The AI Wealth Gap: Why 40x Deflation Changes Everything w/ Dave Blundin, Salim Ismail, Dr. Alex Wissner-Gross | EP #208
Summary
AI intelligence costs are reportedly deflating 40× year over year, turning capability into the episode’s “nuclear core” for broader price declines. Moonshot AI’s Kimi models—described as within roughly 1% of the top of SWE-bench—reportedly cost $4.6 million to train, about 30–40× below earlier OpenAI and Anthropic models. If that curve persists while demand grows 1,000× annually, advanced intelligence stops being reserved for hyperscalers and starts pulling healthcare, food, energy, and other costs down with it.
The near-term risk is that abundance arrives after AI has already widened the wealth gap. A 60,000-person survey across 32 countries put cost of living first globally, followed by unemployment and social inequity; Dave relayed an Iranian intern’s example of her parents spending one-third of annual income on an iPhone and data plan because “you can’t live without information.” Peter expects long-run demonetization but sees the next two to seven years as the danger zone; Dave warned that adding AI could make “the gap … really, really wide.”
Anthropic’s enterprise momentum is both a business-model signal and a bet that code generation unlocks recursive self-improvement. Dave said banks trust Claude with sensitive data and called AI management tools “the goldmine of all goldmines”; Peter cited projections of $70 billion revenue and $17 billion cash flow in 2028. Alex kept the technical call explicitly unresolved: code may rewrite models into an accelerating flywheel, or broad intelligence may still require “visual chain of thought” and physical grounding.
Power infrastructure is becoming the most legible picks-and-shovels trade in the AI buildout. One-gigawatt data centers are expected in 2026, while an $80 billion public-private nuclear initiative may build roughly ten 1.1-gigawatt AP1000 reactors—too late to answer the cited need for 92 gigawatts by 2030. Dave expects as much as $1.2 trillion annually to flow into data centers and power by 2030: “There’s nothing even close in the history of the world to that scale of money movement.”
AI science is moving from assistance toward bulk discovery, with timelines collapsing from decades to years. Sam Altman described GPT-5 as showing “tiny glimmers” of new science and said GPT-6 might deliver a GPT-3-to-4-like leap; Alex expects many or most grand challenges to start falling within three years. Edison’s Cosmos reportedly compresses four to six months of expert research into 12 hours, while disease-cure programs are shifting their stated horizon toward roughly 2030.
Regulation and labor resistance could determine which regions capture the productivity dividend. Peter cited European AI audits averaging €260,000 and 8–15 months, with European models reaching market 6–12 months later; Boston unions, meanwhile, want human safety drivers inside Waymos. Alex framed the missing technology as a system that achieves “acceleration and social cohesion at the same time,” because blocking deployment may preserve specific jobs temporarily while pushing investment elsewhere.
The same deflationary engine is opening physical and biological frontiers faster than governance can absorb them. World models can create traversable synthetic environments, coordinated drone swarms could act in the physical world, AI-guided weather control could steer storms, and embryo editing is moving from selection toward alteration. The episode repeatedly returned to legitimacy; in the weather-control discussion, Alex said deployment was “a social problem, less a technical problem.”
Deep dive
1. Anthropic’s enterprise wedge doubles as a bet on recursive AI
The episode’s Anthropic market-share chart led Alex to the decisive technical question: “whether code generation is the critical path to recursive self-improvement.” Anthropic has concentrated on code while OpenAI pursues modalities such as Sora; if code is the bottleneck, Anthropic’s enterprise gains may also indicate strategic proximity to a much larger capability breakthrough.
Alex’s honest non-answer preserved both cases. Code could rewrite “the core algorithms” and post-training architectures until “the flywheel spins faster and faster”; alternatively, superintelligence might require visual chain of thought or grounding in the physical world that cannot emerge from source code, text tokens, and limited imagery alone.
Peter took the aggressive side of that uncertainty, arguing Dario Amodei or Demis Hassabis could reach the next regime by directing enormous internal compute toward generating “the next test, the next test, the next test.” His marker was Humanity’s Last Exam: he said reports that it may soon be saturated matter because solving those questions and innovating in AI appear highly correlated.
Dave’s commercial read was simpler: banks and other enterprises use Claude because they trust Anthropic with sensitive data, while OpenAI leans consumer; “using AI as a management tool is the goldmine of all goldmines.” Peter cited Anthropic’s 2028 projections of $70 billion revenue, $17 billion cash flow, and a 77% margin, versus OpenAI’s $100 billion revenue target and losses through 2029.
2. Alignment becomes capability, while capital strategy determines reported profit
Alex described a “perverse duality between alignment and capabilities”: aligning a model closely with human intent is itself a valuable capability, while financing stronger alignment requires raising capital and training frontier systems. In his formulation, “every alignment project almost inevitably ends up as a capabilities project,” as happened first with OpenAI and then Anthropic.
Salim kept the institutional counterexample: Facebook began with strong privacy commitments before monetizing user information, so a safety-oriented origin does not prevent a company from becoming extractive. Dave nevertheless argued Anthropic has captured the moral high ground once associated with OpenAI—“if you want it to be guaranteed to be good for the world, come to Anthropic.”
Dave interpreted OpenAI’s prolonged losses as deliberate expectation-setting, not proof that its core service lacks margins. Sam Altman is effectively telling investors he intends to keep spending ahead of the curve; once a public company switches to profitability, asking shareholders to fund a trillion-dollar data center becomes far harder.
Peter compared that posture with Jeff Bezos’s early warning that Amazon would buy customers and revenue before “flipping the switch.” Dave expects Anthropic could similarly advertise a 77% gross margin, then launch an Anthropic-scale Stargate project that consumes it—sensible capital strategy, provided shareholders accept the bargain.
3. World models create a compute frontier beyond entertainment
World Labs’ Marble demo showed a photorealistic, traversable environment that Peter found both magical and troubling. His concern was not image quality but substitution: when users become “the god of your world,” an immersive simulation may compete directly with work, food, relationships, and ordinary reality rather than merely with existing games.
Alex called the consumer competition the beginning of the “Holodeck Wars,” but distinguished two architectures. Google Genie-3 generates every pixel on compute-intensive server GPUs; Marble generates 3D Gaussian splats—transparent blobs that accumulate into a 3D scene and can be rendered dynamically on the client with far less compute.
That creates an “efficient frontier” between versatility and computation, analogous to lightweight edge models versus server-intensive frontier models. Alex expects multiple tiers of generated worlds rather than one winning architecture, with economics determined partly by where rendering and generation occur.
Entertainment was “chicken feed” beside the addressable market Alex sees: synthetic environments for training vision-language-action models, robots, and scientific systems. Fei-Fei Li’s ability to demonstrate Marble on her phone in seconds reinforced Dave’s adjacent lesson—transformative products become easier to recruit, excite, and explain when their value is instantly visible.
4. Forgetting knowledge may make general intelligence radically smaller
Goodfire’s proposed unlearning method advances Alex’s vision of a “diamond-like perfect micro model” containing reasoning while externalizing almost all factual knowledge. The technique tests how small changes to weights affect overall loss: weights essential to generalization cause dramatic degradation, while weights holding memorized, possibly incorrect facts can potentially be pruned.
Peter initially framed the result as an enterprise privacy tool that could forget proprietary healthcare or financial records without losing the intelligence learned from them. Alex called that “half correct”: privacy may benefit, but the paper’s primary objective is separating general capability from memorized knowledge inside models containing hundreds of billions or low trillions of weights.
The holy grail would be a generally intelligent model with fewer than one billion parameters—or, aspirationally, around one million. Dave noted that 90% parameter reductions through distillation are already common: a 10× reduction could turn an assumed 100-gigawatt requirement into 10 gigawatts, making experimentation with open weights potentially worth trillions of dollars.
Google’s nested-learning paper attacks the complementary problem of continual adaptation. Alex described higher-order meta-learning—“learning to learn to learn to learn”—and a possible unification of models and optimizers as different layers of one information-compression process. Dave added that AI can connect its “brain” to notes and tools at immense bandwidth, making it functionally superhuman in learning workflows even before every internal faculty is superior.
5. Kimi puts frontier model development inside ordinary corporate budgets
Dave called Moonshot AI’s Kimi release “the biggest thing that happened in the last month.” He placed the models near the top of SWE-bench, roughly 1% behind Anthropic, while emphasizing their ability to run at high speed on Groq hardware—“G-R-O-Q, not Elon”—and their availability as open weights.
With Meta and OpenAI no longer providing their best models as open weights, Dave said the strongest editable systems are now coming from China. That matters to researchers and companies that want to “play with the guts” of a frontier model rather than merely buy API access.
The reported $4.6 million training cost was the key economic discontinuity: Dave estimated it was 30–40× cheaper than what OpenAI and Anthropic spent on their original models. Some savings come from drafting behind earlier innovations, but the result is that almost any company can now finance a serious training run.
Peter drew the geopolitical conclusion: if a trillion-parameter model costs about $5 million to train, efficient US capital markets cease to be a prerequisite for frontier work. Dave agreed the implications were “absolutely massive,” while adding a practical hedge: he had checked and was using Kimi, but users should still conduct their own spyware and security checks.
6. Forty-fold deflation is the engine pulling every other cost downward
Alex attributed to Sam Altman a 40× year-over-year decline in “cost of intelligence per unit of intelligence,” spanning training and inference. Salim contrasted that with demand rising roughly 1,000× annually, which explains why hyperscalers can keep increasing capital expenditure even as each unit becomes dramatically cheaper.
Alex called cheap intelligence the “nuclear core” of abundance: if sustained, it should drag down the prices of everything intelligence can materially improve. His forecast was that grand challenges in mathematics, science, engineering, and medicine would begin falling within two to three years under this compounding pressure.
Peter’s warning to skeptics was unusually categorical. People extrapolate from yesterday’s awkward model and diminishing benchmark increments, but two consecutive 40× improvements create a 1,600× change in economics; betting that this produces only a small capability gain is, in his words, “crazy.”
Salim’s counterpoint was historical rather than bearish: deflationary curves are normal in scalable technology, although 40× is faster than he expected. Carmakers dismissed electric vehicles because batteries were too expensive, then lithium-ion costs fell 90% over a decade; the lesson was to “go where the curve is pointing you.” Peter retained the condition: wider costs fall only where intelligence can solve the underlying problem.
7. Europe’s compliance burden is becoming a strategic tax on AI
Peter cited venture funding in Europe falling by as much as 30%, European AI models reaching market 6–12 months behind US counterparts, and mandatory audits averaging €260,000 plus 8–15 months. Those reviews reportedly delay 40% of projects while examining data, transparency, bias, documentation, and safety.
Salim argued Europe’s constraints are not merely cultural. His example was a postwar German constitutional restriction preventing one media organization from covering the entire country; regional fragmentation then left room for Google to aggregate the market. Undoing protections rooted in legitimate history is difficult even after their competitive side effects become clear.
Alex reduced the issue to sovereign choice: countries presently define “how much they want to participate in the superintelligence explosion.” Peter told European leaders they need energy plans, data-center sites, and deployable infrastructure within five years; Alex’s pushback was sharper—“more like five months than five years.”
8. The wealth gap may widen during the transition to abundance
Peter’s FII Priority Global Survey covered more than 60,000 respondents in 32 countries, representing roughly two-thirds of the world’s population. Cost of living ranked first globally, unemployment followed closely, and poverty and social inequity came third; Africa’s top concern was unemployment, while every other listed region led with living costs.
Dave grounded the charts in an Iranian intern’s example: her parents spend one-third of annual income on an iPhone and data plan. The intern’s explanation was that “you can’t live without information,” while an unusable currency made Bitcoin necessary and the phone essential for managing it. The spending then exits the country toward technology centers, concentrating wealth before AI adds another paid layer.
Salim argued that jobs are disappearing as a durable organizing principle while education still trains children into their early 20s for a labor market nobody can describe five years ahead. His preferred bridge is UBI; Peter leaned toward universal basic services and warned that the two-to-seven-year transition, before healthcare and education become abundant, is his greatest concern.
The disagreement was about execution, not the fear’s legitimacy. Salim wanted far more positive narratives because people are “10X more likely” to attend to fear; Alex wanted measurable targets—a cost-of-living benchmark, plus benchmarks for healthcare and crime—so 40× intelligence deflation can optimize toward outcomes rather than slogans. He separately suggested that social cohesion may need its own benchmark.
9. Coherent power is the trillion-dollar variable in frontier AI
With one-gigawatt facilities expected in 2026, Alex asked whether coherent training power eventually peaks. Distributed training innovations could cap clusters at a few gigawatts and then lower required density; further intelligence “phase changes” from ever-larger compression could instead drive facilities toward his extreme metaphor of a “desktop black hole computer.”
Peter cited an $80 billion US government, Brookfield, and Cameco partnership involving Westinghouse AP1000 reactors. Each Generation III+ plant provides about 1.1 gigawatts and may cost $7 billion, allowing perhaps ten reactors—but early-to-mid-2030s delivery does not answer the cited requirement for 92 gigawatts by 2030.
Alex positioned mature Generation III+ reactors inside a bridge from natural gas to nuclear fission and eventually fusion, with solar-plus-battery throughout. Unlike more experimental SMRs, of which Alex said there are perhaps only two or three, AP1000 is a relatively mature format; he cited at least six already built. The irony is that Westinghouse went bankrupt in 2017 building reactors before AI made dependable power strategically scarce.
Dave highlighted the financing template: private capital secures government-guaranteed loans, captures the upside if projects work, and limits downside if they fail. He projected $1.2 trillion annually flowing into data-center construction and power by 2030; with roughly 20 million GPUs manufactured in the cited year and every chip likely sold, the infrastructure requirement becomes unusually calculable.
10. Coordinated swarms may matter more than humanoid robots
China’s 16,000-drone display demonstrated what Dave thinks science fiction has underrepresented: perfectly coordinated swarms acting “down to the millimeter.” A swarm can assign two drones to a light object or 50 to a heavy one, making construction, yard work, gutter cleaning, and other physical tasks modular in a way one humanoid is not.
Salim said the Ukraine war is being prosecuted by roughly half a million drones “on either side,” while about 10,000 drones cross the Mexico-US border monthly: “drone technology beats wall technology.” Peter said Ukraine could become Europe’s drone-manufacturing capital after the war.
Tesla’s rumored Roadster demonstration was carefully bounded. Peter said gossip points to SpaceX cold-gas thrusters and perhaps 30 seconds of hopping or hovering, but stressed this is not an eVTOL; Archer, Joby, and EHang address actual multicopter transport, with Archer cited as holding a contract in Los Angeles for the 2028 Olympics.
Dave treated the Roadster spectacle as corporate strategy even if it never becomes mass transport. Tesla spends zero on conventional marketing and instead funds technically audacious projects that attract customers, employees, and attention: “What can I do to be inspirational and cool?” Peter added that the same formula makes elite engineers want to join.
11. Climate control and reusable launch systems turn engineering into governance
Elon Musk’s proposed solar-powered AI satellite constellation would make small adjustments to incoming sunlight. Peter’s sunshade formulation would reflect roughly one-quarter of 1%, acting as a planetary thermostat; reversibility is crucial, but conflicting national interests make any deployment a tragedy-of-the-commons problem.
Salim invoked Mount Pinatubo, whose atmospheric ash he said lowered global temperature by two degrees, and suggested existentially exposed regions might eventually launch interventions unilaterally. His rebuttal to anti-geoengineering objections: civilization is already geoengineering unintentionally through carbon emissions, while nation-state climate conferences have not solved the problem.
Alex’s broader proposal was a “global weather grid” that changes cloud cover through satellites, microwave heating, or other mechanisms, using AI weather models to mitigate or steer storms. Peter raised the liability problem—“Oops, we steered the hurricane in the wrong direction”—and Alex said deployment is “a social problem, less a technical problem.”
Blue Origin’s first New Glenn booster landing supplied the enabling-space counterpoint. The ESCAPADE mission launched toward Mars and the booster landed on Jacklyn; Peter cited SpaceX as launching more than 90% of US spacecraft and perhaps 70% of global launches. Alex welcomed a second reusable “railroad” to orbit because solar-system development cannot depend on one route.
12. Automation’s missing technology is social cohesion
Boston unions’ campaign against Waymo would require a human safety driver inside each vehicle. Dave called the policy “utterly insane” yet expected similar populist resistance everywhere; if major technology hubs cannot form workable transition policies, companies and their investment will simply deploy elsewhere.
Peter insisted the resistance begins with a rational survival calculation: “I need to feed my kids. I need to be able to afford my home.” Until society provides a credible safety net through UBI, universal basic services, or another mechanism, demanding that workers accept displacement as aggregate progress will not resolve their immediate loss.
Alex called regressionism one of the issues that keeps him awake: organizations may block technologies that save lives, generate wealth, and improve quality of life. The required “meta technology” must maintain cohesion while radically accelerating deployment—an optimal trajectory delivering both, not a choice between permanent stagnation and unmanaged disruption.
Salim framed institutional resistance as an “immune system problem.” He said his group has used a 10-week process around 100 times inside companies and a 16-week version in the public sector, then open-sourced the methodology; the harder next layer is adapting it to sector-wide immune systems in healthcare, journalism, and education.
13. AI science is shifting from isolated assistance to bulk discovery
Sam Altman’s framing set the progression: GPT-3 offered a first “spiritual Turing test” glimmer, while GPT-5 shows tiny instances of generating an idea or contributing usefully to a paper. GPT-6, he said, has a chance to produce a GPT-3-to-4-like leap specifically for science.
Alex’s stated maximum is three years for many, if not most, grand challenges in mathematics, science, engineering, and medicine to start falling. The anticipated output is “centuries of human capital” solved in bulk; Peter suggested tracking AI-generated patents and agent-to-agent transactions, including automated licensing, to see the exponent emerge.
The Chan Zuckerberg Initiative illustrates the timeline compression. Its 2016 ambition was curing most disease by the end of the 21st century; with Biohub, virtual cells, and a planned 10× compute increase by 2028, Alex read the new implied horizon as perhaps 2030, similar to timelines he attributed to Anthropic.
Virtual cells could expand into virtual organs and organisms, letting AI search intervention space rather than testing diseases artisanally one by one. Alex’s counterintuitive possibility was that “sometimes it’s easier to solve the more general problem than the more specific problem”—including curing all diseases before completing every separate cure program.
14. Cheap GLP-1s and AI scientists preview healthcare abundance
Peter cited a proposed GLP-1 price of $149 per month and results suggesting the drugs can cut repeat-stroke incidence by as much as half over three months. Alex called their broad benefits biologically mysterious: the “elephant in the room” is why a metabolic drug class appears to improve so many different forms of dysfunction.
Alex connected that affordability to universal basic services—roughly $150 monthly begins to resemble broadly available healthspan medicine. Peter supplied the critical warning: GLP-1 drugs are not a panacea; patients can lose muscle with fat, then regain fat without restoring muscle after stopping, making resistance exercise essential because “your muscle is your longevity organ.”
Edison’s Cosmos offered a different abundance mechanism: an agentic scientist using knowledge graphs and scaffolding to simulate context far beyond today’s millions-of-token windows. The ideal system would ingest trillions of tokens—the internet and every paper—then answer questions such as “What’s the solution to Alzheimer’s?”
The reported operating metrics were four to six months of expert research completed in 12 hours, 1,500 papers read, and 42,000 lines of code run per experiment. Peter argued the larger opportunity is feeding models raw experimental records, allowing neural systems to analyze interacting variables together instead of isolating one chemical reaction and barely clearing statistical thresholds.
15. Embryo editing marks a regime change from selection to alteration
Preventive and Manhattan Genomics were presented as companies developing CRISPR embryo-editing capability, potentially outside the United States and perhaps in the UAE. IVF already permits sequencing embryos and selecting which to implant; Peter called direct editing a “regime change” because it moves reproduction from choosing among existing genomes to altering one.
Salim’s premise was that “the human genome is essentially software” and a human therefore becomes a software-engineering problem; to him, editing embryos is inevitable, leaving “what do you want to design for?” as the central question. Peter argued parents already seek the best healthcare, education, and inherited traits, but acknowledged that extending this logic immediately raises eugenics fears.
Alex traced modern caution to the 1975 Asilomar guidelines and relayed a historical argument that fresh memories of Watergate influenced scientists’ desire for transparent limits. Fifty years later, he said he could not find one US federal statute categorically banning germline editing—only a patchwork of federal and state laws and regulations that strongly deter it.
The warning specimen remained He Jiankui’s 2018 CCR5 editing, intended to reduce HIV susceptibility, followed by global condemnation and imprisonment in China. Alex challenged the default dystopian reading of Gattaca, while Dave said the field urgently needs credible thought leaders; his final pushback was demographic: South Korea’s cited birth rate is already only 0.7 children per couple—“no one’s having any babies at all.”