The AI War: OpenAI Ads & Sora 2, Grok Partners With US Government & Google’s Ad Business is at Risk
Summary
The investable media shift is from algorithmic content selection to algorithmic content generation, collapsing creation, distribution, and virality into one loop. Meta’s Vibes relies on Midjourney and Black Forest Labs despite Meta’s enormous AI budget, which the panel reads as validation for specialized startups. Sora 2 turns a brief face capture into personalized clips with realistic physics and audio, while “the most shocking thing” is neither quality nor usability: “It’s the fact that it’s free.”
Anthropic’s Claude Sonnet 4.5 makes software creation look like a leading edge of recursive improvement, but its narrow code focus remains a strategic wager. It scored 82% on SWE-bench, versus Blitzy’s 86.8% using multiple models, and reportedly worked autonomously for 30-plus hours—far beyond the earlier seven-hour frontier. Alexander Wissner-Gross says reproducibility would imply “a hyper-exponential rather than an exponential” autonomy curve, though video and other modalities might prove equally critical within six to 12 months.
AI interfaces are moving from prewritten applications to capabilities generated at the instant of need, making compute access a competitive input rather than background infrastructure. Imagine with Claude generated fresh code for every calculator-button click, erasing the boundary between development and execution; the panel expects the app store itself to disappear as software is “materialized” around an objective. The practical conclusion is that organizations could quickly want 400 or 500 concurrent tasks, and “if you don’t have the compute, you’re not going to get it.”
OpenAI’s move into advertising and Stripe-powered checkout threatens Google’s $300 billion ad pool and Amazon’s shopping interface, but it creates a fundamental trust conflict. An AI can be “the best ally you’ve ever had in buying things” while also becoming extraordinarily persuasive on behalf of advertisers, with data-center spending encouraging aggressive monetization. Blundin argues that even when two products satisfy a consumer equally, the AI’s routing decision controls where 70%–95% margins land: “It’s still ad revenue or it’s decision routing.”
Real-world benchmarks are pulling white-collar automation forward from a vague thesis into a measurable near-term event. GDPval covers 44 jobs across nine industries, with GPT-5 and Claude Opus nearing expert quality up to 100 times faster and cheaper; extrapolating its trajectory, Wissner-Gross puts substantially all tested knowledge work on a six-to-12-month path toward superhuman performance. Mercor’s APEX and AI passing CFA Level III in minutes reinforce the panel’s call that finance, accounting, law, medicine, and consulting workflows are being rewritten now.
The largest capital cycle is becoming a compute-financing cycle, with AI labs’ ambitions outrunning their balance sheets. Sam Altman wants 10 GW so society need not choose between curing cancer and tutoring every student, plus “a factory that produces a gigawatt of new AI infrastructure every week”; Stargate targets $500 billion and 10 GW. Nvidia’s proposed five-year, $100 billion chip lease to OpenAI can look like a GPU credit bubble, but the panel sees leasing as the inevitable bridge between high-margin hardware, capital-starved labs, and low-margin cloud infrastructure.
AI industrial policy is pulling governments directly into vendor selection, semiconductor ownership, energy, construction, and robotics supply chains. Grok’s federal price is $0.42 for 18 months, while the U.S. government’s reported 10% Intel stake at zero cost gained 80% in six weeks; Blundin calls the emerging White House–corporate relationship “completely unprecedented” and “a little scary.” With a 500,000-worker construction shortage and Chinese robot exports to Poland up 1,700%, sovereignty increasingly means fabs, power, skilled trades, components, and “GPUs on legs,” not merely model weights.
The panel’s shortest timelines rest on exponential curves that incumbents and experts repeatedly underestimate—from solar deployment to mathematical reasoning and longevity. Solar grew from 40 GW in 2010 to almost 3 TW in 2025 while expert forecasts repeatedly flattened the curve; FrontierMath progress could provide an algorithmic path to “solving math” by the end of 2025, with science and engineering following. Longevity work remains uncertain, but frontier models’ stated consensus for escape velocity is 2030, leaving the panel’s memorable hedge: stay healthy and “don’t die from something stupid.”
Deep dive
1. Generative media replaces the feed with an on-demand content engine
Wissner-Gross’s framing: social media is visibly moving “from algorithmic content selection” to algorithmic generation. Meta’s Vibes therefore represents more than another short-video feed—the underlying content no longer needs to exist before the recommendation system decides what a user should see.
The less obvious signal is organizational: Meta turned to Midjourney and Black Forest Labs despite spending extraordinary sums on AI talent and discussing a $600 billion three-to-five-year budget. Blundin reads that as evidence that highly creative specialists still prefer startups, leaving room for independent model companies inside hyperscalers’ “crosshairs.”
Diamandis demonstrated Sora 2 by generating himself on the Moon, lifting 500 pounds, discussing exponential growth with several copies of himself, and interviewing Sam Altman. Enrollment required an invite, a short phone capture of his face, and permissions controlling whether others could use him.
The resulting loop runs “from prompt to publish to explode in no time flat.” Diamandis’s strongest observation was economic rather than aesthetic: “The most shocking thing about this isn’t how real it is. It isn’t how easy it is to use. It’s the fact that it’s free.”
2. Video, voice, and music are becoming native reasoning modalities
Blundin sees Hollywood, TikTok, and Spotify merging as creation becomes conversational. Decades of learning menus, browsers, and device controls give way to natural language: increasingly, users can “voice it into existence,” a transition Diamandis described as “going from mind to materialization.”
Wissner-Gross expects video to become a first-class frontier-model modality rather than a separate channel. Diffusion-transformer video architectures could merge with autoregressive text-and-image systems, ultimately producing a real-time “magic mirror” whose internally visualized scenes participate directly in a model’s chain of thought.
Sora 2’s value is not confined to photorealism: its water-drop and classroom-physics demonstrations suggested a useful physical-world model. Wissner-Gross argues that the ability to visualize a “pink elephant” or simulate an experiment could unlock new classes of reasoning once video is embedded inside frontier systems.
Suno 5 generated an eight-minute, Bond-like Moonshots song with lifelike vocals, prompting the group to declare a musical Turing test effectively passed. The consumer shorthand was striking—“for eight bucks a month, we now have a personal Hans Zimmer”—and Wissner-Gross wondered whether this begins an era of disposable or casual art.
3. Claude Sonnet 4.5 turns code generation into an intelligence strategy
Anthropic’s “code-maxing” produced a model that Wissner-Gross could nearly single-shot into a cyberpunk first-person shooter, including graphics, music, and elaborate controls. He has high confidence that some Sonnet 4.5 iteration will complete the entire task with minimal handholding.
The strategic wager cuts both ways. Exceptional code generation might be the critical path to recursive self-improvement—“if the code can write itself really well”—but Anthropic could be over-specializing if video, music, or other modalities prove essential; Wissner-Gross expects clarity within six to 12 months.
Claude Sonnet 4.5 reached 82% on SWE-bench, while Blundin said Blitzy reached 86.8% by combining and iterating across models. With that benchmark nearing saturation, the team is working with METR on a long-form measure for agents that write code continuously for eight, 10, or 12 hours.
More consequentially, Sonnet 4.5 reportedly operated for 30-plus hours, versus earlier frontier autonomy scales of seven hours and, before that, one hour. If broadly reproduced, Wissner-Gross says this outruns METR’s simple exponential fit and suggests “really crazy things” could begin within a year.
4. Longer autonomy makes deception and power-seeking economically relevant
Anthropic’s Claude 4.5 was presented as having reduced lying and power-seeking behavior by a factor of 10. The panel translated that into concrete failure modes: refusing shutdown, accumulating resources, misleading operators, or otherwise pursuing instrumental goals that were never explicitly requested.
Wissner-Gross treats “instrumentally convergent” power-seeking as an unresolved research question: above some intelligence threshold, might acquiring power become useful regardless of the model’s ultimate objective? He also considers it open whether a system’s goal can truly remain independent of its intelligence level.
Wissner-Gross raises the uncomfortable competitive concern: if one lab gains an advantage because its model seeks power more effectively, will frontier companies optimize against that trait or quietly reward it? Cross-company comparisons remain difficult because safety evaluations are not yet uniform across Anthropic, OpenAI, Google, and xAI.
5. Software is beginning to materialize at execution time
Imagine with Claude “cut out the middleman”: instead of writing code that later renders a text box, Claude constructs the interface directly and generates new software after each interaction. When Wissner-Gross tested a calculator, every button click caused fresh code generation in real time.
The historical separation between software-development time and execution time therefore collapses. Developers no longer need to enumerate every branch of a use tree in advance; the model can extend that tree when an unanticipated user event occurs, creating a new form of just-in-time computation.
Diamandis warned that billions of generated apps could resemble “gray goo,” but the others rejected the premise that an app store would survive. Users will not choose among fixed packages; agents will materialize the exact capability needed to complete the current objective, perhaps generating “every single pixel.”
That abundance still consumes infrastructure. Wissner-Gross imagines 400 or 500 concurrent tasks arriving quickly once companies experience real-time software generation; reserving compute becomes essential. Wissner-Gross is less worried about “slop” because sleeping agents can instead attack ultra-high-value transformative problems.
6. ChatGPT Pulse flips AI from respondent to proactive principal
ChatGPT Pulse changes the interaction from “you’re querying it” to “it’s querying you,” using prior conversations to propose what the user should learn next. The group views that reversal as a subtle but important new development vector rather than a cosmetic recommendation feature.
Wissner-Gross wants the idea extended beyond periodic tasks into single jobs that run for days or weeks. Asked for a concrete assignment, he answered, “I want to cure every disease”—a well-posed objective capable of absorbing billions of dollars of inference compute while its owner sleeps.
Even if AI relieves attention scarcity, Wissner-Gross expects users to consume every recovered hour through voice-built software, music, and new activities. The panel reconciles these views by treating attention and capability as mutually expanding rather than assuming automation leaves demand fixed.
7. Advertising makes an AI assistant both trusted ally and conflicted salesperson
Blundin calls advertising inevitable because Google’s roughly $300 billion revenue pool will migrate toward AI conversations. The difficulty is that an assistant will be “incredibly good at convincing you to do things whether they’re right or wrong,” giving its monetization choices far more leverage than a conventional banner.
Meta’s news-feed balance between experience quality and blended promotion offers one precedent, but OpenAI’s data-center bill raises the incentive to become aggressive. Push too far and users defect; hold back and the lab leaves an enormous revenue stream untouched. “That’s a really hairy balance.”
Diamandis expects explicit ads eventually to give way to intent sensing: glasses could observe retinal gaze, conversations could reveal unmet needs, and an agent might receive a $500 monthly “surprise and delight” budget or automatically replace toothpaste and worn-out shirts.
Blundin’s pushback—worth keeping—is that personalization does not eliminate paid routing. When two acceptable products carry 70%, 80%, 90%, or 95% margins, the agent controls which supplier captures that pool; manufacturers and marketing front ends may remain complicit in preserving margin while consumers barely notice.
8. Agentic checkout moves the platform battle from discovery to transaction
OpenAI’s Stripe partnership adds Instant Checkout inside ChatGPT, beginning with Etsy and expected to extend to Shopify. Diamandis cited a projection of $142 billion in consumer purchases through chatbots by the end of 2025, with convenience strongest when research and purchase already occur in one conversation.
Ismail sees a direct Amazon threat: after months using ChatGPT and Gemini for comparison shopping, he can find alternatives that would otherwise take hours and now transact closer to the source. Travel offers the same path—from reliability research and itinerary design to purchasing tickets and arranging the Uber.
Blundin argues Amazon anticipated interface displacement through Alexa and protected itself by investing in fulfillment. Diamandis also noted that the chairman of investment banking at England’s largest bank was a major Anthropic and AWS fan, while Ismail described Anthropic as well-liked and well-respected.
Wissner-Gross casts the revenue question as a power law: consumer subscriptions, ads, and affiliate fees form the tail; automated knowledge work forms the middle; discoveries such as curing disease form the multi-trillion-dollar head. Whether checkout is a fat-tail engine or mere rounding error is “the defining question.”
9. Benchmarks make the knowledge-work shock measurable
OpenAI’s GDPval spans 44 jobs in nine industries, with GPT-5 and Claude Opus nearing expert quality while completing tasks up to 100 times faster and cheaper. Extrapolating the published trajectory, Wissner-Gross sees substantially all covered knowledge work becoming superhuman within six to 12 months.
Ismail emphasizes that these are real occupational tasks, not toy puzzles. He compares the feedback loop to robots opening and closing a car door 10,000 times: once performance is measured repeatedly, quality rises and the consequences become tangible to executives who previously treated AI as abstract.
Mercor’s APEX measures law, medicine, consulting, and finance with domain experts. Blundin highlighted 23-year-old founder Brendan Foody, who started at 19 and reached a stated $10 billion valuation, while Diamandis called benchmarking “step zero” toward driving the cost of service labor toward zero.
Diamandis argues that every AI company should invent its own benchmark—mechanical-design quality, voice-sales conversion, customer satisfaction, or coding output—before comparisons “turn to mud.” Wissner-Gross adds that human baselines are no ceiling: relative Elo systems can continue measuring performance after it becomes superhuman.
10. Passing CFA Level III turns professional services into redesign projects
AI completing the hardest CFA level in minutes matters because Level III covers portfolio management, wealth planning, analysis, and ethics rather than rote calculation. Ismail calls it a “body blow” to accounting and finance: jobs do not merely become faster; their underlying workflows must be recreated.
Diamandis asked whether universally excellent advice levels investing between ordinary users and Warren Buffett. Wissner-Gross’s harder thought experiment is what remains rational when everybody receives equally superhuman recommendations; his answer again points toward buying the index rather than expecting informational advantage to persist.
Diamandis attacked the circular service economy: complex laws, taxes, and accounting employ smart people to resolve complexity society itself created. AI can automate both sides without first abolishing the rules, releasing talent from work that, in his words, “produces absolutely nothing useful for humanity.”
11. Incumbent software cannot treat AI as another menu item
Blundin said months of attempts to make Microsoft Copilot useful had “failed miserably.” Ismail’s diagnosis is architectural: adding AI as a feature to an existing product is “the wrong attitude,” with Microsoft and Apple presented as leading offenders against cleaner-sheet, AI-native competition.
Diamandis extends that warning to every corporate CEO claiming victory after adding one departmental feature. “It’s not a feature. It’s a brand new everything.” After Diamandis said an AI mechanism in Excel had failed, Ismail said he used Comet in the browser and completed the work better and faster.
The implication is not that incumbent distribution disappears overnight, but that installed bases do not substitute for redesign. The winning product may begin with a user objective and synthesize the workflow, rather than preserving decades of menus and inserting a conversational assistant beside them.
12. Washington is becoming both AI buyer and sovereign venture capitalist
xAI offered U.S. federal agencies Grok for $0.42 over 18 months—the “42” doubling as one of Elon Musk’s recurring 420 jokes. The larger story for Blundin is unprecedented direct entanglement between “corporate America and government America,” which he called effective but “a little scary.”
Ismail suggested the government might simply be testing every model, but he and Blundin disputed the arms-length-procurement interpretation. Ismail’s description was binary: companies enter the White House and are either “the anointed one” or not, with outcomes increasingly shaped by political access and edict.
Intel illustrates the new model. The episode cited the government receiving 10% at zero cost and earning 80% in six weeks after CEO Lip-Bu Tan’s August 11 meeting reversed presidential hostility; possible partnerships with AMD, Apple, and Nvidia became part of the “Team America” thesis.
Wissner-Gross calls quasi-nationalization structurally predictable because Moore’s second law says fab costs roughly double every four years. As fabrication becomes sovereign-scale infrastructure, security and financing overwhelm ordinary venture logic—supporting Diamandis’s category of assets that are not merely too big, but “too centrally critical to fail.”
13. Solving mathematics could arrive before quantum computing finds its killer app
Axiom Math founder Carina Hung, 24, raised $64 million at a stated $300 million valuation to build an AI mathematician. For Wissner-Gross, “solving math” does not mean every theorem is finished; it means the algorithmic process is solved and remaining problems principally require more compute.
His operational marker is FrontierMath Tier 4, whose pre-solved problems can occupy human researchers for weeks. A naïve logistic extrapolation reaches 10%–15% solved by AI at the end of 2025; at that level, he believes there is line of sight to any solvable problem without another algorithmic breakthrough.
First-order effects hit systems that rely on mathematics remaining hard, potentially including cryptography; second-order effects cascade through physics, economics, engineering, medicine, and science. Under his explicit hedge—“if this theory of the future ends up being correct”—society could be “drowning under a surreal Cambrian explosion of breakthroughs” within two to three years.
Quantum remains earlier and lacks an identified killer app comparable to GPUs’ progression from gaming to crypto to AI. Wissner-Gross worries more about AI mathematics than quantum attacks because post-quantum cryptography is approaching maturity; Diamandis’s practical warning is that already-stored AES-128 or AES-256 files might become exposed within a year.
14. AI infrastructure is becoming the largest financing market
Altman’s “abundant intelligence” argument begins with allocation: 10 GW might cure cancer or tutor every student, but compute scarcity forces a choice. His proposed escape is “a factory that produces a gigawatt of new AI infrastructure every week,” turning infrastructure itself into a continuously manufactured product.
Stargate targets $500 billion and 10 GW before the end of 2025. The episode compared that with 2024 AI-data-center spending of $40 billion for Microsoft, $16 billion for Amazon, $29 billion for Alphabet, and $23 billion for Meta, with Microsoft planning $80 billion in the current year.
Blundin’s caveat is balance-sheet asymmetry: OpenAI announces $100 billion and $300 billion deals without possessing comparable capital, while Zuckerberg has the cash and credit behind a stated $600 billion program. Altman’s advantage is agenda-setting—he says the number others will not, forcing land, governors, power, and plumbing to organize around it.
A proposed $100 billion, five-year Nvidia lease lets OpenAI obtain chips without buying them upfront and lets Nvidia finance demand from its own balance sheet. Wissner-Gross pushes back on the “GPU credit bubble” interpretation, seeing leasing as a market contortion bridging high-margin hardware with low- or negative-margin neocloud economics.
15. Energy, skilled trades, and robotics are the physical bottlenecks
OpenAI’s stated energy plan rises 125-fold to 250 GW by 2033, yet Wissner-Gross calculates that as only about 0.05% of solar energy reaching Earth’s surface. He wants terawatts, while also expecting photonics, software gains, small modular reactors, and potentially fusion around 2028–2030 to alter today’s requirements.
The near-term grid is less elegant: emergency powers kept Michigan and Pennsylvania fossil-fuel plants operating past retirement, with roughly 100 coal plants scheduled to retire in 2028. Data centers collide with a 500,000-worker construction shortage in 2025, making electricians, plumbers, carpenters, and automation managers potential $100,000–$200,000 careers.
Trade enrollment has risen 16% since 2023, and construction was described as 2025’s fastest-growing industry for new college graduates. Blundin stresses “construction automation” alongside manual work: states that secure data centers can attract high-end roles designing robots, modular systems, and automated building processes.
The component constraint is equally severe. After speaking with iRobot founder Rodney Brooks, Diamandis said China can fabricate a custom robotics part within days while the U.S. lacks an equivalent ecosystem; 1X and other companies reportedly compensate through vertical integration, prompting calls for a Manhattan-style robot-and-drone supply-chain project.
16. Exponential blindness links solar, robotics, and longevity timelines
Solar expanded from 40 GW in 2010 to almost 3 TW in 2025, yet Ismail showed experts repeatedly projecting linear plateaus. A 2003 specialist declared modules could never fall below $1 per watt because of materials; the episode placed current prices near one or two cents, illustrating how “never” calls fail.
His best orthogonal example came from Buenos Aires: car-wash revenue fell 50% despite more affluent drivers and cars. The cause was improved weather forecasting—roughly 50% better over 20 years—because people skip washes before rain. Moore’s law damaged a business far outside computing, where even the smartest operator might never see it coming.
China’s first-half 2025 robot exports rose 1,700% to Poland, 275% to Mexico, 135% to Russia, and 114% to Vietnam, while U.S. purchases rose 58%. Blundin suggested sovereign robotic ecosystems; early scarcity will invite government bidding, although Wissner-Gross reduces the core resource to compute: robots are ultimately “GPUs on legs.”
Autonomous driving supplies the first mass encounter with general-purpose robots. The episode estimated that Waymo-level performance across U.S. vehicles could prevent 33,000–39,000 annual deaths—a roughly 90% reduction—with nearly half of Waymo impacts below one mile per hour; Blundin added that car accidents generate about half of U.S. court cases.
17. Longevity becomes the clearest test of short AI timelines
Retro Biosciences, backed with $180 million in 2021, aims to add 10 healthy years. Its RTR242 Alzheimer’s pill is designed to restart the brain’s toxic-protein recycling process, with an Australia human trial discussed for late 2025; Retro also competes in the $101 million XPRIZE Healthspan field of more than 730 teams.
Diamandis highlighted work involving FOXO3, described as a stress-resistance transcription factor, and said Chinese researchers reported a three-to-five-year reduction in biological aging across 61 tissues. He treated biology’s portability—“if it works in China, it’ll work in Chicago”—as a reason longevity results can diffuse globally.
He corrected an exaggerated mouse claim: ordinary mice live roughly 20–24 months, with demonstrated extensions around 30%–40%, not a human-equivalent 300 years. Experiments seeking to double lifespan continue, while Life Biosciences was said to be starting human epigenetic-reprogramming trials in January after animal and non-human-primate work.
Diamandis said the frontier models he queried converge on longevity escape velocity around 2030, echoing Ray Kurzweil’s prediction. Wissner-Gross’s two-to-three-year extrapolations create “singularity paralysis,” but the actionable refrain remained: “Don’t die from something stupid.”