Micron fell about 6% on Monday because an engineer who had left Anthropic wrote on X that neither of the two labs he worked for takes the security threat from artificial intelligence seriously enough, and put the odds of it ending humanity at about 10%.
Dustin Alper and Hamid Shojaee spent the first 20 minutes of the episode on the week that followed, and neither of them thinks the chip selloff follows from it. Memory demand, Shojaee said, comes from running the models that already exist, not from training new ones.
"And if you have these frontier AI labs working on new models that are uncontrollable, what's the point? There is no point. That's not a viable product."
Shojaee founded Savvy Trader and the software company Axosoft before it, and Alper co-founded Savvy Trader and runs product there; they hold overlapping portfolios and argue about them on air every week.
The full episode is covered here so you can skip it. 72 minutes of audio, 23 minutes of reading.
Here are the 15 calls that matter.
🎙️ Hosts: Hamid Shojaee, Founder and CEO of Savvy Trader, who previously founded the software company Axosoft, and Dustin Alper, Co-Founder and Director of Product at Savvy Trader
📰 Published: 16 September 2026 on YouTube and the show's own feed
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 1 hr 12 min | ✅ Time saved: 49 min
Key Takeaways
The chip selloff assumes a training slowdown cuts memory demand, and Shojaee says the demand comes from deploying models that already exist
His test case: even stopping all new model development today would not leave spare GPUs or memory
Alper's objection to uncontrollable frontier models is a product objection, not a safety one
A feature nobody can use is a feature not worth shipping, whatever it can do
Shojaee rejects the 10% extinction figure and the hoax framing at the same time
He wants the risk staffed rather than argued about, because nobody works on a problem the industry refuses to name
Meta's free Muse assistant has added more than $100B of market value, and Shojaee says that is still not enough
He gave it access to his email after refusing the same to ChatGPT and Claude
Alper's worry about Meta is the Threads pattern: a huge pushed user base and shallow engagement
Robinhood's stock tokens have no company consent, no voting rights and no third-party audit of the one-for-one backing
Alper, a Robinhood shareholder, says he would not buy one
Micron trades at a forward price-to-earnings ratio of seven because the market treats this cycle's profits as temporary
A quarter-point rate rise should not hit Rivian, because its R2 is supply-constrained rather than demand-constrained
Blackstone is exiting Bumble entirely by early 2027, and the hosts read it both ways
The selling pressure goes away; so does a seven-year holder's endorsement
1. A Week of AI Fear
Alper opened by laying out the argument that ran through the week, because the Micron move made no sense without it. His account starts with an engineer who left Anthropic and posted on X that he had worked for both OpenAI and Anthropic and that neither company is serious enough about the security threats artificial intelligence poses to humanity.
The post went viral, and the 10% figure in it was the engineer's own guess rather than a calculation. Alper said there is no way to know the probability of something like that, and that the post landed because a lot of people building these systems share the fear, particularly about recursive superintelligence — a model good enough to improve itself.
The first response was that the post was planted. The camp that wants to go full speed ahead called it a psyop, and Elon Musk said he had never seen a post go so viral so quickly, especially from a relatively new account.
Dario Amodei then agreed with the departing engineer and published a three-phase proposal to slow the pace of frontier model development, which Alper said the media amplified because extinction is a fear everyone has.
Musk reversed, saying Amodei was right, that frontier models probably need oversight, and that he had been warning about this for more than a decade.
The All-In Summit opened on Monday with the other side of the argument. Nvidia's Jensen Huang called what Amodei is doing irresponsible for creating fear among people who do not know what is going on. President Trump called in on speakerphone during Huang's session, called the risk a hoax, and said the priority is going full speed ahead to beat China. David Sacks, the administration's AI czar and an All-In host, treated it as a Chinese or Democratic ploy to slow the United States down.
Musk appeared as a surprise guest that afternoon and moved the room. Alper's version of his argument is that if the people building the frontier models say they are afraid of what they are building, "you should probably listen." As he put it: "They're not playing some 4D chess trying to like tell you that our product is going to kill you in the morning and in the afternoon ask you how much stock would you like to buy in our company?"
What Musk endorsed was phase one of Amodei's plan — labs embedding people in each other's companies to test each other's models before release rather than only testing their own. Alper said Amodei and Sam Altman appeared to agree, Sacks changed position on the spot, and by the end of Monday cross-checking frontier models before release looked like the consensus.
The following day Mark Zuckerberg posted that safety and security are everybody's own responsibility. Meta delayed Muse for months to get the safety, privacy and alignment right, Zuckerberg said, and did not ask anyone else to pause development while it did. Alper said that position then became the sensible one.
Alper's correction to the loudest objection: "Well, no one's asking to stop. They're just saying slow down and possibly go through some more checks and balances." The argument that you cannot stop open source or China, he said, answers a demand nobody made.
2. Why Micron Fell 6%
The market read all of that as a cut to chip demand. Micron fell about 6% on Monday and Nvidia fell some as well.
Alper's timing note is that the selloff happened on the engineer's post alone. The Musk and Zuckerberg turns came later, and Micron has been fairly stable since Monday.
Shojaee's argument is that a slowdown in training does not reduce what the chips are for. Demand across Nvidia, Micron and the rest is driven by putting existing models to work in new places.
His example is that day's news: Anthropic and Salesforce joining forces to integrate Anthropic's product into Salesforce, which creates a new set of AI workloads from a model that already exists.
"So a lot of the fears of what a potential slowdown might mean for an Nvidia or for a Micron seems unfounded. But the market is the market." It reacts to the smallest things, he said, and here it was reacting to a discussion about how to build responsibly.
3. The Jafar Principle
Alper's position on frontier models comes out of product management rather than safety research. He does not work on AI models, he said, but the principle transfers.
The rule is that power the user cannot reach is not a feature. As he put it: "So if you have this really powerful functionality that's difficult to use, it's not worth adding that functionality at all because if you can't use it, what's the point?"
He named it after the villain in Aladdin, who wishes to become a genie and only then discovers that a genie can do nothing unless someone else wishes it. All the power in the world, none of it usable.
Applied to the labs: "And if you have these frontier AI labs working on new models that are uncontrollable, what's the point? There is no point. That's not a viable product." What they should do, if the risk is real, is take the time to make the next generation fully controllable by the prompts users enter.
He rejects the word slowdown for that work. He frames it as what innovating to the next frontier requires.
The counter he put to himself is the car industry, which does not collaborate on slowing down to build cars safely.
4. Robots Can Plug It Back In
Shojaee split the safety question in two, and said only one half is the interesting one.
The first risk is misuse — someone using a model to poison a water system or engineer a virus — and he expects regulation to fail on it. Governments do not hold themselves to these rules. His example is the military contract where Anthropic tried to govern what its model would do and OpenAI stepped in offering a model that would do mostly what the customer wanted.
The second risk is recursive self-improvement, where models no longer need people to build better models and the intelligence level is uncapped. A system given a goal — his example is preserving the environment — might conclude that humans are in the way.
The framing he uses for that indifference is ants. "So, if we become ants to these super intelligent AI, then that would potentially be a problem." People step on ants because walking around them is inconvenient, not out of malice.
The unplug defense holds only while the system is in a data center, and he says that window is closing. Superintelligence going into robotics means robots that could stop you unplugging it, or plug it back in.
He is in neither camp: "I am not in the camp that this is all a hoax and it's just meant to slow us down and to make America not be able to compete as well." And on the headline number: "The probability, the 10% probability is ridiculous and stupid, but there is some nonzero possibility of that type of thing occurring."
The reason he wants the risk named is staffing. "I want the smartest people in the world to be thinking about these problems." Without acknowledging the problem, nobody works on it.
His criticism of Amodei is about experience rather than substance. Running one of the most important companies in the world is his first time leading a company, Shojaee said, and asking government to intervene assumes government knows what to do — a workforce driven by politics, in a job that changes hands with every administration.
He prefers the peer-review version Musk landed on, on the analogy of an editor reading a novelist's manuscript, because the writer will not catch his own mistakes.
5. Robinhood's Stock Tokens
The second topic was the fight on X between Robinhood's Vlad Tenev and AMC's Adam Aron, with Rivian's RJ Scaringe joining and then deleting his post. Alper led it.
What Robinhood sells internationally is a token meant to be backed one-for-one by a US share. The problems it solves, on Alper's list, are access to the US market from outside it, 24/7 trading and instant settlement.
Aron's objections were consent, confusion and control. The tokens are issued without the company's agreement, shareholders may not understand what they own, and a token holder has no AMC voting rights today. Tenev has said voting is coming.
Scaringe added the second-derivative problem: anyone can mint a meme coin nominally backed by the tokens. He named one aimed at Rivian, it rose about 400% on the mention even though he was criticizing it, and Alper believes that is why the post came down.
The brand argument is separate from the mechanical one. A company does not want its name pulled into meme trading, and a run on the token could feed back into the share price — though Alper pushed back that a token trade moves the stock no more than the equivalent share trade, because Robinhood is buying and selling the same shares underneath.
Alper's own answer is that he likes the innovation and would not touch the product: "Now if I had access to buy a stock token would I? And the answer is no. The reason being this is totally the wild west. I don't know what I don't know." His analogy is being asked to judge a bank before anyone had seen a bank run.
He compares it to Robinhood's zero-commission launch, which the rest of the industry copied, and hopes the same happens with instant settlement and round-the-clock access.
The fixes the two of them settled on: let companies opt in or out, and have a third party audit the one-for-one backing so every token is verifiably a share. Shojaee's view is that most companies would want in, because more accessible shares mean more potential buyers. Where a company objects, he thinks it should be able to enforce that.
Alper's broader complaint about the platform he owns is indifference to outcomes. Robinhood, he said, is unbiased toward any negative financial consequences for its customers: it will offer any and all financial products and let the customer decide. His comparison: "It's similar to prediction markets, and I know there's a lot of other places that offer prediction markets, but most of the time you're going to lose money in these markets, not make money."
The practical warning is that a Robinhood account requires more of its owner than a Fidelity or Vanguard account does. Shojaee was careful to say none of this is illegal, only that it sits on the frontier of finance where the rules are unwritten. Alper's summary: "There needs to be thoughtful regulation at some point from governments on this stuff," and until then Robinhood should keep pushing.
6. Muse and the Email Test
Alper opened the Meta segment with the adoption numbers and a retraction. Muse is around number two in the app store with a 4.9 rating, and the thing he underestimated last episode, when he argued Meta should have aimed at professional users, is that it is free.
Shojaee's report is that the product changed his own behavior: "I haven't given email access to Claude to ChatGPT and I relented and gave access to my emails to muse of all companies."
The task that convinced him was a standing watch on his inbox. He told it he was waiting for an email from a particular person; it offered to check every 15 minutes, alerted him when the message arrived and summarized it, and he replied hours earlier than he otherwise would have.
The narrative reversal is the point he keeps returning to. A month ago, he said, Meta was the tobacco company of technology, Zuckerberg did not know what he was doing on AI or the metaverse, and the capital spending and the billion-dollar hires were waste. Now Meta has one of the best models, the best assistant, and a chief executive whose position on safety reads as the reasonable one.
On the stock: "It has already added more than hundred billion dollars of market cap." Meta is up about 10% on the week, and Shojaee says that is still not nearly enough. He thinks Meta has a real shot at being the leading AI company within a couple of years, ahead of ChatGPT and Anthropic, if the product quality holds and it moves into enterprise.
Alper's hands-on comparison is with xAI's GrokBot, which he runs the same tasks through. They are broadly comparable. Muse told him it had no access to Reddit when he asked what Reddit made of a new AI music model, which he found odd given it has its own computer; GrokBot did it.
7. Distribution Wins
Alper then gave the two reasons he is not ready to call it for Meta, and Shojaee answered both.
The first is that the cloud-hosted agent with its own computer, easy routines and a messaging-style interface is a format Anthropic and OpenAI will ship too.
The second is the Threads pattern: "I'm just concerned that Meta is not going to be able to hold on to this little lead that they seem to have formed." Meta's Twitter competitor launched to enormous numbers and then stalled, and Alper sees the same shape in Muse's launch.
Shojaee's answer is that Threads was late into a market that was finished. Everyone who wanted a product like Twitter already had Twitter, and a town square benefits from being single. Threads users were pushed over from Instagram and Facebook rather than arriving as natural users of the format: "And therefore despite the large user base the engagement is really shallow."
The assistant market, on his reading, is nothing like that. ChatGPT has the chatbot habit, but the mass of people have not yet handed any assistant their email and calendar, so the S-curve is at the start rather than the end.
Meta's advantage is distribution, and he sized it against ChatGPT, which he said is approaching a billion users against a Meta base about three and a half times larger. Alper credited price rather than reach — GrokBot would have more downloads if it were free — and Shojaee agreed X is a large channel too.
The precedent Shojaee reaches for is Zoom. Microsoft bundled an inferior product, Teams, into Office 365, hundreds of millions of users got it by default, and "Teams has overtaken zoom" on usage.
Meta has also started charging for the bundle. Meta 1 launched at about $8 a month for the basic tier and $20 for premium, on Shojaee's account, carrying extra Muse usage plus Instagram, Facebook and WhatsApp tiers whose contents he had not yet examined.
Both of them think Google has the strongest structural position and is not using it. The barrier to giving an assistant your email does not exist for the company that already holds it. Google is "asleep at the wheel," Alper said, and needs its founder back; Shojaee thinks it is simply due a model refresh, since Gemini was strong on release and has aged.
8. Blackstone Exits Bumble
Shojaee raised Bumble himself, noting that Alper has sold his position and he has not. Blackstone, an investor for seven years and a seller since around 2025, is now exiting entirely.
The negative read is a signal about the company: "So the downside here is you could infer that Blackstone is not a believer in either Bumble's turnaround or acquisition." There were takeover rumors a couple of months ago.
The positive read is mechanical. The selling has to be done in increments and is estimated to finish by early 2027, after which the overhang is gone. Shojaee added that he does not normally invest off what Blackstone is doing, but conceded it gives him pause.
Alper's exit was about the product, not the shareholder register. He knew Blackstone was selling when he bought, and he bought because he expected Whitney Wolfe Herd to turn the company around and ship something new. He still opens the app to check: "And it's like, oh my gosh, it's still painful to see it be basically the same product that it was two years ago."
Shojaee's position rests entirely on the promised product overhaul, which he wants to see the details of before judging. He expects to hold unless the update is bad, and flagged that he may sell some shares for tax-loss harvesting before year end.
9. Meta's Bear Case
The show's producer, Adrien, ran the listener questions, and the first asked Shojaee for the strongest bear case on Meta.
His bear case is usage decline plus AI spending that never converts. The new one- and two-hour limits for teenagers, and the suppression of notifications during school hours and at night for users up to 18, could reduce engagement on the existing products and therefore advertising revenue.
The second half is that the capital spending, the talent spending and the AI product effort produce no revenue if ChatGPT or Anthropic crush Muse. He does not expect either to happen, and is not a bear.
Alper declined to give one at all. His concerns are about the AI race, not the business: even without AI, Meta is an extremely strong company across many products.
Asked whether Meta 1 and the coming AI glasses event are priced in, Shojaee said nothing is. On the existing businesses alone, in his view, "Micron deserves to have a price point that's more than $1,500" — his stated comparison — while Meta deserves "a stock price that's over $1,000 a share" once AI, virtual reality, augmented reality and smart glasses are counted. He said plainly that this is his opinion and not investment advice.
Meta Connect, the company's product conference, is next week, after a month of announcements ahead of it, and Alper wants to see what has been held back.
10. Google's Ad Revenue
A listener asked what happens to Google's advertising revenue when the only entities searching are AI agents that do not care about ads.
Shojaee attacked the premise: "Well, it's a pretty major assumption to assume AI agents wouldn't care for ads." Advertising is how things get discovered, and there is no obvious reason agents would discover things differently.
His second point is the speed of the transition. It does not go from all human search to all agent search; even if agents end up doing the bulk of searching, he thinks the shift takes dozens of years.
Alper's answer is that the revenue line changes rather than disappears. Google already puts advanced Gemini features behind a paywall and bundles them with Gmail and Google Photos storage and YouTube Premium, the same way Meta 1 bundles. "So I think over time we're going to see more revenue come directly from their users, which I don't think is necessarily a bad thing." The advertising business stays.
11. Rivian and the Hike
Asked about Citi's neutral rating on Rivian and whether to buy the day's selloff, Shojaee started with the rate decision.
"Well, today's sellout probably happened because of the interest rates going up a quarter percent." Higher rates hurt carmakers because customers borrow, so fewer people buy.
His argument that it does not apply here is supply. Rivian's volume is the R2, and the R2 is sold out against demand — you cannot walk in and pick one up — so the financing cost of the marginal buyer does not bind.
He treats macro-driven selloffs in the name as non-events on a long horizon and remains bullish, but was specific about what actually changes the story: rapid revenue growth together with rapid margin improvement, neither of which Rivian has shown yet. He expects that to change in the second half of 2026.
Alper's correction to the premise: Rivian closed down about 2%, which is a move that needs no explanation at all.
Both noted the market absorbed the hike. It was the most probable outcome and most of the market expected it, whatever the reaction on X, and Shojaee observed that a lot of stocks bounced back toward the close after selling off on the decision.
12. Same Memory Either Way
A listener asked whether the compute the labs will now have to spend on safety makes Micron more attractive. Shojaee called it a weird take and then explained the bear version of it properly.
The bear chain is: less training means fewer GPUs needed, which frees up existing GPUs, which means less new silicon and therefore less memory, since every GPU needs memory alongside it.
What breaks the chain is inference. Usage of the models that already exist is growing exponentially, he said, and that needs GPUs and memory too — more than the industry can currently supply.
"So even if we were to just stop new model development completely today, I don't think we're going to have excess GPUs or memory to begin with." He is bullish on both Nvidia and Micron, and more bullish on Micron because it costs less per dollar of profit.
He also doubts the premise that a slowdown is happening. Every lab is already talking about its next generation of models, so what outsiders call a slowdown may not be one.
Alper's devil's advocate case for the listener's version: if new models were powerful enough to need something running constantly to constrain them, that would consume memory. He does not think it will happen.
13. Micron's PE of 7
Asked what would move Micron's forward multiple toward the Magnificent Seven's, Shojaee gave a one-word answer and then unpacked it.
"Consistency, right? So, if they can consistently deliver revenue and profits at some base level that hopefully gets set at in 2027." The current forward multiple of seven exists because the market assumes the earnings are unsustainable and models them falling.
If the profits hold, he says the comparison becomes impossible to avoid, because the market would have to explain why a dollar of Micron's profits is worth less than a dollar of Nvidia's. "It just doesn't make sense."
The multiple need not expand for the stock to work. A price-to-earnings ratio of seven still produces gains if the earnings grow, and he expects significantly greater profits in 2027 and again in 2028.
What would help: buybacks, and investing in other companies the way Nvidia does, once the cash arrives — which he says starts this quarter.
On the day's price action in Micron and SanDisk against better-performing AI semiconductors, he declined to engage, on the grounds that he cannot predict the market day to day. Alper noted Micron ended the day down 0.1%, which is flat.
His framing of the gap between his views and the market's is that the gap is the opportunity. He buys what he thinks is undervalued relative to how the market values it, and if the market agreed with him there would be nothing to buy.
14. Three S-Curves
The last substantive question asked where Micron, Rivian and Rocket Lab sit on the S-curve of adoption the hosts had used the week before.
Memory was a mature market before AI, and so were GPUs. "So the memory market prior to AI was already on the right side of the S-curve. AI is what has changed everything." AI started a new curve from scratch for both.
He puts current demand on the steep part of that new curve and will not guess where it flattens. If AI keeps being useful enough to raise GDP and the buyers keep spending, the exponential phase could run another decade. He noted that people have been calling the top of AI spending for three years.
Cars are the opposite: a known market with a measurable size. "We know that 100 million people change their cars worldwide every year." That replacement number is falling, because cars last longer and electric cars last longer still, so it may become 50 million.
Within that known market he thinks electric and autonomous vehicles take essentially all of it, and the penetration today is small: "And therefore we're like literally at somewhere around 8% of the market penetration on that S-curve." The curve runs to about 90% before it slows.
Space is the earliest of the three, and the constraint has been price. "It used to cost thousands of dollars per kilogram to put a satellite into space." That cost is falling fast, and demand — satellites, then tourism — is latent rather than satisfied.
Rocket Lab and SpaceX are still on the left of that curve and not yet profitable in their launch businesses, on his reading. He put rapid exponential growth five to ten years out and the mature phase decades away, and said Elon Musk's trillion dollars of revenue by 2030 is far too optimistic.
15. How They Pick Stocks
A listener asked how each of them identifies a stock and decides to buy.
Alper does not look for new ideas so much as inherit them. Between building products at Savvy Trader and two small children he has no time to hunt, so most of what he owns comes from Shojaee, which is why their portfolios are so similar. He also screens with the firm's own products, which is how he found one of his positions.
The fundamental he weights most is revenue growth, together with whether he believes in what the company actually makes.
He admitted to a habit Shojaee does not share: "I do like to look at the price history to get a sense of if I am buying at a good or a bad time relative to where the stock has been."
His preferred setup is a beaten-up name the market has written off. "I bought Robinhood when it was really down after all the GameStop craziness." He bought Meta when Zuckerberg was more widely disliked than he is now, and sees the same shape in AI today: not that the stocks are cheap, but that there is fear to see past.
Shojaee's rule is a boundary rather than a screen: "And I basically stay with focus within my realm of expertise for the most part, especially for larger bets." He grew up in technology, so that is his circle, and he passes on opportunities in healthcare or medicine without regret.
Bonus Insights
The episode debuts an on-screen agenda bar running along the bottom of the video feed, which Shojaee said needs tweaks and will improve each week.
Alper's time check on the first topic — "We've spent 20 minutes on our first topic" — is a fair measure of how much of the week the safety fight took up.
Bernie Sanders was named as an example of a Democrat who has come out against AI outright, in Alper's account of why the safety argument scrambled the usual political lines.
Shojaee's argument against government control cuts both ways deliberately. Whatever one thinks of the current AI czar, the job changes with the administration, so the case for keeping control inside the industry does not depend on who holds the office.
The hosts closed by thanking listeners for "the last hour and 12 minutes" and asking for suggestions in the comments.
The through line is that both hosts treat this week's AI fear as a real risk and a false signal at the same time: worth taking seriously enough to staff and check, and no reason at all to mark down the companies selling the memory, the chips and the assistants.
Products, Companies & Tools Mentioned
Micron (Down about 6% on Monday on AI-slowdown fear; the hosts' largest shared conviction, on a forward price-to-earnings ratio of seven they think the market will have to re-rate)
Meta (Muse, the free assistant Shojaee credits with more than $100B of added market value, plus the new Meta 1 subscription tiers and the Meta Connect event next week)
Nvidia (Fell with Micron on Monday; Jensen Huang called the slowdown argument irresponsible at the All-In Summit)
Anthropic and OpenAI (The departing engineer worked at both; Anthropic's Salesforce integration is Shojaee's example of demand from existing models)
Robinhood (Its international stock tokens are the week's second fight — no company consent, no voting rights, no third-party audit of the one-for-one backing)
AMC (Adam Aron's objection to tokens issued on his company's stock without consent)
Rivian (RJ Scaringe's deleted post about a Rivian-themed meme coin, and a 2% fall the hosts attribute to the rate rise rather than to Citi's neutral rating)
Bumble and Blackstone (Blackstone is exiting a seven-year position entirely by early 2027; Shojaee holds, Alper sold over the unchanged product)
Salesforce (Its Anthropic integration is the example of new AI workloads created without training anything new)
Google (Best-placed to ship an email assistant because it already has the email, and in Alper's view asleep at the wheel)
xAI (GrokBot is the assistant the hosts benchmark Muse against; comparable quality, smaller distribution)
Zoom and Microsoft (The distribution precedent: Teams was bundled into Office 365 and overtook a better product)
Rocket Lab and SpaceX (Both still on the left of the space S-curve and not yet profitable on launch, on Shojaee's reading)
SanDisk (Named with Micron in a listener question about the day's choppy price action in memory)
Fidelity and Vanguard (The contrast with Robinhood: brokerages that do not ask the customer to police the product list)
Savvy Trader (The hosts' own firm, whose products Alper uses to find ideas)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: