Frontier AI Labs Face an Enterprise ROI Test

The immediate question for Anthropic and OpenAI is not whether they have found demand. It is whether public investors will believe that demand can remain economically durable as customers’ token bills rise, cheaper models improve, and enterprises learn to route work more selectively.

Chamath Palihapitiya framed the risk in terms of customer economics. He said his CTO had reported that token costs at his company were doubling every 45 days while downstream productivity gains were perhaps 5% at most. The explanation, as Palihapitiya relayed it, was that further performance gains were requiring substantially more tokens because the company had already reached an effective plateau.

Palihapitiya did not argue that enterprises will stop using AI. His concern was that the current willingness to spend may reflect an experimental period in which companies are still astonished by the technology and have not yet been forced to defend the spend in earnings terms. Once CFOs, boards, and public-market investors demand a clear return, he said, enterprise AI revenue could prove more brittle than consumer revenue. Large enterprise buyers are fewer, more demanding, and capable of reducing expenditure if the savings or incremental revenue do not materialize.

He offered a rough, self-described attempt to test that proposition against public-company earnings. Palihapitiya said Claude 3.5 Sonnet initially told him that AI accounted for 50% of S&P 500 earnings-per-share growth since 2024. He rejected that answer because it included AI suppliers’ gains, including chip sales. When he instead asked about the “S&P 493,” excluding the largest AI-linked companies, he said the answer was 9% EPS growth. He attributed most of that growth to pricing power on top of inflation and buybacks, and estimated that direct AI ROI in publicly available data was between zero and 2%.

That estimate was not presented as a formal study. It was the basis for Palihapitiya’s broader warning: frontier labs may be enjoying extraordinary revenue while their customers have not yet demonstrated commensurate financial gains.

Brad Gerstner agreed that much enterprise AI spending remains experimental and may not have a directly measurable ROI today. But he treated that as a timing issue, not evidence against the market. His argument was that the addressable market is not a conventional software category. It is every worker, function, and company that uses intelligence to make decisions, produce work, or design products.

Gerstner said the relevant evidence is that millions of customers are independently choosing frontier models despite cheaper alternatives. In his account, businesses will increasingly use AI for more than labor substitution: life-sciences research, product innovation, and engineering work where a marginal intelligence advantage can matter more than the cost of inference. He cited Nvidia’s use of AI to design future generations of chips as an example of the “machine building the machine.” A company in that position, he argued, cannot simply choose an inferior model because it costs less.

The disagreement is therefore not over whether lower token prices will change behavior. They almost certainly will. It is over whether lower prices will commoditize intelligence or unlock more demanding workloads that preserve the value of the best systems.

Gerstner made the most aggressive version of the frontier-lab bull case. If a company exits the year at more than $100 billion in revenue, he said, it could potentially grow three to five times again the following year. He acknowledged that adding $200 billion in incremental annual revenue would be unprecedented. His explanation was that AI reaches across the organization rather than remaining confined to one department, as earlier software products often were.

Palihapitiya’s practical inference was nearly the opposite. If public markets are willing to finance these businesses at extraordinary valuations before the ROI question becomes a constraint, he suggested, the companies should list while that window remains open.

The panel treated SpaceX’s IPO as relevant not because it settled the AI valuation debate, but because it showed that public markets could absorb an offering at a scale once thought implausible.

Gerstner said SpaceX raised $75 billion at a $1.75 trillion valuation and subsequently traded around $2 trillion, on roughly $35 billion of forward revenue. He described the offering as a successful template for a company too large and economically significant to fit neatly within conventional IPO assumptions.

The mechanics he highlighted included a large raise, staged lockups, and early index inclusion. Index inclusion raised a legitimate concern, Gerstner said, because newly public companies can experience sharp post-IPO volatility. He cited a typical peak-to-trough drawdown of roughly 50% within the first six months. The risk is that passive investors are pushed into buying at an elevated point before the market has had time to establish a durable price.

But Gerstner argued that existing rules had largely been designed for younger, smaller, less-tested issuers. In SpaceX’s case, the argument for inclusion was that the company was already too large and important to remain outside major indexes for long. In his view, the offering’s performance made it easier for Anthropic and OpenAI to consider comparable approaches.

He said Anthropic was rumored to be trending above $100 billion in revenue by year-end and OpenAI around $70 billion. Neither figure was presented as confirmed. Gerstner’s point was comparative: even the lower OpenAI estimate would be roughly twice the forward-revenue figure he had cited for SpaceX.

Gerstner said he had heard that Anthropic would like to list this year and that OpenAI may face more complexity because of its corporate restructuring. He would be surprised if OpenAI listed before Anthropic, though he said he did not know the companies’ actual timetable. He added that Altimeter would be a buyer “at scale and at size” in either IPO.

A valuation above $1 trillion, Gerstner argued, should not be understood as a retail-investor shortcut to an immediate double.

The investment case, as he described it, would instead be continued compounding from companies already operating at enormous scale. Gerstner expects their revenues to grow more than 30% annually for years, but he did not expect them to be priced so cheaply that investors would reliably receive an immediate 50% to 100% gain after listing.

The operational pattern described across Uber, DoorDash, Coinbase, Databricks, and Decagon is consistent: sophisticated adopters are trying to reserve expensive frontier models for difficult or high-consequence work while shifting routine, well-defined workloads toward cheaper models. But routing is not the same as model fungibility. Context, memory, workflow history, tool integrations, and evaluation systems can make substitution technically difficult.

Jason Calacanis described AI as “intelligence on demand,” a product that reaches every function rather than one specialized department. That breadth helps explain the revenue ramp, he argued. Employees can adopt a low-cost subscription or API service without a major procurement cycle, and organizations can test AI simultaneously in engineering, finance, legal, operations, marketing, HR, procurement, and customer support.

Uber provided the clearest example of an organization trying to turn broad adoption into measurable operating change. Praveen Neppalli, Uber’s CTO, said 99% of Uber engineers use AI tools, more than 70% of pull requests are attributed to local or cloud agents, and engineers have built more than 2,500 agent skills across the software-development lifecycle.

Uber’s answer was to create “Agentic Pods”: AI-proficient engineers paired with domain experts in particular business functions. The engineers shadow the work, document actual workflows, identify opportunities, build agents alongside the people doing the job, test whether the tools generalize, and then ship.

Palihapitiya’s response was that operational gains should eventually be translated into earnings impact. Time savings, improved workflows, and employee adoption may be meaningful, but they do not by themselves establish that an expanding token budget creates returns above the company’s cost of capital.

Calacanis argued that falling inference costs complicate the calculation because lower unit costs can produce far more usage. He described reducing his own token costs by 95% through a configuration using OpenRouter, GLM 5.2, and other models. Instead of running agents daily, he began running them hourly; instead of assigning a single task to one agent, he split work among several. One use case was a recurring system that scanned All-In and This Week in Startups archives for emerging technology trends.

Gerstner characterized that behavior as an instance of Jevons paradox: lower costs can increase total consumption. He said token prices had fallen roughly 90% in each of the previous two and a half years. If the pattern continues, the relevant question may not be whether companies spend less on AI, but what new work becomes viable when intelligence is cheaper.

The companies shown in the discussion were not simply imposing usage caps. They were building systems to direct tasks toward the cheapest model that meets the required quality threshold. DoorDash CTO Andy Fang said the company used internal benchmarks to put the hardest code-review work through Anthropic’s Fable and lower-level work through Kimi K2.6. His summary was “better quality, cheaper cost.” Coinbase CEO Brian Armstrong said the company was experimenting with open-weight defaults, routing, and caching rather than friction and usage caps; he said 91% of employees had not been reaching their existing caps.

David Sacks said enterprises plainly want that outcome. They want to restrain rapidly rising token costs, reduce dependence on a frontier provider, and avoid giving a lab access to strategically sensitive information that could eventually become competitive intelligence. They also want some degree of AI sovereignty.

Yet, Sacks argued, the average enterprise cannot easily build the required middleware. Coinbase and DoorDash were presented as unusual because they had the technical capacity to route tasks across models. Even then, a routing layer does not solve the problem if an agent’s useful context, memory, tools, and workflow history remain attached to the original model stack.

Sacks summarized the gap between enterprise intent and capability as “the spirit is willing but the flesh is weak.” Organizations may want to diversify away from closed frontier models, but many cannot yet deploy a reliable hybrid architecture.

Databricks CEO Ali Ghodsi’s reported findings extended the point beyond model selection. Ghodsi said that, across Databricks’ own tasks, codebase, and infrastructure, the choice of harness around the same model could reduce costs by roughly 2x while retaining quality. He said Databricks used Omnigent to multiplex different models and harnesses for different tasks.

For Calacanis, the harness is part of the emerging operational layer: prompts, skills, memory systems, tool calls, evaluation methods, and routing policies. These are mostly hidden in consumer products such as Claude or Perplexity. More technically capable organizations are beginning to expose, test, and optimize them.

The practical result is a division by use-case maturity. Sacks cited Jesse Zhang of Decagon, which said it runs roughly 90% of its workloads on open-source models. In customer support, latency matters: a model that takes eight seconds on each conversational turn is not a viable product. Once a company knows the distribution of inputs, desired behaviors, and failure modes, it can fine-tune a smaller and faster model for the task.

At the discovery stage, however, Sacks argued that enterprises still need the most capable general-purpose system available. They do not yet know the shape of the problem or which workflow will prove valuable. Frontier intelligence may be expensive, but it can be the appropriate tool for discovering what the specialized system should eventually do.

This distinction also complicates revenue-share data. Gerstner cited a figure showing open-source models falling from 19% to 11% of enterprise LLM spending over the prior year. Sacks and Calacanis noted that self-hosted open-weight use may not appear as model revenue at all. It may surface as compute, cloud, or hardware spending instead. Paid frontier-model tokens are more visible in vendor revenue.

Gerstner nonetheless treated the current spending distribution as evidence that quality still earns a premium. A document summary may take 20,000 cheap tokens and can often be routed to a lower-cost model. A long-running agent doing the equivalent of two hours of work by a $200-an-hour consultant may require 2 million tokens. If it fails after consuming the context and compute, the cost of redoing the work may overwhelm the savings from selecting a cheaper model.

Palihapitiya expects a “good enough” threshold to emerge. He compared it to consumers eventually keeping an older iPhone because the incremental value of upgrading no longer feels obvious. In an earnings miss, he argued, executives may find it easier to cut AI costs than to reduce headcount. Governments building sovereign stacks may also decide that a domestic system at 95% or 99% of frontier performance is sufficient.

Gerstner accepted that open models, sovereign stacks, commodity inference, and premium closed systems will coexist. His disagreement is with the assumption that model intelligence will necessarily converge. If leading labs gain more revenue, use it to acquire more compute, and use that compute to build stronger models, the advantage could become self-reinforcing as agents take on more complex tasks.

Sacks said the visible revenue market already resembles a duopoly led by Anthropic and OpenAI. Calacanis countered that open-source tokens are “dark” in revenue data. The unresolved commercial question is whether specialized systems will gradually erode the frontier premium, or whether the complexity of deploying agents will make the best integrated model stacks more valuable.

The panel’s view was that the open-versus-closed model debate is no longer only commercial. Governments increasingly see AI systems as strategic infrastructure, and they may accept less capable domestic models in exchange for control over data, deployment, and national security risk.

Reuters reported that Chinese authorities had met with Alibaba, ByteDance, and Z.ai about potentially restricting overseas access to China’s most advanced AI models, including models not yet released. The report said officials were also considering national-security consequences for AI leaks or theft and possible restrictions on funding domestic AI startups.

David Sacks cautioned that the report did not establish that every Chinese open model would become closed. But he argued that a shift toward proprietary access follows an understandable commercial sequence. A lab that trails the frontier may have little chance of selling a weaker closed model, while releasing weights can attract developers and improve the system’s utility. Once the lab approaches frontier performance, the incentive to capture the value of the model becomes much stronger.

Sacks described ByteDance’s leading model as already closed and said Alibaba’s Qwen and Zhipu’s GLM line appeared to be moving toward proprietary offerings after earlier openness. He compared that pattern to OpenAI’s own evolution from a nonprofit-oriented, open project into a proprietary commercial provider.

Brad Gerstner put the issue in national-security terms. After meetings in Washington, he said he saw broad agreement in the White House and Treasury that the United States must remain ahead of China. He argued that constraining U.S. frontier labs while allowing Chinese models to circulate freely would be strategically incoherent.

Gerstner also said GLM 5.2 had watermarks from Mythos, which he viewed as evidence of model distillation, and he argued that the U.S. government should take action against it. Sacks agreed that the administration’s stated objective is to win the AI race, but worried that lower-level agencies or Congress could still make blunt policy choices in response to political pressure around safety, jobs, or AGI.

Palihapitiya said sovereign demand extends far beyond the U.S.-China rivalry. Following his participation in a United Nations AI commission with Marc Benioff, Jensen Huang, Brad Smith, and others, he said every country was trying to form its own AI strategy. Many countries, he said, do not consider dependence on a closed American model an acceptable long-term arrangement.

He cited the UAE’s Technology Innovation Institute and Falcon model, Saudi efforts around Arabic-language models, and Japan’s reported $6 billion investment in the Neo Terra consortium around physical AI and robotics. The appeal is not necessarily superior performance. A government may prefer a system that is somewhat less capable but domestically operated and more controllable.

The panel also contended that software policy will ultimately run into physical limits. Palihapitiya said an analysis by his team found the United States could be short roughly “three entire Californias” of electricity demand by 2050, even before fully accounting for AI data-center growth. He said the estimate included ordinary load growth from devices, cars, refrigerators, televisions, and computers.

Energy generation, transmission, storage, and permitting therefore constrain every AI deployment strategy the speakers discussed. Calacanis added Taiwan’s energy vulnerability to the strategic picture, saying the island’s dependence on liquefied natural gas could become acute during a blockade. More models, more inference, and more data centers all require more power.

Trump Accounts were presented as a vehicle for making tax-advantaged ownership available from birth, funded through a combination of government seed money, family contributions, employers, states, and philanthropy.

Brad Gerstner described the accounts as the result of a four-year effort to create privately owned investment accounts for children. As he described the program, an account is created after a child receives a Social Security number. The initial $1,000 is invested in the S&P 500, and the account is free to the recipient for life. The enabling legislation is the Invest America Act, although the accounts are officially called Trump Accounts.

Gerstner’s basic illustration begins with the $1,000 seed, additional matching money, and $10 per week in saving. He said that could produce $50,000 by age 18. The larger objective, in his telling, is to establish a visible asset early enough for compounding to matter.

The app launched on July 4, according to Gerstner. He said more than 1.5 million accounts were created in its first 24 hours and deposits exceeded $1 billion. A screen shown during the program placed Trump Accounts: Official App first among free App Store downloads. The app includes a QR code intended to let relatives and friends deposit directly into a child’s account.

The program’s scale depends materially on philanthropic funding. Gerstner cited more than $6 billion from Michael and Susan Dell, intended to provide $250 to each of 25 million children, primarily from lower- and middle-income families. He also cited $350 million in SpaceX shares from Gwynne Shotwell for lower-income communities, $250 million from Micron for employee-related contributions, and his own $100 million commitment for children under five in Indiana.

Donors, Gerstner said, can direct gifts by zip code and age. He said organizers had told President Trump they believed they could raise $100 billion in the first 12 months. If donations exceed immediate per-child limits, he said, pooled funds could be allocated over time to future birth cohorts.

David Sacks focused on the mechanics that he thought made the accounts unusually valuable to families. As Sacks described the rules, friends and family can contribute up to $5,000 annually to a child before age 18, while employers can contribute up to $2,500 tax-free. He argued that the accounts offer tax-advantaged compounding before a child has earned income, when ordinary IRA participation is generally unavailable.

Sacks described a possible strategy in which an account compounds until adulthood and is then converted into a Roth IRA after the recipient is no longer a dependent and has little taxable income. He said the conversion would need to be managed carefully to avoid the kiddie tax, but cited CPA commentary describing the structure as a “backdoor Roth for children.”

The speakers described the account as more constrained than a conventional cash transfer. Gerstner said he had initially wanted recipients to wait until roughly age 30 for full access, but the eventual political compromise was that people old enough to vote or serve in the military should control their own money. As he described the eventual restrictions, recipients can withdraw up to 25% for specified purposes including a home purchase, starting a business, or college. The remainder rolls into an IRA, where early withdrawals face ordinary penalties.

The age restriction leaves a practical concern: whether recipients will spend the money badly at 18. Gerstner’s answer was financial education. A child who can see a balance and ownership interests in companies such as Nvidia, Apple, Nike, or Microsoft, he argued, has a reason to learn how investing works. He said 37 states require financial literacy and that about 25 states were considering putting money into children’s accounts through existing programs.

Chamath Palihapitiya raised the branding risk. Some parents, he said, may reject an account because it is called a Trump Account; social posts shown in the program included calls to distrust or close the accounts. Gerstner said he was seeing participation across political, geographic, and income lines. He cited Cory Booker, Gavin Newsom, Wes Moore, and John Fetterman as Democratic supporters.

Gerstner’s political argument was that direct ownership should be preferred to government-directed redistribution. The account is meant to be privately held, visible to the recipient, and funded directly rather than through an intermediary determining how resources are allocated. Sacks made the related case that the system is a new channel for family savings, employer contributions, and philanthropy, rather than a conventional transfer program.

The comparison with Social Security was central to Gerstner’s case. He argued that Social Security contributions do not create personally titled or inheritable assets in the same way. Trump Accounts, by contrast, are intended to be owned by the recipient from the beginning.

Gerstner said Americans typically miss that period because meaningful saving often begins only in their mid-twenties. His projections of $50,000 or $200,000 did not assume that families maximize annual contributions, he said; they assumed more modest additions such as $50 per month. His stated focus was people who would otherwise begin adulthood with no savings at all.

He said the model could eventually be extended to adults as a supplement, while calling Social Security a “sacred promise” that neither party intends to remove. The decision-useful question is less whether the program can replace existing retirement systems than whether its contribution rules, employer participation, and philanthropic pipeline can reach enough children early enough to produce meaningful assets at adulthood.