Netflix Angst, the End of the Internet, and Meta’s VR Glasses | Sharp Tech with Ben Thompson

Summary

• Netflix’s weakening engagement puts more weight on its last defense: low churn. A listener frames the problem around U.S. subscriber saturation, engagement 10% below targets, and subscription increases of 16% in 2025 and 12% in 2026. Thompson reads management’s reassurance about cancellations as saying, “Our last line of defense has not collapsed yet”—an increasingly fragile foundation for continued price increases.
• Netflix’s advertising reset makes Meta and short video central to its growth problem. Thompson cites an advertising forecast moving from about $3 billion to $2.7 billion, with expected growth reduced from 100% to 80%. Advertising can expand monetization without charging viewers more, but it requires attention: “TikTok and Reels are swallowing everything up.”
• Hollywood’s persistent streaming ambitions keep Netflix’s licensing endgame out of reach. Thompson expected rivals eventually to supply Netflix with their best content; their continued investment in competing services has delayed that outcome. Sports could strengthen retention and advertising, although leagues capture much of the economics—and Sharp’s $2 billion Sunday Ticket comparison, below 1% of Google’s stated 2027 CapEx, illustrates the purchasing power surrounding these rights.
• Internet security is pushing toward private enclaves defended by agents. The hosts discuss urgent PHP vulnerabilities, compromised certificate trust, and devices whose security foundations might never be updated. Thompson argues that the fear of giving agents access could soon collide with the opposite necessity: having them continuously inspect systems that humans cannot reliably maintain.
• Meta’s VR glasses make portable workspaces more appealing, while fit and immersion remain unsettled. The hosts praise the lightweight hardware and tracking, but Ben still prefers Vision Pro and finds the glasses’ precarious feeling distracting. Their clearest prospective use is travel: carrying a small computer while retaining the screen space needed for reference-heavy work.
• Anthropic’s coherent Claude identity sharpens a dispute over whether AI should be taught or controlled. A listener’s “Claude is always Claude” framing captures its consistent presentation across products. The hosts admire the execution—including praise for “Opus 5.5”—while challenging personification and welcoming early interpretability findings that suggest models are less opaque than previously assumed.
• Agents can make human reading speed the binding constraint on work. A listener finds that output arrives faster than he can understand it, although careful planning feedback determines the final quality. Thompson hopes adaptive interfaces will help; his immediate experiment combines reading and listening, with 3.5x audio described in the context of following the text simultaneously.
• Sharp reverses his dismissal of “You Can See Everything” and recommends a crowded theater. He credits the shared laughter and surprise, alongside a more sincere approach from Nathan Fielder, for a three-hour experience in which he never checked his phone. His admission is direct: “I completely lied on the podcast last week.”

Deep dive

1. Netflix’s reassuring churn figures expose a narrowing margin for error

• A listener, Rav, supplies the bearish setup: Netflix has laid off 6% of its employees, including the head of content and the president of its film division, and reportedly added no U.S. subscribers this year. Engagement is described as 10% below targets, while the base subscription price has risen four times in five years, including 16% in 2025 and 12% in 2026. The question is how much further pricing can carry growth.
• Thompson finds management’s acknowledgment of engagement problems unusually revealing because Netflix tightly controls its earnings conversations. Questions are submitted and read by an employee rather than pursued through open exchanges. Against that background, explicitly conceding the problem carries more weight than the familiar presentation that everything reflects a strategy Netflix always intended.
• The advertising revision compounds that concern. Thompson recalls a forecast of approximately $3 billion becoming $2.7 billion, with a 100% growth ambition reduced to 80%. He accepts the explanations about rebuilding systems and partnerships taking longer. Nevertheless, the business most directly dependent on engagement is also growing less than expected.
• Management says price increases have not produced unusually high churn. Thompson’s objection is that low cancellation rates have supported Netflix for roughly a decade; once that protection fails, the damage is already substantial. He translates the reassurance as, “Our last line of defense has not collapsed yet.” Its survival does not establish that the rest of the business is healthy.

2. Netflix’s global reach cannot compensate indefinitely for weak American content

• The hosts distinguish fixable programming decisions from structural limits. Thompson says U.S. content has not been good enough recently and that Netflix’s brand image has been damaged by the chaos. He argues that pricing power needs one or two popular dramas each quarter that everyone wants to watch, as HBO has. Netflix’s emphasis on quantity over quality has weakened that proposition.
• Thompson points to The Queen’s Gambit, which he recalls as a hit that won many Emmys about six or seven years ago, and says Netflix has not replicated that success. He describes the company as having begun with distinctive, high-quality content, but focusing more on quantity for at least the past decade. He also says the Chronicles of Narnia movies seem to have flopped.
• The discussion of content chief Bela Bajaria highlights the imbalance: Netflix praised her international work, while the rest of the discussion focused on America. Thompson notes that locally popular programs in places such as Korea, France, and South America can do well, but American movies that everyone watches remain strong exports. If the U.S. market is sluggish, he says, that has a larger effect on the overall business than one underperforming business sector in another region.

3. Advertising solves a pricing constraint but creates an attention constraint

• Thompson reduces Netflix’s growth options to more subscribers, higher subscription prices, or advertising. The size of the U.S. market cannot be changed, and subscriptions face price elasticity: eventually people ask, “How much do I have to pay?” Advertising permits monetization to rise without increasing the viewer’s bill. He calls that potential “indefinitely” expandable, then explicitly concedes that the word exaggerates the case.
• Meta supplies his mechanism: better targeting increases an advertisement’s value even after audience penetration matures. Netflix must make its own advertising proposition work, but the conversation identifies Meta as a direct entertainment competitor. The household example is nightly Reels consumption, not a deliberate choice among television series. “TikTok and Reels are swallowing everything up.”
• That makes Netflix’s changing emphasis on podcasts, games, and other internet-oriented programming conspicuous. The hosts recall these initiatives receiving attention six months earlier; this call instead stresses spending on important priorities, with about 5% left aside. They question whether those experiments can materially repair engagement. Lower subscription prices internationally make profitability in those markets harder.

4. Hollywood’s delayed retreat keeps Netflix waiting for its preferred end state

• Thompson considers Sony’s production-and-licensing strategy rational: make desirable entertainment and sell it to the highest bidder. He also praises Fox’s free, advertising-supported approach. His surprise is that other studios continue pursuing competing streaming businesses despite unclear economics. Netflix’s original programming was supposed to bridge the period before excellent outside content returned; that transition has taken longer than he expected.
• He interprets Netflix’s Warner Bros. pursuit as an attempt to force the outcome that competitors would not voluntarily accept: “Okay, then we’ll buy you, so you have to sell to us.” He still believes consolidation toward that end state should happen. The unresolved variable is timing, while Netflix must keep producing growth sufficient to support its stock-market expectations.
• Rav’s NBCUniversal suggestion offers sports rights—Sunday Night Football, the Big Ten, the NBA, the EPL, and the Olympics—plus theme parks. Sharp asks whether the acquisition would be reasonable, and Thompson accepts the strategic rationale. Even one major sports event monthly could help retention and advertising, but leagues capture much of the value. Sports does not offer a comforting future in which competition disappears and content suddenly becomes cheap.
• Fox’s NFL investment is the best analogy: Thompson does not know whether the rights package itself generated much profit, but says it helped turn Fox into a major network. Later, Sharp notes that Google acquired Sunday Ticket for $2 billion, less than 1% of Google’s total 2027 CapEx. The hosts accept the comparison as useful context for technology companies’ potential sports spending.

5. Skydance’s financing introduces another layer of uncertainty into streaming consolidation

• The hosts riff on a list of declines: Oracle down 53%, Skydance down 59%, and a college-football joke about Bryce Underwood down 99%. Sharp identifies the Underwood reference as a specific college-football joke Rav made.
• Thompson discusses a Hollywood Reporter item asking whether a major deal between David Ellison’s Skydance and Elon Musk’s SpaceX is imminent. John Malone predicts a deal involving $8 billion to $10 billion and global distribution. His proposed mechanism is a SpaceX investment that reduces Skydance’s debt. Malone points to Larry Ellison’s connections with Musk, including Twitter financing and Tesla stock. Thompson notes that Malone says it is unclear what SpaceX would gain.
• On Thursday, Sharp says the news is that Musk is investing $1 billion in Skydance, adding that it is not $10 billion but might turn into that. Thompson then refers to “the investment in SpaceX” as almost unbelievable. The hosts leave the ultimate scale and structure unresolved.
• Thompson recalls praising the Model 3’s brand-driven demand, then recoiling from a related solar-company transaction the following week. That experience explains his reluctance to interpret every Musk transaction as a coherent strategic master plan.

6. Internet trust can fail through ordinary implementation mistakes

• The encryption discussion begins with a comparatively reassuring account. The hosts reason through a possible mundane explanation for elliptic-curve seed values: a standard pseudorandom number generator run at a particular time. Thompson takes this as a reason to dismiss suspicions of deliberate NSA interference, saying the seed could simply have been generated and passed along. “This might be the safest news among this week’s updates.”
• The PHP story is more urgent. The hosts discuss an NSA warning about two serious vulnerabilities and advice to take affected systems offline if immediate updates are impossible. Sharp says the PHP Foundation has been distributing patches for versions since 8.2, recalls that version 8.2 was released in 2022, and emphasizes the millions of websites and services still running older software. Thompson describes the potential compromise as exceptionally severe.
• A separate certificate issued by a certificate authority in India illustrates the fragility of foundational trust. Thompson describes a weak key, citing 4,127 during a discussion of 256-bit security, and says the error makes decryption much easier. The essential point is that a trusted issuer’s mistake can permit impersonation. Their explanation focuses on the consequences of defective implementation rather than identifying a breakthrough against encryption itself.
• Browser updates can remove the affected trust, but Sharp worries about connected equipment with embedded certificates that may never receive equivalent updates. That creates a potentially permanent population of vulnerable devices. Thompson’s verdict is categorical: “The incompetence of one company will leave permanent damage to the internet.” Internet-wide trust inherits the failures of individual institutions.

7. AI strengthens both the attackers and the case for continuously monitored fortresses

• Thompson separates AI’s role from ordinary incompetence: he does not blame the certificate failure on AI and says only that he thinks AI helped discover the PHP vulnerabilities. His broader prediction is that models will find many more weaknesses. The same capability supports defense, making continuous inspection increasingly important as the volume of discoverable problems grows.
• Thompson says a tool called Quad discovered a vulnerability; Sharp presents this as an example of an AI agent protecting a network. Separately, Thompson says his Linux version had an old certificate that he needs to fix. His rack is kept as offline as possible and hidden behind Tailscale because he does not want anything exposed on the public internet.
• The hosts’ point is uncomfortable for capable enthusiasts: Sharp says, “You need to stop believing you can handle it yourself.” Regular updates alone do not guarantee that someone is tracking every emerging issue.
• Neal Stephenson’s Snow Crash supplies the metaphor: protected, pleasant enclaves separated by hostile territory. Thompson imagines Tailscale-style private spaces as “my own little fortress,” with agents patrolling their boundaries. Self-hosted email already shows how difficult independent participation has become: Thompson says emails are not delivered unless they come from a major company. He regards the earlier open internet as irrecoverable: “Moving forward is the only way.”
• Sharp remains cautious about declaring catastrophe from roughly six or seven current incidents, but worries that leaked developer credentials can propagate into other services unless painstakingly replaced. The hosts distinguish that problem from an alarming Axios account of agents accessing bank accounts: much of the latter appears to involve access explicitly delegated by users. Authorized automation should not automatically be classified as another mass compromise.

8. Meta’s VR glasses trade headset bulk for a less settled sense of immersion

• The hosts find Meta’s unreleased glasses technically impressive. Ben still prefers Vision Pro, praising the resolution achievable in lightweight glasses but finding the fit precarious. The objection concerns sensation rather than actual slippage: resting on the ears makes it harder to relax into the experience. “It felt like being in the uncanny valley.”
• The pushback is immediate: “You just said they weren’t heavy enough.” Ben’s concern is that the glasses feel as if they might fall off, even though they do not. A headband might address those concerns, and the discussion suggests it could provide fuller blackout. That remains an expectation about an attachment, rather than an established result from their demonstration.
• The display is described as slightly below Vision Pro’s resolution, with a noticeable difference, although Ben says even Vision Pro does not have enough resolution. He considers its resolution sufficient for reading text. Eye and hand tracking receive stronger praise: interactions register more reliably and feel like technology at least two years newer. The glasses are unreleased and will “probably come out next year.”
• Travel is the compelling application. Vision Pro’s volume is a bigger packing problem than its weight, while writing with many references pushes Thompson toward a large laptop. A small computer paired with portable virtual screens could change that compromise. The hosts also expect a richer sports-content offering; neither presents walking around while watching video as the desired use.

9. Claude’s consistent identity contrasts with the appeal of a personal agent

• Listener Joseph identifies a product distinction: “Claude is always Claude,” whether encountered through Slack, Claude Code, or chat. He contrasts that unified presence with more specialized or customizable agents. His Muse is a robot dog named Muff, complete with a call-center headset and animated tasks—a concrete example of personalization extending beyond a name or conversational tone.
• The hosts themselves have kept the default Muse name. Their preference is to experience what the creator designed without first inventing a character: “I don’t want to feel pressured to create this from scratch.” That does not reject personalization’s usefulness; it identifies the setup burden that customization can impose on people who primarily want a working service.
• They concede that Anthropic’s branding is unusually consistent while other companies ask consumers to navigate numerous products. Separately, they discuss reports of an S-1 submission, a possible Friday announcement, and an IPO potentially “next week or the week after.” Those remain recording-time expectations. Thompson wants the actual disclosure and figures rather than selective reporting about the filing.

10. Admiration for Claude coexists with a fundamental disagreement over alignment

• Thompson dislikes treating model output as evidence of a continuous person: “Every token feels like a completely new experience.” His framing is that language feels intensely human, but models lack the instincts, desires, and bodily drives that animate people. He prefers presentation that reminds users they are dealing with technology, and sees Anthropic’s personification as moving in the opposite direction.
• He characterizes Anthropic’s alignment philosophy as teaching Claude to behave correctly, while advocating tools humans can understand and control. Superior capability alone does not imply divinity: “The explosive power of a nuclear bomb is much greater than my fist.” He also criticizes Silicon Valley’s identification with intelligence, suggesting that encountering something smarter can provoke either nihilism or a turn toward godlike explanations.
• Sharp says some people in AI safety dismiss an interpretability paper because it challenges their perspective. Thompson agrees that the paper overturns the claim that we cannot see inside LLMs at all. The hosts describe promising early results suggesting models are “not as much of a black box as we thought.” Sharp says the tools and capabilities to understand them were previously lacking. The proposed consequence is better investigation and control, rather than accepting opacity as permanent.
• The product concession remains emphatic: Thompson says “Opus 5.5 is truly amazing.” He also praises Quads for agent-based tasks, saying multitasking works better and they can track different situations simultaneously, while ChatGPT tends to fixate on one thing. Those practical strengths do not answer Thompson’s corporate-alignment question: whether Anthropic’s account of what these systems are, and how people should relate to them, serves humanity’s interests.

11. Siri’s damaged reputation does not make “Hey Apple” an obvious improvement

• Listener Kyle proposes attaching the assistant directly to Apple’s respected brand, arguing that Siri has become synonymous with frustration. Thompson challenges the proposed insulation from failure: associating weak AI software with the entire company can damage the broader brand. Sharp’s joke captures the assumption inside the proposal—Kyle may be the greater Apple enthusiast for wanting that association.
• Thompson then admits a change of mind. During Google consulting more than a decade earlier, at a Napa Valley meeting, he strongly favored using Google’s own name for its assistant. Actual use changed his judgment: saying “Okay, Google” feels like addressing a multitrillion-dollar corporation. “I think ‘Okay, Google’ is a mistake.” He also explains that conflicts subsequently led him to stop consulting for companies he covered.
• Sharp says Siri’s name is good, easy to pronounce across language backgrounds, and distinctive. Ben agrees that it is easy to read and better than Alexa. They see the product’s performance, not the name, as the problem. Sharp says Siri has been terrible for the past six or seven years, though he allows Apple might improve. Neither treats a corporate wake word as a substitute for execution.

12. Agent productivity exposes a human comprehension bottleneck

• A listener who works in U.S. IT and is fluent in English, though it is not his native language, describes a shift from relying on mathematics, statistics, coding, and domain expertise to spending the day reading plans, reasoning, and summaries. Agents produce material faster than he can absorb it. Skimming saves time, but he finds that careful planning feedback determines the quality of the eventual work.
• Thompson reads and writes unusually quickly, yet recognizes the problem. Reading closely exposes prerequisites that later errors may reveal were wrong; successful agent supervision increasingly requires diagnosing flawed assumptions rather than requesting a superficial correction. “The typo-correction stage has already passed.” Faster production therefore does not remove the need to understand what the agent believes it is doing.
• His hoped-for answer is a “generative user interface” suited to the task, with information and interactions expressed in a more appropriate form. Voice could also help, although speaking and listening take time. These are possible interface improvements, not a claim that current agents have solved the comprehension problem.
• Sharp describes listening to Ari Emanuel’s audiobook while driving to O’Hare and reading along on the plane. He says the experience felt like having the book “administered intravenously” and that he listened at 3x speed. In response to a listener’s question about 3.5x audio, Thompson clarifies that he was reading along while listening; he might have listened at 2x in the car, and thinks 3.5x works in that read-along context. He allows that audio alone might not have stayed with him.
• Sharp prefers 1x and retained pauses for entertainment. Thompson’s sound-quality rationale accommodates both: clearer audio lets people who want speed listen faster.

13. Sharp’s film reversal rests on the experience of watching together

• Sharp admits he had dismissed You Can See Everything because he was not a Nathan Fielder fan and did not want more Elizabeth Holmes. After 48 hours of enthusiastic reaction, he went on a rainy Saturday in Washington, D.C., expecting to find the consensus wrong. Instead, he loved it: “I completely lied on the podcast last week.”
• His recommendation is specifically theatrical. He distinguishes the big-screen spectacle that usually draws him to cinemas from the shared laughter and surprise of this audience. Fielder’s approach felt more sincere than earlier work, and Sharp thought genuine irritation with Holmes helped sustain three enjoyable hours. He never checked his phone. The communal setting made the experience especially memorable.
• Spoiler restraint becomes a final disagreement about community obligations. Sharp appreciates people preserving the pleasure of discovery online; Thompson recalls managing delayed sports viewing in Taiwan by avoiding Twitter and muting conversations himself. They joke about how long the protection should last and whose viewing date starts the clock. Sharp’s practical recommendation is to see the film while a lively theater audience remains available.

Verification Notes

The Thursday investment details remain unclear in the transcript: Sharp says Musk is investing $1 billion in Skydance, while Thompson immediately refers to an investment in SpaceX. The digest leaves the structure unresolved.