Intro
Ranjan Roy of Margins joins the Friday edition to take apart the Anthropic-Salesforce partnership and the software rebound behind it, Reuters' account of Meta's failed plan to run its teams on AI agents, and the leveraged retail blowup in South Korea's stock market.
Guest: Ranjan Roy, writer of Margins, who works in enterprise AI at Writer
Host: Alex Kantrowitz (Big Technology)
Published: 28 August 2026 on Big Technology Podcast · 59 min
Key Takeaways
The Claudeforce announcement is marketing, not product
Roy: "you could already do this in Salesforce and Claude" — his team at Writer already connects to Salesforce and runs queries against it
Roy: "God bless Mark Benioff, because that man can market like no other"
Amodei's friendly face is a pre-IPO decision, not a strategy change
Roy thinks investors told him the Death Star framing was hurting the offering
The partnership makes sense for neither side on its own logic
Anthropic needs to eat Salesforce to justify the valuation it is going out at; Salesforce is handing away the interface
Whose data it is decides how much the governed data layer is worth
Roy: "But in the end, it's the customer's data"
Capability moving fast does not mean adoption moves fast
Kantrowitz: "there's a real gap between the rise in capabilities and how they get implemented"
The software round trip is complete and the stock is roughly where it started the year
Roy: "it is kind of a perfect round trip of software is dead"
Meta's AI-native reorg failed on the technology, not the idea
Roy: "this is to me exactly what companies should be testing and doing and figuring out what the new operating model is and that teams should look different"
Measuring people on AI usage is how you get more code and less product
Kantrowitz: "You write more code, you're not necessarily more productive, you ship more, and when you ship more, you have bigger security flaws"
Cost cutting at a company that pays billions for individual researchers is a judgment about the people
Kantrowitz: "The only reason you don't do that is if you don't think the people are good"
Meta's settlement puts two of Roy's 2019 social media fixes into practice
Like counts off by default for minors, and a chronological feed available to teens
Roy reads the settlement as aimed at YouTube and TikTok
Meta reportedly offered to go to a one hour daily limit if the other two matched it
Social networks are gone; what is left is a shorts feed
Kantrowitz: "social media was a moment, but it's now just all entertainment"
The Korean blowup is what a bubble looks like before anyone agrees there is one
Retail traders levered two and three times on memory chip names, wiped out on the drawdown
Benioff Got Amodei to Sit Down and Smile
Roy opened by describing waking up to an image of Benioff and Dario Amodei seated next to each other promoting the new Anthropic-Salesforce partnership, with someone in his mentions addressing both him and Kantrowitz over having taken the SaaSpocalypse warning seriously
The announcement itself, in Roy's reading, is a repackaging rather than a launch: "you could already do this in Salesforce and Claude", and at Writer, where he works, people already connect to Salesforce and run all types of queries — he calls it one of the most valuable things he has found for his day to day
The marketing, not the product, is what impressed him: "God bless Mark Benioff, because that man can market like no other"
What surprised him more was Amodei describing his own chief commercial officer demoing the tool to ask which accounts Anthropic is trying to close and to talk through the risks on each — "These are not things Dario ever says or talks about, like the most pedestrian, mundane ways of using AI"
Kantrowitz noted Salesforce is an investor in Anthropic; Roy waved the size of the stake away, putting it at five billion against the valuation Anthropic is heading to market with — "You're going out at two trillion. This is not your lead here"
Kantrowitz's own first read was more mundane: "I just thought it's the last week of summer, nothing's happening" and this was the week to get Amodei in a room with Benioff. Roy's verdict on that theory: "That is a very generous interpretation"
The Pre-IPO Theory
Roy's explanation is that Amodei is being managed for an IPO for the first time: he thinks investors have told him that being the company that subsumes everyone and scares everyone is bad for the offering
The pitch Roy imagines being made to him is that "you're the last private company to be standing" is a liability now rather than an asset, so the job is to sit next to another technology leader and show he is not a Death Star
Roy dates the moment when the destroy-all-software framing played well to about six months ago, when people liked it and found it fun
Kantrowitz landed on the same read from the other side: "what you saw with the two of them is almost like the anti-Alex Karp", against a run of CEOs who have gone on air to call Amodei evil
Roy's summary of the message being sent: he is friendly, he is not attacking Benioff live on television, and the IPO market can trust him
Roy kept coming back to how out of character the material was — "When's the last time you heard Dario talk about, like, commercial stuff publicly, something so pedestrian and mundane and every day, like, closing a deal"
Kantrowitz pushed back that Amodei is business minded and knows the numbers, but agreed the setting was new — sitting with a software CEO to walk through why a partnership is good, "even if it is Benioff, it's probably the most famous of them"
Neither Company's Own Story Supports the Deal
Kantrowitz put the investor's case for the opposite strategy: if he held Anthropic he would want the vision where the multi-billion-dollar software companies are the market, not the partners — "the whole vibe code your CRM"
Roy agreed on the Anthropic side without hedging: "It's you need to destroy Salesforce to get to two trillion dollar market valuation", and walking away from that story in the medium and long term does not make sense to him
On the Salesforce side, the risk he sees is the interface: giving up the UI layer while the customer relationship moves into the assistant
The central question underneath it, in his framing, is whose data it actually is — the customer's or Salesforce's — with Amodei himself opening by saying Salesforce has all the data
He does see real value in the layer Salesforce owns: "there is a tremendous value, in fact, in terms of being the governed data layer and the structured data layer" as a secure, governed, structured place for customer data
His doubt is whether that is worth traditional software margins, and he says he does not think so, while allowing that it might be
"But in the end, it's the customer's data" — which is why building the whole story around the data being Salesforce's strikes him as weird
Roy called the one-day move the second biggest in the company's history: "Salesforce stock jumping 22% in a single day, the second biggest jump it's ever had in its existence" — for a company he says is going to market as a smart database rather than a dumb one while handing over the interaction points
Capability Is Not Diffusion
Kantrowitz argued the market has been playing the SaaSpocalypse with a certain level of craziness: even granting that customers may eventually build their own CRM, or that Anthropic moves up market, none of it happens overnight, in a quarter, in a year or in two years
"there's a real gap between the rise in capabilities and how they get implemented" — people watched capabilities improve exponentially and expected business disruption and diffusion to arrive at the same speed
A company that has run on Salesforce for years does not move to Claude overnight even when the capability exists; on his estimate a genuinely better Claudeforce is a three-, four- or five-year switch
His forecast for coverage is oscillation: headlines saying the SaaSpocalypse is on, then off, then on again, when the honest answer needs much longer time horizons to measure
Roy agreed almost uncomfortably, calling it a timeline issue, and added the consequence for incumbents: "it gives the sales forces of the world plenty of time to figure out a new model or a better model or what's their winning strategy"
He still does not think either company should be this excited about the announcement itself once the mechanics are unpacked
The Software Bounce, and the Round Trip
Kantrowitz read out the CNBC framing of the quarter: investors had feared increasingly capable AI models would let businesses do more with fewer traditional software subscriptions, undermining the industry's pricing model, and Salesforce shares had fallen 22% this year through Wednesday's close before the rebound
The host's own numbers on the move, offered with a not-investment-advice caveat: Salesforce up 72% from June 22nd, up 41% over the past month; the IGV software ETF up 19.68% over the past month; Adobe up 16.49%; Microsoft up 29%
Kantrowitz's read of it is investors cooling on the idea that disruption happens overnight, and an opportunity in the swing between headline-driven selling and earnings-driven buying
Asked to guess where Salesforce sat for the year to date, Roy said maybe 3%; Kantrowitz put it at down 1.94%, almost flat
"it is kind of a perfect round trip of software is dead" — Roy walked the year from Claude code hype in February, March and April, through vibe coding your CRM, to nobody knowing what happens next and the stock back at flat
The Not-Investment-Advice Bit
Roy needled the disclaimer: "Has anyone who ever said not investment advice not followed up with investment advice in the history of podcasting?"
Kantrowitz defended the caveat on its merits: listeners should "trade based off of what your financial advisor tells you to do" rather than off observations made on a podcast, and he does not want anyone trading these segments
Roy conceded the point while noting how much investment advice now gets handed out in ordinary conversation, on shows like All-In and elsewhere, and settled on being pro caveat
Asked whether he was going to go buy IGV after all that, Roy played it out: "I was already, Robinhood's already open. It's already done", holding Kantrowitz responsible
Kantrowitz's reply was that he was too late, the bounce already happened, the opportunity window closed — which Roy scored as still more investment advice, with a lawyer replaying the tape for a client in shambles after shorting and then going long
Project OT, and the Meta Reorg That Did Not Take
Kantrowitz read Reuters' account of a plan hatched at Zuckerberg's Hawaii compound at the January leadership retreat, code named Project OT, for organizational transformation, envisioning an AI-native future for Facebook and Instagram's owner
AI taking over much of the daily work of thousands of employees, with virtual workers overseen by smaller, talent-dense groups of human staff
In scenario planning, executives explored cutting the size of many teams by as much as 60%
The pilot ran small tech pods of two to three engineers and a designer, all equipped with AI tools, dispensing with the fixed six-month planning cycle in favor of four-week prototype sprints, with traditional middle management layers eliminated
Roy's first reaction was that the design is right: separated from the layoffs and the organizational questions, "this is to me exactly what companies should be testing and doing and figuring out what the new operating model is and that teams should look different"
His only explanation for failure was the name: "any indication of failure is solely attributed to the worst code name for a project ever" — Project OT, which he said sounds worse than McKinsey
Kantrowitz ran a by-golly joke about how the news was announced internally, and the two set out the by-golly bylaws on the spot: a trillion-dollar market cap or higher, which is why Roy's suggestion of the Salesforce move did not qualify
The Internal Data That Killed It
Kantrowitz walked through the numbers Reuters reported from Meta's internal data, and said he was surprised the story had not gotten more play
"Employees' use of AI had resulted in a vast increase in the code they generated, but with questionable impact on productivity"
"code changes made to the internal software platforms and infrastructure employees use on the job were up 220% year over year" while "But changes that led to new or upgraded features reaching Meta users were only up 36%"
Infrastructure teams warned in April that unchecked agents were performing large-scale disruptive actions humans are unlikely to execute
Major technical and security incidents such as service disruptions and possible data leaks were up 40% on the prior year, with time spent firefighting them up 70%
Kantrowitz called it the epitome of every criticism of AI coding: "You write more code, you're not necessarily more productive, you ship more, and when you ship more, you have bigger security flaws" — and the cleanup costs more than the old way did
Zuckerberg, on the host's account of the story, admitted agent technology is not at the point he expected
Roy put the failure on incentives rather than capability: "I think employee incentive is kind of like what's probably at fault here"
"So if you tell people you're going to be judged on how fast you move, how many tokens you consume, how much AI you do and use" — of course the output is volume, rather than the task or project actually getting executed on
Compute Goes Up, People Go Down
Roy asked how CEOs at that level think about their employees, and made the point that these are the same founders who chose the post-COVID hiring and growth trajectories, at companies heavy on headcount relative to the cash-generating parts of the business
"I still wish one of these CEOs would stand up and be like, I hired too many people" — Roy wants the founder to own the bureaucracy rather than blame the moment
Kantrowitz pointed out that Zuckerberg did exactly that with the year of efficiency, and that Musk has done it too; Roy's distinction was "his always feels a bit more performative. When Zuck says it, I believe it"
On the outcome: Meta went ahead with a 10% cut, but a larger cut planned for November was called off, with Zuckerberg telling employees not to expect further company-wide layoffs this year and expressing a desire to give them stability
The framing Kantrowitz took from Zuckerberg's recorded remarks is a two-line budget: compute on one side, people-related expense on the other, and when compute goes up the other has to come down
Roy's answer on whether that is the right approach was that if your people are good, it is the wrong one — "I always find it a little bit rich when a ultra-profitable publicly traded $1.5 trillion dollar company with massive margins is like we need to find cost savings"
Kantrowitz took it further: a company spending billions on AI talent and billions more on compute could spend some of that on people, so "The only reason you don't do that is if you don't think the people are good"
Roy said that judgment is implicit in the decision, pointing to what Meta paid for Alexandr Wang — he guessed 14 billion, Kantrowitz guessed 15 — as proof the wallet opens when the person is considered good enough
The $18 Billion Settlement, and a 2019 Wish List
Kantrowitz noted Meta had flagged on its most recent earnings call that lawsuits over the addictive nature of its products are a major liability, then read Reuters: "Meta agrees to pay $18 billion to settle U.S. lawsuits over children's social media addiction", payable over the next decade, with strict limits on teenage use of Facebook and Instagram under an agreement with nearly all U.S. states
On the size of the climbdown: "They were seeking $200 billion in civil penalties. They get 18"
Roy's reaction was personal rather than financial. He wrote a five-point plan to fix social media for The Margins on August 2nd, 2019, and two of the things he has been shouting about since are in this settlement
Demetricating — removing like counts and the numbers that push ordinary people to obsess over their performance. The settlement turns like counts off by default for minors
"I honestly think all feeds should just be reverse chronological order. No algorithm" — the settlement at least gives teens the ability to use a chronological feed
He credits TikTok with ending the argument by doing away with the feed altogether in favor of one continuous algorithmic injection of content
He still runs his own accounts that way: "I still use Twitter slash X with the following feed more than the For You feed"
Other terms he read as good: a default two-hour daily usage limit for under-18s, no use from midnight to 6 a.m., and no notifications during school hours
"But I honestly think this should apply to everyone, not just teens" — a position Kantrowitz called interesting and said would never go forward
Parenting, Rabbit Holes, and Who the Settlement Is Really Aimed At
Asked whether he feels better as a parent, Roy said he has a seven-year-old and "I'm hoping we kind of solve some of this by the time he's like 13 and 14 getting a phone"
He rejected putting the whole burden on parents: for high-school-age teens, restricting a child while everyone else is comparing like counts is a fight against a tool designed to produce exactly that behavior, which he says is what Instagram is built to do as a business and a product
Roy's theory on why Meta settled starts from a reported willingness to go down to a one-hour daily limit if TikTok and YouTube followed
Meta has diversified into AI and has other opportunities, and teen advertising already carries protections, so the business damage is containable
YouTube and TikTok skew far younger — his line on YouTube is that everyone pretends it is not an algorithmic social network when the addictive dynamics are identical
His household rule is about the sidebar: his son can watch a video but cannot click the recommended videos on the right rail, because that is step one of the rabbit hole, twelve videos deep
"I think this was just a straight shot at TikTok and YouTube"
Kantrowitz agreed on YouTube's reach — eating into cable television, into Netflix and into the social networks
Social Networks Are Over, Shorts Won
Kantrowitz's stronger claim is that the category is gone: "social media was a moment, but it's now just all entertainment", with Meta, TikTok and YouTube all competing on effectively the same product, shorts
Roy tested it against his own household: none of his friends post, so his Instagram feed is entirely entertainment, while his wife's friend group still posts families, trips, kids and everyday life, leaving a social network element in her feed and none in his
Kantrowitz's counter, in Meta's own vocabulary: "I would be stunned if it was more than 10% of your total Instagram time" is spent on friends and family content, because the whole experience is built to get you into the Reels feed
Roy's verdict: "Social network is dead"
The two compared their YouTube habits — Kantrowitz watches horizontal video because his shorts feed is broken, serving him nothing but creator-growth hustle clips with the same sound and the same title, which he said he would trade for a full day of a hustle-guru creator's content
"I feel like watching a horizontal video on YouTube kind of feels like listening to music on vinyl" — twenty minutes, classic, and in Roy's word "It's nourishing"
Roy's own shorts feed is cooking content, music and video editing, which he called reasonable; Kantrowitz asked for an hour of it to reset his algorithm
South Korea's Ants and the Levered ETF Trade
Kantrowitz set the segment up against the bubble argument: even granting that AI is not a bubble, "there could be, and there are already little bubbles and investing behavior that is putting people at risk"
He read the Wall Street Journal's account — "the world's craziest stock market has turned into a fright ride"
The Kospi has more than tripled in value on faith in the AI boom, with Samsung Electronics and SK Hynix reaching a trillion-dollar valuation and becoming twin forces behind record gains
The index then fell around 40% over six weeks in June and July, burning hundreds of thousands of investors, and has since rebounded about 20% from its lows
Individual investors account for 60% to 70% of daily trading volume and are known as ants — "weak individually, but capable of collective power"
The leverage is the problem, not the index move. Kantrowitz walked through the mechanics of double- and triple-levered single-stock ETFs: a 5% gain becomes 10% or 15%, a 5% loss the same in reverse, "And so when it goes down 40%, you're wiped"
Roy's most fascinating detail was the vocabulary: "the most fascinating part of the story is actually first this term ants", which he found dates to the 2020 COVID trading boom and took on its own form in this cycle
He connected the Korean retail army to the concentrated, levered US positioning around Leopold and situational awareness as two ends of the same trade, and noted the timing — the double- and triple-levered ETFs arrived at the end of May
"as you said, up 40% then down 20% net might not seem that bad, but I mean, you're getting cleared out if you're levered on this kind of stuff"
What he wants to know is the distribution mechanism: how these trades come into existence, how the information is shared, and who is convincing everyone they are missing out on SK Hynix and the levered ETFs
The People Who Got Cleaned Out
Kantrowitz tied the behavior to a conversation the show has had before about fantasy football and online sports gambling being sold as the last place a person has agency, when housing and retirement feel out of reach — and said the same dynamics are at play here
He read the Journal's individual cases. Yun Kyeong-min, a 44-year-old sound engineer, invested half his severance pay in semiconductor stocks after quitting his job and watched $7,200 evaporate in a week, confessing to his wife only a bit of bad luck
A 25-year-old software developer seeded a trading account with around $14,000 from savings and crypto gains after concluding from her first paycheck that she would never earn enough to buy a home, put all of it into SK Hynix and then into single-stock levered ETFs, and started an Instagram account to document the ride to riches
By May her account was up more than 58% and her posts were drawing an average of 4 million views
Her own summary, as read on air: it felt like it was all a dream, and she got greedy, kept thinking about climbing higher even after she cashed out
Roy's response to the pairing of the Instagram account with the trade was that this is the whole episode in one place: "the same addictive mechanisms and the same move away from friends and family content to entertainment content and that entertainment content being investment content", algorithmically pushed at people who then go and buy the stock
"It is kind of an amazing, terrifying way of seeing it all come together in one place"
Little Bubbles Come Before Big Ones
Roy's answer on whether this is worth covering starts with definitions: whether something is a washout, a downturn or a bubble is relative, and mostly a question of how cleanly the narrative packages it afterward
He places these stories at the start of a cycle, not the end of one: "having lived through the 2008 on a trading floor, like 2007, there was Countrywide and Bear Stearns and long before Lehman" — bits and pieces coming out before the thing itself
The difference he sees this time is who is exposed. In 2008 the average household's exposure was home equity; "whereas whatever is happening in this cycle, way more people, thanks to Robinhood and democratized investing, are exposed" — many more people with skin in the game
Kantrowitz's close tied the three stories together as one belief: that the AI line moves immediately and exponentially. Software was about to be annihilated, Meta's work was about to be replaced, and the memory crunch was never going to let up
"There's a Goldilocks zone between going all in too late and all in too early" — and the real risk is the early side, because of the human side of things and the turbulence along the way
Roy's verdict on the wrap-up was that Kantrowitz stuck the landing, which the show never does, and his sign-off for the fall was "The rest of this year is going to be absolute chaos, and I can't wait for it"
Roy's bottom line is that the direction of travel in AI is not in dispute and the timeline is, which is why software got a reprieve rather than a reversal, Meta's agent-native reorg broke on implementation rather than on the idea, and the people getting hurt right now are the levered individuals at the edges rather than the system as a whole.
Products, Companies & Tools Mentioned
Salesforce and Anthropic, and the Claudeforce partnership (The announcement the episode opens on — Claude querying Salesforce records for insights like which accounts to close. Roy says the capability already existed and the news is the packaging)
Writer (Roy's employer, where he says connecting Salesforce data to an AI entry point is already routine and one of the most valuable things in his day to day)
Meta, Facebook and Instagram (Project OT, the failed AI-native reorg; the 10% cut that went ahead and the November cut that did not; the settlement over children's social media addiction)
Reels, YouTube Shorts and TikTok (The three products both men say are now the same product, and the reason they think the social network is finished)
YouTube (Roy's example of an algorithmic social network everyone pretends is not one, and the right-rail recommendation his household bans)
Samsung Electronics and SK Hynix (The twin memory chip names behind the Kospi's rise, both reaching a trillion-dollar valuation, and the underlying stocks for the single-stock levered ETFs Korean retail bought)
Single-stock double- and triple-levered ETFs (Launched at the end of May, per Roy's timeline, and the mechanism that turned a 40% index drawdown into a wipeout)
IGV, the software ETF (Kantrowitz's proxy for the software rebound, and the running joke about giving investment advice)
Robinhood (Roy's punchline for how fast a listener can act on a podcast observation, and, seriously, his marker for how many more people are exposed to this cycle than to 2008)
Adobe and Microsoft (The other software names Kantrowitz cites as having rebounded alongside Salesforce)
X, and its Following feed (The chronological feed Roy still uses in preference to For You)
The All-In podcast (Roy's example of how routinely investment advice now gets given inside ordinary conversation)
Countrywide, Bear Stearns and Lehman Brothers (Roy's 2007 sequence, cited as what the early part of a bubble's ending looks like)
Books & Resources Mentioned
A five-point plan to fix social media – Ranjan Roy, The Margins, 2 August 2019 (The piece he wrote seven years ago; demetrication and reverse-chronological feeds are the two ideas he says the Meta settlement partially delivers)
Reuters' report on Project OT (The source for the Hawaii retreat, the pod structure, the internal productivity data and Zuckerberg's remarks on compute versus people expense)
The Wall Street Journal's report on the South Korean market (The source for the Kospi's fall and rebound, the ants, and the individual investors who were cleaned out)
Get the latest market chatter and takes as they happen:
X | Threads | Instagram | YouTube | TikTok | Facebook