AI Insiders Reveal Elon Musk's Master Plan to Win AI w/ AWG & Dave Blundin | EP #192
Summary
xAI is betting that frontier-model leadership is a physical scale race, and Elon Musk is treating second place as worthless. The panel said Colossus 2 would bring 1 gigawatt and 500,000 NVIDIA Blackwell GPUs online in Memphis, double to 1 million in 2026, and train Grok 5; at the stated $30,000 per Blackwell, chips alone imply roughly $30 billion. Dave Blundin’s formulation was blunt: “You don’t build the second-biggest data center. You either win the race or you don’t win the race.”
Grok Code Fast 1 looks less like commodity pricing than a heavily subsidized land grab for coding workflows. Diamandis cited prices of $0.20 per million input tokens and $1.50 per million output tokens, versus $1.25 input for GPT-5 and $3 input for Claude Sonnet 4; Blundin rejected the “race to the bottom” framing because demand is effectively unbounded, comparing it to “a crack dealer giving out the first hit for free.” Distribution may matter as much as benchmarks: consumers reach the model through coding environments such as Cursor or Windsurf, not directly through a browser.
Horizontal AI models are beginning to erase the interface and feature moats that sustained application software. Google’s Gemini 2.5 Flash Image, or Nano Banana, was presented at 3.9 cents per image with character consistency, multi-image blending and language-based editing; Google Translate now attacks language learning directly, alongside a reported 10% Duolingo stock drop. Alexander Wissner-Gross generalized the threat: “That chatbot is going to devour you” unless vertical SaaS companies raise their ambition by “100x, a thousandx.”
Real-time multimodal models could become the single interface through which consumers transact with every industry. OpenAI’s GPT Realtime API was shown searching Zillow around an $824,000 buying ability, while Diamandis described AI as an overlooked management layer that can turn cameras and unstructured operational data into schedules, purchases and corrective actions. Wissner-Gross proposed “streaming interactive models,” with every voice segment, image and interface generated just in time.
AI profits currently appear concentrated in chips and infrastructure, but export restrictions are encouraging alternative stacks. NVIDIA was described as a $4 trillion company with revenue up 56% year over year and shares up 700% since ChatGPT’s 2022 release; Wissner-Gross therefore leaned toward the “pyramid” thesis in which value pools at the stack’s base. Yet Cambricon, Huawei and other Chinese architectures could weaken the NVIDIA/CUDA monoculture, especially because software gains of 100x–1,000x might overwhelm a nominal 7-nanometer-versus-2-nanometer disadvantage.
Record AI capex and equity valuations are either an ASI-era economic signature or fertile ground for a violent, temporary repricing. The episode cited $375 billion of AI infrastructure spending in 2025, roughly $500 billion in 2026, Jensen Huang’s expectation of $600 billion annually, and a roughly one-percentage-point GDP contribution. With Nasdaq capitalization at 176% of M2 and 129% of GDP, Wissner-Gross asked what the ratio should do before superintelligence; Diamandis answered “approaching infinity,” with Blundin agreeing, while Diamandis still allowed that panic could drive major drawdowns.
Embodied AI is the route by which data-center intelligence enters manufacturing, households and transportation. Jetson AGX Thor was described as delivering 2 petaflops of FP4 compute and 10x Orin’s performance, while China’s humanoid sales were projected above 10,000 units in 2025, up 125%. Tesla’s vision-only training, 1X’s fleet learning and Apple’s supplier-automation mandate support the same thesis: Diamandis said “every job is training the centralized version,” and manufacturers that do not automate may not survive.
Health AI is moving diagnostics into inexpensive, continuously monitored environments, while the longevity evidence discussed remains far more experimental. The panel highlighted a 15-second AI stethoscope and an AI-guided handheld ultrasound, then discussed a mouse psilocybin study reporting 80% versus 50% survival and up to 57% longer cellular lifespan. Diamandis also described his own stem-cell “re-education” procedure and dramatic patient videos, but his personal biomarker results were still pending at one, three, six and 12 months.
Deep dive
1. A two-to-three-year superintelligence horizon breaks conventional career planning
Wissner-Gross’s advice to incoming freshmen was conditional, not a blanket command to quit: if the goal is a startup, there is a strong incentive to start now and perhaps leave school. His governing assumption was that superintelligence could arrive “in the next two to three years,” making career projections inherited from 20 or 30 years ago unreliable.
Diamandis offered a more durable compass: identify a problem that produces genuine commitment, then keep applying the newest intelligence to it. Technologies will change constantly, but a founder’s “why” can persist; his advice was to pause conventional progress long enough to find that driver.
MIT’s new 6E entrepreneurship option provided the institutional counterexample to supposedly irrelevant curricula: students can spend semesters inside a startup or venture fund, learning how companies operate. Wissner-Gross said he would teach advanced algorithms in that program, illustrating a hybrid between formal study and immediate venture execution.
2. Colossus turns the bitter lesson into an infrastructure strategy
The stated plan for Colossus 2 was a 1-gigawatt Memphis facility fitted for 500,000 NVIDIA Blackwell GPUs, doubling again to 1 million in 2026 and becoming the birthplace of Grok 5. Colossus 1, Diamandis emphasized, went from zero to completion in 122 days despite claims that it could not be done.
Wissner-Gross called this “the bitter lesson as applied to hardware scaling.” Richard Sutton’s lesson, as he summarized it, is that decades of “artisanal solutions” for vision, language and speech were steamrolled by general algorithms combining enormous datasets with compute; even the once-newsworthy cat detector became a trivial emergent capability.
Blundin framed Colossus and OpenAI’s Stargate as a winner-take-all footrace: trained models can compile into fast, easily tailored systems, but customers will not knowingly begin with model number two or three. Musk, Sam Altman and their capital providers are therefore “all in.”
Training and inference should not be conflated. Wissner-Gross compared training to development or compile time and inference to execution: frontier training remains relatively concentrated in the United States, while civilization is beginning to tile the planet with inference capacity in India, the Middle East, Norway and elsewhere.
3. Power and capital, not just GPUs, determine who can train the frontier
Diamandis described Musk’s fundraising access as “basically infinite capital,” with family offices and sovereign wealth funds prepared to oversubscribe his rounds. Using Blundin’s estimate of $30,000 per Blackwell, 1 million chips represent $30 billion before racks, networking, cooling, construction and electricity.
Colossus 1 was described as using on-site natural-gas cogeneration rather than materially relying on the grid. Wissner-Gross expects a “pocket economy” in which data centers colocate with nuclear plants, gas generation or other dependable energy because utilities and permitting cannot expand quickly enough.
Blundin’s generator anecdote made the secondary bottleneck concrete: data-center developer Jeff Markley had reportedly bought every available 3-megawatt generator after discovering that the 5-megawatt units were already sold out. Even access to fuel or grid supply is insufficient if equipment cannot turn it into usable electricity.
Diamandis consequently shifted the investment question toward generation and electrical infrastructure, recalling Eric Schmidt’s formulation that AI is “energy limited,” not chip- or intelligence-limited. The panel proposed examining Leopold’s disclosed holdings in a future episode rather than making a fully formed trade call here.
4. Grok Code Fast 1 uses price and workflow distribution to seize demand
Diamandis quoted Grok Code Fast 1 at $0.20 per million input tokens and $1.50 per million output tokens. Against $1.25 input for GPT-5 and $3 for Claude Sonnet 4, he characterized the model as roughly 15 times cheaper on input and 10 times cheaper on output in the comparison shown.
Blundin disputed the obvious conclusion: “It’s definitely not a race to the bottom, even though it appears to be.” His argument was that cheap initial access builds habits and stimulates effectively infinite demand, while users soon want an unlimited supply of the capability.
Wissner-Gross highlighted an unusual distribution choice: consumers could not simply open Grok Code Fast 1 in a browser; they had to find it inside environments such as Cursor or Windsurf. Coding tools are therefore becoming entry points to intelligence that may compete with browsers for strategic control of users.
The economics remained unresolved. Diamandis questioned how the price could cover the underlying FLOPs and said billion-dollar losses could not be sustained indefinitely. Wissner-Gross’s answer was that labs use differing accounting schemes, and that routing optimization may reduce costs; Jevons’ paradox may nevertheless cause total spending to rise as generated code approaches zero marginal cost.
5. Mission, equity and sprint speed form Musk’s recruiting system
The panel interpreted xAI’s recruitment of 14 Meta engineers as a purpose-and-equity offer rather than a pure cash auction. Diamandis argued that builders gravitate toward missions with a “pure signal,” while Musk’s prior companies give candidates confidence that the effort can produce valuable equity.
That contrasted with reports of researchers leaving Meta’s new superintelligence operation despite nine-figure packages, including several returning to OpenAI. The hosts repeatedly acknowledged that they did not know Meta’s internal facts; their inference was narrower—that money alone may not capture a recruit’s “time and attention and heart.”
Wissner-Gross welcomed the competition as evidence against an AI “monoculture” in which one laboratory captures the future. Multiple frontier labs can embody different cultures, including xAI’s “manic focus and intensity,” and humans benefit from those cultures competing rather than one gaining permanent hegemony.
Musk’s “sustainable abundance” vision joined abundant energy and materials with abundant intelligence and autonomy. Blundin paired that long-run abundance with short, 9-9-7-style execution sprints; Wissner-Gross noted that China’s term was 996 and that people were reacting against that culture. A million-GPU data center completed a year late could be worthless, although both Diamandis and Wissner-Gross stressed that seven-day labor is not a sustainable life or necessarily a permanent human requirement.
6. Nano Banana replaces interface expertise with language—and hints at a world model
Google’s Gemini 2.5 Flash Image, nicknamed Nano Banana, was presented as fast and API-priced at 3.9 cents per image. Its standout capability was consistency: it could preserve people, pets and objects through edits, blend as many as 13 supplied images, and generate a plausible view from a selected map location.
The immediate disruption is to interface-based expertise. Diamandis summarized it as “edit through language, not layers”: users no longer need years of familiarity with masks, menus and proprietary controls, removing the switching costs that helped products such as Photoshop and Canva retain subscribers.
Blundin compared the transition with the move from command lines to graphical interfaces, but predicted a larger step change: software will simply act on a spoken request. His mother, in her 80s, could create images conversationally despite never mastering Photoshop—an example of capability democratization and incumbent-interface decay.
Wissner-Gross thought “Photoshop killer” understated the advance. Asking Nano Banana to show the same scene from another perspective implies some internal representation beyond pixel editing; he interpreted it as a possible “tendril from a much more monstrous model,” foreshadowing world models merging into Gemini-class systems.
7. Synthetic media makes visual trust an explicit technical choice
Diamandis’s concern was not merely better editing but accelerated misinformation and the erosion of visual trust. His default reaction to new video is already “Is it real?” and he expects many people’s second reaction to become “No, it’s not real,” ending the old presumption that seeing is believing. Blundin summarized the result as “deepfakes on an industrial scale.”
Wissner-Gross proposed a cryptographic chain from cameras through browsers that could attest an image was unaltered, analogous to secure online payments. He immediately hedged the proposal: his baseline expectation was that verified-only media would not become popular enough to restore universal trust.
Google’s SynthID watermark suggests an ongoing generator-versus-detector arms race, but authorship remains conceptually unsettled: if a person describes a moon scene and software renders it, is the author the prompter or the model? The panel left ownership and human-in-the-loop creativity as open questions.
Wissner-Gross also pushed back on the “moral panic.” Once smart glasses place real-time augmented overlays across ordinary vision, “everything becomes photo edited” by default; he expects today’s anxiety to look quaint relative to the scientific gains and the broader shift toward continuously synthesized reality.
8. Translation shows how a general model cannibalizes vertical SaaS
Google Translate was said to process roughly 1 trillion words monthly for 600 million users across 243 languages, or 58,806 language pairs. Gemini 2.5 adds low-latency conversational translation, bringing transformers back to the machine-translation application for which their encoder-decoder architecture was originally developed.
Wissner-Gross posed a second-order question: universal interoperability might protect low-resource languages rather than finish them off. If speakers no longer need to converge on English to communicate, AI could promote linguistic diversity instead of accelerating the collapse toward one dominant language.
The market warning came from adjacent products. Blundin cited Chegg falling from about $90 to $1.40 after ChatGPT offered a better way to cheat on homework, though Diamandis noted that cheating is not really what Chegg does. Google’s language-practice mode was linked to a 10% Duolingo stock drop on August 29. Duolingo was described as a $13 billion business with 130 million active users, only 10% paying.
The broader call was existential: every SaaS product risks becoming one use case inside a general model. Installed users and regulation can buy time, but Wissner-Gross urged threatened companies to “step up your ambition by 100x, a thousandx”—for Duolingo, perhaps moving toward a brain-computer interface that loads a language in a minute.
9. Real-time models become both interface and operating manager
OpenAI’s demonstration had a Zillow assistant use an $824,000 buying ability, locate a Wallingford property with skyline and Mount Rainier views, and offer to schedule a tour. Diamandis’s desired endpoint was not another search box but an agent that finds the house, buys it, finances it, arranges the moving trucks and reports when to arrive.
Wissner-Gross described GPT Realtime as an API version of low-latency Advanced Voice Mode, suitable for third-party customer service and other applications. His proposed category, “streaming interactive models” or SIMs, extends beyond voice to interfaces, imagery, simulated worlds and perhaps brain-computer inputs generated continuously on demand.
Diamandis emphasized management over spectacle. Cameras already produce vast unstructured datasets; intelligence can convert them into decisions, schedules and interventions—such as catching builders installing Tyvek with the overlaps reversed before water is funneled into a wall and the structure rots.
Blundin’s operating example was Vocara, a voice customer-service company in Link Studio that had doubled annual recurring revenue during the prior week and planned to grow tenfold by year-end. He said consumers preferred it to human call centers, especially once multimodality lets an agent create images and retrieve thousands of examples during a conversation.
10. Infrastructure captures today’s profits while national stacks multiply
NVIDIA was described as defying bubble fears with revenue up 56% year over year, a $4 trillion valuation and a 700% share-price gain since ChatGPT launched in 2022. The panel also reported that NVIDIA had given 15% of China sales to the United States to preserve export access.
Wissner-Gross contrasted a normal pyramid, where profit accumulates in chips, fabrication and data centers, with an inverted pyramid, where applications capture the value and models become commodities. NVIDIA and China’s chip enthusiasm currently pushed him toward the first view: rents are still gathering near the infrastructure base.
Cambricon and Huawei nevertheless suggest the beginnings of a post-NVIDIA, post-CUDA monoculture. Diamandis’s objection to export restrictions was causal: denying China products forces domestic substitutes, as earlier satellite restrictions did; he added that 100x–1,000x algorithmic gains could matter far more than a 7-nanometer-versus-2-nanometer fabrication gap.
India supplied another route to diversification. Diamandis expected Mukesh Ambani’s Reliance Intelligence to repeat Jio’s 2016 playbook—free voice, extremely cheap data, rapid activation and a national 4G leapfrog—while Wissner-Gross emphasized India’s unusually large pool of 20-to-40-year-old talent and AI’s potential to reach people directly despite institutional friction. Blundin added that poor transportation could create opportunities for aerial delivery.
11. AI enters government through interfaces and drafting before elections
Airbnb co-founder Joe Gebbia’s appointment as US chief design officer was framed as a practical entrepreneurial intervention: make government services “as satisfying to use as the Apple Store.” Wissner-Gross identified the open-source US Web Design System as a high-leverage place to improve common components across many federal sites.
Cloudflare’s Matthew Prince topping TIME’s AI list raised a different governance issue: whether autonomous agents should access websites with the same rights as humans. Diamandis’s explanation was that publishers value Prince’s efforts to protect attribution and the economics of internet content from indiscriminate model scraping.
On voting AI into power, Blundin expected AI to draft laws because legislatures cannot match the number or speed of emerging problems, but predicted human politicians would attach their names. Diamandis wanted more machine competence yet doubted entrenched bureaucracies and courts would permit it; corruption might simply migrate to engineers, companies or states controlling the system.
Wissner-Gross rejected the premise as “nonsensical.” His third position was that humans and AI will merge and perhaps speciate, reducing the question to whether people will elect people—still yes, but now biologically and computationally coupled rather than cleanly separated.
12. ASI expectations make conventional bubble comparisons incomplete
The panel cited Jensen Huang expecting $600 billion a year of AI spending, infrastructure investment reaching $375 billion in 2025 and about $500 billion in 2026, and the buildout adding roughly one percentage point to US GDP. Capital from Silicon Valley, sovereign funds and family offices is flowing into the physical economy through AI construction.
The alarming comparison was Nasdaq capitalization at 176% of the US money supply and 129% of GDP, both above dot-com peaks. Blundin cautioned that Nasdaq now represents a larger share of the economy because technology itself became larger, while broader-market price-to-earnings ratios looked less extreme than the chart implied.
Wissner-Gross’s thought experiment was the episode’s cleanest challenge to bubble framing: immediately before artificial superintelligence, what should Nasdaq-to-M2 do? Diamandis answered “approaching infinity,” with Blundin agreeing; Wissner-Gross expected rents to concentrate temporarily among listed infrastructure providers, then perhaps plateau as intelligence-driven profits diffuse through the economy.
Diamandis still allowed for a severe market decline driven by panic. His dot-com analogy was that the internet remained real even while Amazon fell more than 90%, and 9/11 followed as what he viewed as a great buying opportunity; technological validity does not eliminate valuation volatility or timing risk.
13. AI leaves the data center through medicine, robots and autonomous mobility
In healthcare, the panel highlighted a UK AI stethoscope detecting major heart disease in 15 seconds and Exo’s handheld ultrasound, which tells a user where to move or rotate the probe before interpreting the image. Diamandis expects continuous home, wearable and toilet sensors to shift insurers from reimbursing sickness toward financing prevention.
The longevity material was explicitly more experimental. A 2025 mouse study was described as extending cellular lifespan by up to 57%; dosing old mice from roughly week 20 produced 80% survival versus 50% after 28 weeks. Wissner-Gross said the paper used psilocin and hoped its telomere-related benefits could eventually be separated from central-nervous-system psychedelic effects.
Diamandis also reported cycling about 12 liters of his blood, extracting roughly 300 cc of immune cells, co-incubating them overnight with cord-blood stem cells and reinfusing 1.27 billion “re-educated” cells. He showed dramatic type 1 diabetes and ALS videos and made strong therapeutic claims, but his own evidence remained prospective: biomarkers were scheduled at one, three, six and 12 months.
NVIDIA’s Jetson AGX Thor was the embodiment bridge: 2 petaflops of FP4 compute, about one-tenth of a Blackwell or 30 iPhone 16 Pros, with 10x Orin’s performance. China’s humanoid sales were projected above 10,000 in 2025, up 125%, while the longer-range forecast was 10 billion humanoids by 2040.
14. Fleet learning turns human activity into robot-training infrastructure
Tesla’s Optimus strategy was described as moving from motion-capture suits to vision-only learning from worker videos, echoing Musk’s rejection of lidar for autonomous driving. Wissner-Gross called billions of smart-glasses wearers the “bitterest lesson” for robotics: passive footage could train models across nearly every manual trade.
Diamandis said that at 1X, fleet learning was compulsory for the first roughly 10,000 units: household telemetry flowed to a central system with no opt-out. The payoff is collective memory—if one robot knocks over a coffee cup, the other 9,999 can learn not to repeat it because “every job is training the centralized version.”
Apple was said to be requiring tier-one suppliers to automate wherever possible, increasing consistency and reducing cost. Wissner-Gross extended that logic to sovereign supply chains: sufficient robotic inference could let countries domesticate manufacturing that previously depended on cheap offshore human labor.
Waymo’s scale still trailed Uber—700,000 trips per month against Uber’s 30 million per day—but one Waymo reportedly completed more daily trips than 99% of Uber drivers because it could operate almost continuously. Diamandis predicted autonomous rides would become four times cheaper than car ownership, reshaping access, parking and where people can afford to live.
15. Today’s electricity shock may contain tomorrow’s efficiency surplus
US consumer electricity prices were shown rising from 2021 amid the data-center demand discussion. The near-term response is more generation colocated with compute, but Wissner-Gross warned against extrapolating a straight line if recursive self-improvement produces another DeepSeek-like algorithmic shock.
Diamandis contrasted the human brain’s roughly 20-watt consumption with his estimate that frontier models require 100,000 to 1 million times more energy. Better chips and algorithms could therefore yield an efficiency gain of 10^5–10^6, turning the same installed energy base into vastly more intelligence.
Wissner-Gross pushed the theoretical ceiling higher: the brain is not an optimal computer, and reversible computing could move beyond conventional Landauer-limit assumptions. The episode’s final energy thesis was therefore dual-sided—power is the immediate bottleneck, but intelligence may eventually redesign computation quickly enough to reverse the apparent scarcity.