Inside Anthropic’s trillion-dollar AI gamble

The host opens by describing Anthropic’s IPO filing as the most detailed public look yet at the economics of a frontier AI lab. The filing could support a valuation between $1.5 trillion and $2 trillion, potentially making Anthropic one of the world’s most valuable companies. Joey Brookhart of SemiAnalysis says the S1 was exceptionally strong: Anthropic has scaled rapidly to more than $70 billion in annual recurring revenue, with strong gross margins, compute margins, and free cash flow. He says the numbers explain why AI infrastructure companies, neoclouds, and compute stocks rallied after the filing. Brookhart and the host immediately focus on the harder question: where could enough revenue come from to support Dario Amodei’s vision of eventually reaching $1 trillion in revenue? Brookhart says the current market is still narrow. The top 1% of customers spend as much as $50,000 or $100,000 per employee, while most Fortune 500 and Global 2000 companies spend relatively little. Anthropic is also still close to a single-vertical company because coding accounts for much of its usage. Brookhart expects expansion into cybersecurity, biology, financial services, legal work, and other forms of white-collar labor, but says reaching a trillion dollars requires AI to augment or replace a substantial amount of work.

The host points out that Anthropic is not competing in a two-company market. OpenAI, Google, Meta, open-weight models, and increasingly specialized or on-premises systems are all competing for the same future spending. Brookhart agrees that the industry became more competitive during the year: Anthropic generated roughly 90% of net-new industry ARR in the first half, but OpenAI took the majority of net-new ARR in the third quarter. Open-source AI revenue, in his estimate, grew from about $1 billion in the first quarter to more than $10 billion in the third quarter. The discussion turns to whether coding has low switching costs and therefore cannot carry the whole business. Customers can move between coding models, and enterprises increasingly customize open-weight models and run inference internally. Brookhart says new verticalized use cases are essential because coding alone will not triple the industry from its current scale. Frontier models will remain valuable in adversarial or highly competitive fields: a cybersecurity company needs to defend against frontier models, and a defendant would presumably want a lawyer using the best model available if the opposing side is doing the same. But tasks such as email summarization do not require frontier intelligence. Anthropic is already testing specialized offerings for financial services and legal work and offers $25,000 per month in credits to participants in its cybersecurity program.

Brookhart explains SemiAnalysis’s “tokenomics” work as an attempt to model the economics of AI from the macro level down to individual inference requests. Anthropic had roughly negative 94% gross margins on inference two years earlier; the S1 now shows positive margins of about 85%. Those gains come from both hardware and software efficiency: better use of Trainium chips, Google TPUs, and GPUs, as well as improvements in how models interact with hardware. Brookhart says token throughput on Trainium can rise fivefold to tenfold in a year. He estimates that Claude can generate about $65 million in revenue per inference megawatt, up from the mid-teens a year earlier, and says a future model on Blackwell hardware could reach $100 million to $120 million per megawatt. Compute costs are now only a little over $10 per megawatt in his framing. Only about 30% of Anthropic’s compute serves revenue-producing inference; the remainder supports experimentation, research, and training. Anthropic reported roughly 11% non-GAAP EBIT margins, but Brookhart says the company may prefer lower margins—or even losses—if it can redirect cash toward training. The host compares this to Jeff Bezos’s Amazon strategy: accept lower profits while building toward a much larger future opportunity. Brookhart says Anthropic has told investors not to assume its current margin leverage will continue because it is still pursuing AGI.

The host questions Anthropic’s assumption that it must always remain at the frontier. If current models can already automate meaningful work in law, medicine, research, and finance, a potentially more profitable strategy would be to stop constantly training larger models and instead deploy existing models into more industries. The host compares this with Apple continuing to sell older iPhones and reuse production tooling while developing the next generation. Brookhart says frontier status remains strategically important in some areas, especially where competitors or adversaries are also using the best models. He also notes an important S1 figure: 80% of Anthropic’s revenue in the preceding three months came from models released in the previous six months. That suggests customers are still moving toward the newest models rather than treating older systems as durable, high-margin products. At the same time, open-source models are becoming good enough for lower-stakes work, and enterprises sometimes restrict frontier models to coders while giving non-coders cheaper alternatives. Brookhart expects the eventual division between frontier and non-frontier models to depend on whether a task is agentic, whether humans remain in the loop, and how much value the model’s additional intelligence creates.

One of the most visible S1 figures is an $80 billion loss. Brookhart says this is primarily a non-cash accounting charge caused by Anthropic’s sharply increased valuation and the warrants or convertible notes it issued. As the company’s valuation rose, it had to reflect the changing value of those instruments on its balance sheet. He says the headline number looks alarming if read as a simple cash loss, but it should be backed out when analyzing operating performance. Brookhart estimates that Anthropic has actually burned about $10 billion in cash and holds more than $120 billion in cash and marketable securities after its financing activity. The host asks whether Amazon and other investors will ultimately be diluted as Anthropic issues more equity to finance its growth. Brookhart says all technology companies dilute shareholders, but he is not especially concerned if Anthropic becomes a multitrillion-dollar company and grows faster than the dilution. He also argues that Anthropic’s compute arrangements are more manageable than the headline commitments suggest: the company rents compute or uses rent-to-own and special-purpose-vehicle structures rather than immediately making equivalent capital outlays.

The S1 reportedly includes more than $800 billion in long-term commitments to hyperscalers and neoclouds. Brookhart says the headline number becomes less extreme when converted into the amount of power and hardware it represents at today’s prices. He estimates current compute spending at roughly $11 million per megawatt, or $11 billion per gigawatt, and argues that Anthropic could generate enough revenue to support the commitments if enterprise adoption expands and inference efficiency keeps improving. The central uncertainty is how much of the available compute will be used for profitable inference versus training and research. The host raises another risk: AWS, Microsoft Azure, and Google Cloud are suppliers but also competitors, and could eventually stop leasing so much capacity to Anthropic. Brookhart says Anthropic is already moving toward more direct arrangements, including buying chips through an SPV and then renting the resulting capacity. About 48% of Anthropic’s business comes through cloud marketplaces such as AWS Bedrock and Google Vertex, where customers use existing cloud relationships to access Anthropic models. Those marketplaces benefit too, because AI workloads raise cloud utilization and margins. Brookhart expects only neoclouds backed by major labs, hyperscalers, Nvidia, or Broadcom to be resilient. He says data-center shells can be overbuilt without wasting the most expensive component—the chips—but warns that rapid efficiency gains could produce a compute glut by 2028 if revenue does not grow enough to justify roughly $2 trillion in capital spending.

The host contrasts Anthropic with OpenAI, which historically had a much larger consumer business. OpenAI was approximately 60% consumer and 40% enterprise at the end of the first quarter, only becoming enterprise-focused later. Brookhart says OpenAI captured most net-new ARR in the third quarter and had particularly strong enterprise growth, but its business does not monetize consumption at Anthropic’s rate. He expects OpenAI’s eventual S1 to show weaker financials, potentially including negative 100% EBIT margins, even if its enterprise momentum improves. For Anthropic’s road show, Brookhart expects investors to ask about a weak third quarter, when Meta reduced its Claude Code seats from about 60,000 to 30,000 and shifted usage toward internal models. He says the token volume did not disappear, but the customer relationship changed. Anthropic also said roughly 5% to 7% of revenue came from usage that violated its terms of service, including model-distillation activity. Brookhart believes October accelerated after the release of a new Opus model, with net-new ARR potentially recovering toward $8 billion to $10 billion per month after falling from roughly $10 billion in the second quarter to $4 billion in the third. The company will need to sell investors on new verticals—especially cybersecurity—while managing expectations. It must simultaneously present a long-term AGI vision, acknowledge extraordinary risks such as models potentially eliminating jobs or causing catastrophic harm, and avoid promising so much near-term growth that it becomes impossible to beat public-market expectations.