Anthropic’s Growth Case Hinges on Enterprise Returns and Compute Capacity

The figures presented from Anthropic’s S-1 made a striking case for the company’s scale. Revenue reached $16.3 billion in the first half of 2026, the hosts said, up fourteenfold year over year. Second-quarter revenue was reported at $11.6 billion. The company was targeting an IPO that could raise $100 billion at a valuation somewhere around $1.5 trillion to $2 trillion; those numbers were discussed as targets and market expectations, not settled outcomes.

For David Sacks, the combination made Anthropic a candidate for the largest venture-backed outcome yet—and possibly the fastest value creation of that scale. The company was only six years old, he noted, while SpaceX had taken more than two decades to reach its own proposed public-market scale. At a $2 trillion valuation, estimates shown on screen suggested that hundreds of early employees could make hundreds of millions of dollars and thousands more could make tens of millions. Sacks described the potential payout as an unprecedented wealth-creation event for San Francisco and Silicon Valley.

He also challenged the press coverage of Anthropic’s reported $80 billion loss. In his reading, that figure was not a cash loss from operating the business: it resulted from convertible notes issued to infrastructure partners. As Anthropic’s valuation rose, the notes were recorded as a loss on the balance sheet and flowed through the income statement. Had the company issued equity instead, he argued, the accounting loss would not have appeared. Sacks pointed to roughly $110 billion raised, only about $9 billion or $10 billion spent, and more than $100 billion still on the balance sheet as evidence that the company had been unusually capital-efficient.

The operating figures seemed to reinforce his case. Anthropic’s calculated gross margin, according to a company chart shown during the discussion, rose from 12.4% in the first half of 2025 to 85.4% in the first half of 2026.

Sacks attributed the improvement to gains in model efficiency. For a given task, he said, Anthropic was making its models roughly ten times more cost-effective year over year. Lower costs could be passed on to customers, while efficiency gains not passed through could improve the company’s margins. He described this as a combination of scaling raw compute and developing “compute multipliers”: algorithmic improvements, better-curated data, and architectural changes that let a model do more with the same resources.

That reasoning directly answered one concern about the business: that open models and falling prices would push commercial providers toward zero margins. Sacks argued the S-1 showed the opposite so far. Anthropic was lowering prices while improving margins, he said, not simply discounting to defend share. He also suggested that the company could differentiate prices by task: cheaper tokens for simpler work, higher prices for more demanding capabilities.

There was a second, related case for growth: the customer base remained small relative to the potential market. David Friedberg highlighted the increase in customers spending at least $100,000 annually, from around 200 at the end of 2024 to about 1,500 a year later and 11,900 by September 2026. He estimated that roughly a million businesses globally could spend that much on Anthropic’s products while keeping the expense below 1% of their total annual spend. On that basis, he put the company at about 1% of its potential customer base.

Friedberg also laid out an illustrative model for how value might travel through the AI infrastructure stack: Nvidia sells hardware, a compute provider rents capacity, Anthropic sells model access, and customers use the service to create business value. His chart suggested that a dollar of Nvidia manufacturing cost could underpin much larger annual revenue figures downstream. But he was explicit about the weakest link: the customer-value estimate assumed that customers got back two to ten times what they paid Anthropic. “The last box I made up,” he acknowledged when Chamath Palihapitiya asked how that value had been measured.

Friedberg said his own experience using Claude made him believe it helped his company get more done in less time, but he did not present that as a measurement of aggregate returns. The distinction matters to the investment case. Customer counts and Anthropic’s revenue show that companies are paying; they do not, by themselves, establish how much value those customers capture. Chamath saw a related gap in the spending distribution. About 10,000 customers were spending between $100,000 and $1 million, around 1,650 between $1 million and $10 million, and about 250 at least $10 million, according to his analysis of the S-1. He thought the large middle tier could expand as more teams experimented, but the move to multimillion-dollar contracts required a stronger, cross-functional return-on-investment case. He estimated that only a few hundred large companies beyond AI-native firms had yet made that case convincingly.

That leaves a company with fast growth and a large claimed runway, but with an important question still open: how quickly can experimentation become durable, large-scale enterprise spending? Chamath considered hundreds of billions of dollars in revenue possible within three to seven years if customer spending expanded and existing customers increased usage. He also said the IPO valuation around $1.4 trillion to $1.6 trillion looked fair rather than obviously cheap. To buy, investors would need a view on the future growth, not just the current figures.

Chamath Palihapitiya began his S-1 analysis with what he considered the company’s most revealing disclosure: its risk factors. His team grouped words in the section by topic and found that compute appeared more often than any other theme, narrowly ahead of regulation. Within compute, he said, the document emphasized data centers, financing commitments, and power. Customer traction appeared much less frequently. For Chamath, that distribution showed a company whose central problem was not finding demand but building and financing the infrastructure to serve it.

The distinction shaped the discussion of both Anthropic and the wider AI economy. Chamath argued that a small group of companies was driving a significant spending wave: once-profitable technology businesses were committing cash to data centers and other infrastructure, while taking on more debt. Charts he presented showed free cash flow under pressure across several large technology companies and a substantial increase in their balance-sheet debt. He framed the borrowing needs of AI companies through 2027 as large enough to compete with US Treasury issuance for investor demand.

That comparison led him to a market risk he said he had not seen before. If AI companies had to offer higher borrowing rates to attract capital, he argued, financing costs could rise across their suppliers and increase default risk. If Treasuries instead had to offer higher rates to compete, borrowing could become more expensive throughout the economy. In either case, the same buildout that might produce substantial future returns could create financial-system exposure while it was being financed. Chamath was not saying the AI business case had failed; he was saying its size might make its financing conditions matter well beyond the companies themselves.

Sacks took the capital spending as a reason for optimism. He did not think sophisticated companies would make such investments without expecting a return, and he pointed to the economic activity generated by the buildout. Jason Calacanis argued that investment might still be inadequate: if office workers gained access to agents, token use could rise dramatically, requiring far more compute than companies were currently building. Chamath agreed that demand could grow but stressed the obstacles to meeting it: permitting, utility staffing, equipment, and the ability to bring power online.

Friedberg offered a political explanation for why those obstacles might persist. Governments, he argued, are organizations competing for resources and influence, just as companies are. AI could threaten government power by concentrating capital and improving private organizations’ ability to compete. In that view, governments might try to control access to compute—not necessarily the chips themselves, but the compute layer—and could obstruct data-center construction as part of a broader contest over who directs the technology and benefits from it. Friedberg expected each side to argue that its own approach, government or private-market AI, served the public better.

The hosts disagreed about how that conflict would run through party politics. Chamath suggested that Democrats could benefit from slowing data-center and power development if it weakened economic growth ahead of the 2028 election, while Republicans would have reason to accelerate it. Sacks said that was consistent with his view of the current Democratic Party, which he described as increasingly hostile to superintelligence development. Calacanis pressed a different possibility: that Democrats might support data centers and Anthropic because the company could become a major source of wealth and political influence on the left. The exchange did not resolve whether either party would favor enabling the buildout over constraining it.

That dispute connected to Anthropic’s unusual ownership and governance structure. The hosts discussed a filing in which seven founders would hold 50.1% of the voting power through a Class F share, and the allocation of a large share of founder equity to a nonprofit entity. They also debated the company’s political activity. Calacanis argued that Anthropic’s leaders should openly champion candidates if they wanted to shape policy. Sacks replied that the company had already built a substantial government-affairs operation, funded political entities, and recruited former officials from both the Biden and Trump administrations. He distinguished his own public political activity from corporate lobbying, saying he was acting as an individual rather than on behalf of a company.

For Sacks, an IPO would make Anthropic’s scale more visible and weaken any claim that it should be treated as a small, benign startup. He described it as a powerful monopoly or duopoly in the creation of AGI and argued that a public-market giant should be treated with the same suspicion as other large technology companies. In his view, it should not receive a regulatory moat; the market should be kept competitive.

The debate about Anthropic’s public image sharpened around its pursuit of recursive self-improvement. A chart shown during the discussion, attributed to Anthropic, indicated that Claude was leading 26% of the company’s model R&D tasks by August 2026, up from 1% a year earlier. Sacks saw a tension between Anthropic’s safety-oriented rhetoric and its push to use its own models to improve the next generation of models. The same process, he argued, helped explain the efficiency gains underpinning its business. But he also said that recursive self-improvement was among the risks that AI “doomers” feared most, and that Anthropic might find it difficult to maintain its lead without continuing to push in that direction. He speculated that a split between the company and its safety-focused allies could follow; he did not claim that such a break had already occurred.

For potential investors, Chamath separated the risks he thought could be tracked from the ones that were harder to price. One was whether Anthropic could demonstrate enough return on investment to move more customers into multimillion-dollar contracts. The other was the financing risk created by an AI buildout large enough, in his account, to compete with government borrowing. Sacks and Calacanis were more confident that falling token prices would expand demand rather than erode the business. They invoked the Jevons paradox: cheaper use could produce more use. Sacks added that agents and continuous cybersecurity monitoring could consume far more tokens than simple prompts or occasional tests. The claim was plausible within their thesis, but the scale of future usage remained a forecast, not a demonstrated outcome.

Alphabet’s sharp share-price fall after earnings exposed two competing readings of the company. One focused on resilience: search revenue was still growing, Google Cloud was expanding quickly, and newer businesses were making Alphabet less dependent on advertising. The other focused on what might happen to search if users increasingly turned to AI systems that answer questions directly or complete tasks for them.

Jason Calacanis pointed to the pressure in search advertising. Searches were still increasing, he said, but revenue growth had slowed, and people were clicking on ads less. Search revenue had grown 12%, down from 17%, while Google Cloud revenue had doubled year over year to about $31 billion. Overall Alphabet revenue was still growing at 25%, according to figures shown during the discussion, but the company was spending heavily enough that free cash flow had become a concern. The market’s anxiety, as Calacanis framed it, was about whether AI would weaken the business that had historically funded the rest.

Friedberg called it a complicated story. Google Cloud looked strong to him, particularly because the infrastructure business had moved from being unprofitable to generating substantial operating income. The company also had models it could sell alongside its compute capacity, potentially capturing more value than a cloud provider that only rented infrastructure. But he saw greater uncertainty in Google’s core ad business. He raised the possibility that recent revenue growth reflected changes in ad pricing that could prove difficult to sustain if advertisers concluded they were paying more for lower-quality clicks.

The discussion turned on a specific allegation circulating among advertisers. A post by Max Anderson, described by Chamath as someone who had spent at least $500,000 a month on Google Ads, claimed that Google had moved away from a second-price auction model. Under the previous approach described in the post, a high bidder would pay just above the next-highest bid. Anderson alleged that Google had begun charging advertisers closer to their own bids and budget caps, while also reducing the precision of keyword targeting. He characterized the changes as extractive and damaging to the advertiser ecosystem.

Chamath said he could not verify those claims himself. But he argued that, if accurate, the auction change could mean Google was “over-earning”—pulling revenue forward in a way that might ultimately reduce advertiser trust and spending. The allegations had circulated for months, he said, and Google’s investor-relations team should address them. The visible replies to Anderson’s post included advertisers who said they had seen similar changes.

Sacks emphasized a different part of the numbers. Search revenue had risen from about $56 billion to $63 billion year over year, he said. It had not yet deteriorated in absolute terms; its share of Alphabet’s revenue was falling because Google Cloud and other businesses were growing faster. He viewed this diversification as a long-sought improvement, not evidence on its own that search was collapsing. He considered the market reaction excessive, while stopping short of recommending that anyone buy the stock.

The disagreement was not simply whether Google could continue to grow. It was whether the company could replace or adapt its consumer search business quickly enough if AI changed how people looked for information. Friedberg said the key question was what consumer AI product could generate meaningful revenue against a search business that might flatten or decline. He also described uncertainty inside Google about whether its future should center on search, cloud rental, or models—and whether the company’s researchers had the resources they wanted when renting compute to others could be more lucrative.

Calacanis, who said he had bought shares during the show, put weight on Alphabet’s broader assets and the continued increase in search volume. Friedberg added that Alphabet held stakes in SpaceX and Anthropic. Those points supported the case for diversification, but did not remove the question at the heart of the selloff: whether AI can turn into a substantial consumer business for Google before it weakens the economics of the company’s existing one.

The hosts’ election predictions rested on sharply different judgments about both the evidence and the political mood. Chamath Palihapitiya said he had begun paying less attention because of misinformation. He expected Democrats to gain the House narrowly while Republicans retained the Senate, but he also described early signals as confusing. His larger point was that control of Congress could make data-center policy and investigations into the administration consequential issues heading into 2028.

David Sacks urged caution about polls. He cited a polling firm that had described itself as a demonstration of weaknesses in the industry, and a study by Patrick Ruffini that he said found polls had overestimated Democrats in three-quarters of the cases reviewed. He also referred to earlier high-profile polling errors. His conclusion was not that Republicans were sure to win; he said an opposition party commonly gains ground in midterms when its rivals hold the White House. Rather, he objected to treating predictions of a Republican “bloodbath” as settled.

He also disputed the description of the economy as a disaster. Sacks cited GDP growth estimates, household income, poverty, and real wage growth to argue that the economic picture was stronger than critics said. Calacanis instead emphasized dissatisfaction among Trump voters and independents, including concerns about inflation, immigration, and war. He argued that Democrats would benefit from anti-Trump turnout while parts of the Republican coalition felt let down.

Calacanis trusted Polymarket’s odds more than the polls, citing a 68% chance of Democratic control of the Senate and a 92% chance for the House. Sacks countered that prediction markets reflect the views of people willing to bet and can be wrong. He agreed that Democrats might win but rejected the implication that the outcome was already certain.

Their forecasts diverged further on what a Democratic victory would mean. Calacanis predicted investigations and subpoenas, followed by a chaotic two years that could push Republicans toward a new populist candidate. Chamath described the prospect of subpoenas and disclosure as a way to create uncertainty around the administration and its allies. Sacks agreed that a Democratic-controlled Congress could produce a confrontational agenda, but expected the public to be reminded of what he saw as the party’s more extreme tendencies. The disagreement was less about whether a change in congressional control would intensify conflict than about which party would pay the political price for it.