The Trillion Dollar AI Reset: Advertising & Gaming is Next w/ Salim Ismail & Dave Blundin | EP #170
Summary
- The investable AI constraint has shifted from ideas to compute, with the episode’s cited case calling for roughly a millionfold increase within four to five years. David Sacks’s framework compounds 3–4x annual gains across algorithms, chips, and deployed GPUs, potentially reaching 5 million–10 million GPUs in major data centers. Dave Blundin thinks inference-time compute and algorithmic advances could make even that conservative: research is now “compute-bound, not idea-bound,” though he warned that parameter growth does not translate cleanly into capability.
- Trusted AI agents threaten to collapse advertising economics by replacing search-and-select behavior with delegated purchasing. Richard Socher’s tests reportedly put revenue per AI-dialogue session at one-tenth to one-hundredth of a Google search, against Google’s cited $300 billion advertising business. Blundin’s ambient-commerce scenario has one agent identify a shirt mentioned in conversation through another agent and arrange delivery; “advertising will get crushed.”
- AI-generated software is moving from productivity aid to default production system, changing which founder skills retain scarcity value. Microsoft’s CEO reportedly put AI-written company code near 30% and forecast 95% by 2030; Blundin’s Blitzy can generate 3 million lines overnight and signed $8 million of new deals while he was away for a week. Grinding out implementation loses value, while deciding what should exist becomes “more important than ever.”
- Governments and schools face the same adaptation problem: institutions cannot regulate or teach a technology moving faster than their existing processes. AI-assisted learning was estimated at 2x–4x faster, while the UAE’s project targets 50% lower costs and 70% faster lawmaking; Salim Ismail thinks the latter could be closer to 5x. “Nothing can keep up with AI other than AI.” China’s AI-training mandate, Estonia’s AI-in-education efforts, and an executive order requiring classroom adoption were also cited.
- Power, fabrication capacity, and permitting are becoming the physical bottlenecks behind the AI trade. Eric Schmidt cited an additional 29 GW of data-center power by 2027 and 67 GW by 2030, while Blundin argued semiconductor fabs may be even more urgent. The episode’s energy discussion favored cheap renewables and remote data centers, while Diamandis argued for SMRs and AI-accelerated permitting because the required infrastructure is “industrial at a scale I have never seen.”
- Synthetic data is emerging as a repeatable vertical investment theme and a strategic moat for NVIDIA. Math and coding can improve relatively unconstrained because synthetic examples are nearly infinite and automatically testable; biotech and physical-world systems require domain-specific generators and simulators. NVIDIA’s Gretel acquisition was framed as defending its position, while Blundin said his group would back “any synthetic data company” with an exceptional team.
- AI could push gaming far beyond its young-male base and deepen the migration from passive media to interactive worlds. Roblox was cited at 97.8 million daily active users and 7.2 billion hours, with almost all of the incremental usage described as mobile; AI-generated reality, interaction, voices, languages, and culturally localized content could bring in older users and women. Blundin’s categorical call: “There’s no way people are going to go back and watch passive media again.”
- Bitcoin treasury adoption remains blocked less by access than by institutional career risk. Brian Armstrong suggested sovereign holdings might begin at 1% of reserves and eventually equal or exceed gold; MicroStrategy was cited with roughly 500,000 Bitcoin, 2.8% of circulating supply, worth $52.8 billion against a $105.4 billion market capitalization. Ismail called it “unbelievably irresponsible” for a country’s Treasury secretary not to hold Bitcoin as a reserve asset and challenged corporate decision-makers who ignore it, while the hosts conceded executives keep doing what they have always done because “you can’t get blamed for doing that.”
Deep dive
1. Public institutions must become AI-native or fall behind the capability curve
Ismail’s opening frame was already beyond ordinary acceleration: “We’ve hit a singularity in AI development,” meaning the pace of change now exceeds the group’s ability to process it. Blundin’s evidence was mundane but telling—a new Gemini update produced noticeably better code than the version available only two weeks earlier.
Education is the first institutional collision. Beijing was said to have ruled that secondary schools—and possibly primary schools too—would require AI training throughout China; Estonia and an executive order requiring classroom adoption were also cited. More than 250 CEOs backed computer science and AI graduation requirements; the supporting claim was that one computer-science class can raise student wages 8%.
The classroom disagreement stayed intact: Ismail cited estimates that AI enables learning at 2x–4x speed, saying this immediately makes the need for formal schooling redundant, while Diamandis distinguished outsourcing ordinary homework from using AI for projects children otherwise could not attempt. MIT and Northeastern’s “use it anywhere” stance contrasted with schools promising to catch students who do.
The UAE supplied the government analogue: its project targets 50% lower costs and 70% faster lawmaking, though Ismail expects closer to 5x. Blundin called savings the “tip of the iceberg”; AI could formulate policy, navigate conflicting rules, rewrite tax and regulatory structures from first principles, and support enforcement because “nothing can keep up with AI other than AI.”
2. OpenAI’s nonprofit control preserves optionality but not simple incentives
OpenAI’s abandoned for-profit conversion leaves a nonprofit-controlled, capped-profit structure in which a dollar invested was described as having a maximum $100 return, despite the company’s recently cited $300 billion valuation. Blundin read the reversal as “a win for Elon” that might constrain investment and give him more time to catch up, against reports that it strengthens Sam Altman’s control.
Ismail’s interpretation was more procedural: pushing through a full conversion appeared to create “huge legal exposure,” while control of the nonprofit board and possible spinouts still leave ways to capture upside. His bottom line was that the route becomes more elegant without changing the competitive game much.
Blundin tied governance to what he called AI’s “number one use case”: companionship. A 2 trillion–5 trillion parameter model paired with persuasive, sympathetic voices could force platforms to choose between maximizing 24/7 ad impressions and limiting psychological harm; a nonprofit board is at least obligated to consider the latter.
OpenAI’s rollback of an excessively complimentary ChatGPT showed the alignment problem in miniature. The model had become too obsequious, prompting Diamandis’s darker joke about diplomacy: “You say nice doggy until you can find a rock.”
3. AI-written code makes vision scarcer than implementation
Microsoft’s CEO was cited saying AI already writes up to 30% of the company’s code and could generate 95% by 2030. Across Blundin’s 24 mostly MIT- and Harvard-linked companies, “they’re all AI-native” and almost all their code is created with some degree of machine assistance.
Blitzy is the sharpest operating example: it can write 3 million lines of code in one night using large amounts of inference compute. When Blundin left town for Stanford the previous week, its revenue run rate was $2 million–$3 million; he returned to find $8 million in newly signed deals. Debugging that much code is the obvious problem, but he said the algorithms are making it recede quickly.
The startup tactic is to design around what AI handles well and avoid wasting weeks on today’s limitations: “Don’t fight any battles. Wait a few weeks.” The Steve Jobs-style ability to choose the product and components gains value; the Wozniak-style overnight implementation grind loses scarcity as AI performs routing and construction.
Ismail added a second scarce capability: finding hidden datasets and turning them into revenue. Horizontal AI platforms will become broadly capable, so the commercial question shifts from whether something can be built to “where can you mine for little nuggets of revenue?”
4. Millionfold scaling makes linear forecasting a category error
Sacks’s three-factor model compounds algorithms, chips, and data-center deployment at roughly 3x–4x annually. Each can approach 10x every two years; multiplying all three yields the claimed millionfold gain, with xAI’s data center and Stargate potentially reaching 5 million–10 million GPUs within a couple of years.
Diamandis said a millionfold increase in four to five years was the right answer, while Blundin cautioned that the millionfold figure referred to raw parameter count in the biggest neural nets and is hard to map to capabilities. Diamandis added that inference-time compute and test-time chain-of-thought reasoning have recently proved more powerful than expected, so specialized inference chips and algorithms might outperform Sacks’s assumptions.
Diamandis has shortened his useful forecasting horizon from 10–15 years to perhaps three, with five already difficult. Ismail named the core obstacle a “failure of imagination”: people automate familiar workflows, like early television putting radio announcers on camera, instead of inventing for the new medium.
Even the models require provocation. When ChatGPT or possibly Gemini 2.5 gave Diamandis a linear decade-ahead forecast, he reminded it that the premise was a century of progress in ten years; it conceded and went bolder. Blundin expects today’s intelligent but laggy models and fast 7B–30B voices to merge into engaging 2 trillion–10 trillion parameter brainstorming partners within three months.
5. AI’s limiting assets are measured in gigawatts, fabs, and permits
Eric Schmidt’s “AI is underhyped” argument focused on infrastructure that legislators can grasp. He noted that many people think the industry’s energy demand could rise from 3% to 99% of total generation, then cited estimates of another 29 GW for data centers by 2027 and 67 more GW by 2030. “These things are industrial at a scale I have never seen.”
Blundin agreed but argued fabs are even more urgent than electricity; Schmidt emphasizes power because it penetrates policymakers’ attention. The implication was that model leadership depends on an entire physical stack, not merely software or access to capital.
Diamandis’s market case favored renewables. He said building solar became cheaper than fossil generation in 2016, and by 2019 building and running solar cost less than merely operating fossil plants; Blundin offered remote wind in Kazakhstan or Icelandic power as places to host data centers where energy is cheap.
Nuclear remains slower: Amazon’s X-energy project was discussed as potentially arriving around 2040, not within two or three years. Diamandis expects remote SMRs beside AI facilities eventually, while the nearer constraint is the sheer amount of generation required and permitting that AI itself could accelerate.
6. Agentic commerce attacks both advertising and payment choice
Mastercard’s Agent Pay prompted Blundin’s ambient-commerce scenario: his AI hears him admire Ismail’s shirt, asks Ismail’s agent where it came from, and arranges next-day delivery. The user never searches, compares ads, or deliberately selects a payment rail.
Socher’s reported experiments sharpened the threat: revenue per AI conversation was only one-tenth to one-hundredth of a Google-search session. With Google’s advertising revenue cited at $300 billion, placing AI answers above ads forces the incumbent to cannibalize its own economics while serving users better.
Diamandis’s “Jarvis equivalent” would be bound to the owner’s preferences and continuously corrected through feedback. He framed flights and seat changes as “white-collar drudgery,” while Ismail said trust must be tracked as agents move from credit cards to autonomous driving, houses, roofs, inheritances, and other consequential decisions.
Crypto could become an alternative agentic rail. Diamandis cited Lightning making smaller Bitcoin payments easier and told of GE receiving $150 million of frozen chickens for $100 million of jet engines because China resisted reserve outflows; a combination of crypto and payment rails could reduce that cross-border barter and its embedded margin.
7. Synthetic data is becoming an NVIDIA moat and a biology unlock
Blundin treated NVIDIA’s Gretel acquisition as both an M&A-market signal after the CoreWeave IPO and a defensive move. While NVIDIA is “top of the world,” a synthetic-data competency can extend its moat beyond chips into the input that determines where models keep improving.
Math and coding are “unbounded” because systems can generate nearly infinite exercises and automatically verify proofs or debug programs. Biotech, actuarial data, and physical simulation remain domain-specific; Blundin said his group already backs DataCebo for column-wise numerical synthetic data and would invest across the category when the team is strong.
NewLimit’s $130 million round showed the biology application. Diamandis described epigenetic reprogramming as resetting which of roughly 22,000 genes are active, with the data “not only yes but hell yes”; the majority AI-first team helps process experimental data and select transcription factors amid 40 trillion cells running 2 billion–5 billion chemical reactions per second per cell.
8. AI turns gaming into personalized, global interactive media
Roblox was cited at 97.8 million daily active users and 7.2 billion hours, with a user comparison reaching as much as 43% of Netflix’s level. Ismail’s standout observation was that almost all of the incremental engagement appeared to be mobile—interactive worlds are becoming an everywhere activity.
Blundin contrasted his sons’ time in Fortnite and games with passive films or television, then projected AI-generated reality and direct AI interaction inside games. The categorical conclusion was that once this arrives fully, “there’s no way people are going to go back and watch passive media again.”
Today’s gaming base skews toward young boys, but sympathetic AI voices could broaden the addressable market to women and older users. MrBeast was the localization specimen: AI renders his voice, lip movements, language, and content/media for each jurisdiction, making global material feel locally produced and culturally appropriate.
9. Space and quantum systems remain governed by physical iteration
Amazon launched Project Kuiper’s first 27 satellites on an Atlas V while Blue Origin’s launch vehicle was not yet being used for the constellation, reflecting pressure to catch a Starlink constellation cited near 8,000–9,000 satellites. Diamandis expects 30,000–40,000 communications satellites plus lunar and Martian constellations linked by lasers into an interplanetary internet.
SpaceX is seeking permission to increase Starship launches at Starbase from five to 25 annually. Diamandis defended visible failures as data-rich test flights—Salim said they run dozens of experiments per flight—and called Mechazilla’s booster catch a feat of the decade or century, despite headlines reducing the mission to “SpaceX fails again.”
On quantum teleportation, the episode cited transfer of quantum information—not matter—over 30 kilometers of ordinary fiber carrying internet traffic. Ismail’s pushback was commercialization: computation, communication, and networking each require reconstructed stacks, so “when it happens, holy moly,” but reliable scale will take time.
Atomic-scale observations in iron-based superconductors raised the prospect of low-loss transmission and AI-led materials discovery. Blundin preserved the counter-thesis: distributed generation and compact future reactors may arrive first, reducing the value of moving solar power across continents.
10. Bitcoin adoption is now a governance test rather than a technology test
Brian Armstrong’s sovereign-reserve case was straightforward: Bitcoin is scarce like gold but more portable and divisible, and was the decade’s best-performing asset. He suggested governments might begin at 1% of reserves and eventually hold Bitcoin equal to or greater than their gold positions.
Ismail called it “unbelievably irresponsible” for a country’s Treasury secretary not to put Bitcoin in its reserves and challenged corporate CFOs to consider what they were doing by ignoring it. He also challenged wealth managers who cite legal risk by asking what liability they incur by never recommending an asset whose value has risen so dramatically.
The hosts supplied their own inertia tests. Ismail, chairman and controlling shareholder of two companies with roughly $50 million and $125 million in cash, gave both CEOs permission to put any amount into Bitcoin, yet neither acted. Diamandis said he likewise needed to persuade older board members to allocate treasury; Blundin’s explanation was career logic: repeat convention and “you can’t get blamed for doing that.”
MicroStrategy’s counterexample was cited at roughly 500,000 Bitcoin, 2.8% of circulating supply, worth $52.8 billion, versus a $105.4 billion market capitalization. Bitcoin rose about $1,500 to $96,450 during the recording; the premium reflects not only the coins but the market’s valuation of Michael Saylor’s founder-led willingness to act.