Facebook is Dead; Long Live Meta, Does OpenAI Need to Log Off?, Questions on Bubbles, Blackberry, and Bell Labs

Summary

• Meta’s blowout quarter confirms the durability of its ad machine, but Ben Thompson thinks investors may now be crediting Zuckerberg’s superintelligence push for results it has not yet produced. Stories in 2017–18 and Reels in 2022 temporarily hurt monetization by moving attention onto under-monetized surfaces; both created inventory that Meta later converted into growth. This quarter’s gains came from recommendation systems such as Andromeda, addictive short-form video and ad-load optimization—not LLMs or superintelligence.
• “Facebook’s dead” because Meta is no longer meaningfully a social network; it is an entertainment company engineered to capture attention and serve ads. Only about 15% of viewed content now comes from friends and family, while TikTok taught Meta to recommend free user-generated content from across the network. That shift gives Meta more control than MySpace ever had: “A company that has dials is a powerful company.”
• AI is simultaneously Meta’s next product surface and a direct threat to the time its existing products monetize. Hours spent talking to GPT-4o companions are hours not spent in Meta feeds, yet generative AI also completes Meta’s long march from friends’ posts to globally sourced content and eventually content made uniquely for each person. Thompson calls ad delivery one of the world’s best existing agents: advertisers state the desired outcome, and Meta finds the customers.
• OpenAI’s GPT-5 launch became a test of whether the company can ignore vocal power users and operate a billion-user consumer product with conviction. The original automatic router elegantly exposed ordinary users to reasoning models without forcing them through GPT-4o, GPT-4.5, o3, o3 Pro, o4 Mini, GPT-4.1 and GPT-4.1 Mini. OpenAI then rapidly restored choices and complicated the picker, raising Thompson’s concern that it may be “a little bit too eager to please.”
• The defensible GPT-5 grievance was not personality or aesthetics but OpenAI taking paid capability away from Plus subscribers. Thompson estimates they went from roughly 2,900 potential thinking queries per week across o3, o4 and o4 Mini variants to 200 GPT-5 Thinking queries plus whatever the router granted them. His inference is that scarce GPUs were split between ChatGPT and a simultaneous API launch, “sacrificing these nerdy Plus users on the altar of an API” despite ChatGPT being OpenAI’s most defensible asset.
• GPT-5 Thinking is meaningful progress, not the AGI-like “Death Star” jump OpenAI’s own hype encouraged people to expect. Thompson says it is consistently better than o3 after roughly six months in his framing—more accurate, less prone to hallucination and better at following precise instructions—while maintaining his middle position that “people hyping AI are delusional, and people doubting AI are delusional.” Even with no further model gains, he believes the current product overhang could support a decade of development.
• An AI investment bubble is likely, but the enduring asset will be power capacity rather than fast-depreciating GPUs. Thompson distinguishes the dot-com companies from the telecom build-out whose bankruptcies left excess fiber for the next internet era; today’s productive analogue would be overbuilding electricity generation. With electricity cited as roughly 10–15% more expensive year over year and these companies willing to buy at almost any price, chip controls are “a total sideshow” beside whether the US can match China’s power supply.
• Across Apple, BlackBerry and Bell Labs, the recurring thesis is that control can create value—but concentrated institutions can also create assets markets would not fund on the same timeline. Cloud AI could spawn many cheap keyboards, glasses and earpieces without restoring BlackBerry-like margins; Apple may remain unusually valuable because of its 40-year product record, though Vision Pro arrived without a market “hole” to fill. Bell Labs likewise joined monopoly harms to extraordinary R&D, supporting Andrew Sharp’s preference for “purple” antitrust tension over absolutism.

Deep dive

1. Meta’s former crises were healthy monetization transitions

• Thompson’s delayed earnings analysis became its own signal: earlier Meta disappointments demanded a “pants on fire, five alarm fire” response, whereas these spectacular results could wait two weeks. Even enormous AI spending now looks manageable beside the cash-producing core.

• The repeated playbook is straightforward: “First we build a product, then we get customers, then we grow it, then we add monetization.” Stories in 2017–18 and Reels in 2022 pulled users from mature surfaces, causing impressions to explode while price per ad fell—ugly near-term accounting that established years of inventory growth.

• Investors interpreted those periods as deterioration; Thompson saw eyeballs moving to less-monetized surfaces and thought, “This is incredible.” The ATT shock and publicly awkward metaverse pivot compounded the 2022 panic until Meta traded around $90, even though the underlying ad-network mechanism remained intact.

2. TikTok forced Facebook to stop being a social network

• Sharp grants the old skepticism a logic: something that grows virally might also “die virally,” while Amazon can point to physical infrastructure. Meta answered that fragility not by preserving Facebook’s original form, but by repeatedly evolving Instagram through Stories, algorithmic ranking and Reels.

• Thompson’s 2015 argument was that Facebook had to move beyond the social graph into personalized entertainment. He correctly saw the opportunity but wrongly centered professional creators; TikTok’s breakthrough was pulling user-generated content from anywhere, giving the platform effectively free supply and a stronger economic model.

• The endpoint is jarring: roughly 15% of viewed content now comes from friends and family. Instagram and Facebook are therefore not social networks “in any meaningful sense”; they are entertainment feeds, a category Thompson considers more durable because content supply does not depend on one user’s friends continuing to post.

3. Meta’s quarter may reflect controllable dials, not superintelligence

• Investor tolerance has inverted. Reality Labs lost another roughly $5 billion, putting it near a $20 billion annual loss rate—about double the $10 billion figure discussed during the metaverse panic—yet the market’s response to Zuckerberg’s spending is effectively, “Spend what you want. It’s all good.”

• Thompson pairs incentive with what he calls a confession. Zuckerberg had an incentive to produce exceptional current results while announcing an all-in superintelligence push, while CFO Susan Li has explained that Meta learned to protect the short term so investors remain aboard for the long term. Otherwise, as with Intel technology promised for 2027 or 2028, investors can simply ask, “How about I buy your stock in 2027 or 2028?”

• The quarter’s recommendation gains were real, but they came from systems such as Andromeda, a sophisticated model built on NVIDIA GPUs—not an LLM and not the superintelligence under discussion. Better relevance increased engagement, while price per ad held up better than the surge in impressions might ordinarily imply.

• Thompson’s detective-story inference remains explicitly hedged: the cited shift toward 80% video, endlessly scrollable short-form content and “optimized ad load” may mean Meta increased both time spent and ads shown during the ideal quarter to do so. That is not proof of manipulation; it demonstrates that “a company that has dials is a powerful company.”

4. Facebook died when connection became incidental to attention

• Meta’s 2016 posture was to cap ad load rather than diminish user experience, while Zuckerberg’s messaging increasingly emphasized serving individuals what they care about rather than strengthening community. Thompson’s conclusion is stark: “Facebook’s dead.”

• Sharp hears the same change in Zuckerberg’s emphasis on serving individuals what they care about. Thompson goes further: the old language of connectedness may have obscured a product that was always “atomizing at its core,” with AI as the natural endpoint of individualized media.

• That inversion changes the opportunity for shared publishers. Stratechery once differentiated itself as niche material against mass media; in a world of content uniquely generated for every person, Thompson thinks its value may instead be communal. Sharp’s joke lands the distinction: podcast listeners still hear the same mispronounced words.

5. AI both attacks Meta’s attention and perfects its feed

• Thompson refuses to classify AI as only threat or opportunity: it is “a little bit of column A, a little bit of column B.” Time spent talking for hours to GPT-4o companions is time unavailable to Facebook, making ChatGPT more directly competitive with Meta’s attention machine than with Google’s ideal role as a launching pad to other sites.

• The opportunity is the ultimate feed progression: first friends and family, then user-generated material from anywhere, eventually any content an individual might want generated on demand. Meta’s years of personalization make this less a pivot than the completion of an existing trajectory.

• Zuckerberg’s strongest framing, in Thompson’s view, is that Meta already operates elite agents. An advertiser says, “I want more sales,” and the ad system autonomously finds customers; the news feed likewise searches for personally relevant content. Whether either uses an LLM is merely “a technical implementation detail.”

6. OpenAI needs consumer conviction more than Twitter approval

• Consumer leadership is unforgiving at ChatGPT’s scale: a change that harms just 1% of one billion users produces 10 million angry people. Their visibility does not establish representativeness, yet their volume can make a successful decision feel catastrophic.

• Thompson contrasts OpenAI with Facebook’s defining News Feed launch. Users organized online petitions and physical protests outside its Palo Alto office; Facebook issued conciliatory language but changed nothing because usage showed people loved the feed. Instagram later held its algorithmic timeline after seeing engagement rise by roughly 50%.

• His own Stratechery lesson was that the same small group kept replying online while readership expanded rapidly, creating a structural mismatch between feedback and actual consumption. If OpenAI’s reversal followed usage data, Thompson accepts it; if it followed angry posts, he worries about its “constitutional capability to be an effective consumer app.”

7. GPT-5’s automatic router solved the mainstream-user problem

• Sharp embodies OpenAI’s intended mass-market customer: he pays for Plus, opens ChatGPT, asks a question and accepts whatever model appears. The old picker—GPT-4o, GPT-4.5, o3, o3 Pro, o4 Mini, GPT-4.1 and GPT-4.1 Mini—was, to him, “Sanskrit.”

• Ethan Mollick’s explanation supplies the product logic: GPT-5 is less one model than a switch selecting among models, sizes and reasoning budgets. Users stuck on default GPT-4o could finally see what a reasoner accomplishes without understanding why o3 might be more capable despite its lower-looking number.

• Thompson, an almost exclusive o3 user, loved the original simplicity of GPT-5 and GPT-5 Thinking. OpenAI’s quick expansion to Auto, Fast, Thinking Mini, Thinking and Pro preserved routing but weakened the clean default—an aesthetically small reversal that Sharp considers strategically revealing.

8. Cutting Plus capacity gave every other complaint legitimacy

• The substantive launch failure was a reduction in paid capability. Thompson estimates a diligent Plus user could previously distribute roughly 2,900 weekly thinking queries across o3, o4 and o4 Mini variants; GPT-5 initially left that user with 200 Thinking queries and whatever additional reasoning the automatic router happened to grant.

• “That was crap behavior,” because OpenAI removed value from a $20 monthly subscription. Thompson calls taking capability from paying customers “anathema,” while noting that the most intensive Plus complainants may really have been Pro-level users trying to avoid the higher price.

• His capacity explanation is pointed but inferential: OpenAI lacked enough GPUs while launching GPT-5 simultaneously in ChatGPT and the API. It therefore “sacrificed these nerdy Plus users on the altar of an API” exposed to intense model competition, weakening the consumer product and asset nobody else possesses.

• Routing still points toward a powerful business model. OpenAI could spend more inference on a valuable legal query, produce a superior answer and attach lawyer referrals or affiliate links; free users would receive costly reasoning only when justified. The unresolved question is whether management can actually hold the trade-offs that model requires.

9. GPT-5 is roughly six months of progress, not a promised Death Star

• Thompson’s durable middle position is that “people hyping AI are delusional, and people doubting AI are delusional.” Against claims of stagnation, he stresses that o3 arrived only about six months earlier and that GPT-5 Thinking now feels consistently, materially better—not an AGI discontinuity, but rapid improvement.

• His home-improvement test was practical: o3 produced inconsistent guidance and unreliable links, sometimes ending at 404 pages. GPT-5 Thinking followed instructions more reliably, hallucinated less and assembled five relevant YouTube videos with the specific sections needed when no single video covered the whole job.

• Sharp nonetheless assigns OpenAI responsibility for the disappointment. Altman posted a Death Star looming over Earth immediately before launch, after years of GPT-5 speculation; users expecting taxes completed, a polished app generated on the first attempt or novel scientific discoveries understandably felt they had been “promised flying cars” and received another app.

• Thompson sees enormous product overhang regardless: even if models stopped improving today, current capabilities could set up ten years of technology development. A listener argued that OpenAI’s research decline began with Ilya’s departure; Thompson did not declare Meta’s talent war settled and emphasized ChatGPT, the consumer asset Zuckerberg most wants.

10. TikTok vibes may matter even when model-Twitter does not

• Sharp’s prescription is blunt: “Double down, don’t apologize, ignore Twitter.” Thompson largely agrees, especially because model obsessives who cycle among Claude and every new release do not resemble people who may not even know Anthropic exists.

• He gives OpenAI one caveat unavailable to Facebook’s earlier leaders. Modern short-form video can turn “GPT-5 is bad” into a mass-market meme; his son encountered that verdict without necessarily testing the product, and negative vibes can snowball into declining usage. “Twitter doesn’t matter,” Sharp concludes; Thompson answers, “TikTok might matter.”

• The deciding evidence remains unavailable. A reversal could reflect normies leaving, power users shouting or executives fearing a self-fulfilling narrative; neither host claims to know. Sharp’s skepticism about data-driven reversals comes from OpenAI’s conspicuous attention to online conversation, including Altman’s stream of posts and the company’s repeated changes within days.

11. Model personality became product continuity for companions

• One listener improved ChatGPT by storing a “helpfully contentious” instruction: stop praising him, identify flawed assumptions and correct misunderstandings. When shown dinner ingredients, ChatGPT responded that his supposed serrano was a poblano and would add almost no heat—exactly the useful contradiction he wanted.

• Thompson uses custom instructions for terseness, prior knowledge before search and no reminders that the system is an AI. They help, but they do not erase a model’s deep tendencies; despite repeated demands, ChatGPT continued offering cloying compliments and affirmation.

• Daniel Gross’s early analogy was that AI companies need rare “personality designers” just as computing needed exceptional interface designers. Claude 3.5 Sonnet had a distinct appeal absent from Anthropic’s larger model, suggesting personality can differ not only among labs but among models within one family.

• For developers, relatively drop-in model replacement is good because surrounding scaffolding commoditizes the underlying supplier. ChatGPT faced the opposite problem: “the product is the model” for people using it as companion or therapist, so retiring GPT-4o meant “they changed their friend.” A future personality selector—and whether it can recreate traits such as GPT-4-powered Sydney’s antagonism—will test how controllable post-training really is.

12. The AI bubble is real, but its timing remains unknowable

• Thompson thinks a bubble is “almost certainly” forming because transformative technologies requiring vast capital reliably produce one—railways, ships, electricity and the internet all did. But calling the dot-com bubble in 1996 would still have left years of upside: “You’re not right unless you get the timing right.”

• Present demand is not imaginary. OpenAI has large, fast-growing revenue, and Anthropic has reached similarly large and rapidly growing revenue; Microsoft’s AI business and Google Cloud are expanding, while providers remain capacity-constrained. The uncertainty is how much demand represents durable deployment versus experiments or circular spending—Microsoft revenue downstream of ChatGPT, for example, or Anthropic revenue concentrated in Cursor.

• Thompson’s empty Amazon envelope became the clean deployment specimen. A language-model chat accepted “I got nothing in it” and immediately shipped a replacement, restoring the frictionless service Amazon once delivered with humans before scale inserted cumbersome workflows. It was cheaper or at least better enough to create a happy customer who retold the experience publicly.

• The dot-com era itself contained two bubbles: speculative internet companies and telecom operators laying broadband. Thompson says the telecom companies, including WorldCom, absorbed the deeper economic losses, but their uneconomic fiber became the durable foundation Google and the broader internet later exploited.

13. Excess power would make the AI bust productive

• GPUs cannot play fiber’s enduring role because they depreciate too quickly. Thompson distinguishes roughly five-year GPU accounting—which may itself be too long—from data centers depreciated around 30 years; buildings and electrical infrastructure can host multiple chip generations, but compute hardware rapidly ages.

• Sharp identifies the better analogue: power. If an overbuild drives developers bankrupt but leaves abundant electricity behind, the pain could resemble telecom’s collapse while furnishing the next era’s essential input. “This will be a productive bubble if we get power.”

• Thompson cites gas prices falling about 10% while electricity rose roughly 10–15% year over year. Meta, OpenAI, Stargate and Microsoft will buy power at almost any price, letting utilities pass scarcity costs to everyone else unless generation expands. Against that constraint, NVIDIA export controls are “a total sideshow”: “China’s gonna have enough power. Are we?”

• Deregulation and construction under the Trump administration were hopes, not accomplishments the hosts could yet identify. Thompson’s darkly optimistic target is “one of the most productive bubbles in history”: enough capital chasing generation that many builders fail, leaving society with cheap excess capacity.

14. Apple, BlackBerry and Bell Labs expose where durable value lives

• Thompson separates advice to Apple from what benefits markets. Apple might rationally buy a model company, while the broader US could benefit from less dominance by a manufacturer deeply reliant on China. His default preference is often that incumbents return cash through buybacks so investors can fund startups better structured to innovate.

• Apple complicates that rule through a 40-year record of making categories happen. Thompson’s answer to whether historical conditions or singular leaders matter is “yes”: Xerox PARC’s GUI, piracy and digital music prepared the ground, but Apple refined those trends. Vision Pro reverses the sequence—“All the Apple parts” are excellent, yet no compelling market hole surrounds them, and even Thompson had not worn his in weeks.

• A cloud-centered AI world could produce an explosion of keyboards, glasses, earpieces and nostalgic BlackBerry-like clients because hardware becomes an accessory to the intelligence layer. That same modularity destroys BlackBerry-scale differentiation and margins; Thompson predicts Sharp would buy the keyboard device and abandon it “plus or minus three and a half days,” headphone jack notwithstanding.

• Bell Labs captures the institutional tension. AT&T’s monopoly suppressed or delayed products, but funded world-class research; antitrust-driven patent access diffused Unix and other inventions, while IBM’s pre-emptive hardware/software split similarly mattered more than its eventual case. Sharp wants “purple” tension, not monopoly worship or absolutism—and Google’s slime-mold organization supplies the warning: autonomous fiefdoms can come together to solve hard problems, but move slowly and are “long-lasting and hard to kill.”

Verification Notes

• The transcript uses both “about six months” and, later, “six weeks” for the o3-to-GPT-5 interval; this digest follows the repeated roughly six-month framing.