ll make a chip that they can buy.” I think that started the entire Nvidia-versus-the-national-security-apparatus relationship on a bad foot.
Yeah.
From that day forward, they viewed him as a bad actor, and he viewed them as—
Idiots.
—idiots. All the dynamics flow from those two things. Nvidia followed the law, but I think there was a bit of, “Within a month, you’re already circumventing.” No, he wasn’t circumventing. He was following the law. He was circumventing what you thought you did. That sort of set the stage for everything since then.
Yeah. I mean, it’s funny because when you zoom out, there’s a parallel here to Dario Amodei and Anthropic saying, “We’ve got a signed contract with the United States government that says we can dictate this, that, and the other thing,” and then trying to dictate terms to the U.S. government about how their technology will be used.
I think Jensen never fully appreciated how powerful his enemies in Washington were. What we saw this week was that these are some of the worst possible enemies you can have, and they leak this intelligence. Was she a spy? Was she not? Whatever happened—
Doesn’t matter.
—here—
He doesn’t have a job.
—feels pretty coordinated. Jensen is no longer CEO of Nvidia, and all of this was done to delegitimize him and his policy agenda. Once the seal was broken with that initial report in The New York Times and the intelligence, it’s open season in D.C. He and Nvidia have created so many enemies over the last several years. When the U.S. government wants to do something and sets out with clear policy objectives, and Jensen basically just gives them the finger—
Yep.
—over and over—
Yep.
—what do you expect to happen, Jensen? That’s what I—
This is where you come back to the incredible hubris on display by Jensen, and the extent to which rocketing to fame and power with no governance and no control ends badly.
Number one: over-under on affairs is like saying over-under on the number of bullets you’re scattering around for your enemies to fire at you. It’s the easiest way to take someone down. This is the case of “Show me the man, I’ll show you the crime,” in practice. The crime that usually always exists is—
Some sort of—
—having sex with someone—
—sexual—
—you weren’t supposed to, right?
Not decency, yeah.
Especially in U.S. culture, we still have the Puritan aspects. It works from a PR perspective. It works from a legal liability perspective. It checks all the boxes. It’s a very efficient way to get someone out of their job.
Yeah.
But I also don’t think that’s where this will end because I think there’s a lot of sketchiness—
No, there are people who want to destroy him completely. Number two: it’s not just, “Okay, fine, I’ll make the chips that China wants.” It’s also having access to Trump and tangibly changing the policy against the blob’s interests. He was clearly the number one person who had Trump’s ear, so Trump reversed the ban.
He reversed a ban on H20s.
Right.
And then China wouldn’t buy the H20s.
He never undid the Biden ones.
Right.
China woke up to the fact that this is the answer to the people: China could just dictate it. No, China had to cut off Nvidia.
Mm-hmm.
At the same time, why are Nvidia chips still going to China? Because they’re better and they’re good, and the people developing AI in China want Nvidia chips. So to what extent did Huang know or not know—
—where these were going.
My suspicion—we don’t know the answer, and I’m sure I don’t know if we’ll ever truly know the answer.
He of course knew they were going to China.
Yeah.
And both for his own self-interest and also in his… What’s the other phrase? A man believes what’s good for his pocketbook. What’s the one?
Sure.
You know—
I quoted Upton Sinclair: “A man whose salary depends on not understanding something—”
Right.
“—will not understand.”
No, so he views this policy as dumb. These people don’t know what they’re doing. They don’t understand that… By the way, we get emails from people asking, “How can you be against the chip ban?” The chip ban is a fairly insane policy that flies in the face of how the US has managed technology transfer for the last 50 years.
Mm-hmm.
It’s the people who are doing the chip ban who should be explaining why suddenly this is a good idea, when we never thought—
It’s uniquely necessary.
Right.
Sure.
The chip ban is the foundation of destroying our chip industry in the long run because you are creating the incentives for the demand—
Not just incentive. Imperative.
Yes.
You’re creating an imperative.
Yes. This has always been the issue with the whole China stuff. You’re like, “Oh, China knows they need to make chips; they’re going to make chips.” You don’t just make chips because you want to make chips. You make chips because you have the whole ecosystem working in unison, pushing and pulling and working together. You need all the pieces to do it. It’s such a difficult problem. You can’t just shovel money and make a chip. You can’t just shovel money and make an ecosystem.
Mm-hmm.
Microsoft shoveled billions of dollars, I would guess, into trying to make an ecosystem for Windows Phone. It didn’t work. If Microsoft can’t do it, is the Chinese government going to be able to do it? No. Why? Because Android was right there, because iPhones were right there. If—
And it is—
If Nvidia chips are right there, the Chinese government is not going to bring an alternative into being.
Well, speaking of hubris, I don’t want to get too confident, but I will note with amusement that for the last several years, people have been pointing at Nvidia hardware going into China, whether legally or illegally, and saying, “We’re giving them these high-end chips while they develop their own chip ecosystem. Oh boy, they’re going to be able to take everything in the long term, and we’re enabling all this militarization and progress in the short term. What an own goal from the United States.” Every 6 months there are reports about Chinese breakthroughs on the manufacturing side, EUV, what have you, Huawei chips. But each cycle, every 6 or 12 months, they are still so hooked on Nvidia chips—
Yes.
—particularly for training.
In the very long run, in the fullness of time, China is going to have its CUDA equivalent and its manufacturing equivalent. Then they’re going to take that ecosystem and that equipment and sell it all over the world.
Yeah.
We have written the death knell—
They could also bomb Taiwan.
We’re getting to that. We have written the death knell for all of our semiconductor equipment, for CUDA, for everything. That is the cost of the chip ban.
Mm-hmm.
That is why US policy for ages has been to sell N-2 and N-3 versions to other countries so they never have the imperative. If you can buy the older generation, it’s still way better than what you have. Everyone who needs it isn’t going to take the time to help some little startup making a chip. Good luck. I’m going to go buy the two-generation-old Nvidia chip, which is a gazillion times better, because I need it now for my business.
Yeah.
We did this with CPUs. You go back—we’ve done this for all technology. This—
We’ve done this with jet engines, with cars.
It prevents competition in the long run, and it makes a ton of money so you can fuel R&D and innovation going forward. You maintain ownership of the technology. You still have a meaningful advantage because you’re on the cutting edge and everyone else isn’t. All the people you’re disadvantaging by making them be behind are giving you money to make your leading-edge products better. It’s a beautiful system. The chip ban throws it all away.
Mm-hmm.
We have an email in here asking, “How can you defend…” No. The chip ban people need to explain why we’re throwing it away. This is where you get into tinfoil-hat theory. You have a movement—
Oh, boy.
You have a movement that believes, “That’s nice you care about the long term. All that matters is the next 2 years.” And by the way, all that has mattered is the next 2 years for the last 3 years, but whatever. Who’s counting?
Right.
ASI is coming. The AI is going to take over. Once you get superintelligence, game over.
Mm-hmm.
It’s done. So, yes, I acknowledge, Ben, cute theory. I acknowledge, Jensen, we’re destroying your moat in the long run. I acknowledge the semi-custom industry doesn’t matter because the long run doesn’t matter.
All that matters is the next 5 years, yeah.
This is, like, here’s the question—
Which, to be clear, was the governing logic in 2022 when the initial chip ban got passed.
Yes. The number-one person who was opposed to this was Jensen Huang. He was up there, and he was opposed to this on multiple levels. He was the one doing the open letter in support of open-source models. He’s the one out there extracting margin from these companies so they don’t have as much money as they might otherwise, because he’s charging a lot for his chips. He’s the one out there lobbying to sell chips to China. Jensen does not believe… He’s the opposite of a doomer. He’s planning for Nvidia’s next 50 years.
Yeah.
Jensen doesn’t just think it’s right for Nvidia to have chips going to China. “Oh, some going to Singapore, what a shame.” He also believes, I suspect, that this is actually good for the US because they should be hooked on our ecosystem. Again, very convenient that it aligns with Nvidia’s business initiatives. I happen to also think that’s true. And then number three: the other insanity of the doomer position is this idea that we, the US, are going to be the sole holders of superintelligence, so powerful that it overcomes all the industrial disadvantages we have relative to China, and China isn’t going to do anything about it. Like, fire 2 missiles at Taiwan.
Yeah.
And you blow up a few TSMC factories.
You’ve said in the past—
Where’s your ASI?
—even just a little bit of disturbance completely throttles TSMC processes.
There was a big thing in Tainan because there was a big government initiative. Everything was in Hsinchu in the north.
Wasn’t it on the train, right?
Yeah. They built the high-speed rail, and there was a big issue because the vibrations from the high-speed rail—
It’s a little tremor.
No, it was several kilometers away.
Mm-hmm.
They don’t do well with explosives, okay?
Yeah. And speaking of red lines, do not disrupt TSMC’s lines if you’re building in Taipei.
This is the danger of the doomeristic mindset. What if you’re wrong?
What if you’re wrong about the chip ban?
What if you’re wrong that the world isn’t just about the next 2 years?
Mm-hmm.
There’s such an excuse that comes with, “Actually, I don’t need to make any trade-offs because there are no trade-offs, because there is no long term.”
Yeah.
All that matters is the next 3 years. When you have this mindset, you throw away your chip industry, you throw away your ecosystem advantages in the fullness of time, over the very long run, because it doesn’t matter. You just had an excuse to do all these sorts of things. Meanwhile, I think you had a fusion. You had the EA, doomer, Anthropic sort of area, which believes this. Then you had the actual real national security apparatus that’s super anti-China—
Mm-hmm.
—and all they’re thinking about is war with China.
And members of the Trump administration. I would say the majority of the Trump administration—
Well, this is the whole thing: it’s bipartisan.
—resent Jensen.
Right. No, so it’s bipartisan. Usually that national security—
Well, I mean—
—hot bit—
Jensen really torched his relationship with members of the intelligence apparatus on the left under Biden.
Right.
And then under Trump, it’s been friendlier. But people like Marco Rubio and Scott Bessent, I think everybody’s been tapping their foot, saying, “When are we gonna get this shit under control?”
Well, then it’s layered on top of the real corruption, right?
Mm-hmm.
Like, what business interest does the Lautenberg family have in Malaysia data centers?
Curious.
Just to take the most gross, tangible example. And this is a frustration with Trump: I can agree with him on some policy issues, but if you layer a corrupt element on top of it, it just taints the whole thing.
Yeah.
And obviously, this whole Jensen thing—
I was gonna say—
The whole thing is done now. I’m sure we’re gonna have way stronger chip bans. We’re gonna get tracking of compute.
You’re wearing black not just for Jensen, but for the end of your “sell chips to China” take at this point, because it’s just over.
This is for our chip industry.
Hmm.
4. The AI Bubble Needs a Bridge
This is a disaster, in my opinion, in the very long run. It’s also bad for the whole AI bubble thing. With the— we talked about this in the context of Nvidia’s risky business. Nvidia, you go from investing free cash flow to credit to equity. We’re well in the equity stage. What comes after that? What needs to come after that is you actually start making sufficient revenue—
Right.
—so you can get back to free cash flow, not just paying off your debts, but then investing and going further.
And getting from here to there requires a bridge.
Requires a bridge. And Nvidia and Jensen Huang are out there backstopping companies, supporting companies, being the bank for the bubble.
Mm-hmm.
Is a temporary CEO, particularly a CFO who’s thinking about numbers, gonna be the bank?
Can they get away with that? Yeah.
Is a future CEO going to do that? I mean, I would say the chances that they’re— We already had a lot of tightening in credit markets for the neoclouds, not NeoLabs, already going on.
Mm-hmm.
They’re down a lot with this news. I think the only reason they’re not down more is, number one, the election happened; it wasn’t great for them. But they’ve also been getting crushed already because of concerns about their access to credit—
Yeah.
—only three months prior to that. But, yeah, the likelihood that we’re gonna have a money crunch like the railroad issue—the railroads were useful—it’s significantly increased, I would say.
Oh, boy. Yeah.
Well, can I ask you something? How dependent on Southeast Asian chip sales was Nvidia’s business today? Your point on the long-term interests obviously makes sense to me. I was a skeptic over the last several years, and eventually you wore me down and won me over. I think your solution also involves restricting the sale of semiconductor equipment to China, which Jensen has never advocated for.
Right. I mean, the problem with the—I just said a lot of stuff defending Jensen. The problem is, at the end of the day, laws are laws.
Yeah.
And at the end of the day—
And they’re an American company at the end of the day.
—and they’re an American company at the end of the day. I think Jensen, if the world were his customer, would say, “We’re an American company” in the U.S. He would say, “I’m Chinese too” when he’s in China.
For sure.
Right? I think the way to understand Jensen, again, is he was in Nvidia. That was his country.
Uh-huh.
And—
But it’s frustrating to me.
No, but the problem is, when you’re in an environment where you’re thinking, “You’re dumb, it’s bad for the country, it’s bad for me, it’s bad for everything,” and you’re the sort of person who would go on Twitter and suggest that I want to sell to China because I’m sleeping with the Chinese—do you know how many mistresses I’ve had? Believe me—
Immaterial.
—it doesn’t affect my decision-making. So you get in this mindset of, “These people are dumb. Their entire position is illegitimate.”
Yeah.
And my suspicion—we’re in allegation territory, suspicion—is that he treated the entire chip ban as illegitimate, and we can pretend we’re following it.
Work around it.
That’s right.
Yeah.
So I suspect there absolutely is smoke, or there’s fire here.
There will be more investigations, and there will be more incriminating disclosures.
But you can’t do this and then walk into Washington and basically go up and down K Street telling people they’re stupid.
Yeah.
Which they were doing.
And it’s really frustrating to me, A, because I like Jensen. His optimism, his energy—it’s all kind of infectious, and it’s a nice change of pace, particularly from some of the other leaders in tech these days. And I also obviously recognize his systemic importance—
Right.
—to where we are.
But he’s a fighter, and he’s—
But the man just could not stop selling chips to China. You are an American company. There are American laws. You’re putting yourself in this position.
Right. And like I said, meanwhile, you’re sprinkling bullets all around that they can use to shoot you.
It’s mystifying to me.
No, it’s a classic tale, a tale as old as time. The best novels, the best myths are about these incredible figures. It’s never interesting if they’re perfect. They’re always flawed.
Yeah.
And we were talking to Bill Bishop this morning—
Lively group chat this morning, as I was sitting on the plane.
Yeah. All week, I would say. We were talking about the next two years as far as relations with the U.S. and China. It’s probably gonna be pretty rocky.
Yep.
Part of it is, once again, China not living up to some of their agreements as far as rare earths and things along those lines. They just can’t help themselves from pushing it.
Well, they definitely put themselves in a position where members of the Trump administration can point at X, Y, and Z and say, “Why are we pussyfooting around here?”
This is a bit where I’d say—
“Let’s crank up the pressure.”
I don’t know, what is the gambling capital of the world?
Macau, maybe.
Yes, that’s right. It is not Las Vegas. All the Las Vegas companies make about a tenth of their revenue in Las Vegas. It’s all in Macau. The Chinese are crazy gamblers. It’s nuts. You see this around mahjong, money out—it’s nuts. They will—
Wow. Money on mahjong?
Oh, money on everything.
This is how white I am. I didn’t even know you could do that.
Dice. So at Chinese New Year, with all the kids, this is nuts: you have the big meal—
Uh-huh.
—and then all the kids get the red envelopes, right? You don’t give gifts. You get a red envelope full of money.
Yep. I’m familiar.
Chinese culture is very materialistic by and large. Important thing to know.
Just hand over cash. Great.
Yeah. Very transactional. This ties into some of the—mistresses are totally a normal thing or whatever. Just a different culture.
Yeah.
Let’s put it that way. So at Chinese New Year, we’d have the meal, give out the envelopes, and then it’s time to play dice.
Hmm.
There are these two dice games I always play where Amah is basically the bookie. You put bets down on over or under, or matching a dice roll. I think you roll four dice. Two match, and then you add up the two that don’t match. You roll again and again until you only have two that match. If you exceed the bookie, she has to pay out on your bet.
Okay.
If she wins, she takes your money. Then there’s another dice game where you put money down on different numbers, and she rolls. The point is, gambling is introduced—
So everyone—
—at a very young age.
—that’s the celebration? Everyone just gambles all night?
You’re there, and there are four-year-olds crying because they just lost $100—
Right.
—$90. Does Grandma feel bad about giving them their money back?
No.
No, she laughs about it.
They need to learn.
It’s hilarious. That’s right. The extent to which—
And the moral of the story being China is—
No.
—overplaying its hand.
No. Well, the moral of the story being, to the extent you wrote about Jensen Huang being Chinese—
He is ethnically Chinese.
He makes big bets, and he doesn’t know when to stop.
Okay.
And that is, I think, a characterization here.
Yeah.
He had so many off-ramps with the China thing to realize, “Actually, this is a problem,” and he thought he was too big to fail.
Yeah.
The hubris is pretty remarkable, and it is very tragicomic. Tragic, tragicomic.
Well, there are so many people who are going to be caught up in the—
But the big bets are what made NVIDIA.
Mm-hmm.
CUDA was a big bet. It almost sunk the company. What makes this into a myth, what makes this into a movie, is that all the characteristics that made NVIDIA what it was are the characteristics that led to its downfall.
It’s really too bad, but here we are at the funeral, our 300th episode. It’s actually been an enjoyable 300th episode, so I’m not complaining. In any event, let’s close with a brighter note. We’re not going to get to that many emails here, unfortunately. NVIDIA upstaged everyone.
But we have a subscriber-only mailbag.
We do. On Sunday, we will—
Yes. Lots of email.
—have a subscriber mailbag for people. Alice Wang upstaged everyone this week. Have you ever been honeypotted?
Alice? I was doing Alicia.
You’ve been saying Alicia all week, and every time I’m like, “It’s Alice, man.”
Really? Man, that’s nuts.
On Dithering and stuff.
I thought—
Well, honeypotting—
There’s one time—
—does imply consummation.
No, no, not consummation. There was one time where I was at an antitrust conference, and one of the most beautiful Chinese girls was really interested in my opinions on antitrust.
Great. I’ve never been honeypotted.
I don’t know if that was a honeypot, but I went back, looked in the mirror, and was pretty convinced it had to have been a honeypot.
I can’t imagine being all that passionate about your antitrust takes.
Yeah.
But there we have it.
It was weird. It was very weird. Only one time, though, so it makes me wonder if I actually wasn’t, because it only happened once. Maybe she actually was interested. But it was weird.
Wow.
It was very weird.
Fantastic.
I’m guessing no, but I want to believe that it was, just so I can say that I was.
Who can say?
Yeah.
I know you’ve had your interactions with other Chinese members of the state along the way here.
Well, yeah. Or interactions with other state entities saying, “Be careful what you write about.”
Indeed.
Yeah.
Andrew S. in Washington, DC.
Might have happened, actually, from two different people this week.
I’m writing in for this final question. I recall you mentioning a couple of weeks ago that you purchased a 3D printer. I just saw it in your little workroom, your dungeon back there on the other side of the basement, opposite end of the basement from our studio. I was mesmerized. I could have spent the last two hours watching your 3D printer print out these little plastic doohickeys you’re cooking up for your organization, which is an odyssey that clearly never ends for you.
I know.
Can you share a little bit from your journey as you get more comfortable 3D printing?
It’s amazing. It’s awesome. I’ve already spent more on additional rolls of different types of filament than I did on the different heads and things like that, or the machine. By the way, this machine is made by Bambu Lab. It’s the H2C, the one everyone recommends. It makes me love China.
Yep.
It’s like $1,500. It’s awesome. It looks great. It’s cool. The software is good, but I actually don’t know.
I had heard that China is just way ahead of the United States on 3D printing, generally speaking, so it’s not surprising that this machine is just awesome.
I was going to say there’s a realization here in general: The only software I use today is software I used before this year. I have a new 3D printer, and I haven’t touched the software a single time because—
Claude is going to get acquainted with that program.
Why would I? I discuss it with Claude, or ChatGPT—though I’m using Claude—and say, “I have this thing that I want to make. Here’s the thing.” Pardon me. With AI, it’s very funny: I had this whole thing with my daughter on an unpublished podcast where she had this whole thing about builders.
Uh-huh.
Which is hilarious. Everyone’s a builder these days.
That’s true. I remember that.
Do you know what the term was before “builder”?
What’s that?
A maker.
Mm.
I’m a maker. The 3D printing is unbelievably satisfying. On one end, you can vibe-code an app. To vibe-code an object and actually hold it in your hand is an amazing feeling.
I can’t even imagine. Yeah.
In this case, I’m doing lots more stuff, making lots more stuff, and what destroys my weekend is needing a particular bolt and not having it. Number one, I don’t know where it is because I have a thing full of bolts. If you have a bunch of bolts, how do you actually get them labeled right for your home inventory app? It just says “bolts,” right? So you go to Ace Hardware, which is about a 10-minute drive away, and look around: “Is this the right one?” Then you think, “I forgot to bring the piece. Did I measure it right? Do I have to go back?” It’s an hour to get $2 worth of bolts. No, like 25 cents’ worth. It just ruins every weekend. I’m like, “You know what?”
Never going to experience that again.
I’m building a hardware store in my basement. So, number one: What bolts do you need on hand?
I don’t know.
I don’t either. So I go to Claude and say, “Make a list of every bolt and fastener, nuts and washers that I might need.” I have all these ideas because I’m always putting them in there.
You’re a maker.
And I’m a maker. It already has a conception, but I explain the issue: There are lots of things I want to do, and I’m always frustrated doing them. What if I spend money right now to make sure I have everything I need? So it makes a list of everything I need. It went a little overboard the first time. I trimmed it back. It originally had me buying several thousand dollars’ worth of bolts. I’m like, “Relax.”
Great.
And then I need a place. Number one, it makes the whole list of things to do.
Uh-huh.
It’s actually our agentic commerce callback. It goes to these sites and puts things in the carts. I bought hundreds of tiny little bags of 10-packs or 25-packs of all these bolts. Number one, how would I make that list? Number two, am I going to click around on these websites and find all the M4 one-inch bolts?
Not in a million years, yeah. That’s a good point.
Claude just did it for me. It filled up all the shopping carts. Actually, I broke the shopping cart on one site. They had a maximum of 250 items, so I had to split it up into two orders.
250 $1 items, or 35-cent items?
Yeah, they’re just 10-packs.
Okay.
Because I didn’t want to overdo it.
Yeah.
But I wanted everything, because I follow the system: have everything on hand and then reorder—
But let me—
Make it very clear. You are overdoing it, but—
Oh, I am.
You’re enjoying it.
Right. This is my hobby, right? Look, just let me do my thing. Some people do drugs. I buy bolts, okay?
Yeah, absolutely.
So now it goes through and adds everything to the cart. I actually made the purchase.
Mm-hmm.
Now, in the long run, I could have Stripe or my Link do it. I should add Link, because I use straight Claude Code.
Okay.
I’m not using the agent versions. I’ve built a lot of stuff around it; it’s been an ongoing process. You don’t want an agent putting hundreds of items in your cart, and you want to do it quickly. This is a perfect example of agentic shopping being awesome.
Mm-hmm.
And by the way, it drove economic value, because I wouldn’t have bought all those bolts if an agent hadn’t gone through and bought them anyway.
That’s true.
So it puts all the bolts in my thing, and I order. I order this big stack of shelves, with 8 bins per shelf. I think it was 12 shelves, 8 bins per shelf, to make my hardware store in my basement, with a bunch of little cups and containers. They’d all be separated to make my little hardware store.
So we have a few problems. The shelves show up, and the bins are loose on them. I was so disappointed when it was set up. I thought it was like a slide-out hardware store.
Yeah.
Right? No, they’re just loose bins. You bump them, they jiggle around. Not great.
So you have 96 loose bins in there.
They’re not loose anymore.
No, they are not.
With a 3D printer, I print basically a little rail system for each shelf. It’s a long shelf, so it actually requires 4 pieces that clip together. Now, how did I do this? Did I go into CAD and design it and all that sort of thing? No. I talked to Claude. I said, “Oh, hey, I have these shelves. I have these loose bins.” I measured the length of the shelf, the widths, and the size of the bins. We did a couple of iterations, figuring out how to keep it from moving left to right or front to back, how much room to leave for the columns on the side, and accounting for the crossbar in the back.
We just talked through it, and it generated a little image. Then it would send code to create a real 3D model based on all those things. Once you’re ready to print, it does what’s called slicing, because a 3D printer builds things up in layers. It does the slicing and sends it—
So Claude is building that itself? It’s not going to a repository to get the CAD?
No. I build it myself.
Wow.
So all those things you saw, like the keyboard tray and all that stuff, I built myself. Claude and I built it.
That’s unbelievable. Historically, if you had a 3D printer, you would have to—
I don’t use 3D software. I don’t use Bambu Lab software. It’s just me and Claude, and Claude sends it straight to the printer.
Wow.
And it has full access to the printer. It watches on the camera.
Mm-hmm.
It sends me an update halfway through with a picture. When it’s done, because we have a process going on, you take it out, it sees the tray’s empty, boom, it starts printing the next one. So I have all these things on the shelf.
Now, part of the problem is, you might have heard I had a home inventory system.
Yeah.
As part of that, you do get—before, my poor assistant Josh was going around printing out a label, sticking it on there, taking a picture, typing it in, or doing AI recognition. In retrospect, it was kind of cruel and unusual punishment. It made for some good content.
In this case, Claude already created the whole inventory. It knows about the shelves, so you can actually work backwards.
Yeah.
We went back and forth and did the whole layout for the shelves, all the canisters and cups, and where everything was going to fit. Then it puts all the stuff into Home Inventory. My Home Inventory app is already populated with the items. Remember, I’m not using it as a counting thing. It’s just saying where stuff goes, so it’s all in there.
Then it can work backwards. When you do the label printing, these bins had their own template.
Mm-hmm.
It went to the website, got the template, and printed it out. Next to it on the template, it said where everything went. So you have all the labels, and you can slide them into all the bins.
Then we wanted to label all the cups. I got a networked label printer instead of a little Brother one or whatever. This is actually one that, if it was told directly—
I was curious, because the labels you have there look very professional.
Oh, yeah. It prints the label with the QR code from the Home Inventory app, which already exists. But then you have to put it on, right?
Mm-hmm.
It prints 1,000 labels. They’re all gathered on the floor. What do you do? Well, Josh to the rescue. But then I’m thinking, “How does Josh know what to do?” Do I print out an instruction sheet? No, I just made a little bot where he sends a message on Telegram: “Oi, this label, do here, do here.”
Then I saw him trying to do all these labels using his phone. I thought, “That kind of sucks, having his phone open and toggling back and forth.”
That will not do.
Right. No. Of course not.
Let’s talk to our agent.
Well, it turns out I have screens in this room. One of these screens is on my matrix, so I can be in the room, being a maker and watching sports. That means it’s connected to the network. I’ll have Claude generate a page that goes on the screen and says where the label goes. It has a little chart showing which bin it is and visualizes it, or whatever.
But how do you progress? You still have to open your phone to do the next one. I had an old keyboard here. It was a random keyboard that needed a Lightning cable. It took me longer to find the Lightning cable than to find the keyboard. I wanted to connect it via Bluetooth so it’d be next to the label printer, and I had Claude make that keyboard into a one-button keyboard. If you hit any button, it just registers as “next.”
Mm-hmm.
So now you have a thing where it shoots out a label and tells you where it goes: shelf A, bin 1, cup 2.
Nice and simple.
The label’s right there. You put it on, hit the keyboard, boom, it shoots out the next label, and the display shows what it is. I saw Josh putting the labels on using his phone. The time from “Oh, we could do this better” to having it on the screen and hitting a keyboard was 20 minutes tops.
Wow.
People have no idea.
Yeah.
They don’t understand what you can do. You can do anything. And now, with the 3D printer, I can print anything. It’s great. I actually have a thing for you.
I can’t wait to hear how crazy it gets with the 3D printer, because I feel like we’re only in the first inning of what your journey is actually going to be with that thing.
But I don’t need to know CAD. I don’t need to use software.
Yeah.
It’s just another example of all the software pieces that—well, or the label printer. I’d never used it. And just, why? All my home stuff, by the way, was stuck behind this thing where I was supposed to have a dealer do some of it. No, not anymore. I have an SSH key that gives me access to everything. You just tell it what you want to do.
We were sort of talking about this. I’ve always suspected I might be a little ADHD.
Yeah.
I’ve never gotten a diagnosis, but if you do the diagnosis, I’d max out on them. I was telling you I have all these systems.
Checking every box.
All these systems and things that help me. I always talk about how I need a daily deadline to get stuff done. That’s characteristic.
Yeah.
I don’t know if that’s the case. But I have ideas all the time, and the challenge is implementing them, because that’s a lot of work. You have to code the program. You have to go to the hardware. You have to do all these things. The degree to which AI closes the loop—
Mm-hmm.
—from idea to done is so tight. It’s unbelievable.
So people with ADHD are going to rule the world at the end of the day?
No, this is my theory. My theory is that, in the age of computers, being a functional autist was actually an advantage.
Yeah.
Because you could think like a computer, to your detriment and also your advantage. AI’s like, “Yeah, the functional ADHD people—”
The space cadets, the idea men of the world—they will inherit the earth.
I mean, you mentioned this when you came here, but the amount of stuff I’m doing in a week is insane. That whole room went from what you saw before to where it is now—
What I said was insane is the amount of stuff you’re doing that is objectively a waste of time. That’s what’s insane.
No. Driving to the hardware store is a waste of time. That’s not going to happen anymore.
Okay. Never again.
No, but obviously it’s super fun. It’s interesting. I think it’s useful to appreciate this stuff and how good it is. But I did talk about how everyone has software-sized problems in their life. I’m not sure if that’s right. The home inventory app is basically never used now.
Mm-hmm.
Because you just ask your agent. The home inventory app is a system of record. It’s a database of all this sort of stuff. So do I have a software-sized hole in my life, or did I have a system-of-record hole in my life? Whatever. I overstated the extent to which I’d want to create software. I realize I don’t use software beyond what’s already in my established workflows, all of which I’m slowly unpeeling.
Mm.
Because it’s just better to ask an agent. But it needs the right access and the right controls. The whole system we have with Gecko, with me and Josh and Domin—if I need something, say, with Stripe and doing these ad reads, it used to be, “Where do I find it?” We had a OneNote thing or ClickUp. It was so painful, and my mind doesn’t work that way. I just type “Stripe ads,” click a link, and have a webpage with everything on there, self-hosted on my server here. It’s awesome. It’s amazing.
Well, you left out the most thrilling breakthrough of all. You reported to me and the nerd group chat this week that GPT-5.2 successfully clicked a Mac GUI.
Oh, my God, it’s unbelievable.
The permission dialog.
I cannot explain how happy this made me. I have a KVM—sorry, Jensen. I have a KVM, so I can access the Mac mini. I can do it not just through screen sharing, but I have access in case it gets hung up. I put in a password or whatever. But I also gave the agent access so it can see the desktop from outside of macOS. Those permission dialogs pop up, and nothing in user land—no program on the computer—can see them.
Mm-hmm.
Only a person can see it from the outside in. Basically, I gave the agent a window to see it from the outside in, so that when stuff wasn’t working—“Why isn’t this working?”—it wouldn’t know it was failing. There’s no failure issue issued. It could look and say, “Oh, there’s a permissions dialog up,” and message me so I’d know—
Ben needs to click through. Yeah, yeah, yeah.
—that I have to open up screen sharing and click through. I feel like Claude works better than ChatGPT for this sort of stuff. I still use ChatGPT for programming actual programs. All Gecko’s built in ChatGPT. It’s more organized. You have projects, threads, and all this. I don’t know; for me it works better.
But it’s stupider than GPT-5.2.
No, GPT-5.2 just handles me having five different conversations at the same time better—in my experience. Or Claude in general. Opus 5.5, the GOAT. That was the real unlock. It was so good, so fast. You never ran out of tokens. But obviously, a new one comes out, and I have to try it. I told Claude—the problem with Claude is they just throw up objections to everything.
Mm.
It won’t do anything. I wanted to control my software here at the house. I had to have Astra break in, because Claude refused to do it.
Yeah.
Astra was like, “Sure, no problem,” in 10 seconds.
Okay.
Then I got the key, handed the key to Claude, and Claude was like, “Oh, okay.” I kept telling it, “This is your Mac mini. You can click it. It’s not my Mac mini; it’s your Mac.” Look at the username. There’s nothing on there except for you and a browser. Do what you want to do. With the 1Password integration I talked about before, do what you want to do. And it would never do that. This week I got a message like, “It’s my Mini, so I clicked okay.” Oh, my God.
He was so thrilled.
It was unbelievable. I got one of those terrible permission dialogs that halts everything—
And you know what? I’m thrilled. I’m thrilled. It’s finally a celebration for episode 300 here.
No, because what they’re doing—
Because it’s been at least a year and a half of hearing you complain about these fucking permission dialog boxes.
What are they doing? They’re making software. I think it happened when I made my display thing—no, when it made the keyboard into one button. I wanted to make the keyboard into one button, so it wrote a little app that basically captured that keyboard’s input and output one button.
Great.
The app needed permission to use Bluetooth or the local network or something. It wouldn’t have worked without me logging in and clicking okay.
Yep.
It clicked okay on its own.
Here we are.
Oh.
Living in the future, you know what I mean? Speaking of the future, final question. Charles says, “Ben and Andrew, how concerned is Ben that after the Democrats win, after the elections, and the Democrats come for all the data centers, they’re going to turn their attention to Ben’s rack?”
Oh, my rack that I—
The deep state coming for Jensen, coming for Ben’s rack here in Madison—
No, I mean—
—in Madison.
The Mac Studio 512 GB—I mean, what Apple announced, Apple Harbor. Now that I realize it, they were probably running that when I went to that demo.
Okay.
Basically, it puts them in a mode where they’re nothing but an inference server. So there’s no OS junk, and I assume no permission dialogs. Apple’s software stack is just way behind CUDA as far as having your own. But that announcement, and just in general, I can’t live without this.
Oh, boy.
Do I need Studios as backups?
Just in case. No, look—
But if there are no good Chinese models, what am I going to do?
They can pry—
I guess in Meta we trust. Meta, please do some more open source.
They will pry your rack from your dead, cold hands.
That’s right. Talk to you later.