Emad Mostaque

Google’s memory and visual-reasoning work points to architecture—not another benchmark bump—as the next major unlock. Titans and MIRAS use “surprise” to decide what enters long-term memory, potentially moving beyond transformers’ quadratic context cost; the panel framed 2 million tokens as roughly 3,000 pages or 16 novels. Visual tokens added to chains of thought produced stated reasoning gains of 3% to 6%, supporting the prospect of models that understand everything from screens and medical images to an always-on user’s surroundings.

Chinese openness is filling a research vacuum created as U.S. frontier labs stop publishing. NeurIPS drew more than 29,000 registrants, nearly 50% above the prior year, while Alibaba had 146 accepted papers and Mandarin was reportedly the language heard most in the hallways. Chinese-affiliated first authors at ICLR rose from 9% in 2021 to 30%, versus a U.S. decline from 52% to 36%; open weights become both a distribution “land grab” and a route to deeper social, economic and industrial integration.

The frontier-model contest has become a capital-intensive weekly leapfrog, with safety exposed to the logic of “a rat race.” An explicitly unverified leak put GPT-5.2 as arriving as early as the following week and at 67.4% on Humanity’s Last Exam, versus Gemini 3 Pro at 37.5% without tools and nearer 50% with them; Mostaque questioned the chart because it included Video-MME despite GPT-5 lacking video understanding at the time. OpenAI’s “code red” was framed simultaneously as employee mobilization, investor signaling and evidence that “a good crisis is a terrible thing to waste.”

Falling token prices will not make aggregate intelligence spending disappear because models are consuming vastly more tokens, iterations and parallel agents. Gemini 3 Deep Think was presented as the template: fleets of agents collectively becoming “countries of geniuses in a data center,” with answers already taking three to five minutes under load. DeepSeek V3.2 reportedly used 2 billion tokens per problem to obtain IMO gold, illustrating why cheaper inference can coexist with acute compute scarcity and trillions of dollars of agent revenue.

Public listings could determine whether retail capital participates in AI hypergrowth or remains outside the intelligence explosion. Anthropic was discussed at a possible $300 billion valuation against projected next-year revenue of $26 billion—roughly 10 times sales, compared by Ismail with Palantir at 111 times—and an IPO potentially as early as 2026. The financing need is physical: the panel cited surging HBM and copper prices and a report that OpenAI reserved 40% of global memory supply; “dollars are the best benchmark” once agents can autonomously earn returns.

China’s Nvidia substitution is initially a domestic supply story, but the panel sees a credible global competitor emerging within several years. Cambricon plans to triple 2026 output to 500,000 accelerators; Mostaque compared its 5090 with Nvidia’s A100 and its 6090 with the H100, at roughly half the cost. Peter added that the chips were more power-efficient. The strategic variable is trust: Chinese systems may be cheaper and more open, yet adoption by countries such as those in Africa could turn on whether users trust China, the U.S. or particular vendors—“scarcity equals abundance minus trust.”

Human-capital policy is struggling to match AI’s speed, creating both near-term labor opportunities and long-term substitution risk. Europe’s planned 2026 AI gigafactory was judged insufficient without a roughly 100-fold increase in institutional “metabolism,” while U.S. data-center construction currently pays skilled workers $100,000 to $225,000 amid a stated shortage of 450,000 people. The same panel expects humanoids to threaten that opportunity in five to ten years and argued that education should shift toward “show me what you have built and done with AI.”

AI’s physical buildout is pulling capital into commercial space and humanoid robotics before either market’s economics or governance is settled. Orbital-compute projections assumed launch costs approaching $100 per kilogram, but Diamandis rejected a move from terrestrial power at $12 per watt to $6–$9 in orbit as too small for the complexity; heat, radiation and security remain open constraints. Meanwhile China installed 54% of the world’s robots, and Engine AI’s T800 prompted the concrete question: “Do we actually want to have regulations around the maximum joint torque of humanoids in the street?”

Genesis Mission recasts U.S. basic science as national AI infrastructure, combining Department of Energy supercomputers, federal datasets, and laboratory tools to compress research from years to days. Alexander Wissner-Gross called it a “1939 moment” in which America becomes “one big AI factory”; Peter Diamandis highlighted biotech, fusion, and quantum, while the stated goal is to double U.S. scientific productivity over a decade. The upside depends on proper funding and execution. Wissner-Gross’s key condition: open science could make the impact “truly exponential,” whereas closed public-private work with strong IP protections would have much lower impact.

Claude Opus 4.5 is presented as a possible threshold for recursive self-improvement, not merely another benchmark release. The headline claims were 76% fewer tokens for equivalent results, leadership in seven of eight programming languages, and AI performance exceeding incoming Anthropic performance-team employees on key assignments—the “canary” for Wissner-Gross. Mostaque reported 52% on SWE-bench Pro without reasoning tokens versus 45% for his Intelligent Internet framework, a 67% cost reduction to $25 per million tokens, and a path to one-shotting typical 100,000–200,000-token codebases next year.

The cost of capable intelligence is falling as evaluation shifts from puzzle scores toward dollars earned. Opus reportedly scored 75% in a same-model multi-agent setup and 88% when orchestrating Haiku or Sonnet, while Wissner-Gross led with, “We’re driving the cost of intelligence to zero.” Mostaque expects benchmarks such as Vending Bench and trading tests to measure real economic output next; he would be surprised if a single entrepreneur could not build a billion-dollar business within two years, “probably next year,” while Wissner-Gross argued an altcoin-pumping “baby AGI” could do it now-ish.

AI-native companies could invert the labor-and-capital model by making nearly every operating input variable cost. Mostaque’s mechanism is unusually concrete: enterprises can charge customers upfront, pay AI providers one or two months later, and automate compliance, forecasting, tax, and payments—potentially allowing a complete business to launch “in minutes” in about a year. Salim Ismail connected this to near-zero acquisition and supply costs, while Wissner-Gross argued agents are “neither capital nor labor” and humans may become investors in fleets of AI entrepreneurs.

The compute trade is broadening from Nvidia scarcity into a heterogeneous market spanning Google TPUs, AWS Trainium, memory, power, and interconnects. Google’s seventh-generation Ironwood TPU was described as four times faster than its predecessor; Mostaque emphasized Google’s chip interconnects, million-to-2-million-token context, and an environment where DRAM prices had risen about fivefold. Amazon, meanwhile, plans up to $50 billion of U.S.-government AI infrastructure and 1.3 GW of new capacity starting in 2026, while its $11 billion, 2.2-GW Indiana facility runs 500,000 Trainium2 chips largely suited to inference.

Shopping agents turn control of user intent into the next distribution battle. ChatGPT’s shopping research, using ChatGPT Mini, claimed up to 64% accuracy, but Mostaque contrasted that with Amazon Rufus’s reported 250 million users, conversion rates up to 60% higher, and an estimated $10 billion of incremental sales next year. The winning layer may be the secure, charming “Jarvis” beside the user—observing requests, conversations, and eventually gaze—then routing work across specialist agents while disintermediating search, affiliate media, and recommendation engines.

The panel sees severe labor disruption arriving before abundance, making coordination and economic growth the binding policy problems. Mostaque’s forecast is that most keyboard-and-mouse work becomes “negative value” within at most 900 days, though he explicitly did not predict every job disappearing; he also cited Grok 4.1 Fast scoring roughly 95% on TaoBench at $0.50 per million words and predicted no customer-service jobs within two years. Proposed bridges included universal AI, AI social scientists, UBI, universal basic services, and universal basic equity—but Wissner-Gross’s prerequisite was to grow the economy faster than conventional human labor loses value.

Brain-computer interfaces and falling launch costs remain the episode’s high-upside physical-world bets. Paradromics was said to reach 200 bits per second versus Neuralink’s roughly 10, with approval to begin human testing in about two months; Ismail, formerly a “hard no” on high-bandwidth BCI, conceded, “Oh, shit. He’s right again.” Diamandis separately traced launch costs from roughly $50,000/kg for the shuttle to $2,500 for Falcon 9, a projected $100 for Starship, and potentially $0.10/kg for lunar mass drivers—cost curves that would expand usable land and material supply far beyond Earth.

OpenAI’s reach is becoming a two-sided control point: 800 million weekly ChatGPT users on one end and massive compute demand on the other. Developers doubled from 2 million in 2023 to 4 million, while API throughput jumped from 300 million to 6 billion tokens per minute. Alexander Wissner-Gross annualized that to 3 quadrillion tokens and projected 30 quadrillion next year—approaching humanity’s estimated 50 quadrillion spoken annually—while GPUs, and separately energy, remain binding constraints.

The everything-app contest is fundamentally a battle for finite attention, with an app-store phase that may be transitional. OpenAI’s Apps SDK puts Booking.com, Figma, Coursera, and Zillow inside ChatGPT, while Meta, Google, and X pursue the same conversational real estate for their agents. Mostaque’s progression is investable shorthand: consumption became cheap, creation is becoming cheap, and “the valuable thing is curation and attention.”

Agentic software development is crossing from code assistance into recursive production. OpenAI said Agent Builder was completed in under six weeks with Codex writing 80% of its pull requests; Mostaque noted the Codex CLI receives two updates a week and interpreted Dario Amodei’s claim that 90% of code would be written by AI as meaning it can be written by AI. Visual workflow boxes are viewed as transitional because “code is just a human translation layer”; the end state is a voice-and-image Jarvis that can explain its own continuous changes.

Sora 2 turns generative video into a product-design API, with pricing already exposing the labor-substitution curve. Mattel’s demo converted a sketch into a photorealistic toy video with apparent physics, while Alexander Wissner-Gross called it “mechanical design getting solved.” At $0.10 per generated second, or $360 per hour, his assumed 10x annual cost deflation would make API-based design dramatically cheaper—provided compute supply catches demand.

OpenAI’s $20 subscription faces compression even as its installed base expands. Mostaque said breakthroughs from DeepSeek, Grok 4, and others have cut token costs 20–30x, reducing a basic chat experience from roughly $200 a year to “a couple of bucks a year.” That forces OpenAI toward advertising, commerce, likeness-driven media, and economically valuable agent workflows while it conducts a global user “land grab.”

The AI capex trade is spreading from GPUs into the entire industrial stack. OpenAI’s 6 GW AMD agreement follows 10 GW with Nvidia; at Mostaque’s estimate of $50 billion per gigawatt, that is roughly $800 billion of buildout. BlackRock’s reported $40 billion pursuit of 78 data centers totaling 5 GW, plus Corning optics, liquid cooling, valves, power, and fab inputs, shows the breadth—but Blundin says the calculable ceiling remains TSMC, Intel, and Samsung manufacturing capacity.

Digital computer use and embodied autonomy are converging into one labor platform. Anthropic was projected to reach superhuman OSWorld performance within months; FSD 14.1 adds 10x more AI parameters and neural-network routing, while Gemini Robotics-ER 1.5 and Optimus point toward common vision-language-action stacks. Mostaque put the tipping point in “the next like six months,” and Wissner-Gross supplied the recursive endgame: robots building robots, then data centers, which produce digital superintelligence that improves the materials and energy efficiency of the whole system.

Mostaque’s near-term thesis is that “useful intelligence,” not AGI, breaks labor economics first. The decisive systems will be reliable “cooks” that follow instructions across keyboard-video-mouse jobs, not frontier “chefs” inventing new science. He expects virtual workers costing roughly a dollar an hour, an 8-billion-parameter medical model running on a smartphone, and agents operating for hours without supervision to make the next thousand days radically different even if capabilities soon plateau.

AI’s cost curve creates an “abundance trap”: intelligence becomes plentiful while a scarcity-based economy records the result as unemployment and poverty. GPT-3 input reportedly cost $60 per million tokens versus roughly $1.25-$1.50 for GPT-5, while a website that once cost thousands can already be generated for $20-$40. Because GPUs “don’t need to eat,” buy housing, or consume their wages, rising AI output need not recycle demand, tax revenue, and employment through the economy.

The “intelligence inversion” leaves human labor with no obvious higher rung to climb after machines outperform both muscle and cognition. Mostaque traces economic advantage from land and serfs, through physical labor and industrial or software capital, to intelligence itself; once AI becomes the marginal producer, “there’s nowhere left to pivot.” Nathan Labenz’s electricity calculation sharpens the threat: a laptop can run for hours on less than two cents of power in his Detroit market.

The distributional upside therefore accrues first to GPU owners, frontier laboratories, and data flywheels—not automatically to workers or consumers. Mostaque warns that OpenAI, Anthropic, and xAI need not keep their best models available through APIs; once internal systems outperform public ones, vertically integrated labs could compete with every application and ultimately “take on the entire economy themselves.” He sees Grok 5 versus Grok 4 as a near-term scaling test, while stressing that even a plateau around today’s frontier would still disrupt employment.

Human care, community, and creativity remain meaningful, but they do not solve the transition’s monetary arithmetic. Mostaque accepts that people may intrinsically prefer human teachers, nurses, friends, and performers, yet estimates that paying every American adult the roughly $16,000 poverty threshold would cost $5 trillion—the entire US tax base—while corporate taxes contribute only about $0.9 trillion. His distinction is that “computation and consciousness are different”: AI can supply the how, but humans still create the why, provided they retain income, identity, community, and attention.

His alternative scorecard replaces GDP-only optimization with four multiplicative “MIND” capitals: material, intelligence, network, and diversity. Material goods such as apples are rivalrous and depleted by consumption; knowledge can circulate without being lost, networks determine trust and coordination, and diversity supplies resilience and optionality. Together with laws of flow, openness, and resilience, the framework argues that maximizing profits or GDP while degrading social connectivity and redundancy makes systems look efficient precisely as they become brittle.

The proposed counterweight to digital feudalism is collectively controlled AI infrastructure financed through a Bitcoin-like “FoundationCoin.” Verified deployments of open models in health, education, finance, and government would secure the asset, while sale proceeds would fund public-interest compute such as multilingual diagnosis checking and organized cancer knowledge. Mostaque is explicit that the second currency—a cash-like issuance tied to being human—“we haven’t quite figured out yet,” making this a direction and incentive-design experiment rather than a completed monetary system.

The thousand-day urgency comes from defaults hardening into control: whoever owns users’ memories, data, interfaces, and GPU capacity may become almost impossible to dislodge. Mostaque’s preferred outcome is neither one corporate singleton nor nationally fragmented AI, but a loosely coupled swarm whose objective is human flourishing. His closing choice is stark: use AI “to increase the nature of our agency versus replace us with agents,” so that machines taking jobs frees people for family, care, art, exploration, and “our real work.”

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.

GPT-5’s launch lost the theater but won on distribution, price, and coding parity. Dave Blundin called the anticipation “up there with the top three product launches of all time,” yet the folksy presentation and familiar coding demos helped invert Polymarket’s roughly 80% odds of OpenAI retaining the best model toward Google. Beneath the disappointment, the panel’s compact verdict was consequential: OpenAI cut AI costs at least in half, caught Anthropic in coding, and moved 700 million weekly users toward frontier intelligence.

OpenAI appears to be raising the consumer floor while keeping its most capable intelligence inside the lab. Emad Mostaque described GPT-5 as a router selecting among Thinking, Mini, and Nano, after expectations of routing across models from Mini up to Pro, rather than exposing one expensive “mega AI.” He argued that OpenAI already has better internal models and may increasingly offer “decent models for everyone” while reserving its strongest systems to outcompete everyone else.

The durable economic story is intelligence hyperdeflation, not a single benchmark crown. API pricing fell from GPT-4.5’s stated $75 input and $150 output per million tokens to GPT-5 at $1.25 and $10, while GPT-5 Mini and Nano established a new cost-performance frontier on ARC-style tests. Alex Wissner-Gross framed the decisive comparison as unaffordable superintelligence versus intelligence “too cheap to meter”; cheaper inference also permits 10 times more search across mathematical and scientific completions.

The models are crossing from impressive demos into dependable economic work, forcing an AI-native operating decision. Emad highlighted longer unsupervised performance across law, logistics, and sales, with fewer hallucinations; the uncertain outcome is “either a productivity boom or the inverse,” including layoffs. Salim Ismail’s advice was categorical: “Just go all in and start turning your business into an AI native business,” while Dave warned that increasingly opaque benchmarks can paralyze executives precisely when experimentation matters most.

GPT-5’s most material frontier result may be a slow-motion automation of mathematics. Alex’s straight-line extrapolation from Frontier Math Tier 4 suggested AI could solve 15–20% of hard problems by the end of 2025, 35–40% by the end of 2026, and 70% by the end of 2027—what he called a “slow motion solution to math.” Emad added that extended reinforcement learning had already produced an IMO gold-medal system and predicted that the breakthroughs will be elegant theories found by running “a million different things at once,” not merely brute-force calculations.

Coding has become the immediate commercial battleground, with price and distribution threatening Anthropic’s strongest franchise. The launch demo itself looked months behind what users already did with Claude, but Dave’s investor reading was that GPT-5 had nevertheless caught Anthropic “in their wheelhouse.” Emad said OpenAI and Anthropic each had roughly $3 billion of API revenue, with about $1.4 billion of Anthropic’s tied to Cursor and Microsoft Copilot, while GPT-5 was priced roughly 40% below Sonnet; closer Cursor–OpenAI alignment could now redraw the application stack.

Healthcare’s constraint is shifting from model intelligence to longitudinal patient data. Sam Altman said GPT-5 scored higher than previous models on HealthBench, built with 250 physicians, while Emad cited doctors scoring about 20% against newer models at 60–70%; Peter Diamandis’s counterpoint was that even “the best AI” remains only as useful as the scans, biomarkers, wearables, and history supplied to it. Dave saw life-saving cases as regulatory protection for continued acceleration, whereas Salim viewed the launch segment as incremental PR until models are deeply integrated into routine care.

Cheap open-weight models and sovereign compute broaden the opportunity while intensifying infrastructure and valuation risk. Emad estimated OpenAI’s new open-weight model cost about $4 million to train, said one laptop-capable version used only 5 billion active parameters, and predicted GPT-5-level training below $1 million within two years; Alex cautioned that synthetic data may conceal the fixed cost of a larger teacher model. Meanwhile OpenAI pursued a roughly $500 billion valuation and a Norway site with 100,000 GB300 chips and 230 megawatts expandable to 520, as Google, Grok, national governments, and constrained power supplies turn the race into a literal land grab.

Grok 4 has pushed academic benchmarks close enough to saturation that the competitive question is shifting from raw intelligence to usable agency. It scored 100% on AIME 2025, while Grok 4 Heavy reached 44.4% on Humanity’s Last Exam versus 26.9% for Gemini 2.5 and 21% for o3. Emad Mostaque’s crucial distinction: the model “is reasoning, but it’s not planning” — leaving planning, memory and agentic integration as the next bottlenecks.

xAI’s lead is as much an infrastructure and execution story as a model story. Founded only 28 months earlier, it scaled to a cited 340,000 GPUs after Elon Musk delivered on a seemingly implausible promise to operate 100,000 H100s; the panel put the installed hardware near $10 billion and said xAI was targeting 1 million GPUs. Dave Blundin recalled that experts said coherence at that scale was impossible, then reacted, “Oh, god dang, he did it.”

The training-cost mix has flipped, creating a new flywheel around synthetic, structured reasoning data. Mostaque said post-training once consumed roughly 1% of compute, rose to 10% with DeepSeek and is now approximately equal to pre-training, partly because frontier models can generate data for their successors. That does not guarantee a “more sane” Grok — mode collapse remains possible — while Peter Diamandis said model competition is becoming “an engineering and quality challenge” rather than pure brute force.

Inference is rapidly demonetizing even as premium intelligence may command higher prices in high-value workflows. Grok 4 was quoted at $3 per million input tokens and $15 per million output tokens; Mostaque estimated about $20 for “a million very good words” and projected equivalent intelligence could get 5–10 times cheaper annually, potentially reaching $1 per million words. Diamandis argued developers will pay materially more for marginal gains that compress engineering time, while Mostaque suspects the $300 SuperGrok Heavy plan is a loss leader for enterprise conversion.

Enterprise value will initially come from augmentation, error reduction and data assimilation, not instant wholesale replacement. Emad said the Arc Institute was testing Grok 4 across millions of experiment logs and CRISPR workflows; Peter cited approximate medical-study results in which AI alone beat both physicians and physician-plus-AI combinations. Mostaque nevertheless stressed that replacement remains “way off” on liability, while Salim Ismail emphasized processing scans and sensor data no human could integrate.

Coding, games and video expose the same opportunity: generation is arriving before planning, feedback and distribution are solved. xAI showed a first-person game produced in four hours and said a specialized coding model was weeks away; the release speaker forecast the first really good AI game and first watchable AI movie next year, with a half-hour of watchable AI television potentially this year. Mostaque’s warning for incumbents: lower production costs benefit companies, but “for the individuals working in the industry, this is terrible.”

If frontier models converge on one capability plateau, the durable bottlenecks become interfaces, agents, chip access and distribution. Mostaque expects Grok 5 to coordinate anywhere from 60 to 6,000 agents, use professional tools and behave like a remote worker that “just gets the job done and it doesn’t sleep.” The panel discussed Google’s roughly 3 million chips, million-chip ambitions at xAI and Meta, and supply constraints across the sector; efficient edge models such as Liquid AI could provide the everyday intelligence tier.