Anthropic’s Trillion-Dollar Question: Can Enterprise Value Keep Pace With the Frontier?

1. Guest Background

Joey Brookhart approaches the AI business through the relationship between demand, computing capacity, and financial returns. He is an analyst at SemiAnalysis, a research firm focused on chips and artificial intelligence, and a co-lead of its Tokenomics model. That work covers AI demand and unit economics across major cloud providers, emerging compute providers, and model laboratories.

His team had already built an outside financial model of Anthropic before this interview. That experience is relevant because the conversation repeatedly moves between the economics of serving an existing model and the expense of developing its successor. It also establishes an important boundary: Brookhart brings an analytical model, rather than complete access to the company’s internal decisions. His estimates and forecasts should retain that status.

In this approximately 42-minute episode of Decoder with Nilay Patel, Patel presses the customer-side question: what useful work will generate the money laboratories expect to earn? Brookhart explains how throughput, consumption, and research spending might support that ambition. Their exchange examines both Anthropic’s reported financial story and the broader assumptions required to turn frontier AI into a durable business.

2. What the Episode Covers

The episode opens with an unusually optimistic financial framing. Patel says Anthropic has filed an S-1 for a Nasdaq listing, with an expected valuation between $1.5 trillion and $2 trillion. Brookhart says its annual recurring revenue run rate exceeds $70 billion and praises its capital efficiency. These are statements made in the episode, rather than independently verified findings of this article. The discussion’s value lies in examining the assumptions behind them.

Brookhart describes a market that remains narrow despite its scale. Spending is concentrated among heavy users and coding customers; he says the top 1 percent of customers can spend $50,000 to $100,000 per employee. Reaching the trillion-dollar revenue ambition discussed on the show would require expansion into additional industries and a transition from assisting workers toward saving or replacing labor. He points to testing in financial and legal services and a cybersecurity program offering $25,000 in monthly credits, correcting his initial figure during the conversation. Such programs indicate a direction of travel, rather than proof of recurring commercial success.

The financial discussion also separates accounting losses from cash consumption. Asked about an $80 billion loss, Brookhart attributes it to noncash charges associated with warrants and convertible notes after valuation increases. He acknowledges a separate debate over excluding stock-based compensation from non-GAAP measures. He says cumulative cash burn is about $10 billion, cash and marketable securities exceed $120 billion, and an IPO could add roughly another $100 billion. His confidence that growth can outrun dilution rests on an explicitly bullish scenario in which Anthropic becomes an AGI leader worth more than $10 trillion.

Recent growth provides a less settled picture. Brookhart estimates that Meta once represented more than 10 percent of revenue and cites reporting that its Claude Code seats fell from 60,000 to 30,000 as consumption shifted to internal models. He also identifies enforcement against prohibited distillation, associated in his account with roughly 5–7 percent of revenue, as a factor in the slowdown.

His estimate of monthly net new ARR falls from $10 billion in the second quarter to $4 billion in the third, followed by a projected recovery toward $8 billion to $10 billion in the fourth. He cites substantial vulnerability-scanning contracts with existing customers as a cybersecurity opportunity. These remain estimates, reported explanations, and expectations. They lead to the episode’s closing tension: an IPO narrative must combine sweeping claims about AGI and work replacement with credible near-term results and guidance for 2027.

3. Core Views: Reasoning, Examples, and Limits

The central disagreement concerns the connection between technological progress and commercial advantage. Patel asks whether a laboratory could earn more by deploying its current models across underserved industries than by continually training replacements. Brookhart responds with a claim that approximately 80 percent of revenue comes from models released within the previous six months. However, he changes the revenue measurement window from six months to three months while speaking. That ambiguity matters: the observation may support demand for newer models, but it cannot establish the return on every subsequent training program.

Brookhart’s more concrete defense of frontier investment is task-specific. In cybersecurity, an attacker using a stronger model may create an incentive for defenders to upgrade. He offers litigation as another adversarial setting where the other side’s capabilities could influence tool selection. Email summarization presents a different purchasing problem. He says some enterprises permit only programmers to use frontier models while restricting other employees to Sonnet. The implication is that a frontier premium must be justified by the work being performed. Competitive pressure in selected tasks does not demonstrate that every employee needs the most capable model.

Improving inference economics strengthens the laboratory’s options without determining which option it will choose. Brookhart says Anthropic’s inference gross margin moved from negative 94 percent two years earlier to positive 85 percent, and that Trainium token throughput rose fivefold to tenfold year over year through hardware and software improvements. Yet he also says only about 30 percent of compute serves revenue-producing inference. The balance supports research, experimentation, and training. In his account, management could accept zero or negative operating margins to obtain more training capacity, rather than preserve the roughly 11 percent non-GAAP EBIT margin discussed in the episode. Better service economics can finance a larger research budget as readily as higher reported profit.

Efficiency also complicates the infrastructure investment case. Brookhart argues that more revenue per inference megawatt means less capacity is required for a given revenue level. He therefore raises the possibility of excess capacity in 2028 and asks whether revenue will justify more than $2 trillion in capital expenditure. This is a conditional scenario, not a prediction established by the interview. Demand could expand fast enough to absorb efficiency gains. Nevertheless, present shortages alone cannot demonstrate the attractiveness of future construction when the productive capacity of each megawatt is changing.

Adoption creates a second source of uncertainty. Patel describes customer concerns about handing data to laboratories and law firms running systems on premises. Brookhart discusses zero-data-retention requirements and custom training on open-weight models. These examples show why capability and procurement move on different schedules. Patel consequently proposes a competitor organized around deployment engineers, open models, and vertical automation. Brookhart agrees that products and deployment must bridge today’s models and the envisioned ability to perform complete jobs. The episode does not establish that this approach will defeat frontier laboratories; it identifies a place where customer value may be captured before the larger technical ambition is realized.

Consumption revenue adds another complication. Brookhart says Anthropic’s share of SemiAnalysis model usage fell from roughly 90–95 percent to 70 percent before returning above 90 percent after Opus 5.5. One organization’s experience is not a market-wide retention study, but it demonstrates that substantial usage can coexist with an ability to switch. He also relays approximately 900 percent net revenue retention while describing a roughly 20 percent decline in token spending during the final two weeks of the previous year, which he attributes to reduced holiday coding. Customer expansion, model competition, and workload sensitivity can all operate simultaneously; a large retention figure does not turn consumption into a multiyear contractual commitment.

Finally, cloud relationships contain incentives on both sides. Brookhart says around 48 percent of business comes through channels such as Bedrock, Foundry, and Vertex. Providers can earn both infrastructure revenue and a share of model distribution, creating reasons to support a laboratory that also competes with them. His illustrative calculation uses a 16 percent AWS share, but the conversation does not establish that as a contractual term. The broader lesson is similarly bounded: shared economics help explain cooperation, while leaving strategic dependence and competitive risk unresolved.

4. Learning and Application

For enterprise buyers, the most useful application is to budget by task. Select a workflow whose outcomes can be checked, compare a frontier model with a less expensive model and an open-weight alternative, and record quality, rework, and human review requirements. Brookhart’s account of different permissions for programmers and other employees supports this differentiated approach. Define data-retention and deployment requirements before expanding the pilot: a performance advantage has limited practical value if the proposed service cannot satisfy the organization’s operating conditions.

Also distinguish customer coverage from depth of use. Brookhart says Anthropic serves about half the Fortune 500 while accounting for only approximately 0.1 percent of those customers’ operating expenses. He contrasts average annual spending of $18,000 among the top 5 percent of Claude Code users with monthly spending well below $75 among non-Claude Code users. A buyer can use that contrast to examine its own distribution: are a few developers driving consumption while other workflows remain disconnected? Low spending identifies a question to investigate. It does not prove latent demand, since limited task value or deployment difficulties could explain it.

For financial analysis, keep inference economics, training expenditure, noncash charges, and cash consumption in separate schedules. Brookhart’s team estimates approximately $65 million in revenue per inference megawatt and projects $100 million to $120 million for Opus 5.5 on Blackwell, alongside compute costs slightly above $10 million per megawatt. The passage does not specify the relevant measurement periods. Before deriving margins or payback, establish comparable periods, hardware, utilization, pricing, and the share assigned to inference. Otherwise, precise arithmetic can conceal incompatible inputs.

Company comparisons require similar discipline. Brookhart describes OpenAI’s first-quarter-end revenue mix as roughly 60 percent consumer and 40 percent enterprise, versus token consumption representing about 80 percent of Anthropic revenue. He cites a media-reported OpenAI EBIT margin near negative 100 percent against Anthropic’s positive 11 percent. Use the contrast to ask how subscriptions, consumption, monetization, and research burdens differ. Figures from different sources and accounting presentations are insufficient for a complete financial ranking.

Infrastructure decisions should test slower demand, faster efficiency gains, and a larger training allocation together. Preserve the interview’s different commitment figures: Patel cites more than $800 billion over ten years, while Brookhart discusses $500 billion to $600 billion over seven to ten years. Brookhart also describes SPVs, rent-to-own structures, and an Anthropic-related TPU purchasing pipeline of approximately 3.5–4 gigawatts or more. More direct procurement changes the financing arrangement; it does not by itself establish lighter obligations.

When evaluating a compute provider, examine the substance of its financial support. Brookhart favors providers backed or backstopped by major laboratories, cloud companies, or chip suppliers, naming CoreWeave and Nebius. His estimate that hyperscalers represent 65 percent of capital spending and newly deployed compute helps explain pressure on the less-supported remainder. A practical assessment should inspect the support’s form, delivery conditions, and allocation of responsibility, rather than infer a guarantee from a prominent partner’s name. Finally, the option he relays from cloud CFOs—leave relatively inexpensive shells empty and defer chip purchases—preserves some spending flexibility. It does not establish an adequate return on money already committed, and permitting, power, and demand still require separate scrutiny.

Source

• Original episode: Inside Anthropic’s trillion-dollar AI gamble