10 of The Wildest Hot Takes From the All-In Summit 2026

Summary

• Brad Gerstner reads the week’s AI-safety chaos as consensus forming in public, not crisis. After the “10% chance of an extinction event” claims and Dario’s independent-monitor essay, Elon backed independent peer review, Jensen said the labs themselves need to get it under control, and the president called in mid-panel. “The sausage-making’s kind of ugly,” but the destination is a “pragmatic, commonsensical middle way”: company-led safety plus some form of peer review, with Elon floating that China might join open-source review.
• Naveen Rao says Unconventional AI will disclose “real results on real hardware”: nanojoules per image generated, thousands to tens of thousands of times more efficient than existing chips. With AI set to “run out of energy in the next three years,” brain-derived efficiency first keeps scaling alive, then drastically disrupts the price of intelligence, putting it in everything down to “your toaster.”
• Rao’s second call: in 20 years software carries “very little value.” He sees value capture converging into energy → hardware → intelligence; the SaaSpocalypse is “absolutely real” because AI dissolves the interface layer; hardware and energy are clearly high gross margin, while neocloud gross margins are “TBD.”
• Toby Rice, founder and CEO of EQT Energy, sees a 20–40% increase in U.S. natural-gas demand on a roughly 110 BCF/day market. AI is only about 40% of the power-generation pull — more than 174 gigawatts of reliable generation, mostly coal and nuclear, must be backfilled — plus another 10–20 BCF/day of LNG exports. His warning: 80% of the public opposes data centers (“ICE is more popular than data centers”), and the anti-energy playbook is now being used against AI infrastructure.
• Wen Sang of Genspark argues capability isn’t the bottleneck — access is — so “we won’t slow down.” Echoing Jensen’s point that “AGIs are out of here” but not everywhere, he says coding-agent power users are in the “single millions,” $20 agentic-AI users in the tens of millions, and free-chat users in the hundreds of millions; the vast majority of one billion-plus knowledge workers haven’t touched AI. His fix is an “FSD vehicle” instead of a Formula One car, with safety handled through agent telemetry and auditing — “more controls in place, but it’s not about slowing down.”
• Geoff Woo of Anti-Fund doesn’t believe humans have free will, doesn’t think GPT-7 or GPT-8 will be conscious, and offers two responses to the singularity: looksmax/vibesmax or burn “as many tokens as possible” while they’re subsidized. IQ is offloadable to compute, game theory makes acceleration “de facto,” and he says AI has solved Navier–Stokes; with rumors around Hodge, he thinks biology may be computable and that humanity has a chance to live forever — while France and Italy become “Disneyland for rich Americans and rich Chinese people.”
• Kent Draper of AIRIN says execution is the differentiator — and 2026 capacity is already sold out. As a vertically integrated neocloud owning its data centers, AIRIN just handed over the first phase of its large Microsoft deal; demand “far exceeds supply,” supply is getting harder, and the imbalance will only “exacerbate.”
• Michael Brandt says the human body is the next major platform for innovation. He frames the body as an API with improving “read” and “write” capabilities, using Ketone-IQ as his “super fuel” example; optimization is splitting into different goals such as looksmaxing, neuromaxing, endurance, and longevity.
• The contrarian grab-bag: Sweetgreen’s Jonathan Neman insists “AI’s not going to kill us, but ultra-processed food will” — this summer’s Cyclospora scare involved one farm/supplier serving iceberg lettuce to Taco Bell, which he says was used to scare people away from fresh food; Nick Shirley wants California’s high-speed rail killed ($15 billion spent, zero track laid, $200 billion-plus projected); Jake Paul wants the Constitution amended so citizens can bear “more advanced weapons systems” against an autonomous-robot-armed government.

Deep dive

1. Gerstner: the safety fight is “open-sourcing in real time how consensus forms”

• Gerstner’s framing of a chaotic week: Friday brought people claiming “a 10% chance of an extinction event” from AI, which he criticized; Saturday brought the Dario essay proposing independent monitors; then on stage Elon said, “I tend to agree with Dario,” and called for independent peer review; Jensen countered that “the labs themselves need to control it”; and the president called in while Jensen sat on the couch. “Everybody thinks there’s chaos. I don’t think it’s chaos.”
• The synthesis he draws: more power and more risk than assumed, responsibility starting with the companies, probably some peer-review layer — and Elon suggesting China “maybe” participates in open-source peer review.
• His meta-point on discourse: podcasting and X push everyone to “join a tribe,” but the answer here, as with Trump accounts, is “a pragmatic, commonsensical middle way” — “the sausage-making’s kind of ugly.”

2. Rao: nanojoule chips, and doomerism as “coordinated vibes”

• Rao previews Unconventional AI’s public disclosures: “real results on real hardware” at “nanojoules per image generated” — “thousands or even tens of thousands of times more efficient than existing computer chips.” The premise: computers were always built to go faster, never for power efficiency, because it did not matter; now compute is power-constrained globally, and “we’re gonna run out of energy in the next three years or so for scaling AI.” Efficiency extends scaling first, then drastically disrupts the price of intelligence.
• His anti-doomer case: no credible evidence “other than vibes right now” — “it’s literally Hollywood. Go watch Transcendence.” He questions whether the warnings are financially motivated: “If you’re really that worried, stop working on it. Don’t go IPO.”
• The rogue-AI rhetoric about “models copying themselves” is, he says, “a complete fabrication”: they do not do that and need infrastructure — “tons of hardware” — to run. Satya’s bug analogy gets his endorsement: fix the bug, add the guardrails needed, and keep building. He says there is still work to do, including preventing bots from pushing people toward suicidal tendencies.
• His second hot take: he thinks value capture will converge to energy → hardware → intelligence, and “in 20 years there will be very little value to software.” The AI becomes the interface, dissolving the software layer the way a good human hire needs only “please go do this thing.” Neoclouds sit inside hardware delivery with real money flowing through, but “the gross margins are kind of TBD. I just don’t know.”

3. Toby Rice: a 20–40% gas-demand wave — and data centers polling below ICE

• The demand math from Rice, founder and CEO of EQT Energy: a 20% increase in gas demand from power generation, of which AI is only about 40% — the rest is backfilling “over 174 gigawatts of reliable power generation, mostly coal and nuclear” that was shut down — plus another 10–20 BCF/day of LNG exports on a roughly 110 BCF/day market. He sees a 20–40% increase in natural-gas demand, anchored in the Marcellus, “the biggest gas field in the world… not in Texas.”
• The abundance argument: “three billion people in this world live on less electricity than it takes to run a fridge.” His prescription is to triple all forms of energy — solar, wind, gas, and nuclear — which he claims would increase global GDP five times over while lifting billions from poverty.
• The risk he flags: political force overriding market forces already pushed American energy bills up over 40%, and now “80% of the public saying they don’t support data centers. ICE is more popular than data centers.” The same techniques used to shut down energy development “are now being used to shut down intelligence.” His counter: “We’re not just playing for data centers. We’re playing for abundance.”

4. Wen Sang: the gap is access, not capability — “not about slowing down”

• Genspark’s read on the moment, echoing Jensen’s point that “AGIs are out of here” but not everywhere: heavy users who can maneuver coding agents number in the “single millions,” $20 agentic-AI users in the tens of millions, and people using free chats in the hundreds of millions. “The vast majority of the world has not touched AI just yet… not because it’s not there. It’s because it’s hard to maneuver.”
• His self-deprecating proof: an MIT-trained mechanical-engineering Ph.D. who admits, “I don’t even know how to draw the full power out of Phi-305 with my bare hands,” while Copilot Cowork can ask users to download “10 things from GitHub.” Genspark’s answer: “We build an FSD vehicle. You just tell us your destination” — “a fleet of Goldman Sachs analysts in your pocket” so “everybody can work like Jensen Huang.”
• On safety, the same accelerationist posture: Genspark was one of the very initial members of NVIDIA’s Open Secure AI Alliance and is contributing to mapping and auditing agentic behaviors. “It’s about putting more controls in place, but it’s not about slowing down.” Access itself reduces fear: “Don’t leave anyone behind.”

5. Woo: no free will, subsidized tokens, and a “singularity midlife crisis”

• Woo opens, “I don’t believe in free will. I don’t know what consciousness is.” He says the neocortex is “just explaining a story of why you’re making decisions”; he doesn’t think GPT-7, GPT-8, or future models will ever have consciousness, and “I don’t think you need to be conscious to be smart.” His hedge on his own nihilism: not believing in free will can spiral toward nihilism and hedonism, “but we should pretend that we can mold our destiny.”
• The practical corollary: “either looksmax and vibesmax, or all you do is spend a bunch of tokens before the singularity takes us out” — burn subsidized tokens from OpenAI, Cognition, and others “while money is still meaningful and status is meaningful,” because IQ points are already “offloadable to compute.”
• His longevity take, as stated: AI can solve Millennium Prize Problems; he says it solved Navier–Stokes, with “rumors that it solved the Hodge conjecture.” That leads him to think biology is a computable problem and may become tractable, giving humanity “a chance to actually live forever.” On Molly’s pushback that nobody in Italy or France is talking about AI, he says, “France and Italy are gonna be Disneyland for rich Americans and rich Chinese people. It already is.”
• Why acceleration wins regardless: even if all Americans banned it, “there’s other jurisdictions”; his game-theory conclusion is, de facto, “accelerate.” “Let’s not be reckless about it, but how do we anticipate and take advantage of that momentum?”

6. The populist takes: amend the Constitution, kill the train

• Jake Paul says the Constitution needs an amended, updated version — “written in 1781, I believe” — because modern weapon systems change the balance. He interprets the Second Amendment as protecting citizens if the government controls them in harmful ways, but says “a pistol or a shotgun is not going to be able to overthrow” a government armed with autonomous robots. Citizens might need drones, EMPs, RPGs, or “potential anti-helicopter” weapons; otherwise “the Second Amendment is just a hobby.”
• Nick Shirley’s “not even a hot take”: California has spent over $15 billion on high-speed rail without laying a single track, with completion projected above $200 billion, while China built one in 2008 and has since built over 30,000 miles connecting the country. He says more than $500 million went to contractors who could not work because California had not cleared the way.
• His additional receipts: Newsom said the fraud Shirley exposed was not real; hours later, 12 people in his San Diego video were charged in connection with $10 million in fraud. Shirley also says more than 30 fraud-related bills followed his Minnesota video.

7. Operators’ corner: AIRIN sold out, the “lettuce psyop,” the body as platform

• Draper’s take is execution as the differentiator: AIRIN is a vertically integrated neocloud that owns its data centers for control over “time to first token,” just handed over the first phase of its large Microsoft deal, and has sold out its 2026 capacity. Demand “still far exceeds supply… and the supply side is actually getting harder, so you’re only gonna see that position exacerbate.”
• Neman’s counter-doomer take: “AI’s not gonna kill us, but ultra-processed food will.” Cyclospora was, in his framing, something that happens every summer; one farm from one supplier serving iceberg lettuce to Taco Bell was unsafe while “99.999% of places you were eating lettuce were safe.” He says the scare pushed people toward ultra-processed food “which we know will kill you.” Sweetgreen is funding a 30-day, lunch-and-dinner challenge after a customer’s version went viral.
• Brandt’s Ketone-IQ thesis: “the human body is the next major platform for innovation,” and “your body is the most advanced piece of technology that you own.” “If you think about the human body as an API, we’re getting really good at the read, and we’re also developing new capabilities on the write.” Optimization is bifurcating away from one “human.jpeg” ideal into looksmaxing, neuromaxing, endurance, and longevity paths that “might be at odds with one another.”