Apple’s AI Crisis & ChatGPT’s World Domination w/ Dave Blundin & Salim Ismail | EP #162
Summary
At $300 billion, OpenAI is either a cheap claim on a multi-trillion-dollar consumer gateway or an expensive bet on a commoditizing model layer. Dave Blundin says AI could create $20 trillion of impact within a couple of years, but DeepSeek’s disruption and constant benchmark leapfrogging make any single foundation-model winner uncertain: “This is either really cheap or really expensive.” Salim Ismail is personally out, citing competition and talent drain, while conceding that betting against Masa has historically been risky.
The durable foundation-model prize may be distribution rather than technical supremacy. Blundin sees ChatGPT and Grok trying to replace Google as consumers’ first stop online; against a company he describes as worth a couple of trillion dollars with roughly $300 billion in annual cash flow, a model need not remain best forever if it owns the habit. Combining X with xAI adds real-time data, while SoftBank could rationally accept only break-even or 2X on OpenAI if the position wins access to “20, 30, 40 other investments” across its application ecosystem.
For venture investors, the application layer looks more attractive than trying to select one model champion. Blundin calls vertical AI “the best investment theme I’ve ever seen by such a wide margin,” and his MIT-adjacent incubator has increased annual backing from roughly 10 companies to 30 because everything is moving three times faster. Founders use multiple APIs in parallel and keep Llama 3 or DeepSeek locally as a control point because model leadership changes quickly and hosted services can become overloaded.
AI is collapsing the time, headcount, and experience once required to build a multi-billion-dollar company. Mercor, led by recently turned 21-year-old Brendan Foody, reportedly reached a multi-billion-dollar valuation within two years—something Blundin says had happened only five or six times historically—by using AI interviewing for technical recruiting, with Meta cited as a likely major customer. A 15- or 20-person team can now avoid the thousand-person management hurdle, making a two- or three-person multi-billion-dollar company plausible “sometime very soon.”
Apple’s AI failure and education’s resistance share the same incumbent problem: institutional control systems cannot match AI’s release cycle. Ismail contrasts Yahoo’s eight-month feature process with Facebook developers shipping weekly, arguing that consumer internet competition demands “speed and risk”; Blundin calls Apple’s deterioration “pathetic” but sees talent turnover as the opening startups need. In education, Beijing and Estonia were moving AI into schools while U.S. education systems still often treated it as cheating—a legacy “immune system” protecting curricula and jobs.
Coding is shifting from inspecting software to directing its creation, even while the model race remains unsettled. Blundin says industrial-strength work still required conventional coding at the time of recording but predicted “game over within 2025”; Blitzy’s claimed three million lines in one night represented work he valued at $300 million and three years, making manual debugging nonsensical. Gemini 2.5 Pro approached 20% on Humanity’s Last Exam, yet Ismail predicts the leader two years later will be an unknown entrant rather than Google, OpenAI, Meta, Elon Musk, or DeepSeek. On safety, Diamandis relayed Eric Schmidt’s concern about dystopian users and bioweapons, while Blundin recalled Sam Altman warning that a hypercompetitive race would be the worst outcome; Diamandis argued the current roughly $1 billion-a-day race is already unconstrained.
AI is moving attention and scientific discovery from passive consumption and expert silos toward interactive generation and pattern search. ChatGPT’s image feature reportedly drew one million sign-ups in 60 minutes, while Blundin expects interactive AI and companionship—the “number one use case”—to drain audiences from conventional media. In resources, the Goldcorp precedent carried the argument: releasing proprietary geological data for a $500,000 contest helped identify six million ounces of gold and accompanied a market-cap rise from $50 million to $3 billion because “the answer is in the data.”
Scarce compute, regulatory permission, and capital-market liquidity become more valuable as software itself gets cheaper. China’s commercial flying-taxi approval, transferable humanoid-robot skills, and idle data-center capacity all point to infrastructure as the constraint; Blundin even connects Bitcoin mining to monetizing unused power and compute. CoreWeave’s debt-heavy IPO—priced at $40 after targeting $48, initially falling, then moving well above issue price—could be a “Yahoo moment” that reopens IPO and M&A exits for AI companies.
Deep dive
1. OpenAI’s $300 billion valuation prices a revolution and a winner-take-most gamble
Diamandis’s opening question was blunt: in or out at $300 billion? Blundin’s answer separated the sector from the security—the AI revolution “totally justifies the price,” but nobody yet knows whether foundation models themselves capture the value.
DeepSeek’s impact on Nvidia illustrated the fragility of model leadership, while Gemini’s subsequent benchmark lead showed how quickly the hierarchy reverses. Blundin nevertheless expects roughly $20 trillion of economic impact within a couple of years: “This is either really cheap or really expensive, depending on whether it’s the winner.”
Ismail was personally out. His bear case combined intense competition with OpenAI’s talent drain; Blundin sharpened it by noting that Ilya Sutskever and Mira Murati were launching new foundation-model companies, with Blundin correcting Sutskever’s cited valuation to $30 billion.
The counter-case is capital supremacy: if OpenAI has $40 billion, perhaps half could go directly to compute. Blundin framed even a medium-to-low probability of victory as investable when the upside is measured in trillions, while Ismail noted that abundant private funding removes any immediate need to go public.
2. Consumer habit may matter more than owning the best model forever
Ismail read the X–xAI merger as an easy way for Elon Musk to limit downside while combining AI with a massive proprietary dataset. Diamandis, already an xAI investor, called it “a pretty spectacular move.”
Blundin’s central thesis was that foundation-model companies have two paths to fortunes, and Sam Altman and Musk are pursuing the consumer path: displace Google and become “the first place they go on the internet.” ChatGPT and Grok gain strategic value from habit; X adds current information for news, sports, and other real-time queries.
Anthropic, in Blundin’s reading, is not primarily playing that consumer-brand game, while Perplexity is. A model does not need to remain technically best forever if it owns consumer distribution against Google, which he characterized as a couple-trillion-dollar company producing roughly $300 billion a year in near-pure cash flow.
3. Ecosystem access can justify an anchor investment with modest direct returns
Diamandis argued that investing $40 billion at OpenAI’s valuation seemingly requires belief in a multi-trillion-dollar outcome and perhaps 10X upside. Blundin pushed back: SoftBank’s economics may include the deals its OpenAI relationship unlocks, not only the return on OpenAI itself.
His scenario allows the anchor investment to break even or return 2X while creating preferred access to “20, 30, 40 other investments” in the surrounding ecosystem. Diamandis inferred that SoftBank and Microsoft would want the application layer built around OpenAI.
That layer is Blundin’s strongest call: vertical applications represent “the best investment theme I’ve ever seen by such a wide margin.” His incubator’s pace has risen from about 10 investments annually to 30 because “everything’s three times faster and bigger than I’ve seen before.”
Founders are not expressing comparable loyalty to any model. They run multiple APIs, especially for code generation, and debate whether to keep Llama 3 or DeepSeek on their own hardware as protection against latency, overload, and sudden performance leapfrogs; Ismail therefore considers the foundation layer already commoditized.
4. India shows how cheap models turn local applications into the real race
European conversations left Ismail impressed not only by U.S. technology but by “cracks at the piñata” emerging from China and India, particularly open-source and DeepSeek-like work. His emphasis was less on Indian foundation models than on aggressive deployment.
After predicting that a complete AI doctor would arrive toward the end of the year, Ismail received three responses—two from India—saying, “We’ve already got a fully fledged AI doctor live and running. Here’s the link.”
The causal chain was low access cost to experimentation, then adaptation to local needs, then rapid deployment. Diamandis added that a country of 1.41 billion people effectively needs AI to democratize healthcare and education; for investors, adoption velocity may matter more than benchmark leadership.
5. Mercor demonstrates how AI removes the managerial age barrier
Mercor’s $100 million raise mattered because recently turned 21-year-old CEO Brendan Foody had helped build a multi-billion-dollar company in under two years. Blundin’s ChatGPT-assisted research—“maybe not totally reliable”—found only five or six historical precedents, though he expects multiples of that number within two more years.
The old constraint was not youthful intelligence but the experience required to manage hundreds or a thousand people. With AI as workforce and multiplier, Blundin argues that 15 or 20 employees can reach a scale manageable by a gifted 21-year-old; “a company with two or three people” could reach a multi-billion-dollar valuation soon.
Mercor uses an AI face and voice to interview candidates globally, particularly for technical jobs, more effectively in Blundin’s telling than traditional HR departments. Meta was cited as a likely major customer; his contrast was a bank whose HR team would recoil from the idea rather than run with it.
Blundin called the broader mechanism “un-hobbling”: accelerators supply credibility, accounting, legal work, data, networks, capital, and even housing. He wrote a $5.5 million check for an MIT-area apartment building after calculating that saving Mercor one week—while its valuation rose about $20 million weekly—could justify the asset: “Don’t overthink it. Just run like hell.”
6. Generative media is converting passive audiences into participants
ChatGPT’s viral image generator reportedly brought one million users in 60 minutes. Blundin treated the fad itself as temporary but the attention transfer as permanent: each interactive breakthrough takes eyeballs from something passive until an even more compelling interactive product replaces it.
His analogy came from consulting for Washington Post owner Don Graham while the internet consumed newspaper audiences. The same “relentless onslaught” now reaches film and television because people prefer to interact, engage, and create rather than only watch.
Diamandis cited Musk’s expectation that short video movies would soon dominate and another post predicting feature-length work the following year. His envisioned Netflix successor carries vast crowdsourced libraries whose audience rankings compound the disruption to Hollywood.
Blundin called companionship AI’s number-one use case. Ismail said simulated empathy already gets “99.9% of the way there”; Blundin expected highly empathetic voices and top LLM reasoning to converge within one or two months, once inference latency for trillion- or two-trillion-parameter models was managed.
7. Better pattern search turns apparent resource scarcity into data abundance
Earth AI’s reported discovery of copper, gold, and silver in overlooked Australian regions combined historical geological records with targeted drilling. Ismail connected it to a prior figure: roughly 40% of oil and gas deposits found over 20 years had used digital search and new sensors.
The signature precedent was Rob McEwen’s Goldcorp challenge. After internal experts could not locate the next six million ounces, he released normally secret geological data on physical media and offered roughly $500,000 to outside teams.
Remote teams reportedly found the deposits without visiting the mine by transferring pattern recognition from metals such as nickel into gold geology. Gold specialists had not cross-mapped those datasets; Ismail said the company’s market cap moved from $50 million to $3 billion after convention was broken at trivial relative cost.
Blundin linked the story to the Avatar XPRIZE: “You don’t have to go places to do things.” When the valuable observations come from nonvisual sensors, transporting the data to global analysts can matter more than transporting experts to Australia.
8. Apple’s AI crisis is the innovator’s dilemma operating at machine speed
Diamandis welcomed the “ugly and embarrassing” crisis headline with a catalog of Siri failures, including misspelling Kristen three ways inside her own message thread and rendering Peter’s name with an I. Blundin’s household refrain was harsher: “effing Apple” roughly 10 times a day.
Ismail’s defense was procedural, not technological. Large companies must pass features through systems, integration, privacy, and brand testing; at Yahoo, a release could take eight months, while Zuckerberg let developers ship directly and accept personal consequences if they broke the site, producing weekly releases.
His rule for consumer internet companies was uncompromising: “The two characteristics you better have are speed and risk.” Yet he also acknowledged Apple’s uniquely hard timing problem—an AI component might be obsolete within a week, before a careful integration finishes.
Blundin still called the performance “pathetic,” comparing Apple’s decline after Jobs was booted with its recovery after his return. Diamandis located the missing asset in a single creative through-line; Blundin added that a magnetic founder attracts frontier talent, while Apple’s brand has drifted from “it just works” toward monetizing every song and movie.
9. AI education replaces standardized instruction with on-demand capability
The conversation paired a Texas AI-tutored private school reportedly reaching the country’s top 2% with Beijing’s plan to baseline AI in primary and secondary education in September 2025; Estonia had announced a similar move. Diamandis objected to treating AI as “shackles” or cheating rather than a “jetpack.”
Ismail’s son used AI to watch a math problem being solved, understand the method, and then repeat it—the traditional apprentice pattern with unlimited digital masters. His historical sequence ran from one-to-one tutors, to one-to-many classrooms, to internet-enabled many-to-many learning, and now an effectively unbounded supply of instruction.
The destination is pull-based learning: choose a project or massive transformative purpose, then retrieve the knowledge needed to execute it. Teachers’ unions become what Ismail called the institutional “immune system,” while countries willing to restructure learning gain an advantage.
Blundin argued that slide-rule instruction eventually became absurd, and AI is creating the same break far faster. Students now vibe-code with Lovable, Replit, or Cursor, while schools occupy scarce time with curricula he says have barely changed since Newton despite human knowledge expanding perhaps “a million X.”
10. Vibe coding changes software economics from debugging code to steering outcomes
Asked whether traditional coding was dead, Blundin said industrial-strength software still required coding at that moment but predicted “game over within 2025.” His operating advice was to build whatever comes easily through vibe coding and avoid inspecting the generated code.
Blitzy, one of his portfolio companies, reportedly wrote three million lines in a single night. Blundin equated that output to $300 million of R&D over three years; attempting to debug it manually would negate the economic advantage: “You’re dead.”
The replacement workflow is conversational—repair accessible modules through voice, keep prompting, and wait for improving models to solve what remains. The transition rewards specification, judgment, and patience more than ownership of every implementation detail.
11. Open models, shared protocols, and uncertain AGI definitions keep the race fluid
Gemini 2.5 Pro impressed the speakers across domains and approached 20% on Humanity’s Last Exam, a benchmark Blundin described as containing questions most people could not even parse. Still, he saw its lead as another temporary leapfrog ahead of Llama 4 and subsequent releases.
Diamandis would bet on Grok, citing Musk’s capital, X and Tesla data, humanoid robots, chip access, transformers, and electricity. Ismail rejected the offered field entirely: two years later, he expects an unknown entrant to lead Humanity’s Last Exam, repeating the pattern by which Google, Facebook, and OpenAI arrived from outside prior assumptions.
Safety and the finish line were contested. Diamandis relayed Eric Schmidt’s warning that the main risk is a dystopian user, especially through bioweapons, and said AI spending is already unconstrained at about $1 billion a day. Blundin recalled Sam Altman warning that a hypercompetitive race would be the worst outcome, while Diamandis argued that the current race is already all-out.
Blundin said AGI definitions are disputed. Demis Hassabis’s “artificial Einstein” test—using information available through about 1910 and asking an AI to rediscover relativity—offers a concrete benchmark, while Blundin opposed moving the goalposts after the Turing test. He favored systems that are not self-aware but are super-capable across domains useful to humanity.
Anthropic’s Model Context Protocol offered a different competitive layer: a simple open standard through which AI can access tools and information without receiving uncontrolled access to a laptop. Blundin compared its potential durability to HTML—simple, functional, and capable of coordinating an ecosystem.
Ismail celebrated its community lineage and connection to OAuth-style interoperability work. Blundin’s dissent was worth keeping: OpenAI and Microsoft may not be adopting MCP altruistically; their reaction might be closer to “Okay, let’s steal it” than an endorsement of open ethos.
12. Physical infrastructure and capital markets become AI’s binding constraints
China’s commercial flying-taxi approval illustrated how regulation can dominate technical readiness. Ismail argued autonomous flight is easier than autonomous driving but still needs gates, fueling, and licensing; Diamandis pointed to Archer and Joby, with LA Olympics deployment planned as far as he knew.
The adjacent asset trade was unusually concrete: Blundin proposed buying cheap island property whose access problem disappears with electric vertical flight. Ismail’s group counted roughly 10,000 islands and lakes within a 200-mile radius around Toronto and Chicago airports and envisioned placing landing pads on them while waiting.
Humanoid robots offered another scale mechanism: once one unit learns a skill, Ismail said a million others can inherit it. Blundin defended the human form less as engineering destiny than recruitment strategy—it inspires a generation, builds the component and software supply chain, then enables any later form factor.
Wearable haptics extended the avatar thesis to pressure, twisting, and stretching. Ismail pointed to remote-robotics and off-planet applications, while Blundin warned that Earth–Mars latency would kill the use case. Diamandis claimed quantum communications could make distance irrelevant; Blundin explicitly said he was not deep enough in physics to know whether that was real, and Ismail said true entanglement would imply capabilities beyond haptics.
Public-market plumbing closed the episode. The Trump family’s American Bitcoin venture entered mining as U.S. stocks fell and Bitcoin dropped roughly 10%, contradicting its uncorrelated-store-of-value narrative; Blundin blamed indiscriminate ETF flows but linked mining to monetizing idle data-center power. CoreWeave’s debt-heavy $40 IPO, after targeting $48, initially fell and then moved well above issue price; its nearly 42% surge could become the “Yahoo moment” reopening IPO and M&A exits, with 2021 the last comparable year in Blundin’s view.