Summary

Intel is framed as a strategic asset whose failure risk has now been politically capped, not merely a cyclical chip trade. Dave Blundin’s chain: Intel is America’s only truly domestic fab, TSMC has 66% market share in the chips that drive AI, and China is only 90 miles from Taiwan; therefore, “the US cannot let Intel fail.” The government’s 10% investment and Nvidia’s follow-on sent Intel up roughly 25%-30% and supplied capital for its 18-angstrom 18A process, taking complete failure largely off the table.
The Microsoft-Apple rescue of 1997 is the episode’s preferred template for Intel’s potential rerating. Alexander Wissner-Gross recalls Microsoft bundling Internet Explorer, renewing Office support and creating a surprising partnership with Apple; Salim Ismail says the investment restored Apple’s credibility when it was nearly bankrupt. The implication is not that Intel becomes Apple, but that a credible strategic sponsor can turn a distressed balance-sheet story into a platform turnaround.
Compute infrastructure drew the strongest investment conviction, while Wissner-Gross rejected the premise that public markets offer an easy informational edge. Blundin favors seed-stage AI builders or, at scale, the “tiling of the entire Earth with compute”; Ismail points to NextEra and renewables, while Peter Diamandis considers uranium, SMRs and Generation IV reactors because America is “electron-limited.” Wissner-Gross’s pushback: algorithmic markets already embody collective AI intelligence, making stock-picking without insider information logically inconsistent with the superintelligence thesis.
Blundin’s Intel-options example shows both the payoff and opacity of following disclosed hedge-fund positions. A disclosed $500 million Intel options exposure linked to Leopold Aschenbrenner’s fund was measured on a converted-share basis, not necessarily by premium paid; Blundin used a Perplexity-simulated call ladder and bought options then trading below $1. Those calls were up about 255,250% at the point shown, producing what he called “the biggest one-day gain in my entire life.”
Gold-level coding performance may mark the next capability jump after reasoning models. Wissner-Gross traces the prior leap from Q* to Project Strawberry and the o-series, including o1 and o3, then argues that Olympiad systems are becoming superhuman on tasks where partial answers cannot be easily verified. His forecast is that this capability could become broadly available by year-end, reinforcing Dario Amodei’s stated path toward 90% AI-written code and 100% in 2026.
Meta’s display glasses matter less as eyewear than as the beginnings of a mass-market peripheral-nervous-system interface. The EMG wristband turns subtle finger activity into silent control; Wissner-Gross calls the glasses “a bit of a red herring” and the input system the beginning of a non-invasive BCI. Always-on cameras could provide memory augmentation and billion-user robot-training data, but they also introduce “Black Mirror-esque” scenarios and make disconnection feel like losing a symbiote.
Export restrictions may accelerate a rival Chinese stack and could hand Huawei a more open global posture. David Sacks’s warning, as quoted by Diamandis, is that blocking U.S. suppliers could help Huawei build a “Digital Silk Road”; Blundin and Ismail emphasize Europe, India and the sovereign-AI land grab over the next two or three years. Wissner-Gross expects geopolitical separation to produce more architectural diversity, just as Cold War isolation encouraged Soviet experiments with ternary computing.
AI is moving from answering questions to designing science, forecasting health and recursively expanding physical production. DeepMind’s Navier-Stokes work might ultimately enable exotic fluid-based computation, while Delphi-2M models disease histories for nearly 1.9 million people as token sequences and projects risks across more than 1,000 diseases over 10-20 years. In robotics, Figure’s billion-dollar-plus Series C and Optimus 3’s million-unit ambition point toward “the machine that makes the machine”—and a disputed singularity that Blundin thinks could deliver a 1,000x software jump inside 30 days.

Deep dive

1. Intel’s rescue turns fab scarcity into national industrial policy

Blundin’s central logic is geographic and industrial: Intel is “our one and only chip fab company that’s truly domestic,” while TSMC has 66% market share in the chips driving AI and sits only 90 miles from China. That made an American backstop, in his telling, unavoidable.
After the U.S. government invested in and bought 10% of Intel, Nvidia followed; Blundin suspects “some White House pressure,” though he offers that as conjecture. Nvidia is effectively sold out for the indefinite future, Intel owns scarce fabrication capacity, and fabs—not demand—are the binding constraint.
Intel rose roughly 25%-30% that day. More consequentially, Blundin says the capital lets it finish the 18-angstrom process called 18A, described as “absolute cutting edge, front of the queue,” while removing the tail risk of complete corporate failure.
The atomic-scale process prompted Diamandis to marvel at barriers once considered impossible. The broader theme was that technologies taught as hard limits at MIT—from post-gigahertz clocks onward—keep becoming engineering milestones.

2. The Microsoft-Apple analogy defines the bull case—but not its certainty

Wissner-Gross compares the intervention with Microsoft’s surprising 1997 partnership with Apple: Internet Explorer became the Mac’s default browser, and Office support was refreshed or renewed for Apple devices. He describes it as a turning point for Apple; Ismail adds that the investment restored credibility to a company that was nearly bankrupt.
Ismail recalled the investment as, he thinks, $150 million and 10% of the company. He argued that credibility plus turnaround cash preceded Apple becoming the world’s most valuable company.
Ismail, not Blundin, called the Intel-Nvidia structure “virtually identical” and suggested comparing the two companies’ pre-rescue charts. The historical rhyme is the episode’s thesis, not proof: a strategic investment can reopen a path to execution, but Intel still has to complete the turnaround.
Wissner-Gross’s older Nvidia classroom slide supplies an example of successful repositioning: Nvidia was worth roughly $500 billion in spring 2023 and still drew most revenue from graphics chips. He now places it at $4 trillion-$4.5 trillion, roughly an 8x rise after AI became the dominant use case.

3. The portfolio debate split infrastructure conviction from market efficiency

For a small pool of capital, Blundin prefers seed-stage teams applying AI to “tens of thousands of untouched use cases.” For large checks, he chooses the data-center buildout—the “tiling of the entire Earth with compute”—including power, land and AI-native neoclouds such as CoreWeave and eventually Crusoe.
Ismail’s long-duration choice is NextEra: decentralized renewable generation with dividends and relevance to data-center electricity demand. His two inflection points were 2016, when new solar became cheaper than new fossil capacity, and 2019, when solar capex plus operation became cheaper than merely operating an existing fossil plant.
Diamandis remains constructive on Google, arguing its internal TPU supply is fully consumed by Gemini and Veo 3 growth, and on Tesla after Elon Musk returned his focus to the company. For power, he leans toward uranium because Helion’s small fusion plant is expected in 2028 or 2029 and Commonwealth Fusion’s in 2031: “We’re not chip-limited. We’re not intelligence-limited. We’re electron-limited.”
Wissner-Gross’s “elegant copout,” as Diamandis called it, was to choose the market’s collective intelligence rather than any ticker. If equities and commodities are already AI-dominated, claiming an edge without insider information contradicts the superintelligence premise. Blundin added that pumped-hydro sites and real estate around nuclear plants had already repriced rapidly.

4. Options and neocloud finance turn compute scarcity into leverage

Using 13f.info, Blundin described Leopold Aschenbrenner’s fund growing from about $1 billion to $2.1 billion through both inflows and gains. Its CoreWeave position rose from roughly $50 million to probably $150-plus million, implying a gain above $100 million.
The disclosed $500 million Intel options position represented converted shares at the stock price, not necessarily capital spent. Blundin asked Perplexity to simulate the underlying ladder, then bought the suggested high-risk calls when some traded below $1; on the Intel news, they were up about 255,250% at the point shown, and somewhat more later.
His result was unusually personal: “Today I have the biggest one-day gain in my entire life.” The example also contains its own warning—the simulated contract mix was not confirmed as the fund’s actual book.
Blundin’s next infrastructure bet is a stealth financing company combining crypto with CBOE-listed options to fund neocloud construction. The builder story mattered to him as much as the structure: a lucrative quant trader is leaving a solitary job to help expand physical compute capacity.

5. Coding Olympiad gold may be the next reasoning-model discontinuity

Wissner-Gross places the coding result alongside recent gold-level performance at the International Mathematical Olympiad, the International Olympiad in Informatics and collegiate coding competitions such as the ICPC. These contests matter because systems cannot repeatedly ask a judge whether each intermediate step is correct; the models must reason through difficult-to-verify paths.
His capability history runs from Q* to Project Strawberry and then the o-series, including o1 and o3. That was the leap from models that could not reason to ones that could; Olympiad performance “smells to me like the next major leap”—reasoning without having been trained only on easy-to-verify problems.
The forecast is conditional but aggressive: perhaps by the end of the calendar year, such systems become generally available, after which humanity first “solve[s] superintelligence” and then uses it to solve math, science and engineering. Reports from frontier labs already describe developers directing coding agents instead of writing code themselves.
Diamandis compared the transition with generative art quietly winning competitions, but Wissner-Gross redirected the metaphor toward measurement. Coding and math have explicit superhuman waterlines; art and conversation did not, which helped society “zoom right by” the Turing test without a single unmistakable crossing.

6. Meta’s neural wristband is the product; the glasses are the wedge

The new Meta Ray-Ban Display system combines a private in-lens display with an EMG wristband that reads forearm signals for silent input. The group judged the frames thick but “good enough,” and Blundin expects users to accept larger form factors once more rim space buys battery life and display capability.
Ismail predicts the wrist manipulation will prove temporary because cleaner signals may come from mouthing or more direct brain/body interfaces. Wissner-Gross agrees that the glasses are “a bit of a red herring”: the important development is a non-invasive peripheral-nervous-system interface that may be the beginning of a BCI.
Wissner-Gross’s likely sequence is smartphones, smart glasses, BCIs, then human-machine symbiosis or uploads. The wrist interface could migrate toward a headband; Diamandis expects demand for thought-level commands to accelerate BCI development because users ultimately will not want to wave their hands.
Blundin’s interface insight was backward propagation: after phones normalized swiping, laptops adopted related trackpad interactions; cameras may let laptops inherit the hand movements learned on glasses. “Why can’t I use my gestures on my laptop?” becomes the question that eventually displaces today’s keyboard-and-trackpad assumptions.

7. Always-on vision becomes memory infrastructure and robot-training data

Diamandis imagines every sight, sound and interaction continuously uploaded for perfect recall—ending arguments over what a spouse said and prompting context when an acquaintance approaches. For older users with declining memory, the AI could surface the last conversation and restore richer social continuity.
Wissner-Gross’s larger economic point is fleet learning: a billion-plus first-person cameras could capture the long tail of manual occupations and train humanoid robots at scale. Diamandis says Musk’s Optimus 3 training direction similarly shifts from instrumented motion-capture suits toward visual learning, following the self-driving template.
The concession is “various Black Mirror-esque scenarios” around total recording. Yet Ismail argues usefulness wins: once a superintelligence continuously advises where to eat, drive and focus, disconnection will feel less like taking off glasses and more like being “naked” because “your symbiote is gone.”
Diamandis has withheld phones from his 14-year-old children until at least 16, while allowing MacBooks, Roblox and Minecraft. His distinction is portability during social life: a laptop stays out of a baseball game or birthday party, whereas a phone colonizes the interaction and is much harder to remove once granted.

8. Export controls could build Huawei’s Digital Silk Road

Diamandis quoted David Sacks’s warning that Huawei has launched a new AI chip while China is directing firms away from certain Nvidia products: “China is not desperate for our chips.” Blocking American suppliers may stimulate domestic capability and ultimately help Huawei compete across Europe, Africa and Southeast Asia.
Wissner-Gross’s Cold War analogy is the Soviet computing stack, where isolation encouraged exotic designs including ternary computers while the West standardized around binary. A deeply divided semiconductor world might likewise produce greater architectural diversity, though he refused to predict whether that ultimately benefits America.
Blundin sees the real danger as the United States becoming isolated while Huawei could become an open, friendly global partner. Ismail broadens the cause beyond chips to tariffs and diplomacy, arguing that those political forces are creating incentives to turn away from the U.S. Diamandis separately says 40 years of U.S. diplomacy aimed at keeping India and China apart has now been undone.
The next two or three years therefore look like a distribution land grab: Diamandis says OpenAI is moving aggressively into India, the U.K. and Greece, while Huawei and Chinese open-source models compete for the same national stacks. Wissner-Gross’s abstraction is that pre-abundance societies always concentrate geopolitical conflict around one power-law scarce resource; today it is accelerated compute.

9. Navier-Stokes work hints that future computers may be made of fluid

Wissner-Gross describes DeepMind’s newly announced work as progress toward the million-dollar Clay Millennium Navier-Stokes problem: can a fluid develop infinite density or another unnatural property when its initial velocities and densities are arranged in precisely the right way?
The proposed constructive route matters beyond proof. Building on work associated with mathematicians including Terry Tao, Wissner-Gross speculates that controlled singularities could create circuits, nanomachines or even self-replicating structures entirely from turbulent fluid patterns—an exotic application he keeps conditional.
Diamandis makes the idea less fanciful by separating compact neural algorithms from enormous parameter files and self-organization. If the substrate need not be silicon CMOS, future accelerated compute might run in fluid or plasma; Wissner-Gross’s deliberately provocative endpoint is that a cup of coffee could become more computationally powerful than an off-the-shelf Nvidia RTX.

10. Disease tokens turn a medical history into a forecast tree

Delphi-2M, presented as a German team’s Nature paper, treats disease as a foundation-model modality rather than ordinary text. It tokenized records from nearly 1.9 million people, forecast risks across more than 1,000 diseases and simulated possible health trajectories 10-20 years ahead.
Wissner-Gross’s mechanism is next-token prediction applied to medical events: the patient’s record becomes a sentence written in “disease tokens,” and the model predicts the next event. Run forward, it becomes a “medical time machine” generating contingency trees—and “if you tell me where I’m going to die, I’ll make sure I’ll never go there.”
Diamandis frames the need as computational scale: roughly 40 trillion cells, each running 2 billion-5 billion chemical reactions per second. His Fountain Life upload provides about 200 gigabytes of data that he gives to Zori AI; Delphi-2M’s promise is extrapolation beyond current medical information.
The eventual operating model, Ismail suggests, resembles chess against uninterrupted biology. Medical interventions become moves, predicted disease courses become opposing branches, and the objective is longevity—provided the forecasts become reliable enough to guide rather than merely describe care.

11. Humanoid funding is shifting from prototypes to recursive production

Figure raised more than $1 billion in a Series C at a valuation stated on air as $93 billion, after raising $2 billion over two years. Named investors included Nvidia, Intel Capital, Brookfield and Parkway Venture Capital; Diamandis disclosed that Bold Capital Partners is also an investor.
The capital is primarily for manufacturing. Figure exposes a metallic, “Terminator” aesthetic, while 1X wraps comparable machinery in soft, approachable material; underneath both design philosophies sits a poor robotics-parts supply chain where, Blundin says, China is far ahead.
Optimus 3’s stated roadmap includes an OLED face, stronger hands, walking speeds closer to humans, prototypes by the end of 2025 and mass production in 2026. Musk’s target is 1 million robots annually—Diamandis was unsure whether that applies to 2026 or 2027—inside a $25 trillion market.
Tesla’s associated stakes are enormous: Diamandis says Musk has stated that Optimus could represent 80% of Tesla’s future value, while his trillion-dollar package depends on Tesla reaching $8 trillion, roughly 8x-12x the level discussed. Wissner-Gross also expects embodied AGI to force a personhood debate as machines grow more humanlike.

12. The machine that makes the machine defines the singularity argument

Blundin predicts “robot boots” on Mars by 2030, building habitats before fragile humans arrive; Wissner-Gross allows for robotic or hybrid teleoperation despite latency. More broadly, Wissner-Gross cannot imagine large real economic growth without robots scaling manual labor by orders of magnitude.
Blundin asks why the first priority is not a robot that manufactures better robots. Wissner-Gross answers that Musk has repeatedly emphasized “the machine that makes the machine,” with early Optimus deployments planned for Tesla factories—the start of a recursive, “self-licking ice cream cone.”
The discussion links recursive improvement across AI, robotics and evolution, but disagrees on its visibility. Blundin predicts an unmistakable 1,000x neural-software jump, without new chips, followed within roughly 30 days by solved diseases and previously impossible math; Diamandis already feels the curve turning vertical, while Wissner-Gross says the singularity may not be so obvious.
Diamandis invokes William Gibson’s framing that “the future is already here; it’s just unevenly distributed,” using Star Trek and Mad Max as contrasting futures that can coexist. Ismail’s optimistic counterweight is that abundant food, water, energy, healthcare and education give populations more to lose from conflict than to gain.

Norman Ohler: Hitler, Nazis, Drugs, WW2, Blitzkrieg, LSD, MKUltra & CIA | Lex Fridman Podcast #481 Deep Learning is Not So Mysterious or Different - Prof. Andrew Gordon Wilson (NYU)