All-In on Anthropic’s IPO, AI Infrastructure and the Midterm Elections

The All-In hosts welcome David Sacks back and joke about reporting that they had cooled on the term ‘superintelligence.’ Sacks disputes the account, saying he was not on the episode cited. The conversation turns from their introductory banter to Anthropic’s proposed public offering.

Jason Calacanis introduces the S-1 as the filing that describes a company ahead of an IPO. He says Anthropic is targeting a November listing and a $100 billion raise, and discusses a possible valuation in the trillions. Those are the panel’s figures and expectations at the time of recording, not a completed offering.

Calacanis points to $16.3 billion of revenue in the first half of 2026 and describes rapid year-over-year growth. He also mentions an ambitious end-of-year ARR projection, while noting that the term may refer to recurring revenue or an annualized run rate. He asks whether the growth makes Anthropic an exceptionally valuable business.

David Sacks says the potential venture outcome could be extraordinary because of how quickly it was built. He compares its age with the much longer history of SpaceX and says an offering at the expected scale could distribute substantial wealth to employees.

Sacks challenges the interpretation of a widely discussed reported loss. He says changes in the value of convertible instruments produced a large non-cash accounting charge, and argues that this should not be read as equivalent cash burn. The distinction matters, he says, when evaluating how efficiently the company used its funding.

In Sacks’s reading of the filing, Anthropic retained a large cash balance relative to the capital it had raised. He also describes improving gross margins and suggests that producing a given model capability has become much cheaper over time.

Sacks says that lower costs can benefit both customers and the model provider: prices can fall while efficiency improves even faster. He calls the latter improvements ‘compute multipliers,’ including algorithmic changes, data curation and architecture, and says they can be used either to lower costs or produce greater capabilities.

He argues that those trends challenge the idea that model providers must inevitably lose all their margin to commoditization. He also suggests that a lead might grow if models help improve future systems. These are his interpretations and forecasts, not proof of how the market will develop.

Calacanis asks David Friedberg about the effect of a large capital influx and the proposed ownership structure. He raises questions about where the founders’ wealth and influence might go and what that could mean for the technology industry and politics.

Friedberg turns first to customer demand. He cites a steep rise in customers spending at least $100,000 and estimates the number of businesses worldwide for which that amount would represent less than one percent of annual spending. By that measure, he argues, the possible customer base could be much larger than the current one.

To illustrate potential value, Friedberg sketches a chain from chip production through infrastructure to model services. He uses illustrative cost, revenue and margin figures for different layers and asks whether enough economic value reaches the final customer.

Friedberg suggests customers might receive two to ten times what they spend. When challenged, he expressly acknowledges that the last figure is an assumption he made, not a measured result. Continued spending might indicate customer value, he says, but early exuberance could also mean some buyers are not yet getting that return.

He describes his own experience using Claude to help his company accomplish more, while also mentioning investment in an open-model setup. He thinks a high-margin provider may have room to segment pricing: charge less for easier tasks and more for more demanding ones.

Chamath Palihapitiya shares a comparison of Microsoft’s revenue growth and relatively flat headcount over several years. He reads it as an illustration of potential productivity gains, while acknowledging uncertainty about future staff levels. The comparison by itself does not establish how much of the change was caused by AI.

Palihapitiya begins his own filing analysis by looking at risk factors. He says his team counted the prominence of terms in that section and found compute-related issues especially visible, with government policy, privacy and safety also appearing. He treats the relative word counts as a way to investigate management’s concerns, not as a formal ranking supplied by the company.

Within the compute category, he highlights data centers, financing and electricity. In his view, the availability of power and infrastructure may constrain growth across the broader AI industry. He then speculates about how the issue could shape the incentives of the two US political parties.

Calacanis offers a competing political theory: perhaps Democrats would want to support data-center construction and benefit from the success of a company whose founders might fund causes they favor. The discussion is explicitly conjectural; neither host establishes that either party has adopted the proposed strategy.

Sacks disagrees with Calacanis’s political assessment and argues that influential Democrats favor much stricter AI controls. He and Calacanis dispute the motives of political actors and the condition of the economy. Their characterizations are partisan views, not a settled account of either party’s position.

The hosts also disagree about how to interpret economic data and voters’ experiences. Calacanis says some Trump supporters are dissatisfied with inflation and other issues; Sacks challenges his description of those voters. They agree that the election will provide further evidence about political preferences.

Returning to the filing, Palihapitiya divides Anthropic’s customers into spending tiers. He describes a large base of smaller users and much smaller groups spending more than $100,000, $1 million and $10 million. His conclusion is that growth could come both from new customers and from higher spending by existing ones.

He cautions that moving a corporate customer from a modest budget to a multimillion-dollar commitment typically requires broader approval and a clearer case for return on investment. Some large customers may themselves be venture-funded AI companies, so he regards the proof of durable enterprise ROI as still partly speculative.

On capitalization, Palihapitiya says his earlier estimate underestimated the cash Anthropic already held. He revises his view of how much it would need to raise to reach his preferred cash buffer and offers a tentative IPO valuation estimate. Sacks adds his view of how cash and credit facilities could support future compute commitments.

The hosts briefly check Meta’s net-cash position after Palihapitiya expresses surprise at a negative figure. They clarify that net cash subtracts debt from cash and short-term investments. The exchange illustrates why a balance-sheet headline needs its underlying definition.

Palihapitiya broadens the argument from Anthropic to the US economy. A small group of large technology companies, he says, is undertaking an enormous investment cycle in AI infrastructure. The spending can support growth, but it also consumes cash and increases borrowing needs.

He presents a comparison between the financing sought by AI companies and planned Treasury issuance. His concern is that corporate borrowers and the government may compete for investors’ capital, potentially increasing borrowing costs. He stresses that the concentration of activity in a handful of firms is unusual in his experience.

Calacanis offers a more optimistic reading of the investment: the companies making it may expect attractive returns, and enterprise adoption could still have a long way to run. He argues that more agents and applications would use many more tokens and could leave the industry short of compute.

Palihapitiya agrees demand could be strong but points to practical bottlenecks: permits, utilities, staffing and the equipment required to connect new facilities. In his view, even an industry willing to spend may struggle to bring capacity online at the desired speed.

Calacanis asks how the industry can build support for data centers when voters across parties are skeptical of concentrated wealth and powerful companies. The group considers the political narrative surrounding infrastructure, public benefit and the distribution of AI’s gains.

Friedberg frames government and private organizations as groups competing for resources and influence. He suggests powerful AI could intensify that competition because it affects productivity, information and capital accumulation. This is his conceptual framework, not a prediction the other hosts all endorse.

Friedberg speculates that governments may eventually regulate access to computing capacity itself, rather than only controlling chips. He expects political arguments over whether public or private control delivers more benefit. Palihapitiya reiterates his concern about a small group of companies seeking large amounts of financing.

The discussion turns to political influence. Calacanis asks why Anthropic’s founders have not publicly supported candidates in the same way as some other technology figures. Sacks says Anthropic has invested in government affairs and political organizations; Calacanis distinguishes such activity from a founder’s explicit public endorsement.

Sacks argues that an Anthropic IPO at the discussed scale would change how the public sees the company: from a startup into one of the largest technology businesses. He says its influence should be scrutinized and raises his longstanding concern about regulatory capture. Those are his allegations and judgments, not findings established by the discussion.

Sacks then points to criticism from AI-safety advocates who, he says, see tension between calls to slow advanced AI and a race to improve models. He argues that Anthropic’s use of its own models in research could accelerate progress and that continued competition may strain its relationship with some safety supporters.

He speculates that a public appeal to pace AI development could later serve as an explanation for continuing the race if competitors do not agree to slow down. He acknowledges that the outcome of this potential split is uncertain. His interpretation of Anthropic’s motives is contested commentary.

Calacanis returns to Palihapitiya’s investment framework. Palihapitiya separates two questions: whether Anthropic can persuade customers to expand from smaller contracts to much larger ones, and whether the financing needs of the AI sector create a broader market risk that is harder to price.

On customer growth, he says larger contracts will require defensible use cases and evidence that executives can use to approve spending. He suggests watching customer tiers over successive quarters, while recognizing that projections depend on businesses achieving the value they expect.

On market risk, he describes two possible paths. If corporate borrowers pay higher interest to attract capital, AI infrastructure projects and their suppliers become more expensive to finance. If the Treasury must pay more, higher benchmark rates can raise borrowing costs across the wider economy.

Palihapitiya says the concentration of capital flows makes him uneasy even if the underlying AI businesses perform well. Calacanis takes the opposite view, emphasizing the potential strength of American technology franchises and arguing that investors may still have strong reasons to finance both the government and the industry.

Calacanis compares the possible trajectory of Anthropic’s business with the early growth of major technology franchises. His comparison depends on the company reaching future targets, so it remains a scenario rather than a description of revenue already earned.

Sacks asks Calacanis whether he has changed his earlier view that open models would erode Anthropic’s business. Calacanis says the reported growth has made him reassess the near-term threat, although he still sees open models as a competitive risk. Palihapitiya clarifies that his own warning concerns financing concentration, not a denial of customer traction.

The panel discusses rumors about a possible slowdown and subsequent change in demand, while acknowledging that the relevant quarter’s numbers are not yet known. Sacks suggests efforts to limit model distillation and recent price reductions could affect the pattern. The claims are speculation, not verified results.

The hosts argue that cheaper model tiers can be sufficient for many enterprise tasks. If prices fall and more uses become economical, demand could grow even as the price per token declines. They invoke Jevons paradox to describe this possible expansion, rather than treating it as a guaranteed outcome.

They also expect agents and continuous security monitoring to consume more tokens than occasional prompts or periodic tests. Calacanis’s estimates of future usage are forecasts. Sacks says open models could still make certain tasks cheaper and more commoditized, even if they do not overtake the leading models overall.

Sacks identifies another tension: if a provider relies on its latest models to improve efficiency and stay ahead, it may increasingly use techniques that the same company has described as risky. He wonders whether safety advocates would eventually demand a slower pace and whether open competitors might close the gap.

Calacanis points to survey data suggesting that frequent workplace AI use still has room to expand. He argues that security concerns and restrictions on connecting company data can limit adoption. His expectation of much broader future use is a projection, not an observed adoption rate.

As the discussion closes, the hosts praise the business they have been analyzing while Sacks repeats his concern about concentration and lobbying. They also say future filings from other companies could give investors more information with which to compare the sector.

Calacanis turns to Alphabet after a sharp stock decline following earnings. He says search revenue growth slowed while cloud revenue grew quickly, and asks whether the company’s other businesses can compensate if the economics of search advertising change as AI answers more questions directly.

Friedberg calls it a complicated story. He sees strong growth in Google Cloud and the value of its computing infrastructure, while asking how much the company can earn from its own models as well as from renting infrastructure to others.

He is less certain about the traditional consumer advertising business. Friedberg discusses claims that changes in ad pricing might temporarily increase revenue per click but eventually lead advertisers to reduce spending. He treats the durability of that revenue as an open question.

For Friedberg, a central issue is whether a consumer AI product can generate meaningful revenue as search advertising potentially flattens or declines, and how quickly either transition occurs. He also speculates about internal tradeoffs between developing models and allocating compute to cloud customers; he says he does not know the causes of individual departures.

Calacanis notes that Alphabet has spent years diversifying, pointing to businesses including YouTube, Android and Cloud. He asks whether the company should prioritize earnings from search while investing elsewhere. Palihapitiya challenges the comparison Calacanis uses but addresses the concern about search economics.

Palihapitiya refers to an advertiser’s widely circulated allegations about changes to Google’s ad auction. He explicitly says he cannot verify the claims himself. If advertisers and investors believe the system has changed in ways that temporarily boost revenue, he says, the company should explain or refute the allegation.

He describes a second-price auction as a mechanism where the winning bidder pays according to the next bid, and argues that changing auction mechanics could alter advertiser economics. His discussion is conditional on the allegations being accurate; the podcast does not establish that Google made the changes claimed.

Palihapitiya adds that Alphabet’s changing revenue mix could be healthy diversification, but a transition from a dominant search business to several large segments might also change which investors want to own the shares. He emphasizes communication with investors and retaining AI research leadership.

Sacks offers a different emphasis. Although search fell below half of total revenue by the measure he cites, he says its absolute revenue still increased year over year. In his view, the change in share reflects especially rapid growth elsewhere, notably in Cloud, and is not itself evidence that search is shrinking.

Sacks views the development of Cloud, YouTube and subscription businesses as long-sought diversification. Calacanis agrees that Alphabet has valuable assets and discloses that he bought shares during the recording. Their remarks are their own investment views, not a recommendation or a verified valuation.

The hosts move to the coming US midterms. Palihapitiya says conflicting information makes him uncertain, but offers a tentative scenario in which Republicans retain the Senate and Democrats narrowly win the House. He speculates about how investigations and infrastructure policy could play into the next presidential cycle.

He says state-level races and governorships also deserve attention because decisions affecting power and data centers may occur outside Washington. His suggestions about what political parties might try to do are conjecture, not established plans.

Sacks expresses distrust of polling, citing reports of fake polls and a study of historical polling errors. He argues that some coverage may depress Republican turnout and urges voters to participate. These are his interpretations; the panel does not substantiate a general claim that polls are deliberately intended to suppress voting.

Sacks disputes the description of the economy as a disaster and cites growth, income and poverty figures as support. Calacanis has earlier challenged his view of how voters experience the economy. The figures and competing political interpretations are presented as statements made in the conversation, not independently verified here.

Calacanis says prediction markets point toward Democratic wins in both chambers. Sacks replies that prediction markets can be wrong, citing prior elections, and notes that a president’s party often loses ground in the midterms. He expects a potentially more ordinary result than some of the most dramatic forecasts suggest.

Calacanis predicts a Democratic sweep, arguing that dissatisfaction with Trump among different voter groups could affect turnout. He also speculates about investigations, impeachment and the next presidential race. These are predictions about events that had not occurred at the time of recording.

He expects a period of political conflict if Democrats gain control and suggests Republicans may eventually seek different leaders. Palihapitiya and Sacks discuss the possibility of congressional investigations. The panelists differ in their views of both the likely outcome and its consequences.

Sacks argues that if Democrats win Congress, voters will see policies he opposes, while Calacanis wonders who might emerge as future leaders in each party. The hosts close by discussing their planned election coverage and acknowledging that the actual outcome remains ahead of them.