Elon’s Anthropic Deal, The Next AI Monopoly?, “FDA for AI” Panic, Trading the AI Boom
Summary
Elon’s Anthropic deal turns existing compute into a hyperscaler business while relieving the bottleneck throttling Claude. The reported lease gives Anthropic all of Colossus 1, alongside its addition of more than 220,000 NVIDIA GPUs and 300 MW; xAI’s models are now at Colossus 2. Brad Gerstner estimates $45 billion of incremental revenue this year and calls Elon unmatched at “converting electrons to tokens.” Chamath says the terrestrial business subsidizes Grok and de-risks orbital data centers, reinforcing the SpaceX valuation case.
Anthropic is winning now, but David Sacks’s monopoly extrapolation is the disputed bull case, not a settled outcome. Sacks charts ARR from roughly $10 billion on January 1 to $30 billion on March 31 and $44 billion in April, making an approximately $100 billion year-end exit plausible and $1 trillion in 2027 the provocative question: “Dario calls it AGI. I call it the biggest monopoly in human history.” Brad counters that March GAAP revenue merely matched OpenAI, compute remains scarce, and OpenAI, Google and xAI are already responding.
The hosts rejected an “FDA for AI” approval regime while treating frontier cyber capability as an immediate, narrower problem. Brad said Kevin Hassett’s drug analogy referred to government coordination rather than preapproval; Sacks favored “specific solutions to specific problems,” KYC during model previews, faster public-private testing and rapid distribution to cybersecurity firms. Sacks said that within three to six months, all major frontier labs, including Chinese models, would have cyber capabilities; open-source models would have them anyway.
Hyperscaler revenue has answered the first AI-revenue question and is keeping Brad aggressively long compute and memory. AWS reached a $150 billion run rate growing 28%, Azure $108 billion at 39%, and Google Cloud $80 billion at 63%; Brad moved from medium to large exposure in December and January, with more than 80% centered on compute, AI and memory. He argues Meta at 17 times, NVIDIA at 19, Microsoft at 20 and SK Hynix at five times fully taxed GAAP earnings are “not the stuff that bubbles are made of.”
Chamath thinks the AI trade has roughly 500 days before token buyers must prove measurable returns. He remains net long the companies “making the new thing,” but says there is “not a scintilla of evidence” yet that AI caused S&P 500 margin expansion; eventually enterprises must demonstrate, “I spent X and I made Y,” with Y exceeding X. Jason argues that ROI should cascade from infrastructure to models, applications and end users, while he says startup ROI is already “fait accompli.”
AI’s political defenses look weak because its concentrated wealth is more visible than its distributed benefits. Chamath gives tech leadership a “D-minus trending to an F,” warning that communities see a few potential trillionaires while nearly half of the roughly 9 GW planned this year faces protests. Jason proposed IPO allocations to Invest America accounts and more visible investment in health, education and basic science; Brad suggested free household electricity around data-center sites.
The employment and productivity verdict remains contested despite strong headline data. Sacks points to AI supplying 75% of Q1 GDP growth, construction wage gains of 25–30%, unemployment around 4.2% and improving prospects for recent graduates; Brad adds that S&P operating margins rose from 11% in 2023 to 13% this year. Chamath counters that financial engineering may explain the margins and that March labor-force participation was only 61.9%, versus 63.3% before COVID: the downstream proof is promising, but incomplete.
Deep dive
1. Local anti-tax politics is becoming a capital-flight trade
Jason predicted Spencer Pratt could win the Los Angeles mayoral race because his social team and relaxed debating style produce unusually viral attacks. Pratt’s substantive distinction was that street homelessness is primarily “an addiction issue” and “a mental illness issue,” not simply a shortage of housing.
Chamath’s non-consensus California call: Pratt could win while the Retirement Protection and Savings Act passes, constitutionally protecting retirement savings and personal assets while banning a wealth tax. The combined message would be nationally significant; Sacks replied, “From your lips to God’s ears,” but plans to remain in Texas until voters actually send it.
Jason described Mamdani’s video outside Ken Griffin’s home as personally dangerous amid a CEO being shot for ideological reasons and Sam Altman’s house being targeted. Griffin’s reported response—“I’m out,” with future efforts directed toward Florida—fit the pattern that previously pushed him from Chicago; Jason’s blunt conclusion was that New York is becoming “a flyover city.”
2. Colossus turns Anthropic’s compute shortage into Elon’s revenue
The reported transaction gives Anthropic all of Colossus 1; Anthropic also added more than 220,000 NVIDIA GPUs and over 300 MW. Claude subsequently doubled Claude Code rate limits, removed peak caps for paid users and increased Opus API volume, while xAI’s models moved to Colossus 2.
Chamath’s premise is categorical: Anthropic and OpenAI revenue performance has “nothing to do with demand. Zero.” It is entirely constrained by data-center capacity and power; with infinite power, he believes both revenue curves would be “even more parabolic,” making quarterly forecast beats and misses largely beside the five-year point.
That scarcity could worsen. Chamath said roughly 9 GW is scheduled to come online this year, nearly 50% is already being protested, and history suggests much of the contested capacity will be turned off—turning Elon’s early acquisition of sites, power and scale into a “critical asset” with king-making leverage.
Sacks framed the financial fix: xAI bore the full cost of giant training clusters without meaningful coding revenue, the enterprise market where current AI revenue is concentrated. Leasing capacity lets Elon maintain a frontier lab without “massive unpaid-for CapEx commitments,” while Anthropic removes its immediate ceiling.
3. “EWS” expands SpaceX from launch stack to hyperscaler
Brad used Shawn Maguire’s “five-layer cake” to describe the emerging stack: launch, connectivity, compute and hyperscaler, space data centers, applications and models, plus other bets. The ace card is that Elon was building “AWS all along—or EWS,” creating revenue before xAI itself catches up.
Brad estimated $45 billion of incremental revenue this year on top of analyst estimates in the mid-$20 billions. He described the less-connected H100 capacity as well suited to inference, while Macrohard and Macroharder contain 1.2 GW of Blackwell capacity.
Jason extended the thesis from giant factories to distributed compute: Tesla cars, Powerwalls and Starlink-connected homes could contribute capacity before orbital centers arrive. Pulte Homes and SPAN are already putting mini data centers with NVIDIA clusters beside homes, while Base Power was cited as pursuing a related model.
Chamath expects SpaceX to command 40–50 times revenue: at $40–50 billion next year, a $2 trillion offering becomes conceivable because only Elon offers that pipeline and TAM. Sacks said Tesla and SpaceX could merge into “Elon Corp” by the end of the year or middle of next year; the group agreed the premium is for innovation, while disputing whether Apple and other incumbents are penalized or merely fairly valued.
4. Anthropic’s exponential makes monopoly math conceivable, not settled
Sacks said Anthropic has compounded roughly 10-fold annually for three years, then moved from approximately $10 billion ARR on January 1 to $30 billion on March 31 and $44 billion in April. If compute arrives, he sees about $100 billion at year-end; the provocative question is whether it reaches $1 trillion in 2027.
Chamath’s comparison put the audacity in context: 2025 revenue was $420 billion for Apple, $300 billion for Microsoft, $390 billion for Alphabet, $700 billion for Amazon, $190 billion for NVIDIA, $185 billion for Meta and $110 billion for Tesla—roughly $2.3–$2.35 trillion collectively. A trillion-dollar Anthropic would suggest “Mag 1,” not Mag 8.
Brad’s pushback—worth keeping: Dario and Dwarkesh said combined leading-lab revenue might reach $1 trillion in 2029, while the labs expect about 5 GW by this year-end and 10 GW next year. Brad also noted that Anthropic’s March GAAP revenue was only around OpenAI’s level, leaving both startups fragile beside incumbents generating vast free cash flow.
Friedberg portrayed Anthropic as the focused “porcupine” while rivals played fox—spending time on Nano Banana, Sora, image generation and fantasy chatbots instead of coding, agents and co-work. The response is underway: OpenAI’s Codex is improving on GPT-5.5 and its new Spud base model, Google remains formidable, and xAI has tied up with Cursor; leaders still benefit because competitors must change while they merely maintain inertia.
5. “Safe Oil” reframes AI safety as a possible moat
Sacks’s thought experiment recast Rockefeller’s Standard Oil as “Safe Oil.” Because kerosene could illuminate a house or burn down a city, a PR-savvy monopolist could demand licensing, testing and government debates over “proper wick thickness”—appearing altruistic while regulators eliminated dangerous independent refiners.
The modern implication is not that Anthropic already owns a monopoly, but that another 18 months on its current trajectory could give it unprecedented control over the era’s most important technology. Sacks warned that safety policies can become regulatory capture, and questioned what competitive justification Anthropic had for banning OpenClaw from using its models.
Brad rejected the premise as wildly premature: five months earlier OpenAI looked unstoppable, Google and Amazon have substantial AI businesses, and five major labs are competing at the frontier. His fear is Washington “preemptively” choosing winners and losers before the race has left the starting gate; even on OpenClaw, he would only “double click,” not immediately file a case.
6. The “FDA for AI” panic leaves a narrower cyber mandate
The New York Times report said the White House might create a working group and review process after an Anthropic cyber model caused alarm. Hassett compared release to proving an FDA drug safe; Bessent instead described a continuing “calculus” between innovation and government’s duty to maintain safety.
Brad said Hassett personally clarified that he meant showing models to government so agencies can coordinate and harden systems, not requiring Washington approval. Brad supports better technical capacity, faster cyber review and a finite government feedback window, but called a standing model-approval agency “a disaster.”
Sacks said there was a large “fake-news” component to the FDA narrative, pointing to a rebuttal from Chief of Staff Susie Wiles and saying he did not think any senior official supported it. The March 20 national AI regulatory framework he worked on already lists legislation the administration could support: its principle is “specific solutions to specific problems,” not laissez-faire absolutism or a giant federal power grab.
The real deadline is the cyber model the hosts called “Mythos”: Sacks expects all major frontier labs, including Chinese models, to gain comparable cyber capability within three to six months, while open-source models will have those capabilities anyway. Because offense and defense use the same tools, he wants preview-period KYC and rapid access for CrowdStrike, Palo Alto Networks and startups; Brad added that labs already log APIs, fight distillation and flag suspicious activity. Both stressed Anthropic and OpenAI had acted responsibly rather than trying to release the models.
7. AI’s social license depends on visible, distributed gains
Chamath sees a “profound vibe shift” against tech oligarchs moving from Main Street into Washington. The industry talks constantly about AI’s negatives, communicates its upside poorly and fails to reinvest visibly enough in America; his grade for tech leadership is “D-minus trending to an F,” with political “antibodies” now building.
Jason proposed that NVIDIA, SpaceX, Anthropic and OpenAI allocate perhaps 1% or 5% of future offerings to Invest America accounts—an “IPO for kids”—or donate 1% of stock annually for 20 years. Unlike an end-of-life giving pledge, that would let ordinary citizens compound alongside the AI boom while funding health, education and basic science.
Jason’s broader capitalist prescription included gradually lifting minimum wages company by company and finding a universal-health-care structure, because lower-income consumers spend incremental earnings. Brad agreed that health coverage is desirable but challenged the financing with P.J. O’Rourke’s warning: “If you think health care is expensive now, just wait until you make it free.”
Sacks said AI ranks only 29th of 39 voter concerns, while cost of living and the economy rank first and second; he attributed 75% of Q1 GDP growth and 25–30% construction wage increases to the boom. Brad’s hyperlocal answer was even simpler: if a data center lands in Abilene, make household electricity in Abilene free.
8. Cloud growth keeps the US economy in a claimed Goldilocks window
AWS reached a $150 billion run rate growing 28%, Azure $108 billion at 39%, and Google Cloud $80 billion at 63%. Each added roughly $9.5–$10 billion of annualized revenue—about $30 billion collectively—even allowing for Azure and Google Cloud definitions that include software products; Jason paired that strength with Uber and Disney results as evidence of a resilient consumer.
Brad described a “Goldilocks” setup: accelerating GDP, controlled inflation, a 10-year yield around 4.3% and US leadership in “AI, AI, AI, compute, compute, compute.” Jason granted the administration’s business-friendly stance but argued the economy could be stronger without its tariffs and an unnecessary $100 billion war.
Sacks credited Trump with rescinding the prior chip-and-model approval regime, rejecting worldwide GPU-sale licensing and saying the president wanted AI companies to generate their own power rather than drain consumer grids. In his contrast, “Drill, baby, drill” and a blue-collar construction boom beat Bernie Sanders’s proposed data-center moratorium.
9. Brad is buying compute at multiples he says are anti-bubble
With the S&P 500 up only 8% this year, Brad sees neither index euphoria nor stretched AI leaders. His fully taxed GAAP comparisons were Meta at 17 times earnings, NVIDIA at 19, Microsoft at 20 and Google at 24: “This is not the stuff that bubbles are made of.”
About 25% of Brad’s portfolio is in SK Hynix, at five times fully taxed GAAP earnings; he also cited Samsung at six times and Micron at seven. More than 80% of overall exposure sits in compute, AI and memory.
The change came in December and January, when evidence from private model companies and public hyperscalers showed that AI revenue was arriving. Brad moved from medium to large exposure because without Anthropic’s growth and cloud reacceleration, investors would have concluded infrastructure lacked ROI and marked the market down 10–15%.
Chamath’s near-term allocation agrees: own “the people that make the new thing”—NVIDIA, memory makers, Anthropic, SpaceX and OpenAI—because those suppliers still need to be valued and demonstrate their own growth. His disagreement begins farther downstream, where buyers eventually must show measurable benefit.
10. Chamath gives the AI trade about 500 days to prove buyer ROI
Chamath first placed the critical fork two or three years away, then translated the trade into “call it 500 days where you just got to be net long.” After that, token buyers must prove either that OpEx and workforces shrink, or that revenue grows fast enough for margins to expand while OpEx stays flat or rises—the societal consequences differ sharply.
His accounting test is deliberately mundane: Anheuser-Busch must sell more beer, Nike more shoes and a medical-device company more hips and knees. Until a business can trace “I spent X and I made Y,” with Y greater than X and margins lifted, experimentation has not become a self-reinforcing flywheel.
Jason connected the stages: infrastructure ROI was uncertain until model revenue arrived; now model buyers must validate application ROI, which should then reach end users. He remains optimistic because enterprises continue spending month over month on coding tokens and bespoke software could drive a productivity wave.
Jason called startup ROI “fait accompli”: three people inside his 22-person investment firm now ship interfaces it once would have bought as SaaS. He also cited AI-generated Nike and DoorDash imagery replacing photo shoots, with creative costs cut roughly in half and advertising effectiveness improving by double-digit percentages. Chamath cautioned that broader margin gains could still reflect financial engineering rather than AI.
11. Margin and labor data do not yet settle the productivity debate
Brad supplied the strongest aggregate counterpoint: S&P 500 operating margins rose from 11% in 2023 to 11.8% in Q1 2024 and 13% this year, while combined Magnificent 5 headcount grew only about 3% over three years. Chamath replied that the margin increase could reflect the same financial engineering seen over the last decade, and Brad conceded post-COVID layoffs and financial discipline may explain part of it.
Brad noted that efficiency improved while unemployment stayed near 4.2%, within economists’ 4–5% full-employment range. He also said recent college graduates may be “AI natives” who use the tools better, despite predictions that entry-level work would disappear first.
Chamath’s closing counterexample was participation: only 61.9% of adults participated in March, versus 63.3% before COVID, so absent workers never enter the unemployment rate. Jason called the degree-level evidence mixed and said it was early; Brad rejected Jason’s “too soon to tell” framing, while the discussion remained divided over how much of the labor data reflects AI.