The Latest in AI: Job Loss, Elon & Sam Altman Chip Race & the "AI Bubble" w/ Brian (Blitzy) & Emad

Summary

• AI’s demand is real enough that the panel rejects the bubble analogy, even while conceding Nvidia is “priced to perfection.” Unlike Cisco’s dot-com-era price surge without matching earnings, Nvidia’s stock and forward EPS have risen together; Elliott argued that every GPU OpenAI uses will be booked because “AI is useful” and produces economic value. Mostaque’s distinction was latency: internet capex took years to monetize, while AI infrastructure can lift earnings almost immediately.

• Compute, power, and construction—not model demand—are becoming the binding constraints. OpenAI’s proposed 10-gigawatt build represents roughly 4–5 million GPUs, Nvidia’s cited $100 billion commitment equals half a normal year of US venture investment, and data-center capacity is forecast to rise from 44 GW to 156 GW by 2030 even as demand was said to be growing 10x annually. The emerging economy is “converting electrons into intelligence,” with “abundance everywhere except compute scarcity.”

• The labor outcome looks more like smaller organizations and displaced workers than a universal three-day week. Mostaque predicted AI could address roughly 50% of economic labor within a year and said humans will have “negative value in cognitive labor in a few years” when they slow teams of tireless, better-informed agents. He suggested job programs and public-sector expansion might preserve income, structure, and identity.

• Alphabet’s distribution and vertical integration make it the panel’s strongest incumbent contender. Gemini reportedly passed ChatGPT in US iOS rankings while ChatGPT remained far ahead globally, and prediction markets cited on the show put Google at 99% to lead by the end of September and Alibaba’s Qwen at 91% to rank second. Google combines reach, DeepMind talent, cash, and mature TPUs that Blundin estimated are “probably five times more power efficient” than Nvidia chips for relevant workloads.

• Higher education’s economic moat is collapsing toward admission prestige and networks. The share of Americans calling college very important fell from 75% in 2010 to 35%, while tuition was cited as up 180% since 2005 and almost 900% since 1983. Elite endowment-rich institutions may remain insulated, but schools numbered roughly 40–400 face a squeeze as AI education, alternative credentials, and weak graduate hiring expose curricula that can change more slowly than “build a nuclear reactor on campus.”

• The entrepreneurial edge lies in converting proprietary domain knowledge into owned workflows, efficient models, and scalable applications. Blundin warned that merely selling expertise for model training could leave an expert valuable for “a month or two”; Elliott instead favors companies built around regulatory or vertical knowledge, while Mostaque emphasized the human who understands context and “gives a damn.” Task-specific data, distillation, and verifiers could produce the same result with 1% of the parameters and compute—a claimed 100x cost advantage.

• AI infrastructure links the solar, battery, semiconductor, and robotics theses into one industrial race that China currently scales faster. The panel cited China at 880 GW of solar capacity in 2024, growing 45.6%, versus 177 GW and 27% growth in the US; Blundin argued America repeatedly invents technologies but fails to finance their scale. Robot projections ranged from one billion to 10 billion units by 2040, making even the low case worth $25 trillion at $25,000 per robot—far above Morgan Stanley’s cited $5 trillion estimate for 2050.

• Tokenization could repair public-market access while also creating the episode’s likeliest genuine bubble. Nasdaq was described as targeting tokenized trading by late 2026, while Robinhood’s EU platform already offered roughly 200 US stock tokens plus private-company exposure to OpenAI and SpaceX. Mostaque expects digital assets—not generative AI—to display unmistakable bubble behavior as legal clarity brings corporate blockchains, continuous markets, and eventually agent-directed trading.

Deep dive

1. College’s moat has shrunk to admission prestige and networks

• Diamandis opened with a collapse in perceived value: Americans calling college very important fell from 75% in 2010 to 35%, while “not too important” rose from 5% to 24%. Tuition, meanwhile, was cited as up 180% since 2005 and nearly 900% since 1983.

• Blundin’s diagnosis was institutional latency: available knowledge is advancing faster than curricula can absorb it. One MIT insider had reportedly joked, “We can build a nuclear reactor on campus faster than we will ever change this curriculum,” leaving graduates indebted for material that may already be obsolete.

• Elliott argued that elite college has long functioned primarily as credentialing: getting into MIT proves something even if its curriculum is freely available elsewhere. Y Combinator, a strong portfolio, or becoming an MIT dropout can now supply alternative signals—prompting the joke that MIT should accept students without expecting them to attend.

• The economics bifurcate sharply. Diamandis said top schools receive more than twice as much budget contribution from endowment returns as from tuition, while institutions numbered roughly 40–400 desperately need enrollment; Mostaque proposed that endowments instead fund GPU clusters because research quality may soon depend on “how many GPUs you have.”

2. Distribution is moving the model rankings as fast as capability

• Gemini reportedly took the top US iOS position after viral interest around Nano Banana V3. Mostaque initially disbelieved the result, then checked the App Store data directly. ChatGPT remained “miles ahead globally,” but Google can push products through distribution just as Chrome did when it “blew away” Firefox.

• Prediction markets cited by Diamandis assigned a 40% chance that Gemini 3 would arrive by October 31, a 99% chance Google would have the best model at September’s end, and a 91% chance Alibaba’s Qwen would be second. Those prices were observations, not the panel’s own forecasts.

• Mostaque said Qwen was releasing almost daily—six models on the day of recording—and closing the frontier gap through Alibaba’s reach, data, and team. Blundin connected Qwen’s ability to keep up with reinforcement learning and its widespread use. The broader point used Threads’ cited 400 million monthly users: products can become enormous through distribution even when the panel barely encounters anyone actively choosing them.

• The host floated Grok 5 as a possible first-to-AGI model, while the concrete score discussed was Grok 4 at 15.9% on ARC-AGI v2. Blundin stressed that v2 is exceptionally hard; Elliott cared more about the x-axis—how reliably additional dollars improve task performance—than consumer popularity.

3. Benchmarks are useful signals until labs optimize for the scoreboard

• Grok 4 Fast Reasoning was said to rank first on an extended New York Times Connections benchmark: 759 puzzles, more categories than the original, and no allowance for a wrong first answer. Blundin called it a “shockingly good” general-intelligence test.

• His reservation was benchmaxing. Foundation-model companies can train directly on a celebrated test to manufacture PR, whereas Elliott had explicitly said Blitzy topped SWE-bench without tuning to it; the panel suspected optimization around Connections but conceded it could not prove that happened.

• Mostaque cited an Epoch AI projection that every current benchmark could saturate within three or four years. Blundin’s pushback—worth keeping—was that ARC-AGI v1 had already saturated, but saturating the much harder v2 would mean “you’re beyond superhuman intelligence” and “in another universe.”

• The pace itself was the signal: the discussion occurred only about one year after o1 was announced, already described jokingly as “ancient history.” Static leaderboards therefore reveal less than improvement curves, test-time spending, and cost per successful task.

4. Compute scarcity is turning data centers into economic infrastructure

• xAI’s Colossus 2 was presented as a gigawatt-scale facility with 110,000 GB200 GPUs, 119 air-cooled chillers, and Tesla Megapacks. Musk’s stated ambition was to be first to 10 GW, 100 GW, and one terawatt—scales Mostaque compared with states and national electricity systems.

• Sam Altman called Nvidia’s $100 billion commitment a “small dent” in OpenAI’s 10-GW plans. Altman described a multisquare-mile “superbrain”; Diamandis said the deal covered millions of GPUs, while Greg Brockman said a world where everyone effectively has one would require roughly 10 billion.

• Jensen Huang translated 10 GW into approximately 4–5 million GPUs—about what Nvidia sold in the prior year, and twice each of the two preceding years. Blundin noted that a single $100 billion commitment equals roughly half the entire cited $200 billion annual US venture market and could absorb 20–30% of near-term chip output.

• Lab compute was said to have tripled in one year, while data-center capacity was projected to grow from 44 GW to 156 GW by 2030. With demand reportedly rising 10x annually, providers already route queries toward the smallest sufficient model while research and self-improvement consume the premium capacity.

5. Scarce compute will be rationed by economic value per flop

• Elliott said costs would rise dramatically and that leading model providers could charge more. Mostaque framed access around the marginal dollar: an enterprise workload can pay 100x or 1,000x what a virtual-girlfriend interaction or homework query supports, so high-value customers may retain service while lower-value uses face throttling or sharply higher prices.

• Mostaque’s transition marker was experiential: GPT-4o felt like “a very smart goldfish-memory buddy” requiring constant supervision; current systems can be “set it and forget it,” process millions of tokens or lines of code, and act proactively. He estimated addressable economic labor could rise from 1–2% to perhaps 50% within a year.

• Improved data may move the efficiency frontier before exotic compute does. Mostaque cited a Qwen model from Alibaba’s Tongyi team, reporting roughly 22% on Humanity’s Last Exam with three billion active parameters, a self-reinforcing continual-learning system, and smartphone operation; he treated its data quality, not brute-force size, as the breakthrough.

• Diamandis’s entrepreneurial warning was concrete: task-specific datasets, distillation, and the right verifier can make execution 100x cheaper. “Don’t miss your opportunity to reserve your compute”; just as every suitable mountain-and-lake site for pumped hydro had reportedly been bought, scarce capacity may be locked up before builders realize it is no longer a utility.

6. Scale rewards businesses that win when every model improves

• Elliott’s infrastructure posture is model- and provider-agnostic: “You really want to be a player where everybody else wins when you win.” That does not eliminate scarcity risk; it means becoming important enough to suppliers that the company is not their fifth- or sixth-priority customer.

• Diamandis framed specialized efficiency as a defensible barrier: match a frontier result with 1% of the parameters and 1% of the compute inside a valuable domain. The cloud era trained founders to assume capacity would always appear after entering a credit card; frontier compute invalidates that assumption.

• Blundin compared the opening to Dropbox using S3. Many storage companies existed, but Dropbox’s architecture made it roughly 10x cheaper and allowed rapid scale; application companies can similarly compound an underlying infrastructure advantage into distribution and product power.

• Mostaque also pointed downstream: suppliers control a scarce marginal input, AI adopters with pricing power can expand margins by replacing labor, and human attention may become one of the few remaining scarcities. That led him, somewhat counterintuitively, toward media and the attention economy.

7. Frontier-model finance has reached sovereign scale

• OpenAI reached a deal with Microsoft allowing restructuring toward a for-profit structure and targeting a cited $500 billion valuation, while leaving approximately $100 billion in the nonprofit. Diamandis called it potentially the world’s largest nonprofit pool; Elliott noted it would be roughly twice Harvard’s endowment.

• Microsoft’s historical investments were cited as $1 billion in 2019 and $10 billion in 2023, with an unconfirmed estimate of about 30% ownership. Yet Microsoft was reportedly notified of the Nvidia transaction only one day beforehand, illustrating how far OpenAI has moved from depending exclusively on Microsoft compute.

• The panel noted that figures once considered fantastical now barely register because OpenAI may ultimately support a trillion-dollar buildout. Mostaque cited projections toward $200 billion of revenue, including roughly $80 billion from a new AI-agents line and another $20–30 billion from other new activity.

• Zuckerberg’s answer is a committed $600 billion of US data-center investment by 2028 because being late to superintelligence is worse than losing billions. Mostaque’s narrowed field was xAI, Google, OpenAI, and probably Meta; Blundin’s concern was the unprecedented concentration of capital and consequential decisions in very few hands.

8. Recursive AI turns research ideas into another compute workload

• Dario Amodei’s claim was that Claude already plays “a very active role in designing the next Claude.” The loop cannot yet fully close and is “not yet going super fast,” but it has started—the panel’s proposed takeoff mechanism.

• Mostaque cited Tri Dao, creator of FlashAttention, saying Claude Code made him at least 50% more productive; FlashAttention itself had reportedly improved performance by about 30%. AI has also helped design TPUs, pointing toward integrated feedback from silicon and kernels through training and the next model generation.

• Altman’s stated ambition of one new gigawatt of compute each week would support that vertically integrated loop. Mostaque expects “data centers of geniuses checking each other’s work in parallel,” including thousands of Lean provers and systems capable of formalizing difficult mathematical arguments in days.

• Blundin recalled OpenAI’s Noam Brown saying core research was already gated by compute, not researchers: the lab had more ideas than it could test. The remaining idea-constrained window might last a year or two; once AI generates the hypotheses too, the backlog becomes almost purely physical capacity.

9. Alphabet’s full stack is its strategic advantage

• Alphabet had reached a cited $3 trillion market capitalization, with shares up 33% in 2025 and 55% over the prior year. Mostaque’s bull case combined Google’s distribution, DeepMind talent, cash generation, proprietary infrastructure, and ability to avoid the “Nvidia tax.”

• Google had spent years developing TPUs before OpenAI began its reported Broadcom chip effort. Mostaque said Stability AI had used thousands of TPUs; Blundin estimated TPUs were “probably five times more power efficient” than Nvidia alternatives and had better interconnects for large-context models.

• The unresolved question was whether Google would commercialize them. Blundin said they might start selling them soon, while Diamandis noted that Google had effectively pulled them from the market because it was using them internally.

10. Long-horizon agency already exists above the model layer

• Microsoft AI CEO Mustafa Suleyman described current models as one-shot prediction engines unable to plan over time, then forecast action across effectively unlimited horizons by the end of the following year. He called today’s pocket intelligence “magic” whose novelty society already dismisses.

• Blundin’s pushback was that planning had improved so rapidly the clip might already be stale. A Tesla traveling across America would itself demonstrate planning; Diamandis added that a vision system could observe, take notes, and write while the vehicle drives.

• Elliott resolved the apparent disagreement at the application layer: a single model may not execute a long plan reliably, but an orchestrated set of models can create “AGI-type effects.” Users care about the delivered experience, not where cognition resides, and software engineering already exhibits this behavior.

• Diamandis’s Replit example made the accessibility concrete: using Starlink while his airplane flew on autopilot, he built a mindset application during the flight. Elliott separated that disposable prototype from enterprise software, where concurrency, caching, reliability, and hundreds of thousands of users still demand a different system design.

11. Domain experts gain leverage by owning the implementation

• Replit CEO Amjad Masad’s hypothetical was a world-class lawyer who withholds rare knowledge from open datasets, embeds it in a specialized agent, and scales the service. Blundin liked the mechanism but challenged the economics: after a month or two of extraction, what prevents the expert from becoming unnecessary?

• Diamandis’s answer was ownership—start a company around a deeply understood problem, particularly where regulation or specialized vertical knowledge creates a barrier. “There’s no barrier to starting a company”; the opportunity is not limited to 21-year-old founders.

• Elliott was more bullish on the 45-year-old operator. Software expresses business processes, flows, decisions, and market-specific pricing—not merely technical architecture—so an insurance underwriter or financial-product specialist can now create enterprise systems that were previously beyond reach.

• Mostaque’s differentiator was “giving a damn.” He borrowed Nassim Taleb’s “intellectual yet idiot” framing for well-credentialed actors without skin in the game and applied it to AI: models lack lived concern, while experienced humans can translate context, navigate organizations, reassure customers, and carry adoption from early users into the “vast middle.”

12. Ambient AI converts work and government into training data

• Amazon’s AI glasses, codenamed Jayhawk, were described as targeting late 2026 or early 2027, with a consumer product and a workforce version. The company reportedly planned 100,000 pilot units by Q2 across 390,000 drivers, capturing operational data that could train future robots.

• Mostaque called the loop “kind of obvious”: glasses record physical work, while Slack messages and code commits can help create virtual versions of knowledge workers. As glasses become light and useful, Diamandis expects continuous recording to move from discomfort to assumption, with privacy becoming “a long-lost concept.”

• Albania’s AI-made minister was presented as an anti-corruption mechanism for public tenders. Elliott said an 80%-accurate system could beat a status quo in which mistakes deliberately enrich relatives; Mostaque argued that even an imperfect system would not deliberately demand a bribe or kickback.

• Diamandis kept the governance objection intact: whoever chooses the training data or controls the data center may shape the minister. Mostaque highlighted the uncanny launch statement that the AI was “very disappointed” by public reaction—either a human scripted the emotion, or the system’s apparent disappointment creates a different problem.

13. The energy contest is ultimately a manufacturing contest

• Diamandis rejected the US energy secretary’s proposed 50-year bet that solar never reaches 10% of global energy. He cited 18 GW installed in the first half of 2025, solar providing 50% of new US generating capacity in that period and 69% in Q1, plus NREL’s projection of at least 40% of US electricity by 2035.

• Diamandis described intermittency as solar’s historic weakness but storage as solved or nearly solved, with continuing improvement. The harder issue is strategic dependency: America’s strongest scalable answer to Chinese energy expansion still relies heavily on a China-linked solar supply chain.

• The panel cited China at 880 GW of solar in 2024, growing 45.6% annually, against 177 GW and 27% in the US. Elliott warned that China’s lead could widen once robots build the factories that manufacture more robots, batteries, panels, and industrial capacity.

• Blundin’s structural complaint was that US venture economics favor companies needing hundreds of thousands, then a few million dollars—not automotive, solar, or energy manufacturing. America invents, others finance scale and drive down cost, then the products return as imports; a $200 billion annual venture industry cannot close that gap.

14. Robotics forecasts ignore self-improvement and self-manufacture

• Diamandis corrected an earlier error: Figure raised $1 billion at a $39 billion valuation, not $93 billion. Its Brookfield partnership provides access to 100,000 homes and 500 million square feet of offices and logistics space, supplying diverse environments rather than repetitive factory motions.

• Blundin disagreed that embodiment is required for AGI, but accepted that a system must experience dropped toys, collisions, and household disorder to understand daily life empathetically. Mostaque added that video-based physics inference is insufficient for reliable real-world action, motivating proprietary robotics data and foundation models.

• OpenAI’s renewed hiring across teleoperation, simulation, and mechanical engineering signaled its return after pausing robotics around 2021 to focus on ChatGPT. The panel treated physical AI as a multitrillion-dollar extension of the model race, not a separate category.

• Morgan Stanley’s cited $5 trillion market estimate for 2050 looked too conservative beside forecasts of one billion to 10 billion robots by 2040. At $25,000 each, the one-billion low case alone implies $25 trillion; Elliott compared linear estimates to defining Uber’s addressable market as the existing taxi business.

15. Real AI revenue coexists with labor and market-structure shocks

• The anti-bubble exhibit compared Cisco’s dot-com price rise against flat forward earnings with Nvidia’s price moving alongside forward EPS. Elliott’s test was plain: “AI is useful. People pay for it because it has economic value,” so Nvidia-funded OpenAI purchases may resemble circular financing without sharing the old absence of demand.

• Mostaque noted that internet value arrived after capex latency, whereas AI monetization can be immediate. Diamandis said Nvidia was priced to perfection but grounded in real revenue and earnings; the company’s cited $4.5 trillion valuation leaves little room for disappointment.

• Diamandis remembered MicroStrategy near a $14 billion valuation with negligible revenue and Yahoo rising from a $300 million IPO valuation to roughly $110–120 billion before falling 95%. Dave Blundin added that the internet bubble ultimately produced strong investments after confidence returned.

• Blundin’s tradeable specimen was Better Mortgage, ticker BETR: a thinly traded microcap where AI workflows and voices reportedly produced a rapid operational lift. His screen was companies with large consumer flows, then management teams capable of “AI-ing” those workflows before competitors—not a recommendation, as the panel explicitly cautioned.

• Nasdaq’s tokenized-securities proposal could begin trading by late 2026, while Robinhood’s EU service offered roughly 200 US stock tokens and private-company exposure. Mostaque expects digital assets to show what a genuine bubble looks like; the panel nevertheless argued blockchain reporting could shorten the onerous gap between private growth and today’s very large IPO threshold.

16. A shorter workweek does not solve human displacement

• Zoom’s Eric Yuan projected a three-day week; Bill Gates suggested two to three days within a decade, Jensen Huang four, and Jamie Dimon 3.5. Elliott saw no competitive equilibrium between full utilization and elimination: if human input creates value, companies will demand more of it; if AI does better, they need fewer humans.

• Mostaque was categorical that humans will have “negative value in cognitive labor in a few years”: they become the least capable team member beside agents that never sleep, learn from mistakes, and absorb 10 million tokens at once. Because jobs supply identity and structure, he anticipates public programs, expanded government employment, taxation, and social support.

• Diamandis illustrated the negative-value loop with cited diagnostic results: physicians scored 74%, physicians using GPT-4 reached 76%, and GPT-4 alone reached 92%. Blundin added that universal Waymo-level driving could save 40,000 US lives and $1 trillion in social costs annually, making human control the riskier input.

• Duolingo reported four-to-fivefold productivity, no full-time layoffs, and a revenue forecast raised from $996 million to $1.02 billion. Mostaque’s reading was less comforting: a company growing 30–40% should normally expand headcount similarly, so standing still already represents displacement; abundance may arrive, but “there is no mechanism right now for distributing it.”

17. AI health shifts medicine from population averages to individual systems

• Apple Watch’s FDA-cleared hypertension alert addressed a condition said to affect 1.5 billion adults, 30–45% of the population and 60% of those over 60. The panel cited 46% going undiagnosed and only 21% being controlled, making continuous detection valuable before symptoms appear.

• Diamandis described combining Apple Watch, Oura Ring, and continuous-glucose-monitor data inside his own health AI. The intended advantage is longitudinal questioning—whether sleep or glucose changed alongside a specific medicine or supplement—rather than isolated measurements and population averages.

• Mostaque argued AI can integrate personal health data and subjective trial reports from first principles. Beyond discovering compounds, he expects major returns from repurposing existing drugs because models can compress and analyze anecdotes and rich individual data that conventional trials barely capture.

• The episode cited an AI-designed idiopathic-pulmonary-fibrosis drug already in human trials, DSP-1181 reaching trials for obsessive-compulsive disease in 12 months rather than four to five years, 150 AI-first small molecules discovered in 2025, and at least 21 successful Phase 1 completions at an 80–90% rate. Mostaque expects in-silico predictions to enter FDA processes within five years, if regulation permits.