Summary

The OpenAI–Nvidia deal is a balance-sheet problem, not an “infinite money glitch.” A gigawatt of AI capacity rents for $10–15B a year on five-year commitments — $50–75B out the door per gigawatt — and Sam Altman wants “more than 10 gigawatts.” Nvidia’s $100B equity pledge (first tranche: 1GW, $10B) effectively lets OpenAI pay partly in stock: of the ~$50B it costs to build a gigawatt, ~$35B goes to Nvidia at 75% gross margin, and roughly half that gross profit is recycled as equity. Patel: “Nvidia’s lowering their prices without lowering their prices” — and Patrick calls it “about the highest stakes capitalism game of all time.”
Scaling is not diminishing returns — it’s child-labor economics. Each 10x of compute buys one tier of capability, but a tier is a six-year-old versus a 13-year-old: “the amount of work you can get a 13-year-old to do is actually quite valuable.” An intelligence at Google-senior-engineer level is “$2 trillion of software value” — the world’s software wage bill. The unhedged tail risk: “If the models don’t improve, we’re absolutely screwed… the US economy will go into a recession” — and probably Taiwan and Korea with it.
GPT-5 was a serving decision, not a scaling failure. Token demand doubles every two months while hardware doesn’t, so cost per intelligence tier must collapse — GPT-3-quality tokens are ~2,000x cheaper; DeepSeek was ~500–600x cheaper than GPT-4, GPT-OSS cheaper still. After GPT-4.5 proved “quite a bit smarter” but unserveable, OpenAI kept 5 “basically the same size as 4o” to serve everyone and push adoption, moving intelligence into thinking tiers. Capacity/cost, not latency, is the bottleneck he’d fix with a magic button.
Post-training is at “we’ve thrown the first ball.” Roughly 40 Bay Area startups are building RL environments (fake Amazons, data-cleaning drills, math puzzles the models “hill climbed… like crazy”), and post-training “will subsume the majority of the compute at some point.” Pre-training on text is still “quite early,” multimodal scaling remains underway — and the robotics data flywheel “hasn’t even left the dugout.”
“Nvidia’s holding no risk. Everyone in the middle’s got a lot of risk.” Neocloud economics: Blackwell costs ~$2/hr all-in over six years and rents at $3.50–4 short-term — until the next chip is “10x faster for 3x the cost.” The golden goose is long-term contracts with real balance sheets: Nebius’s ~$19B Microsoft deal carries at least $6B of gross profit. Oracle signed $300B with OpenAI against ~$15–16B ARR — if it works, ~$100B of profit; if not, debt-funded exposure to a customer with no balance sheet.
US–China: “if we don’t accelerate we die.” Without the AI boom the US “probably would be behind China and no longer the world hegemon by the end of the decade.” China has dumped at least $400–500B (as heard) into semiconductors, prioritizes an insular supply chain over aggregation theory, and “could build a 10-gigawatt data center in a few years” — “Elon’s slow compared to China.” And the Taiwan doomsday breaks portfolio logic: “you can’t invest in Apple” either if you believe the risk — “so it’s like yolo invest in TSMC.”
Positioning calls: flipped from bearish to “super bullish Google” (lowest cost per token via vertical TPU stack, waking up on every front); Meta “has the cards to potentially own it all” — the only company with the glasses hardware, models, serving capacity, recommendation know-how, and capital for the next human-computer interface; more optimistic on Anthropic than OpenAI because its revenue “is accelerating way faster” against the $2T software market; XAI is “in a real danger of not being able to raise capital” without a business model “besides Pornbot.”
The SaaS reckoning: AI collapses the cost of building software (tilting build-vs-buy toward build) while adding “humongous COGS” to every AI product with customer-acquisition cost unchanged — so “the era of software-only businesses is really really tough in the age of AI.” China is the proof: developers at ~10x lower cost meant SaaS and cloud never scaled there to the same extent. Already-scaled platforms win instead: “he who controls the platform is going to win and win and win.”

Deep dive

1. The OpenAI–Nvidia deal: “who is the balance sheet for this?”

Patel opens by mocking the meme — OpenAI pays Oracle, Oracle pays Nvidia, Nvidia pays OpenAI, “the infinite money glitch” — then dismisses it: “that’s not actually what’s happening.” The real driver is that “the compute precedes the buildup of business”: you need the cluster before you can rent it for inference, and before you can train the model that unlocks the next use cases. Despite 800 million users, “there’s very much a risk of OpenAI being too small to matter” — the competition is Zuckerberg, Google, Elon, “the richest people in the world,” in what he calls the “Pascalian wager” of the tech giants.
The arithmetic behind the headlines: data center capacity rents at $10–15B per gigawatt per year, and OpenAI signs five-year deals — $50–75B of cash out per gigawatt, with Sam wanting “more than 10 gigawatts.” Microsoft soured, and “Oracle doesn’t even have a balance sheet like Google and Microsoft and Amazon” — so OpenAI needs allies willing to spend capex ahead of the curve and trust the rental income arrives.
Oracle’s side: a $300B deal against roughly $15B ARR (“maybe it’s like 16 now,” ~20 by year-end). “If the bet works out they’ve just made $100 billion of profit… pure cash profit.” If not, they’re stuck with the buildout — and they’ve started raising debt.
The Nvidia mechanics, simplified by Patel: 10GW earns OpenAI $100B of equity investment in tranches (first: 1GW, $10B). A gigawatt costs ~$50B to build, of which ~$35B goes straight to Nvidia at 75% gross margin — call it $40B revenue, $10B COGS, $30B gross profit, with roughly half recycled into OpenAI equity. Nvidia books the capex dollars up front and holds stock “in a company that may or may not be worth something… that may or may not be able to pay hundreds of billions of dollars of compute deals.” Patrick calls it “the highest stakes capitalism game of all time.”

2. Scaling isn’t diminishing returns — and the downside is a recession

Patrick pushes directly: are you confident the log-log curve continues? Patel: “everything has shown that it will continue.” Yes, 10x compute buys one tier — but the value step between tiers is a six-year-old versus a 13-year-old: a company of high schoolers refreshed every six months could only “dig trenches and do yard work,” while 25-to-30-year-olds build drastically more valuable businesses. “It’s a drastic value change,” not a diminishing return.
Where we are today depends on domain — for software “we’re really pretty good,” which is why Anthropic went “from a billion or less of revenue to seven to eight already… the fastest revenue ramp we’ve ever seen,” and “it’s basically all code related.” Infinite Google-senior-engineer intelligence is “$2 trillion of software value, because that’s how much the world pays software engineers today” — and it’s a force multiplier, not a pure replacement.
His calibration is more nuanced than “maximum bull”: Sam says AGI “in less than a thousand days,” Dario is “way more bullish,” so are his roommates (an Anthropic researcher and the podcaster Dwarkesh) — yet famous investors think what Patel says “sounds like crazy” talk. On the upper limit he’s “among the most bullish you can get”: digital god eventually — “is that 10 years? 100? 1,000? I don’t know” — but even pausing capability six months from now would be a “godsend in terms of how much efficiency and value can be created.” His favorite bear is likely Yann LeCun: right that “autoregressive pre-training on the internet doesn’t work to get you to AGI,” but “completely wrong” in dismissing RL too — plus the investors who “think this is [nonsense] but are just making tons of money on it anyways,” buying Oracle before earnings on market perception.
Patrick raises the Carlota Perez glut argument — every shortage is followed by overbuild. Patel doesn’t dodge: “If the models don’t improve, yes, we will overbuild… the US economy will go into recession, straight up” — probably Taiwan and Korea too. But the historical comps split: tulips and crypto were “complete Ponzi,” while UK railroads consumed ~6% of GDP for a decade and were real — “we’re nowhere close to 6% of our GDP.” And the strongest balance sheets can pull the plug: Microsoft did, then “plugged it back in” and had to buy capacity from Nebius.

3. Tokconomics: GPT-5 was a serving decision, and commerce is the monetization

He’s coining “tokconomics” — the economics of tokens (“kill off crypto finally, once and for all”). A gigawatt can serve 1,000x the tokens of a bad model, 1x of a good one, 0.1x of an amazing one. Demand doubles every two months, but “I’m not doubling my hardware every two months” — so cost at a given intelligence level must collapse, and it does: GPT-3 quality is ~2,000x cheaper now; DeepSeek spooked markets at ~500–600x cheaper than GPT-4; GPT-OSS is cheaper still and “actually a little bit better than GPT-4 OG because it can do tool calling.”
The GPT-5 decision follows: OpenAI tried the big-step route with 4.5 — “it was actually quite a bit smarter,” but “no one could serve it” at reasonable cost or speed. With only ~2GW of effective capacity by year-end and rate limits already throttling users (Patel runs multiple ChatGPT accounts to fire off deep-research jobs), 5 is “basically the same size as 4o,” roughly the same cost or cheaper, so they serve way more users and put the extra intelligence into thinking/Pro tiers.
On the magic-button question — latency or capacity? — “I’d probably still say capacity/cost is more important than latency.” His own behavior is the evidence: he has access to a likely Claude 4.1 Opus but uses Sonnet far more — “it’s objectively dumber, but [Opus] is slow… my time’s worth something.” Same reason Anthropic’s revenue comes from Sonnet, not Opus: “no one wants to use a slow model.”
The monetization endgame is purchasing: more than 10% of Etsy’s traffic comes straight from GPT (Amazon blocks it, “otherwise it would be really high”); OpenAI’s applications chief built Shopify’s shopping agent. “The models are going to purchase for you… even if it’s 0.1%, 1%, 2% — it’ll be like a credit card transaction. Visa is the most amazing business in the world because of this, and chat could be that too.”

4. Post-training: “we’ve thrown the first ball”

Bigger isn’t the current problem — it’s grokking: models memorize before they generalize, and an overparameterized model “never had the opportunity to generalize.” The real challenge is data in useful domains: “nowhere on the internet does it show you how to fly through a spreadsheet using only your keyboard… it can’t do basic stuff, which is like play with a spreadsheet,” even though it read the whole internet.
Hence RL environments — “there’s 40 startups now in the Bay doing these environments” for OpenAI, Anthropic and others: a fake Amazon where the model must buy the right deodorant among decoys; iterative data-cleaning drills; model-graded medical cases; and math puzzles, which models “hill climbed up… like crazy” from Q4 last year to Q2 this year — largely by learning to write Python that does the math.
Innings check, per Patrick’s framing: text pre-training is “quite early” (learning efficiency can still improve, and any pre-training gain feeds everything downstream); multimodal scaling remains early with V3 and Banana Nano (likely Veo 3 and Nano Banana — Google’s video/image models); and post-training? “I think we’ve thrown the first ball” — and it “will subsume the majority of the compute at some point.” His image for how early: his brother’s newborn calibrating senses by sticking a hand in his mouth — “we’re so early in reinforcement learning because that’s what humans are. We’re reinforcement learners.”
On memory: transformers are amazing at exact recall (needle-in-haystack is now handled well) but “what they really suck at is having infinite context” — humans compress the world into something sparse (your childhood memories are re-remembered pictures, “morphed a little bit”). The model doesn’t have to work like us: deep research already writes notes off to the side, “using language to compress information,” running 45 minutes across millions of tokens — “a lot of memos that you read from people are on par with deep research, at least a junior[’s].” And this is why labs need millions of GPUs: not one giant run, but “I need to try a bajillion different things because I don’t know what will work.”

5. Talent wars: ML research is semiconductor manufacturing

The billion-dollar pay packages are rational: researchers steer experiments on chips costing $100B — wasting even a third of the compute makes their impact enormous. But adding people slows research down; Meta’s pre-superintelligence problem “was that they just had too many people that weren’t led by leadership that was amazing.” His friend likely Roon at OpenAI tweeted: “I get visibly viscerally angry every time I think about how many H100s Meta is wasting” — though “everyone’s wasting compute.”
His favorite recent analogy: ML research is exactly semiconductor manufacturing — a thousand process knobs per tool, a search space you can’t exhaustively test, so intuition picks points, you read fuzzy data, “and then just yolo.” The R&D fab “is producing zero economic value besides that it’s teaching you how to do the next node” — same as burned training compute.
Run’s other idea — make ridiculous offers to acqui-hire process knowledge from Shenzhen — Patrick calls it “a great idea”; Patel says it’s Run’s idea, not his, and ties it to Intel’s decline: the smartest 18-year-olds skipped $200K nanochemistry PhD tracks for $800K at Google, $10M at OpenAI, $100M at Meta — the same skew that sends the highest-scoring doctors into dermatology and anesthesiology. The hard part is selection: “how many people suck at talking and are really freaking good at doing?” And he flags Sam’s cope — “they didn’t get our best people” — delivered while doing internal counteroffers.
The line he keeps from Jensen Huang: “The reason America is rich is because we’ve exported all the labor, but we’ve kept all the value” — Nvidia and Apple outsource manufacturing to Asia and keep the gross profit.

6. Power dynamics: “the most fascinating soap opera ever”

Does Anthropic hold all the cards over Cursor? Cursor is at nearly $1B annualized revenue, sending most of it back to Anthropic (margins “slightly positive,” he thinks) — and Anthropic pours the gross profit into compute, so “the gross profit dollars are going to the hardware layer” either way. But Cursor keeps the data, the users, its own embedding and autocomplete models, and can swap to OpenAI “whenever I want to” — maybe even train a segment-specific model that beats Anthropic. “Everyone’s frenemies.”
Microsoft–OpenAI is “the most crazy power dynamic in the world”: in 2023 “Microsoft’s going to own the world”; by H2 2024 Amy Hood and company pulled back — “maybe we don’t need to be on the hook for $300 billion” — pausing data centers and relinquishing compute exclusivity to Oracle. The deal itself: ~20% revenue share, a 49% capped-profit structure, IP sharing, and the AGI clause terminating API/IP rights — “what the [expletive] does that mean?” The renegotiation MOU was “the most non-announcement announcement ever.” His point on AGI definitions: show ChatGPT to someone 20 years ago and “this is AGI”; the bar always moves — for him it’s the hand-in-mouth sentience moment.
Nvidia’s conundrum: it can’t acquire (blocked from ARM “when they were pretty much a nobody on the grand scheme of things”), and “you’re a loser if you just do buybacks — that’s admitting you can’t get higher returns on your capital.” So the balance sheet becomes the weapon: demand guarantees, backstopping a CoreWeave cluster for short-term rental that would never otherwise be built, and effectively frontloading OpenAI’s first year of compute — “I have a year of a gigawatt to figure out a business model.” Meanwhile, “when venture capitalists fund a company and then 70% of their round is spent on compute — they [expletive] love that.”

7. Neoclouds: Nvidia holds no risk, everyone in the middle does

The neocloud model is “absolutely amazing or terrible depending on how you do it.” A Blackwell costs about $2/hour all-in over six years and rents short-term for “north of $3.50 or $4” — insane margin, until the next generation arrives “10x faster for 3x the cost” and short-term pricing tanks. The golden goose is long-term contracts with balance sheets.
Nebius just signed the archetype: ~$19B with Microsoft, “at least $6 billion of gross profit off of this deal… I would do that all day.” The market literally prices Microsoft’s obligations cheaper than US government debt — “which is insane to me, but whatever.” CoreWeave rode the same trade until Microsoft stopped coming, then found Google and OpenAI — but OpenAI contracts, whatever their stated value, sit on a customer that “doesn’t have a balance sheet. So how can I be sure they’re actually going to pay?”
Google, short on data center capacity, is now backstopping deals with crypto miners (TeraWulf and Fluid Stack among them) — physically selling TPU systems to third parties who deploy and rent them, “and Google still makes all the money.” Inference providers have a real niche (Roblox wanting an LLM in-game, Shopify customer service, fine-tuning and serving open-source models) — but there’s also “yolo, I’m selling tokens to random people trying to build SaaS apps” who may run out of runway. The through-line: “Nvidia’s holding no risk. Everyone in the middle’s got a lot of risk.”

8. The buildout is real: power pansies, grid quirks, and a business AI built

On energy panic: “the first approximation is that we’re being a bunch of pansies — it’s not that much power yet.” Data centers are ~4% of US power (half of it AI): “that’s literally nothing, dude. It’s just we haven’t built power in like 40 years.” The constraints are supply chains and labor: GE doubling turbine production, Mitsubishi too; transformer-coil curing capacity with two-year builds; a company wiring diesel truck engines in parallel because turbines are sold out; Elon shipping power equipment from Poland. Mobile electrician wages have doubled — West Texas “is like 2015 and being a fracking guy.” Still, perspective snaps back: OpenAI’s 2GW site draws “the entirety of the power consumption of Philadelphia,” and a now-yawn-inducing 500MW is “$25 billion of capex once you put in the GPUs.”
The third-order effects are the fun part: training workloads swing so fast they can destabilize grids — skew the frequency and “your refrigerator will break down sooner… and you might not even know it because the data center’s nearby.” Texas and PJM are adopting curtailment rules — 24–72 hours’ notice to cut half a site’s power — which forces on-site generators, which then fail air permits if run “more than eight hours a month.”
Against Patrick’s skeuomorphic-era challenge (aren’t we just doing old things faster?), Patel’s counter is his own P&L: Patel’s second-highest-revenue product — image recognition run on satellite photos of every data center on earth, plus LLM-scraped permits and regulatory filings, sold as bi-weekly spreadsheets (“this Amazon data center’s fans are starting to spin, so we can forecast Amazon’s revenue”) — “this business is not possible without AI,” built with three people instead of the 50–100 it would have taken. Add mainframe migration: Amazon leaving Oracle’s database “took [expletive] 20 years”; now it can be far faster.
His hand-drawn bullishness spectrum, worth keeping: Dario → ML researchers → Patel → Patrick → the New York semis investor → “the Sequoia guy who thinks AI has been a bubble since 2023” → the utility guy who still won’t build power.

9. US vs China: “if we don’t accelerate we die”

The stakes as he sees them: “Without AI, we’re definitely just going to lose” — supply chains slower and costlier, unsustainable debt, overconsumption, social instability amplified by the visibility of income inequality on social media and algorithmic feeds splintering the old monoculture (“you and I are pretty similar and our feeds are completely different”). “AI has to dramatically accelerate GDP growth. Once you start talking about dividing the pie, you’re screwed.” Without the AI boom, “the US probably would be behind China and no longer the world hegemon by the end of the decade.”
China plays the long game it ran in steel, solar, phones and PCBs: at least $400–500B (as heard) dumped into semiconductors over a decade through SOEs, tax policy, provincial land grants and the Big Fund — versus US tariffs and a CHIPS Act that are “drops in the bucket.” China optimizes for an insular supply chain over aggregation theory; the US, ironically, is “kind of doing what China’s done historically — dumping tons of capital into something,” betting ChatGPT becomes the YouTube of intelligence (lose money forever, then own the platform at 3–5 billion users). ByteDance is already the third-largest GPU user in the world; DeepSeek engineers out-earn peers but nobody pays $10M — there’s no poaching culture.
Build speed: “Elon’s slow compared to China — and I think he knows it.” If China wanted a 10-gigawatt data center, “I bet they could build it in a few years” (smuggling chips if needed), while OpenAI’s total capacity optimistically reaches 10GW in the same window. They lack the best chips and memory but have the most power and the fastest construction.
The Taiwan doomsday breaks conventional risk management: a blockade or AI-supercharged subversion (“China could do a billion times Cambridge Analytica into Taiwan”) means “the US economy kind of free-falls — we can’t make refrigerators without Taiwanese chips,” no new cars, no new AI data centers, no cloud growth. So the PM rule against TSMC is incoherent: “you can’t invest in Apple” or Amazon or Google either if you believe the risk — “so it’s like yolo invest in TSMC.”

10. The book: long Google and Meta, Anthropic over OpenAI — and a SaaS reckoning

Speed-round verdicts, as delivered: Anthropic over OpenAI — “their revenue is accelerating way faster because what they’re focused on is more relevant to that two trillion dollar software market,” while OpenAI splits across consumer, science and take-rate bets. AMD: “I love them but they’re pretty mid” — his first multibagger, a soft spot, not a thesis. Oracle “is going to make so much money if you believe OpenAI is successful” — but in most worlds where OpenAI pays out $300B, “OpenAI is like a $10 trillion or $5 trillion company.” XAI is “in a real danger of not being able to raise capital” at the required scale: Colossus 2 (300–500k Blackwells) will be the world’s biggest single data center, but he needs a business model “besides Pornbot” — Patel’s proposed fix being an OnlyFans partnership that Trojan-horses creators into X.
The two OGs get the strongest endorsements. Meta “has the cards to potentially own it all”: the next human-computer interface is telling the AI what you want, and Meta uniquely holds all four requirements — hardware (the new display glasses), models, serving capacity, and recommendation-system mastery — “plus the capital.” Google: “I was pretty bearish Google like two years ago, but I’m super bullish” — TPUs now sold externally, genuinely competitive models, aggressive infrastructure, and positioning to capture both the consumer and professional interfaces, where Meta only gets the consumer.
The closing framework (credited to a colleague, likely Doug O’Laughlin): the SaaS golden age ran on flat R&D, tiny COGS and high CAC amortized at scale. AI breaks it twice — the cost of building competing software tanks (shifting build-vs-buy toward build), while AI features add “humongous COGS” with CAC unchanged, so markets fragment and companies “never hit the escape velocity.” China is the natural experiment: software developers at ~10x lower effective cost meant SaaS and cloud never scaled there to the same extent. Google benefits again — “the lowest cost of goods sold for any token of any company” via the vertical TPU stack — and scaled platforms thrive as content-generation costs fall: YouTube’s “glory days” ahead, because “he who controls the platform is going to win and win and win.”
The kicker, given the show’s stock-curse lore: “We’re popping the bubble right now, because the limit of AI is infinite” — though Patrick notes they checked, and the curse “is just market performance”: Applied Materials rose ~70% in the six months after Patel last talked it up.

Building an AI Physicist: ChatGPT Co-Creator’s Next Venture Science Corner Special! David Friedberg, Cleo Abram, Alex Filippenko, and Keller Rinaudo Cliffton