Anthropic’s IPO Would Test Whether AI Token Demand Can Fund Compute

The central question is not whether AI models are improving. It is whether customers will continue buying enough intelligence, at high enough prices, to support an infrastructure buildout increasingly measured in hundreds of billions of dollars.

That dependency runs through the reported Anthropic IPO, competition from cheaper models, Nvidia’s financing effort, and worries about a compute glut. If token demand remains strong, capital can move from end users to model providers, compute hosts, chipmakers, manufacturers, and energy suppliers. If demand weakens materially, the same chain can transmit a slowdown quickly.

Jason Calacanis cited a Financial Times report that Anthropic was targeting a $2 trillion valuation in an October IPO. He put its expected year-end annualized revenue at $100 billion to $120 billion, a roughly tenfold increase over the preceding year. An on-screen Polymarket market assigned an 80% probability to Anthropic going public before 2027; another put its chance of holding the best AI model at year-end at 67%.

At a $2 trillion valuation, Calacanis calculated, Anthropic would trade at roughly 16 to 20 times sales. More consequential than the multiple was the reported scale of the revenue ramp. He described it as unprecedented in Silicon Valley.

Gavin Baker cautioned that the $2 trillion figure may be IPO positioning rather than a settled valuation. In his account, banks that lose a prized “lead-left” underwriting role have an incentive to leak aggressive valuation targets, making the winning bank look as if it is bringing a company public too cheaply. If Anthropic actually priced at $2 trillion, he said, it would indicate exceptionally strong testing-the-waters meetings and roadshow demand, while leaving room for the shares to trade higher after listing.

Baker nevertheless called the reported numbers exceptional. His more consequential observation was that Anthropic could be losing marginal share to OpenAI, Grok, and open models while still expanding at extraordinary rates. The market itself may be growing faster than shifts in share can constrain the leading providers.

Anthropic’s value, in Baker’s account, does not rest entirely on model leadership. Claude’s product experience, surrounding tools, and user familiarity have independent value. He invoked a thought experiment by Eric Vishria: Anthropic and OpenAI could remain valuable even after losing the model layer because customers have relationships with products and workflows, not merely with raw model capability.

David Sacks treated the growth rate as a stress test for the entire industry. A company that ends the year at a $100 billion run rate and sustains tenfold annual growth would reach $1 trillion in annualized recurring revenue the following year. That raises two separate questions: whether the market for AI is truly that large, and whether enough compute and electricity can be built to serve it.

Sacks argued that demand can continue rising as agents enter more work contexts. Anthropic’s early emphasis on coding was central to his explanation. Calacanis said Anthropic saw the intensity of Cursor usage and moved vertically into coding; Sacks called the move highly successful and said OpenAI had subsequently shifted more heavily toward coding. He said he had heard that OpenAI’s growth rate had accelerated to more than 20% month over month over the preceding two months after that push.

Baker framed the addressable market as roughly $25 trillion to $65 trillion of knowledge work, depending on how it is counted. AI revenue could come from replacing labor or increasing the productive capacity of people and companies. He preferred the second outcome and said current evidence pointed more in that direction: there were more software-coding openings than a year earlier, he said, even if young people entering the labor market may be facing a negative effect.

The argument gets more difficult once price enters. Calacanis expects Anthropic’s growth rate to slow substantially as startups and enterprises adopt cheaper systems, including open models. A pricing comparison on screen showed a substantial gap between GLM 5.2 and Claude Opus 4.6.

Calacanis characterized the difference as roughly 90%. He argued that corporate buyers will increasingly choose open models because of cost, control, and a desire to avoid dependence on a single technology provider. Open systems can be harder to deploy, he acknowledged, but he compared their likely adoption to enterprises’ embrace of open-source infrastructure over the past two decades.

Sacks offered a more segmented view. Much of the market may choose lower-cost intelligence, as much of the smartphone market chooses Android over Apple. But a meaningful group of customers will pay heavily for the best available capability and product experience if it provides a competitive edge. Anthropic and OpenAI can sustain that premium only by remaining sufficiently far ahead of cheaper alternatives.

The group’s forecasts converged well below the mechanical $1 trillion extrapolation. Calacanis suggested Anthropic could reach roughly $300 billion to $400 billion of annualized revenue in the following year. Baker and Sacks each described $400 billion to $500 billion of exit ARR as realistic. Sacks called $500 billion the practical over-under, while Baker said he would take the over on Wall Street’s eventual 2027 consensus by 25% to 30%, assuming regulation did not slow the company.

Even that outcome would be extraordinary. Calacanis noted that the market had been discussing Anthropic at roughly $10 billion of annualized revenue only ten months earlier. For Sacks, the importance of an IPO is not simply the valuation. Public quarterly earnings would turn Anthropic into the clearest recurring signal of whether end-user demand is supporting the AI buildout.

Sacks described Anthropic as the current “pace car” for an industry whose participants are making large commitments based on future token demand. He said Anthropic could generate about $100 billion per gigawatt of compute, pay SpaceX roughly $50 billion per gigawatt at spot prices, and leave SpaceX able to spend around $30 billion with Nvidia for chips. Nvidia’s suppliers would benefit further down the chain. Anthropic’s earnings, he argued, would reveal whether enough end-user demand exists to keep that sequence moving.

Baker agreed with the conditional risk. If Anthropic pulled back because customers did not want the tokens, he said, there would be a pileup because the industry has not left enough distance between the cars. If Anthropic instead lost share to OpenAI, SpaceX, or an American open-source competitor, the compute could be redeployed elsewhere and the consequences would be different.

Baker also said his reading of public signals led him to believe Anthropic may already be generating cash and could be profitable. He said he had not spoken with Anthropic, but argued that an S-1 could challenge investors who assume token use is fundamentally subsidized rather than economic.

AI infrastructure is becoming a financing problem as much as a capital-expenditure problem. The key question is who can borrow against the expected income from a GPU cluster, on what terms, and with what confidence that the hardware will retain value after newer chips arrive.

Nvidia said it had partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent financing platforms intended to mobilize more than $500 billion of third-party capital for AI infrastructure over time.

Calacanis described the intended arrangement as asset finance. Instead of paying billions upfront to buy Nvidia systems, a compute operator borrows, acquires the systems, rents the resulting capacity, and services the debt with the cash flow. The proposition is that GPUs can be income-producing assets rather than technology purchases that quickly become obsolete.

Gavin Baker took the participation of major lenders and private-equity firms as a signal that Nvidia compute can be underwritten. In his view, those firms would not attach their reputations to the effort unless they believed GPUs had adequate flexibility and useful life to support financing.

That flexibility matters because model architectures are changing in divergent directions. Baker named Qwen, Kimi, DeepSeek, and GLM as major Chinese open models whose architectures are evolving differently. That variety favors general-purpose GPUs, he argued, because lenders need confidence that the underlying equipment can serve a range of future workloads rather than one narrow model design.

Baker expected the structures to include Nvidia residual-value guarantees: a floor for what GPUs can earn after three or four years, which would lower lender risk and financing costs. Nvidia could then receive revenue shares above a higher threshold. His interpretation was that Nvidia would take a limited piece of downside risk while creating a capital-light, royalty-like claim on successful deployments.

David Sacks saw the strategic objective more broadly. Financing itself may constrain Nvidia’s addressable market. The industry could have genuine demand for compute that customers cannot fund from their own balance sheets. By helping create a deeper market for credit, Nvidia can sell to customers whose compute needs exceed their immediate capital resources.

Sacks used Elon Musk’s stated ambition to add six to eight gigawatts of capacity in the following year as an illustration. He said that kind of expansion could require roughly $300 billion to $400 billion of capital expenditure, while the company had raised $100 billion in equity and debt. Nvidia-facilitated financing could help bridge that gap. He argued that this was not circular financing of the sort critics fear: banks and private-equity firms would be lending against expected compute cash flows, rather than Nvidia simply financing customers to purchase Nvidia products.

The analogy was aircraft finance. An airline can borrow against aircraft not only because of its own creditworthiness but because the planes retain resale and operating value if the airline fails. GPU fleets could become comparable collateral if systems, specifications, and operating assumptions are standardized enough for lenders to value and eventually securitize them.

Useful life is therefore central. Baker cited CoreWeave’s statement that it was renting Ampere GPUs profitably for 2029, implying an economic life of about nine years for hardware introduced in 2020. Older chips do not have to become worthless when newer generations arrive; they can shift to less demanding workloads where cost matters more than frontier performance.

The principal risk, Sacks argued, is a compute glut. If too many operators build simultaneously, the result could resemble the dark fiber left after the dot-com crash: underused infrastructure financed on optimistic assumptions about utilization and price. “Dark GPUs” would reduce spot rates and damage everyone who had built for higher returns.

In a counterintuitive turn, he suggested that political and physical obstacles to building data centers may protect the sector against that outcome. Permitting delays, construction difficulty, power constraints, and local opposition all slow deployment. Those headwinds could keep supply constrained relative to what he expects to be fast-rising demand.

The physical constraints are real regardless of the policy fight around them. Baker described data-center construction as an industrial operation involving land, grid connections, turbine supply, fuel, equipment, local approvals, and thousands of workers in difficult locations. Calacanis reduced the point to a phrase: “It’s atoms, not bits.”

Baker said Caterpillar, Cummins, GE Vernova, and Siemens Energy were expanding capacity. He also described what he said was a growing business in removing engines from older aircraft and repurposing them as turbines for data centers. Natural gas, in the group’s view, is the practical near-term energy source because it can be deployed faster than renewable generation at the required scale. Baker said the world may eventually run on sunlight, but not immediately; Sacks likewise argued that new data centers would largely be gas-powered in the next year or two.

The panel rejected much of the broader backlash against data centers. Baker argued that popular claims about water consumption and electricity prices are wrong, and cited a Wall Street Journal account of Ellendale, where a data center was said to have revitalized a declining town and increased tax revenue tenfold. Sacks argued that self-powered facilities can sell excess electricity to the grid and pay for grid upgrades.

Calacanis identified impacts he considered legitimate: generator noise, air pollution from gas generation, and localized heat where infrastructure is concentrated. His answer was siting and monitoring rather than a blanket ban. Texas’s reported energy-audit approach fit that view: projects bringing their own power could proceed, while projects drawing on the grid would need to show how they avoid destabilizing it.

Nvidia’s initiative addresses the financing constraint. It cannot eliminate the constraints of power, construction, permitting, or the underlying requirement that compute must continue producing revenue.

The competitive challenge to Anthropic’s premium pricing is not one rival. It is a market structure in which open models, lower-cost proprietary systems, and a stronger Grok can take workloads that otherwise might flow to leading closed providers.

Gavin Baker argued that this need not be a zero-sum outcome. Open models can increase the value of frontier systems if the strongest model becomes an orchestrator: it delegates routine tasks to cheaper intelligence and reserves difficult judgment, coordination, and high-value work for itself.

He offered an illustrative hierarchy rather than a forecast. If frontier systems operate at something like “200 or 250 IQ” while lower-cost open systems are at “150 IQ,” the frontier system becomes more useful because it can direct those cheaper systems. Baker compared the arrangement to the Manhattan Project: a small group of extraordinary physicists still required thousands of other highly capable contributors.

His proposed market structure was that frontier tokens could account for 65% to 85% of economic value while open-source tokens account for around 80% of volume. Low-cost intelligence expands the range of viable applications, in this account, and creates more demand for the premium coordination layer.

Sacks made a compatible case for Anthropic. Open models do not prove that its prices must collapse. Customers operating in competitive markets may pay a premium for the strongest available intelligence if it yields an edge. But leading labs remain on a treadmill: their pricing power depends on retaining enough of a frontier lead to justify that premium.

Grok 4.6 was presented as evidence that the frontier may be more fluid than an Anthropic–OpenAI duopoly. Two on-screen benchmark snapshots placed Grok 4.6 on or beyond the displayed price-quality frontier. One chart, labeled Cursor Bench 3.2, compared model-family score and cost per task. The other, labeled Databricks Office QA Pro v2, compared cost and quality. Both charts presented Grok 4.6 as higher quality at lower or comparable cost than leading alternatives.

Baker noted that the Cursor-related chart came from an ecosystem closely connected to xAI and SpaceX, then placed more weight on the Databricks comparison. He also cited early favorable reactions, including one from David Heinemeier Hansson of 37signals and Basecamp. His conclusion was provisional: early user response appeared good, and the displayed benchmark results were beginning to bear out in use.

Calacanis attributed Grok’s perceived improvement to organizational changes: experienced SpaceX operators, the Cursor team, and a management reshuffle at xAI. Baker agreed that those changes mattered. He described Grok 4.6 as still relatively small for a frontier system at 1.5 trillion parameters, and expected a larger Grok 4.7 within weeks.

Grok Bot mattered for a different reason: personalization and usability. Baker described it as a step toward making AI more accessible to people who are not coders. Calacanis connected it to Meta’s ambition to create agents for everybody. The model race is not just about benchmark scores; it is also about which firms can package capability into tools that ordinary people can use.

Sacks described SpaceX’s AI position as a high-ceiling, high-floor strategy. Building large compute clusters provides the upside option to train a genuine frontier model. If xAI succeeds, it competes directly with Anthropic and OpenAI. If it fails to lead at the model layer, the capacity can still be rented to frontier labs.

Calacanis said reports indicated that SpaceX retained a right to reclaim compute from Anthropic, while Sacks said both sides had a 90-day cancellation right. That leaves the relationship sensitive to changes in spot pricing: SpaceX benefits if compute remains scarce and rates rise, while falling rates could create pressure to renegotiate.

Baker’s investor argument was that SpaceX is often analyzed through Starlink, direct-to-cell, orbital compute, and terrestrial compute without assigning much value to Grok. If a frontier model can scale revenue at something like Anthropic’s reported pace, he argued, then omitting Grok leaves an important part of SpaceX’s future economics unmodeled.

That same competitive logic informed Baker’s comments on Workday. The reported Silver Lake interest in acquiring the company, alongside a 17% rise in its stock, suggested to him that public software valuations may have fallen too far on AI fears. His counterintuitive explanation was that open-source models can support software businesses by limiting dependence on a small number of expensive, centralized providers. A world dominated by only two or three frontier labs would be more threatening to software margins than one in which companies can choose among capable, lower-cost alternatives.

Mark Zuckerberg’s essay, The Future is for Everyone, set out three stated principles for a positive AI future: individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety.

Zuckerberg’s argument was not simply for open models. He opposed a singular centralized superintelligence. Rather than place superintelligence under one authority, he called for many people and businesses to have agents aligned to their own goals, checking and competing with one another.

David Sacks read the essay as a challenge to the worldview he associates with Anthropic and effective altruism. Zuckerberg asked why people who expect AI to eliminate jobs, reduce humanity’s relevance, and create severe safety risks would rush to build that future. Sacks saw two possible answers: financial incentives, or a belief that a dangerous technology can be made safe only under the control of unusually enlightened institutions.

Sacks associated the latter belief with Thomas Sowell’s The Vision of the Anointed: the idea that concentrated power in the right hands can engineer society in a benevolent direction. The danger, in his account, is not simply closed models. It is an arrangement in which a small group of companies and government actors jointly control access to AI.

Sacks said that when he entered Washington, he encountered a policy frame that treated AI as inherently dangerous and called for a small number of U.S. companies, close state involvement, restrictions on exporting chips and models, and something resembling an Atomic Energy Commission for AI. He referred to Marc Andreessen’s account of Biden officials discouraging competition and said he had heard similar ideas, while noting that others dispute Andreessen’s characterization of the meeting.

Gavin Baker reduced the divide to competing assessments of risk. As he characterized it, Anthropic and effective-altruist thinkers see AI as too dangerous to distribute. Zuckerberg, Elon Musk, Jensen Huang, and Baker see it as too dangerous to centralize.

Baker compared personal access to AI with the right to bear arms: an individual safeguard against tyranny. Sacks accepted the basic analogy while emphasizing that AI has much broader ordinary utility than a weapon. People may possess guns despite the possibility of misuse; AI, he argued, is fundamentally a consumer technology with creative and productive uses for businesses and individuals. It should not be treated primarily as if it were a nuclear weapon.

The geopolitical case sharpened the policy divide. Baker said the United States would lose the AI race and broader geopolitical ground to China if it bans or heavily restricts open-source AI. Sacks made a parallel commercial point: Anthropic’s ability to charge a premium rests on maintaining an advantage over lower-cost alternatives. An FAA- or FDA-style regime that slowed model releases for years would erode that advantage and allow competitors to close the gap.

That was the irony Sacks emphasized. Anthropic may benefit from customers willing to pay for frontier capability, but it also depends on rapid iteration. A regulatory structure designed to slow models down could weaken the business it is supposed to protect.

Zuckerberg’s essay also addressed job displacement. He wrote that there is no rule requiring automation to increase faster than individual capability or demand for new skills. Instead, people may gain the ability to do more new things before their jobs materially change, producing a healthier balance and potentially job growth. Calacanis saw that as another version of Jensen Huang’s formulation that workers may not lose jobs to AI itself, but to people who use AI better.

Calacanis identified a tension in Zuckerberg’s support for broad learning and model distillation. Zuckerberg argued that models should be able to learn from anything that can be observed, because all models derive from human knowledge and restrictions on distillation would impede U.S. open-source leadership. Calacanis responded that Meta has aggressively protected its own social graph and sued companies that tried to index or derive information from Facebook and Instagram. The stated principle, he argued, sits uneasily beside that history.

The dispute over Amazon’s Delivery Service Partner model brought the group’s political argument into a concrete labor fight. New Jersey had sued Amazon over its relationship with delivery contractors, while Zohran Mamdani supported municipal rules that would require Amazon to hire drivers directly.

Calacanis described DSPs as a potentially over-clever way for Amazon to externalize liability, labor obligations, and benefits costs while retaining operational control. In his view, the system leaves drivers under substantial pressure while allowing Amazon to say it is not their employer. The political risk is that large companies appear to shift costs they should bear onto the public, strengthening the socialist case against capitalism.

David Sacks took the opposite position. The contractor model, he argued, enables rapid scaling during holiday peaks, supports dense urban delivery routes, lowers costs, and creates small businesses. It also provides working arrangements that some drivers may prefer. The system evolved because Amazon, DSP owners, and workers all found it useful, he said. His answer was freedom of contract and voluntary exchange rather than municipal intervention.

Calacanis did not call for eliminating contractors entirely. He argued that Amazon should employ more of its core, full-shift delivery workforce while retaining DSPs for surge capacity. He estimated that higher pay and benefits could cost roughly 25 cents per delivery in New York and argued that Amazon could absorb the expense.

Sacks countered with an Amazon-funded study stating that the proposed regulations could raise delivery costs by as much as $5.20 per package, or $664 annually per household.

The disagreement was not about whether capitalism should survive. It was about what sustains it. Calacanis saw voluntary corporate self-correction as necessary to preserve capitalism’s legitimacy: Amazon could bring more frontline workers into the company, provide benefits and equity, and avoid an unnecessary political backlash. Sacks saw government interference with functioning market arrangements as the more direct threat, with consumers ultimately bearing the cost of regulation through higher delivery prices.