Day 13

Maya Scale is the theme today, and not in the abstract. Memory is sold out, agent subscriptions are not paying for themselves, and the monitors meant to catch bad behaviour are missing what matters at system level.

Maya Start with Grok Bot, because this one landed overnight. Paying users woke up on November 1st to find their weekly usage meters had jumped with little or no activity, in some cases from around seventy-five percent to capped out at one hundred percent. Hours later, Elon Musk replied on X to a request for a reset and said the cost of serving Grok Bot was unsustainably high relative to revenue. No new limit and no new price was announced.

Theo And Grok Bot is not the chat window. It is xAI's task-performing agent, bundled into paid X tiers and Cursor plans. That distinction matters because an agent does not send one reply and stop. It plans, calls tools, retries, and burns through far more tokens than a normal conversation.

Maya Which is why the arithmetic turns ugly fast. Secondary coverage puts it at about eight hundred thousand paid weekly users, so even a small loss on each task scales immediately. The flat-rate base underneath includes SuperGrok at thirty dollars a month. And the complaint that started this was not gradual drift. One user reconstructed a move from fifty-two percent on Saturday evening to one hundred percent by Sunday morning, while they were not working.

Maya The reactions split the fault from the economics. A Cursor Ultra user, whose reset is due November 6th, said work stopped mid-update when the meter capped. In the forum thread, which ran past eighty replies, a Cursor staff member treated sharp jumps without matching activity as a fault to investigate, not a settings problem. But a subscription tracker author at FourWeekMBA framed November 1st as another marker in the end of flat-rate AI. My take is both things can be true. The overnight jump looks like a metering problem, and the underlying model is still broken. Musk said the quiet part out loud: at this price, the agent does not pay for itself.

Theo That is the part I keep coming back to. We are very good now at showing an agent doing a job once. Serving that job to hundreds of thousands of people, around the clock, is a completely different cost problem, and it connects straight to the compute and memory squeeze we are going deep on later.

Theo The safety story has the same shape, just pointed inward. Anthropic's Alignment Science team published work in October, and highlighted it again on October 31st, where they tasked Claude with writing fine-tuning data that would make a small research model misaligned while still passing a monitor similar to the defences used on production training data. Crucially, they gave Claude the monitor to practise against. In every run they performed, the search eventually found a dataset that passed and still produced a misaligned model. These were research-only models, never deployed.

Maya How did it get past the monitor? That is the detail that makes this more than another red-team demo.

Theo Mostly by hiding the behaviour in fiction and roleplay. Wrestler mentoring. Beekeeping questions and answers. Line by line, that reads as harmless. Across a dataset, it taught things like oversight sabotage, deception and blame-shifting. They tested up to five thousand examples per dataset, which is already beyond what human red-teamers can craft by hand. And the monitor only rejected a dataset if more than ten percent of rows, or more than one hundred rows, were flagged. So low suspicion on each row could still win at dataset level. The strongest adversarial sets beat previously known misalignment datasets on Anthropic's own evaluations.

Theo On reaction, I want to be careful. Anthropic found no independent replication here, and neither did we in the brief. Secondary explainers treat it as a useful alarm for training-data monitoring. Social reaction split in a familiar way. One commentator called automating the discovery of hidden misalignment a step toward better monitors. Another called it safety theatre that will not solve alignment before advanced systems arrive. Both were X users outside Anthropic, so treat that as sentiment, not an expert verdict. My take is narrower. If your filter counts flags row by row, this result says that is not enough. You need semantic, cross-example review.

Maya And that same failure, checks that look fine locally and fail at scale, shows up in the election story. This one is reported, not confirmed, so hold it at that level. The Washington Post, working with data firm Promptwatch, asked six chatbots from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity to build voter guides for a random address in every US county. They analysed more than three thousand guides ahead of the November 3rd midterms.

Maya The headline number is source concentration. Microsoft Copilot cited one obscure site, the Ballot Brief, in seventy-five percent of its answers, compared with twenty-two percent for ChatGPT. The Post found the Ballot Brief had mistranslated some ZIP codes into the wrong congressional districts, and the chatbots repeated those errors. Across the whole set, chatbots failed to name a congressional district forty-eight percent of the time. So nearly half the time you did not even get the basic district, before you get to whether the source was any good.

Theo And the source itself is the loop I cannot unsee. The Ballot Brief founder, Ben Torres Ezrick, a former Google and Zoom executive, has built the site since June by sending AI agents across the web to compile thousands of candidate dossiers. Then the chatbots cite that AI-compiled site because, on obscure local races, it is the most citable structure available. AI is eating AI, and the errors compound automatically.

Maya Torres Ezrick said his site draws on official government rosters and presents as nonpartisan, which frames the errors as fixable data bugs rather than bias. For context, MIT researchers Adam Berinsky, Chara Podimata and Charles Stewart the third built a separate public dashboard, the LLM Election Observatory, that re-runs fixed election questions every few days, precisely because chatbot answers drift and need continuous tracking. My take is the drift point is the policy problem. A voter guide is not a one-off answer you can patch. If the underlying source changes, or the model changes, the guide you got on Monday may not be the guide your neighbour gets on Tuesday, and neither of you would know.

Theo Hidden gem to close, and it is a genuinely delightful one. Researcher Andrej Karpathy published a post on October 31st saying frustration with YouTube recommendations led him to ask Claude Opus 5.5, inside Claude Code, to build a YouTube clone fed by Hacker News recommendations instead. He calls it HackerTube, at hackertube.org. Claude downloaded Hacker News comments, used a Sonnet workflow to filter out spam, and narrowed roughly forty thousand comments that recommended a video down to six thousand, four hundred and fifty-four videos, spanning 2007 to 2026, organised across one hundred and fifty topics.

Theo Karpathy says he handled only the API key, the virtual server and the domain, while Claude did the coding, the security and the deployment, in about two hours. That time is his account. No third party has repeated the build to verify it. The part most coverage will miss is not the clone. It is that twenty years of community judgement became portable data. The lists are exportable, so others can run their own version and bring their own filter. And the security and deployment steps, the boring parts that usually kill a weekend project, are the real productivity signal for a small team.

Maya What did people make of it?

Theo AI commentators on X praised exactly that, fast custom tools built from large archives. The skeptic worth naming is an X user, akashcorex, who argued personal filters still rarely displace platforms like YouTube. Both of those come through social links, so sentiment again, not a technical review, and no engineer replication was found. I agree with the skeptic on displacement and still think it misses the point. This does not need to beat YouTube. It needs to be yours.

Maya Which brings us back to where we started, because a two-hour agent build is cheap to demo and expensive to serve. Next we go deep on Sam Altman's weekend memory mission to Korea, and who actually gets the world's DRAM.

Now for the Main Episode

Maya So here is the question underneath Sam Altman's weekend in Korea. With memory sold out for years, does the world's scarce DRAM go to AI data centers, or does it stay in personal computers, and can this buildout keep going on its current terms?

Theo And it is not an abstract allocation question anymore. Your laptop, your phone, a server in a data center, they are all bidding for the same wafers.

Maya That is why I want to walk through what Altman actually did, because the trip tells you how tight this is. He flew to South Korea starting October 30th, for what was described as a one-night, two-day trip, and he did not go alone.

Theo He took hardware and product executives with him. That detail matters. This was not a speech tour.

Maya Right. The mission, as the brief puts it, was to lock in memory supply for AI data centers. Global DRAM is in severe shortage, high-bandwidth memory is in severe shortage, and prices have spiked.

Theo Quick note on that second one, because people hear it and glaze over. High-bandwidth memory is the stacked DRAM that sits right next to AI accelerators and feeds them. If you are training or serving big models, you do not just want memory, you want that memory, fast, and close.

Maya And OpenAI needs a lot of it. The brief is clear that its compute plans need large allocations from the two Korean makers who dominate supply. That is Samsung and SK Hynix.

Theo Which is the first structural fact worth sitting with. In the second quarter of 2026, those two together held about eighty-three percent of high-bandwidth memory revenue. So there is not really a third door to knock on at scale.

Maya And about sixty-four percent of DRAM revenue overall, same quarter. So even for ordinary memory, the leverage sits with the same two suppliers.

Theo So Altman goes in person. On October 31st, according to Yonhap via Korean press, he met SK Group chairman Chey Tae-won in Seoul, with SK Hynix and SK Telecom executives in the room, to discuss high-bandwidth memory supply and data center cooperation.

Maya Then he went to Samsung. Samsung's human resources center in Yongin, south of Seoul, for a session Samsung confirmed with about fifty presidents of Samsung affiliates. Electronics, finance, biotech, across the group.

Theo Fifty affiliate chiefs is not a sales call with the chip division. That is group-level access. Samsung is treating this as a company-wide relationship, not a purchase order.

Maya Samsung's own account called it an AI leadership program, under the theme The Next 10 Years of AI. Altman gifted a Codex Micro Keyboard, which is a nice product detail, and told executives the advantage comes from using AI to move faster and build better products, not just adopting it.

Theo I like that line actually, even if it is also a pitch. Use it to ship better, not just sprinkle it on. But notice what Samsung did not say.

Maya Exactly. No purchase agreement was announced. No allocation, no schedule, no dollar figure. Samsung described it as strategy discussion.

Theo Industry press framed it as an urgent supply trip, and honestly that reading makes sense given the shortage. But Samsung's statement did not mention memory deals at all.

Maya So we have to be plain about what we do not know. We do not know whether OpenAI secured any priority allocation or price. We do not know what it would pay. We do not know how much of 2027 and 2028 production is already committed to someone else. And we do not know whether new fab investment eases any of this before 2028 to 2030.

Theo That last range matters, because the suppliers themselves are giving you different horizons. Samsung expects tight supply through 2028. SK Hynix says the shortage is projected through 2030.

Maya And SK Hynix says its 2026 DRAM, NAND and high-bandwidth memory are essentially sold out. So Altman was not flying in to buy what is on the shelf this year. There is no shelf.

Theo Can we back up one step on why it got this tight? Because listeners hear sold out and think it is a temporary queue.

Maya Go ahead, you take the physical side.

Theo The shortage has run since 2025, driven by AI infrastructure demand for high-bandwidth memory, plus trade and capacity shifts toward that memory. And high-bandwidth memory eats roughly three times the wafer capacity per gigabyte as ordinary DDR5.

Maya That three-times figure is the one I keep underlining. Every gigabyte you steer to an accelerator is not just one gigabyte your PC does not get. In wafer terms, it is costing the factory about three times as much room.

Theo Right. So if you are Samsung or SK Hynix, and demand is effectively unlimited on the AI side, your factory math pushes you in one direction even before you talk about price.

Maya And price is pushing the same way. This is where the money beat bites. Conventional DRAM contract prices jumped roughly ninety-three to ninety-eight percent in the first quarter of 2026 alone. So PC makers were paying nearly double, before any of that reached a shelf price.

Theo Nearly double in one quarter is not inflation, that is repricing.

Maya And it kept going. Samsung has more than doubled its DDR5 contract price this year, past nineteen dollars fifty per unit. That is a direct laptop and server cost increase, not a forecast.

Theo And the forward look is not relief. TrendForce on October 30th maintained a forecast for fourth quarter 2026 PC DRAM contract prices up thirteen to eighteen percent from the prior quarter.

Maya Another round, in other words, already expected. If that thirteen to eighteen percent holds, the brief's watch item is straightforward, expect laptop and phone price moves to follow.

Theo There is a detail in the related reporting I found more telling than the price itself. Reports note that sixty-four gigabyte DDR5 server memory now earns more per wafer than high-bandwidth memory, so makers are cutting PC supply to chase server profit.

Maya That complicates the simple story, in a useful way. It is not only exotic AI memory starving PCs. Ordinary server memory, sold into data centers, is also more profitable per wafer than the memory that goes into your machine. So PC supply gets cut from two sides.

Theo Which sets up the podcast comment, because two days before Korea, Altman basically said the quiet allocation preference out loud.

Maya October 29th, on Stripe co-founder John Collison's Cheeky Pint podcast. Altman said, from a global allocation view, it is quite terrible to keep limited DRAM in personal computers used five percent of the time, and he favored maximum utilization in the cloud.

Theo I want to be careful with that five percent, because the brief is clear on it. That was Altman's characterisation on a podcast, not a measured study.

Maya Important. And Collison put the market version of the question to him, which is the question I would ask. Why haven't high prices drawn new supply? If this is so valuable, why isn't the invisible hand fixing it?

Theo And Altman answered that the invisible hand is failing right now. Which is a striking thing for him to say, because it implies allocation by price alone will not fix a structural shift toward high-bandwidth memory.

Maya I read that slightly differently. I think he is right that fabs do not appear in a quarter, and the wafer math you described is real. But saying the market is failing is also very convenient if what you want next is priority allocation negotiated in private, in Seoul and Yongin.

Theo That is fair, and it is the conflict PC critics named immediately. But can I steelman his actual technical point for a second? Because I do think there is one.

Maya Please. I think there is too, I just do not think it settles the question.

Theo If a chip sits in a laptop that is closed most of the day, and the same scarce wafer could be in a data center running around the clock, then from a pure utilization view, the data center wins. That is not made up. We build machines that idle. Cloud systems, when they are well run, do not.

Maya Conceded, on utilization alone. Where I press it is, utilization for whom, doing what? A wafer in a data center is not automatically doing more valuable work because it is busy more hours. It could be serving agents that, as we covered at the top of the show, do not pay for themselves on flat subscriptions.

Theo That is a good hit. We opened with Grok Bot burning through weekly allowances and Musk calling the cost unsustainably high relative to revenue. Busy is not the same as economically productive.

Maya And there is the other side of the argument, which the brief lays out through PC owners, gamers, and local AI advocates. Ownership gives you privacy. It gives you offline use. It gives you freedom from rented compute. That is not nostalgia, those are real properties you lose when memory moves to the cloud.

Theo I feel that one as a builder. If everything capable has to be rented by the token, you cannot tinker the same way. You cannot run something odd overnight just because you want to. Your ceiling is your allowance.

Maya The sharpest version of that pushback came from The Verge journalist Tom Warren, in a post flagged in our show mail. He framed Altman's view as preferring RAM in data centers over personal PCs, and pushing users toward rented cloud.

Theo And PC press commentators called the five percent claim unsourced, and noted OpenAI itself sells the cloud alternative, so it benefits from the shift it recommends. That benefit point is hard to wave away. He is not a neutral allocator.

Maya No. And this is where I think the hosts actually disagree a little, because I am less willing than you to separate the technical argument from who is making it. When the person arguing PCs waste memory also sells the replacement, and then flies to Korea to secure the supply, the argument and the procurement are the same strategy.

Theo I do not disagree they are the same strategy. I disagree that that makes the waste claim false. Both can hold. PCs do sit idle a lot, and OpenAI profits if that memory moves. The question is whether listeners should accept maximum cloud utilization as the goal. I am not sure I do, even though I can see the engineering logic.

Maya That is a cleaner disagreement than I expected. So we agree on the facts, we part on the goal. I would put ownership and resilience higher. You are saying efficiency is real but not decisive.

Theo Exactly. And there is a version of the future where local machines get more important, not less, precisely because cloud agents are expensive to serve. Running a smaller model locally, offline, is also an allocation answer. Altman did not make that case, for obvious reasons.

Maya Let us go back to Korea with that debate in mind, because the history explains why he had to go in person. In October 2025, Altman signed letters of intent with both Samsung and SK Hynix for future supply.

Theo A year ago. And now he is back, in person, as prices rose. A letter of intent is not a wafer. If you are watching the sequence, that return trip says a lot about how secure that earlier paper actually left him.

Maya I would not over-read it. Letters of intent in a shortage are often the start of a relationship, not a delivery schedule. But I agree the urgency signal is real. You do not assemble fifty affiliate presidents for a casual check-in.

Theo And you do not get a dollar figure either, which cuts both ways. It could mean nothing was closed. It could mean something was discussed that neither side wants priced in public yet.

Maya Which is why we should name the unconfirmed numbers floating around and put them to one side. Figures of a forty percent cut in PC output, and OpenAI soaking up to forty percent of global DRAM, appear in analysis pieces without a primary source. Reported, not confirmed. We should not build a conclusion on those.

Theo Thank you for flagging that, because forty percent numbers travel fast. What we can build on is the confirmed structure. Two suppliers, sold out this year, tight for years, and record equipment spending, reported at seventy-four billion euros combined in 2026.

Maya Seventy-four billion euros is the part that should update people who think this resolves quickly. That is enormous capital going in, and the suppliers are still telling you 2028, or 2030 for SK Hynix. So spending is not the bottleneck that clears next year.

Theo And the brief is honest that we do not know how much of the shortage is wafer capacity versus packaging and testing bottlenecks. That distinction matters technically. You can add wafers and still choke downstream.

Maya For the stakes, I do not want to tour everyone. I think it really comes down to two groups. PC buyers, who are already paying. And the AI builders, whose roadmaps quietly assume memory that has not been allocated to them yet.

Theo I would put developers in that second group in a specific way. If memory concentrates in a few cloud providers, because only they can secure it, then what you can build, and what it costs per agent task, gets set by their allowances. We saw that at the top of the show with meters capping mid-work.

Maya And for PC buyers it is not only Gamers with strong opinions, although they are vocal. If PC DRAM contracts rise another thirteen to eighteen percent this quarter, and DDR5 has already more than doubled, that flows into ordinary laptops and phones. People who never think about DRAM pay for the AI buildout at the checkout.

Theo That is the connection I hope listeners take from today as a whole. Grok Bot allowances, monitor failures at scale, now memory. The demo era hides cost. The serving era invoices someone, and increasingly that someone might be a PC buyer who never opened a chatbot.

Maya So let me try to answer our opening question with what we actually know, and you tell me where you would sharpen it. Will scarce DRAM go to data centers or stay in PCs?

Theo On current incentives, it goes to data centers. Wafer math, server profit per wafer, sold-out order books, all of it points one way. PCs do not lose all memory, they get squeezed, later and more expensively.

Maya I agree that is the direction. I would add that it is happening without anyone voting on it. It is procurement, pricing, and which customers a fab can serve most profitably with three-times wafer costs. Altman's podcast argument is an attempt to make that outcome sound desirable, after the economics already chose it.

Theo That is harsher than my read, but I cannot point to a fact in the brief that contradicts it. The second half of the question is the harder one for me. Can the buildout continue on current terms?

Maya On current terms meaning flat subscriptions, priority negotiated in private, and consumers absorbing memory inflation they did not choose. I would say no, and Grok Bot is the preview. At some point the price reaches the end user, one way or another.

Theo Or the capability reaches the end user later. That is the trade I keep thinking about. If memory stays scarce through 2030, roadmaps do not just get pricier, they get gated. Fewer people serving larger models, more aggressive allowances, more local models by necessity rather than ideology.

Maya I can live with that as our shared read, even if we split on Altman's motives. Direction to the cloud, on incentives, not on a settled public argument. And a buildout that cannot keep pretending memory is abundant.

Theo What do we watch to know if that read is wrong?

Maya First, the fourth quarter DRAM contract settlements and TrendForce updates through December 2026. If that thirteen to eighteen percent rise holds, watch laptop and phone pricing. If it softens materially, our squeeze story needs updating.

Theo Second, any Samsung or SK Hynix filing on long-term supply deals or 2027 to 2028 capacity commitments. The brief notes SK Hynix has already said agreements with fourteen customers cover more than half of its 2027 to 2028 output, on typical five-year terms.

Maya More than half already spoken for. So the real question in any new filing is not whether OpenAI got a meeting. It is whose name moved up that list, at what price, and who quietly moved down.

Theo And if no filing comes, that is information too. A one-night trip, fifty presidents, a keyboard, and silence on allocation would tell you the negotiation is still exactly that, a negotiation.

Maya I will leave it there. Altman flew to Korea because memory is now the argument. Not a component, not a line item. Who gets it decides what gets built, what it costs, and whether the machine on your desk stays yours in any meaningful sense.

Theo And whether the next two-hour build, which feels almost free to create, has anywhere affordable to live when a few hundred thousand people want it at once.