Anthropic S-1 Shocker, Greatest Business Ever? Google’s $500B Wipeout, Midterm Predictions

Summary

• Anthropic is targeting a $100 billion IPO after producing $16.3 billion of first-half revenue, up 14× year over year. Calacanis cites a November 16 listing under ANT at a $1.5–$2 trillion valuation, while Sacks calls the six-year-old company potentially “the biggest venture outcome of all time.” The $100 billion year-end ARR projection remains ambiguous between recurring revenue and annualized revenue run rate.
• The strongest bull case is that cheaper tokens are expanding margins rather than exposing a business dependent on subsidies. Sacks calls the reported $80 billion loss a “phantom number” arising from convertible-note accounting, against roughly $9–$10 billion of capital consumed. His mechanism is 10× annual efficiency improvement permitting substantial price cuts while retaining savings; Friedberg specifies approximately 85% incremental gross margins, with baseline compute costs still present.
• Enterprise penetration leaves substantial runway, but demonstrated customer returns remain the critical test. Customers spending at least $100,000 reached 11,900, approximately 1% of Friedberg’s estimated eligible global customer base. Chamath wants evidence that experimentation converts into multimillion-dollar commitments, while Friedberg acknowledges his assumed 2–10× customer return is unsupported: “I have no idea.”
• A remarkable business can still arrive at a fair valuation inside an increasingly concentrated funding market. Chamath’s revised IPO midpoint is $1.5 trillion; around $1.4–$1.6 trillion, “there’s no deal.” His larger concern is AI financing demand equivalent to roughly 30% of Treasury issuance through 2027: either higher corporate borrowing costs pressure the AI ecosystem, or higher Treasury yields raise borrowing costs throughout the economy.
• Power availability and government influence could become more binding constraints than customer demand. Chamath reads compute infrastructure as Anthropic’s dominant risk and argues that slowing permits could become a political weapon; Calacanis counters that Democrats could instead champion Anthropic’s expansion. Friedberg predicts government allocation of compute capacity, while Sacks warns that an exceptional business makes regulatory protection especially objectionable.
• Lower prices strengthen the adoption thesis while leaving open-model competition and recursive self-improvement unresolved. Calacanis revises his earlier commoditization call “in the face of new evidence” and imagines 500× token usage as workers acquire multiple agents. Sacks still sees cheaper open models threatening important markets such as coding, while Anthropic’s reliance on recursive self-improvement could collide with its safety constituency.
• Google’s $500 billion selloff divides concerns about search monetization from evidence that diversification is working. Search revenue still rose 12% to $63 billion, while Cloud doubled to $30.7 billion and operating income exceeded $11 billion. Friedberg questions the timing of a consumer AI replacement; Chamath wants auction-manipulation allegations addressed; Sacks sees the stock as somewhat oversold, and Calacanis buys more during the show.
• The central midterm disagreement concerns the Senate, with broad agreement that a Democratic House would unleash investigations. Chamath expects a Democratic “spritz”: a narrow House majority with Republicans retaining the Senate. Calacanis backs Polymarket’s Democratic probabilities of 68% for the Senate and 92% for the House; Sacks distrusts both polling narratives and prediction-market certainty, expecting potentially “a more normal midterm result.”

Deep dive

1. Anthropic’s IPO combines record ambition with concentrated control

• Calacanis says Goldman Sachs and Morgan Stanley are targeting November 16 for an ANT listing raising $100 billion at a $1.5–$2 trillion valuation. That would exceed SpaceX’s $85 billion June IPO; he cites a 78% Polymarket probability of raising at least $95 billion. Sacks calls the six-year-old company potentially “the biggest venture outcome of all time,” emphasizing the speed of value creation.
• Revenue, in Calacanis’s rundown, reached $16.3 billion in H1 2026, up 14× year over year; Q2 contributed $11.6 billion, roughly 30× the level five quarters earlier. Anthropic projects $100 billion of year-end ARR against a $40 trillion TAM, whose derivation he says he does not know. He also leaves unresolved whether ARR means recurring revenue or annualized revenue run rate.
• The seven founders would retain 50.1% of voting power through a single class of F shares. Calacanis also describes an optional arrangement directing 80% of their shares into a nonprofit-type entity, while explicitly acknowledging uncertainty over its form: “Who knows what it’s gonna be structured as?” His question is how that concentration of wealth and control changes technology and politics.

2. Compute multipliers undermine the cash-burning narrative

• Sacks argues the much-discussed $80 billion loss is a “phantom number”: convertible notes issued to infrastructure partners, including Amazon, generated accounting losses as Anthropic’s valuation rose. “If they had just issued equity, that loss wouldn’t be there.” Against roughly $110 billion raised, he says the company consumed only about $9–$10 billion, substantially changing his assessment of capital efficiency.
• His operating explanation separates raw compute from “compute multipliers”: better algorithms, curated data, and architectures such as MoE. Delivering a given capability becomes roughly 10× cheaper annually; cutting customer prices by, say, 50% still leaves substantial efficiency gains for Anthropic. Alternatively, those savings can buy greater capabilities. His rebuttal to commoditization is that falling token prices and improving margins can occur together.

3. The productivity stack looks profitable, but customer returns remain assumed

• Friedberg tracks customers spending at least $100,000 from 200 at year-end 2024 to 1,500 at year-end 2025 and 11,900 by September. He estimates roughly 200,000–300,000 US businesses, and more than a million globally, could spend that amount while keeping it below 1% of total spending. On that definition, Anthropic has penetrated only around 1% of potential customers.
• Friedberg’s illustrative stack turns $1 of NVIDIA cost into $4 of chip revenue at 75% margins. For SpaceX’s Colossus 1, he uses $10 billion of capex, depreciation over five years, and $1.25 billion monthly revenue to derive roughly 83% margins. Anthropic’s approximately 85% margin is incremental, with baseline compute costs remaining; the original dollar ultimately supports $18 of Anthropic revenue.
• Chamath challenges the final box: how much value customers actually receive. Friedberg concedes, “The last box I made up,” explaining that he assumed customers make 2–10× what they spend: “I have no idea.” His evidence is experiential—Claude lets his teams accomplish more, faster—even after installing their own open-source cluster. He allows that early exuberance could instead be masking poor returns.
• Chamath supplies Microsoft as a concrete productivity example: revenue increased from $198 billion to $332 billion over four years while staff increased by only 2,000. He interprets this as more output from approximately the same human input. Depending on uptake of an exit offer extended to 60,000 people, he says headcount could fall below 200,000.

4. Million-dollar budgets separate experimentation from enterprise conviction

• Chamath’s estimated customer distribution comprises approximately 40 million individuals and small businesses spending below $100,000, 10,000 customers spending $100,000–$1 million, 1,650 spending $1–$10 million, and 250 spending above $10 million. Expansion within existing accounts—“negative churn”—plus new customers could support hundreds of billions in revenue over three to seven years without approaching full penetration.
• His sales explanation: directors and VPs can experiment within delegated budgets of around $1 million, signing up for several products. Larger commitments require cross-functional approval and an explicit ROI or total-cost advantage. About 1,900 customers exceed $1 million, but stripping out venture-funded AI natives might leave only a couple hundred Global 2000-type companies. “The ROI case is still a little speculative.”
• The quarterly test is whether the 10,000 customers below $1 million graduate into single-digit millions and then tens of millions, supported by defensible use cases and analyses. Chamath hopes a listing after November 11 will require disclosure of Q3 figures. He reports whispers that growth moderated to around 50%, while leaving Q4 explicitly “TBD.”

5. A stronger balance sheet changes the IPO pricing calculation

• Chamath revises his earlier funding and valuation call: he had assumed only $30–$50 billion of cash and wanted pricing nearer $1 trillion. “I was wrong.” He says Anthropic had $125 billion entering June and estimates an $80–$120 billion raise would bring it toward the roughly $250 billion cash buffer he wants.
• Sacks adds approximately $80 billion of debt-facility capacity and puts prospective dry powder above $300 billion after the IPO. His answer to bears questioning future compute commitments is that Anthropic now has capital to support them, potentially through around 2030. Chamath agrees that the smaller necessary equity raise materially improves the offering’s prospects.
• Chamath’s revised valuation range is $1.3–$1.7 trillion, centered on $1.5 trillion. Raising less improves pricing power and control over the float. Nevertheless, at approximately $1.4–$1.6 trillion, he calls the stock “kind of priced fairly”: “There’s no deal,” although investors receive an entry into the company’s future.

6. Data centers turn power policy into an economic battleground

• Chamath begins his S-1 analysis with a word cloud of risk factors, treating management’s allocation of words as evidence of priorities. Compute dominates, especially data centers, financing commitments, and power. Regulation follows, then privacy, data retention, IP, and safety. Customer traction receives comparatively little attention. His conclusion: “As goes data centers, as goes power.”
• His partisan inference is that Democrats seeking an economic slowdown before 2028 could obstruct power and data centers, while Republicans must accelerate them. Calacanis challenges the incentive: why wouldn’t Democrats champion Anthropic as an ideological ally and prospective source of political funding? Sacks doubts that today’s party would do so, describing proposed model approvals as “a DMV for models.”
• Sacks cites GDP growth of 2.5% in the first half and tracking around 3.7% in Q3, against AI build-out spending near 3% of GDP; he argues the build-out effectively accounts for economic growth. Calacanis accepts AI’s importance but disputes the broader prosperity claim, emphasizing inflation, war, tariffs, and dissatisfaction among people without equity holdings. Sacks rejects calling the economy a disaster.

7. AI borrowing creates a risk that company analysis cannot settle

• Chamath says major AI companies collectively owe roughly $340–$400 billion, with Microsoft an exception for halving its debt. Calacanis checks Meta’s figures and reports $80 billion of cash and short-term investments against $82 billion of debt. The broader issue: anticipated AI financing through the end of 2027 equals approximately 30% of Treasury issuance over the same period.
• If investors prefer government debt, hyperscalers must offer higher coupons. Those financing costs then cascade through neoclouds and other infrastructure providers, bringing default risks forward and potentially creating a negative feedback cycle. This is why Chamath thinks a strong operating business can still face substantial market risk.
• If AI companies obtain attractive financing instead, Treasury may have to pay more. Because borrowing elsewhere is linked to government yields across two-, five-, and ten-year maturities, higher rates then affect the broader economy. Chamath says he has never seen this market risk before; Calacanis says he has not seen anything this close to the GFC in 15–20 years.
• Chamath calls the complex “too big to fail” but says, “I’m generally bullish.” Calacanis counters that exceptional American franchises could attract sufficient global capital. Conditional on Anthropic meeting its targets, he says its cumulative first seven years could exceed the comparable combined revenues of the iPhone, Google paid search, and Amazon retail, excluding AWS.

8. Government influence could become a constraint on private AI power

• Friedberg views companies, nonprofits, and governments as competing organizations accumulating resources for their constituencies. AI’s capital, productivity, and influence threaten governmental power. His prediction extends beyond blocking construction: governments could regulate the actual compute layer after production, allocating who receives how much capacity. He expects the conflict to develop regardless of which party governs.
• Calacanis wants Anthropic’s founders publicly championing candidates as visibly as Sacks supported Trump. Sacks pushes back that Anthropic already has substantial PAC funding and a large government-affairs operation, hiring former Biden officials and, increasingly, former Trump officials. He distinguishes his individual political activity from a corporation pursuing regulatory advantage.
• For Sacks, an IPO would clarify that Anthropic is a major technology company deserving the same suspicion as its peers—an “800-pound gorilla,” rather than an innocuous startup. He pairs his “best business ever” thesis with opposition to regulatory capture: an AGI monopoly or duopoly should receive stronger competition, not additional regulatory moats.

9. Cheaper intelligence expands usage while open models and RSI test the lead

• Calacanis revises his earlier open-model call “in the face of new evidence”: competition is “not slowing them down yet.” Sacks stresses that Q3 remains undisclosed; the claim that Anthropic eliminated 99% of distillation, lost revenue, then reaccelerated after token-price cuts is “scuttlebutt on X,” rather than an established explanation.
• Sacks’s demand argument is “Jevons paradox”: cheaper tokens unlock uses, and agents consume far more than prompts. Calacanis imagines 100–200 million white-collar workers each running five agents and predicts 500× token usage. His adoption forecast is conditional: once cyber constraints ease, daily workplace use could rise from roughly 10% to 60–80%; large enterprises currently restrict connections to email and internal knowledge.
• Sacks still sees open models commoditizing valuable applications such as coding without necessarily matching the frontier. Anthropic’s defense increasingly uses its own models to discover algorithmic efficiencies. He cites an August chart showing recursive self-improvement used in 26% of tasks involved in developing subsequent models; he does not know October’s share.
• He suspects a split between Anthropic and safety “Doomers”: the company must accelerate RSI to stay ahead, while its allies fear exactly that. He reads Dario’s pacing letter as “almost like creating an alibi in advance for themselves.” Sacks does not fully endorse their RSI-risk story, and leaves unresolved whether safety pressure could slow research enough for open models to catch up.

10. Google’s search slowdown obscures a rapidly improving cloud franchise

• Calacanis reports a 10–15% Thursday decline erasing approximately $500 billion of market value after Q3 earnings. Overall revenue grew 25%, but search missed expectations for a second consecutive quarter, growth slowed from 17% to 12%, and free cash flow was negative. Searches themselves are increasing; monetization is the concern as AI supplies answers and agents complete tasks.
• Sacks disputes the implication that search is already deteriorating: quarterly revenue increased from $56 billion to $63 billion. Its share fell below 50% because other businesses grew faster. Cloud doubled to $30.7 billion; Friedberg notes its operating income increased from approximately $3.5 billion to more than $11 billion. Sacks regards diversification as something Alphabet has long wanted.
• For Friedberg, roughly 21× operating income requires reliable future cash flows. The question is whether consumer AI can replace search monetization quickly enough as the legacy engine flattens. He also floats an explicitly uncertain internal resource conflict: Google may earn more renting compute externally than allocating it to its own model researchers.

11. Google’s auction controversy complicates the diversification story

• Chamath highlights advertiser Max Anderson’s allegations, explicitly saying he does not know whether they are true. Anderson describes spending more than $500,000 monthly and claims changes to Google’s ad auction. Chamath explains the traditional Vickrey mechanism: the highest bidder pays the runner-up’s price. Abandoning that mechanism could, in his view, let Google overearn temporarily.
• Friedberg’s concern is that advertisers eventually recognize weaker economics and reduce spending. Chamath urges investor relations to address the allegations before “pulling money from the future into the present” becomes the prevailing interpretation. Separately, a more balanced mix of Search, YouTube, Cloud, and other bets could improve diversification while requiring a transition to investors comfortable owning a conglomerate.
• Chamath’s talent priority is blunt: “How do you keep Demis at Google?” He also says to pay Thomas Kurian whatever he wants. Sacks calls the stock a little oversold; Calacanis buys more during the recording. Friedberg notes Alphabet’s approximately 6% SpaceX and 15% Anthropic stakes, while Calacanis says he thinks it owns 85% of Waymo.

12. The midterm split concerns the Senate and the next two years

• Chamath still expects a Democratic “spritz”: Republicans retain the Senate, Democrats obtain a narrow House majority. He anticipates subpoenas and document releases creating “FUD” before 2028. He also thinks an economic slowdown benefits Democrats, returning the conversation to state and local control over power permitting and data centers.
• Sacks distrusts polls after hoax firms entered averages and cites Patrick Ruffini’s finding that three-quarters of polls across three elections erred toward Democrats. He also challenges prediction markets, citing misses on Bolsonaro—though he allows Bolsonaro may have been robbed—and Netanyahu. His forecast is modest: the governing party normally loses midterms, and the outcome might be “a more normal midterm result” than the advertised bloodbath.
• Calacanis backs Polymarket’s 68% probability of Democrats taking the Senate and 92% for the House. His reasoning combines anti-Trump turnout, disillusioned moderates, and an America First constituency angry over inflation and another war. He expects a Democratic sweep and presents that as analysis rather than his personal political preference.
• Calacanis predicts subpoenas, renewed impeachment fights, and legislation designed to force damaging Republican votes before 2028; he expects new Republican leadership beyond JD Vance and Marco Rubio. Sacks agrees that investigations would return but argues Democratic overreach could revive the political pendulum. Their disagreement is chiefly over the electoral outcome and eventual political beneficiaries; both expect confrontation if Democrats take the House.