Over the past six months, the defining characteristic of the AI trade has not been valuation, it has been surprise. Between December and early June, investors consistently underestimated the speed at which artificial intelligence was translating into revenue growth, capital spending, and enterprise adoption. The Q1 earnings season forced analysts to repeatedly raise estimates, while investors who remained underweight found themselves chasing a market that had already moved. Fear of missing out became the dominant force behind the AI rally.
That phase now appears to be behind us. Expectations have risen dramatically, investors are heavily positioned for those expectations, and many of the strongest companies are no longer being rewarded simply for producing excellent results. Instead, markets have entered a more balanced environment where exceptional execution is increasingly expected rather than celebrated. This does not suggest that the AI thesis is weakening. It means the market is transitioning from a period driven by earnings surprises to one driven by valuation support and the durability of future exponential growth in a technology the world has never seen, backed by a capital expenditure dollar amount never before attempted, but one for which everyone is now positioned.
As that transition has taken place, the conversation around AI has changed just as quickly. The debate is no longer centered on whether demand exists. Instead, investors, now positioned and at risk, are navigating an anxious period focused on falling token prices, rising enterprise costs, open-source competition, AI sovereignty, and whether frontier models will ultimately become commoditized. Every major technological cycle develops new reasons for skepticism once expectations catch up to reality, and this cycle is proving no different.
Sitting in a seat where, despite warning about an AI mid-cycle slowdown, I have had to listen to this anxiety rise over the last couple of weeks, I found Brad Gerstner’s recent appearance on the All-In Podcast remarkably useful for bringing clarity to these worries. It was not his explicit intent to play the role of market stabilizer, and there were no engineered “hot takes” designed to go viral on X, but his responses to these core market fears were exceptional. I highly recommend listening. They were presented from the perspective of a long-term capital allocator trying to understand the underlying economics of frontier intelligence. Rather than simply summarizing the conversation, this framework uses his insights to answer the most critical questions investors are asking about the next phase of the AI cycle.
Question One: Why could frontier AI labs justify multi-trillion-dollar valuations?
The private-market model for the most important technology companies is undergoing a structural transition. Frontier AI labs are reaching a scale at which traditional late-stage venture financing may no longer provide enough capital, liquidity, or price discovery. Public markets are increasingly becoming the only pools of capital large enough to support their infrastructure requirements and provide liquidity to employees and early investors.
Gerstner argued that successful mega-cap offerings are creating a blueprint for companies such as OpenAI and Anthropic. His point was not simply that these companies could go public, but that institutional investors are prepared to absorb them at valuations that would once have appeared implausible based on standard software metrics. Specifically, Gerstner pointed to the textbook execution of the SpaceX public offering as a template that has deeply informed the frontier labs regarding structured lockups, pricing liquidity, and index inclusion:
“The SpaceX IPO… was textbook. It was a hugely successful IPO… I think Anthropic and OpenAI were watching very closely because frankly we had not had an IPO of that size… I think they’ve gone to school.”
When asked point-blank whether public markets would support multitrillion-dollar clearings for OpenAI and Anthropic, Gerstner explicitly committed his own capital:
“Today as I sit here today Altimeter would be a buyer at scale and at size in both of those IPOs at three trillion.”
The relevant question is whether investors should compare these businesses with historical software IPOs or with platforms that could become foundational layers of the global economy. If frontier intelligence becomes embedded in software development, healthcare, financial services, legal work, scientific research, education, and consumer activity, then the addressable market is not limited to traditional software budgets. It begins to include a portion of the value currently captured by human labor, professional services, and corporate decision-making.
Question Two: Is the addressable market really large enough to support those valuations?
“Intelligence is the largest TAM we’ve ever seen in the history of the world.”
Traditional software models are built around licenses, seats, departments, and defined pools of information-technology spending. Frontier AI does not fit neatly within those categories because it is not merely another software tool. It is increasingly becoming an operational layer capable of performing cognitive work.
Gerstner described the addressable market as every small, medium, and large company in the world. That distinction matters because the revenue opportunity is not constrained to replacing an existing software product. AI can expand into tasks that companies currently perform through employees, consultants, contractors, outsourced service providers, and manual workflows. Gerstner emphasized this unique, unconstrained addressable market:
“The total addressable market here is every single small medium large company on the planet and so we’ve never seen revenue growth like this because we’ve never seen a TAM like this. And if you look at the distribution of revenues across these businesses, it’s not like it’s concentrated with four or five customers. There are millions of customers independently economically making the decision that is rational for them every day.”
This is why conventional discounted-cash-flow models can struggle with frontier AI. The challenge is not simply estimating market share within a known category. It is determining how much of global knowledge work can be converted into software-mediated intelligence. Revenue growth that appears unprecedented relative to the history of enterprise software may be more understandable when the market opportunity is measured against labor and economic output rather than software spending alone. Gerstner framed the blinding velocity of this expansion:
“Our minds were blown if a company could go from 100 million to 300 million. We’re talking from a 100 billion to 300 billion. 200 billion of incremental revenue is incomprehensible in the history of Silicon Valley… in the history of the world.”
Question Three: Do falling token prices weaken the economics of frontier models?
One of the most common concerns is that inference prices will continue to decline, pressuring revenue and causing enterprises to migrate toward cheaper models. That argument is valid for low-value and easily verified tasks. Document summarization, basic classification, simple retrieval, and routine workflow automation can often be handled by smaller or less expensive models.
The economics change for long-running, high-value workloads. When an AI system is writing production code, conducting financial analysis, reviewing legal documents, or completing a multistep engineering task, the cost of failure can be much greater than the cost of inference. A model that is slightly cheaper but more likely to make an error, lose context, or fail late in a workflow may produce a worse economic outcome. Gerstner mapped this out by defining the wide chasm between commodity workflows and premium agentic tasks:
“The premium workload… summarizing a document may take 20,000 cheap tokens to do—of course shoot that to a lagging model or an open source model. But if you’re talking about replacing a software engineer for 2 hours, that may take 2 million expensive tokens, and the consequence of using something that’s 95% as good is really high, right? Because you have a long-running task, and if the task breaks early or it breaks in the middle or breaks at the end, there’s a huge cost to that.”
This dynamic is currently being stress-tested by Meta. Recognizing the difficulty of competing purely on frontier capabilities, Mark Zuckerberg has shifted his game theory, aggressively using cost as a weapon against the leading labs. By releasing models like Muse Spark, Meta is actively attempting to commoditize the space by offering comparable performance at a fraction of the price. Gerstner observed this strategic pivot:
“The Meta thing was really intense because I thought, okay, you know we talked about the game theory which was Mark should scorch the earth with open source, I think they flubbed that play but then I think he has now said he’s going to create a price war… He was basically like, ‘Hey guys, I’m going to give you the same quality at like 1/100th of the cost.'”
However, a massive price cut alone is not enough to immediately threaten frontier revenue. Meta still faces significant structural hurdles in building out the complex enterprise distribution and middleware layers required to win over corporate clients.
Ultimately, Gerstner used the example of replacing a consultant who costs $200 per hour to prove why massive price cuts, like Meta’s, don’t necessarily alter enterprise math. In that setting, the difference between spending a few dollars on a cheaper model and $15 on a superior model is economically insignificant when the more expensive system produces a more reliable result. The relevant comparison is not expensive tokens versus cheap tokens; it is the total cost of the AI-generated outcome versus the cost, time, and reliability of the human alternative:
“If an AI agent is replacing a $200 an hour consultant… the difference between spending three bucks on a cheap model or 15 bucks on an expensive model to replace a $200 an hour consultant, it’s just irrelevance. That inference cost difference is irrelevance if you’re getting something that’s bulletproof for 15 bucks.”
Question Four: Will model capabilities eventually converge?
This stands out as one of the most critical segments of the discussion. The convergence argument assumes that intelligence behaves like a mature software feature. Once competitors reach a similar benchmark level, the market should commoditize and pricing power should disappear. That may prove true for certain classes of models and tasks, but it is not necessarily true at the frontier.
Frontier intelligence is not a fixed destination. The leading systems are being used to improve software development, generate synthetic data, support scientific discovery, and accelerate the research process itself. If the strongest models help their developers build even stronger models, then a small lead can become part of a recursive advantage. Gerstner presented this explicitly as the core non-consensus debate over market structure:
“There’s this implied assumption in the world that there’s going to be this convergence of intelligence… One non-consensus argument might be that intelligence is not converging at all, that super intelligence becomes fully self-recursive. And as it becomes recursive, you actually extend the lead because the smarter your model gets, the more revenue you get, the more compute you can buy, the more compute you can buy, the better the model is that you can build.”
Under that scenario, intelligence does not converge. This directly mirrors my own experience since late last year; today, 90% of my baseline daily usage is concentrated in ChatGPT and Claude. Throughout 2025, the competitive landscape shifted multiple times, and the top two models never commanded more than 50% or 60% of total workload distribution. Now, a consolidation is occurring at the tip of the spear.
The frontier can continue moving outward, allowing the most capable labs to preserve or even extend their lead in the highest-value workloads. Gerstner pointed out that this structural gap is already showing up in corporate financials, running entirely contrary to prevailing bearish narratives:
“People are speculating that the intelligence gap between that commodity stuff and the frontier stuff is going to collapse to the point that people won’t pay for the frontier stuff. There is no evidence of that on the field today.”
David Sacks strongly echoed this evidentiary point as well, noting that while enterprises technically want to diversify onto cheaper architectures, “the spirit is willing but the flesh is weak,” the technical barriers to hot-swapping models without losing memory, context, and refinement mean that wallet share continues to concentrate back into the frontier.
Question Five: Does open source undermine the frontier-lab investment thesis?
Open source and frontier intelligence are often presented as opposing outcomes, but the market is more likely to support both. Open models are particularly valuable when organizations require customization, local deployment, data control, lower operating costs, or independence from a third-party provider. Frontier models are more attractive when capability, reliability, ease of deployment, and time to value matter most.
This creates a dual-track market rather than a winner-take-all outcome. Closed frontier systems can dominate high-value workloads where performance is critical, while open models capture sovereign, regulated, localized, and cost-sensitive applications.
The important investment implication is that open-source adoption does not eliminate AI demand. It changes where the revenue appears. When a company runs an open model inside its own data center, the resulting token activity may never appear as revenue for a frontier lab. Gerstner referred to these as “dark tokens”:
“Keep in mind when you’re doing open source, those are dark tokens, those don’t come up as revenue… You don’t see that, it doesn’t come up as revenue, it comes up as free. The only thing you’re paying for there is the hosting cost.”
The usage remains economically real, but the spending is captured through GPUs, high-bandwidth memory, networking, power, cooling, data centers, and cloud infrastructure rather than through a commercial model API. The data shows that the rise of open-source deployment is not cannibalizing the revenue generation of the premier labs:
“Despite all of those arguments, and now we’re 18 months into this… the facts in the field are just the opposite. The share of economic value, the share of wallet, is actually increasing to the Frontier Labs while the share of tokens, these commodity tokens, is obviously going up to the other guys.”
Question Six: Is AI sovereignty a threat or an additional source of demand?
Governments, defense agencies, financial institutions, healthcare systems, and regulated enterprises may be unwilling to place their most sensitive data inside foreign-controlled or externally hosted models. These organizations increasingly want systems they can operate locally, customize, audit, and control.
This does not reduce the overall AI opportunity. It creates an additional layer of infrastructure demand. Sovereign AI requires domestic compute capacity, local data centers, secure networking, specialized models, energy supply, and national technical expertise. Countries that might otherwise have consumed intelligence through a small number of global APIs may instead build parallel AI infrastructure. Gerstner validated this reality, noting that the demand for sovereign architecture doesn’t cannibalize the frontier market, but rather runs parallel to it:
“I think that Chamath’s absolutely right about the sovereign stacks that are going to get built around the world. This is not either/or. We are going to have open source and we are going to have frontier intelligence.”
The result is less software centralization but potentially more aggregate capital spending. Sovereignty fragments the model layer while expanding the physical layer beneath it, creating parallel capital expenditure cycles across different geographies.
Question Seven: How does competition with China affect the investment cycle?
Artificial intelligence is no longer treated solely as a commercial technology. It has become central to national security, economic competitiveness, military capability, scientific leadership, and geopolitical influence. That reality creates a degree of policy alignment in Washington that is unusual in most areas of technology.
Gerstner, fresh off meetings with high-level officials in Washington, emphasized that political disagreements over regulation are secondary to the broader objective of maintaining a structural lead over China:
“Having spent some time in DC this week and talking with both the White House and Treasury, etc. on this topic, what I can tell you is while there may be some debates about regulation of US models, the one thing there’s absolute agreement on is doing everything to stay ahead of China… It is a unifying force in Washington.”
This strategic competition supports semiconductor controls, domestic manufacturing incentives, energy investment, data-center construction, and continued funding for frontier research.
The competition also extends beyond hardware. Model distillation allows smaller foreign systems to learn from the outputs of more capable American frontier architectures. Gerstner highlighted this specific operational issue, pointing directly to evidence found in Chinese open-source releases:
“We know they were distilling… I will tell you GLM 5.2 has watermarks from Mythos all over it, right? So we know they were distilling, etc. And I think the US government’s going to take steps against distillation, which they should do.”
Consequently, future policy may move beyond semiconductor restrictions and place greater emphasis on model access, weights, data protections, and software-layer controls to mitigate intellectual property leakage via public APIs.
Question Eight: What does this framework mean for investors?
The AI trade is entering a more demanding phase. The market has largely recognized the scale of the opportunity, earnings expectations have risen, and strong results will no longer guarantee strong stock-price reactions. The next stage will require greater discrimination across the value chain, including a granular understanding of which companies benefit from falling unit costs, aggregate token growth, infrastructure expansion, and the continued premium placed on the most capable models.
Gerstner’s comments provide a highly useful framework for that transition. Many of the concerns surrounding AI are framed as rigid, either-or outcomes: frontier models or open source, centralized systems or sovereign infrastructure, falling inference costs or durable economics. Yet intelligence represents an addressable market broad enough for frontier labs, open models, sovereign systems, hyperscale infrastructure, and local compute to expand simultaneously while serving entirely different needs across the macro ecosystem.
My own view is that OpenAI and Anthropic have now created a meaningful lead at the frontier, and that advantage is becoming increasingly difficult to dislodge as enterprise agents move into production. AI native companies, especially technology ones, will be driven by revenue and cost conscious behavior so they will be fast to move to the lowest cost productive model option. Bureaucratic enterprises are very different. OpenAI and Anthropic models are improving quickly, their enterprise relationships are deepening, and the pace of new releases gives customers little incentive to interrupt adoption in order to test an unproven alternative. Over time, enterprises will likely use a broader mix of models, routing lower-value workloads toward cheaper systems while reserving the most capable models for complex and high-consequence tasks. In the near term, however, the competition to deploy agents is moving too quickly for many companies to repeatedly reassess the model stack. When reliability, speed of implementation, and competitive urgency matter more than marginal inference savings, the incumbency of OpenAI and Anthropic becomes a powerful advantage.
The first phase of the AI trade was largely about anticipating which companies would beat quarterly earnings expectations as adoption, revenue growth, and capital spending repeatedly surprised to the upside. The next phase will require a longer-term framework. Gerstner’s central point is that intelligence represents the largest addressable market the world has ever seen, broad enough to support frontier models, open-source systems, sovereign infrastructure, hyperscale compute, and local deployment at the exact same time. The core investment question is therefore shifting from who can deliver the next near-term earnings surprise to which companies are structurally positioned to capture value as the global market for intelligence expands.
Periods like this are often when the market shifts its attention from what could go right to everything that could go wrong. Gerstner’s framework is a reminder that rising anxiety does not necessarily imply a weakening investment thesis, it may simply reflect a market searching for the next way to value an opportunity that continues to expand and intelligence is the largest TAM the world has ever seen.
