Trump Brokers Gaza Peace Deal, National Guard in Chicago, OpenAI/AMD, AI Roundtripping, Gold Rally
Speakers
Chamath Palihapitiya
— CEO, Social Capital
Brad Gerstner
— CEO, Altimeter Capital
David Sacks
— General Partner, Craft Ventures
Jason Calacanis
— Angel Investor / Founder, LAUNCH
Summary
The All-In hosts, joined by returning guest Brad Gerstner, discuss Trump's Israel-Hamas ceasefire deal and its implications for Gulf oil producers, then debate the National Guard deployment to Chicago and Trump's falling approval ratings. The bulk of the market discussion covers the OpenAI-AMD warrant deal, Nvidia's dominance and Wall Street's skepticism about the scale of the AI buildout, supply-chain bottlenecks in HBM memory and power, and whether circular AI financing deals are a real concern. The show closes with explanations for gold's rally above $4,000 (Tether Gold issuance, central-bank buying, BRICS de-dollarization) and ICE's $2B investment in Polymarket.
- Trump brokered phase one of an Israel-Hamas ceasefire; Chamath argues Gulf states want stability so they can monetize oil reserves before alternative energy erodes oil's value.
- National Guard deployment to Chicago and Trump's approval ratings debated; Sacks defends the policy while Calacanis objects to ICE tactics and the cost per deportation.
- OpenAI-AMD deal: six gigawatts of MI450 GPUs with warrants for up to 10% of AMD; Gerstner calls it a bet-the-farm move given Nvidia's near-100% share of incremental AI revenue.
- Wall Street models Nvidia revenue flat-lining in 2027-29 versus 100 GW / $4-5T buildout talk; hosts argue there are no dark GPUs and that demand will exceed estimates.
- Chamath flags HBM memory (SK hynix, Samsung) and energy as the gating inputs whose controllers will command leverage and equity participation in the AI buildout.
- Sacks questions whether the GPU market bifurcates, with Nvidia dominant in training but inference open to Groq, Cerebras, AMD, TPUs, Trainium and Broadcom-based ASICs.
- Circular AI deals judged to have economic substance: Nvidia's roughly $450B of 2025-27 cash flow dwarfs its equity checks and OpenAI revenue is ramping toward $50-100B.
- Gold above $4,000 attributed to net new buyers (Tether Gold, central banks, macro funds) and China/BRICS diversification away from the dollar; ICE invests $2B in Polymarket.
Ideas
Oil is a depreciating asset; monetize now
Gulf oil producers face an obvious problem: their core resource, oil, is becoming steadily less valuable. Oil left in the ground adds nothing to a balance sheet, and over a 10-20 year horizon abundant electrons from nuclear, gas and solar will diminish the net long bid for oil. That is why the rational outcome is peace - with stability these countries can focus entirely on monetizing reserves quickly and redeploying the proceeds into services and other assets. The implication is faster monetization of reserves and a structurally weakening long-term bid for oil.
AMD MI450 is a binary bet-the-farm shot
The OpenAI deal is a bet-the-farm move by Lisa Su: AMD granted OpenAI warrants for up to 160 million shares, about 10% of the company, if six gigawatts of compute get deployed, whereas Nvidia gave away none of its equity in its own OpenAI deal and instead got the right to buy into OpenAI. In 2022 AMD and Nvidia both had roughly $25B of revenue; this year Nvidia will do about $210-230B versus AMD's $33B, because Nvidia has captured nearly 100% of incremental AI data-center revenue and AMD's MI350 was simply not competitive. AMD has one shot: if the MI450 gets adopted it could earn roughly $150B of incremental revenue from OpenAI alone for a 5 GW buildout and validate the chip for the rest of the market; if not, AMD is out of the game. It is far from a done deal whether the MI450 can compete with Nvidia's Vera Rubin and Rubin Ultra.
Nvidia dominates AI compute; Street underestimates
Nvidia has captured nearly 100% of incremental AI data-center revenue since 2022, growing from about $25B to $210-230B, because the unit of compute is now the entire data center - an ecosystem of software, networking and extreme co-design across 5-10 chips - and it wins on performance per watt when power is the constrained resource; hyperscaler CFOs say you could price competitor chips at zero and Nvidia would still be the more economic choice. Wall Street does not believe the buildout: estimates flatline in 2027-29 at under 10% CAGR (about $360B of revenue by 2029, roughly 9 GW a year) versus the 100 GW and $4-5T of buildout being discussed. Yet Nvidia trades at a lower price-to-earnings multiple than when the stock was $200, there is not a dark GPU in the world today and there will not be one next year, and Nvidia will generate roughly $450B of cash flow in 2025-27, so its equity checks into customers are tiny relative to the business and not round-tripped credit - OpenAI does not even have to buy Nvidia chips.
HBM suppliers hold leverage over AI buildout
AI growth will be constrained not by the ability to design next-generation silicon but by energy and ingredient inputs, and the companies that control those inputs will rise to power - the Rothschild lesson applied to AI. HBM memory is the clearest example: Nvidia has made a huge architectural bet on HBM and takes the majority of supply, then Google, and AMD's next architecture must also sit on HBM, so AMD now has to step into that supply chain and ask for share. The suppliers that control HBM - SK hynix and Samsung - therefore have leverage. OpenAI's recent handshake deals with SK hynix and Samsung in Korea are effectively Sam Altman buying forward HBM capacity so that he can allocate it and extract a tax in the form of warrants and equity. Second- and third-order input controllers like these will dictate the pace and scale of the AI expansion, and that is where Chamath would start looking.
Energy owners gain pricing power from AI
Energy will be the gating item for AI beyond a shadow of a doubt. Whoever controls electrons - hydrocarbon to electron, electron to electron, photon to electron, it does not matter which - will be able to demand equity, upside and participation from the foundation-model makers and chip companies such as Nvidia, AMD and Broadcom, instead of remaining a low-margin linear member of the supply chain. Those who control such scarce inputs will dictate the pace and scale of the AI expansion, so energy suppliers are where Chamath would start looking for the second- and third-order winners.
Bet the over on AI compute demand
Investors should bet the over on AI compute demand. Capex years out is hard to predict because token demand depends on applications not yet invented - new agents, video-generation models like Sora and xAI's release - but all of the compute being built will eventually be used, just as the 1990s fiber overbuild was absorbed once photos, video, YouTube and social networking arrived. Faster-than-Moore's-law chip efficiency and sparser model architectures are offsets, but the Jevons paradox means cheaper tokens unlock AI in more and more contexts and fuel demand. The flurry of deals (OpenAI-AMD, OpenAI-Nvidia, Nvidia-xAI) is evidence of healthy investment and competition, and Nvidia's customer financing has economic substance because downstream demand is real and OpenAI revenue is ramping from roughly $5B toward $25B and $100B. AI accounts for about 40% of US GDP growth and 80% of US equity gains this year; provided regulators do not sabotage it, the buildout can drive 4-5% GDP growth for years.
No dark GPUs; AI buildout is real
The AI infrastructure buildout is not a repeat of the dark-fiber bubble: fiber sat dark because demand did not exist, whereas there is not a dark GPU in the world today and will not be one next year - tokens are being consumed by every consumer who wants answers instead of ten blue links, by enterprises deploying AI, and by sovereigns. Incremental generative-AI revenue of roughly $100B in 2026 has to grow to well over $1T by 2030, and top-down a 5-10% productivity gain on $125T of global GDP is worth $5-12T a year, which is why Masa, Sam Altman and others are placing heroic bets. Wall Street models far less buildout than the 100 GW and $4-5T being discussed, capital and power have been unlocked in the US, and this creative-destruction wave has not peaked - investors should take the over on demand.
Bittensor can fill decentralized AI compute demand
Beyond the hyperscaler GPU buildout, other solutions that have not yet hit the ground will arrive to fill the Jevons-paradox surge in compute demand. Bittensor, an open-source decentralized AI network, is one worth looking into, and Jason has started a small fund with his own capital to explore it.
Multi-trillion AI TAM justifies trillion-dollar buildout
Sizing the AI opportunity bottom-up: the developed world has about 500 million business users spending roughly $3,000 a year on SaaS products, and about a billion consumers who pay a couple hundred dollars a year for services like Netflix and Disney Plus, which adds up to an AI revenue pool of a couple of trillion dollars a year. Against a prize that size, a trillion-dollar infrastructure buildout - roughly $50B per gigawatt of data center, with OpenAI and Elon Musk talking about 10 GW sites that would cost $500B each - does not look farcical.
Watch whether inference fragments away from Nvidia
Nvidia has by far the best training chips and overall demand should be bet over, but the big open question for its business is whether the GPU market bifurcates between training and inference. Inference will be roughly 99% of the market versus 1% for training, and cheaper inference alternatives are multiplying: Groq, Cerebras, AMD, Huawei as a wild card, custom ASICs built on Broadcom IP, Google TPUs, Amazon Trainium and OpenAI's own planned chip. Nvidia could end up with training locked up while inference fragments to cheaper chips - not a prediction, but the key risk to watch.
Tether, central banks, macro funds buy gold
The gold rally is not the safety trade, anti-dollar trade or money-supply story people invent to sound smart - it is simply many more net new buyers. The most important new buyer is Tether, whose Tether Gold stablecoin custodies physical gold on holders' behalf and whose issuance volumes keep rising. At the same time central banks are rebalancing into gold, and macro funds that have decided central banks cannot be trusted are not long bonds or currencies, so they are long gold. Net new speculation, stablecoin-related issuance and a loss of confidence in central-bank policy together underpin the bid.
China and BRICS keep diversifying into gold
A further big driver of gold is central-bank diversification: China's central bank has increased its gold reserves for 11 consecutive months, adding 40,000 ounces in September alone to reach 74 million ounces worth about $283B, gradually substituting away from US dollars and Treasuries because, as America's chief competitor, it does not want to depend on the dollar complex. The trend began when the Biden administration weaponized the dollar after the Ukraine invasion - cutting off SWIFT access and freezing Russian assets - which pushed BRICS countries to develop alternatives, most likely gold-backed certificates for settling large international trade flows. That is a very long-term project rather than something arriving in the next few years, but it is a structural source of demand for gold.
Prediction markets threaten sportsbooks like DraftKings
Everything is becoming a market: sports, knowledge and eventually equities and debt will be tokenized, fungible and tradeable long or short on-chain, with crypto pools able to run high-frequency trading faster and at near-zero infrastructure cost. With ICE investing $2B in Polymarket and distributing it to thousands of financial institutions, a chunk of the trades that today happen on betting sites like FanDuel and DraftKings can migrate to prediction markets - an open question for the sportsbooks' business models that is worth monitoring.
Up Next
This All-In Podcast video, published October 10, 2025, features Chamath Palihapitiya, Brad Gerstner, David Sacks, Jason Calacanis discussing WTI, AMD, NVDA, 000660.KS, 005930.KS, Power generation stocks, SMH, TAO, GLD, DKNG. 13 trade ideas extracted by AI with direction and confidence scoring.
Speakers: Chamath Palihapitiya, Brad Gerstner, David Sacks, Jason Calacanis · Tickers: WTI, AMD, NVDA, 000660.KS, 005930.KS, Power generation stocks, SMH, TAO, GLD, DKNG