The Anthropic IPO

“For the first time in history, intelligence is something humans can build. The cognitive capability to reason, analyze, synthesize, decide, and create is producible at scale and accessible on demand. We believe this abundance of AI will rapidly transform every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet. AI that truly delivers on its potential—with the risks well managed—could significantly accelerate global growth and dramatically improve quality of life for people around the world.”

We would place the advent of super intelligence amongst the biggest economic revolutions in recent history—i.e., the industrial revolution, electricity, the transistor, and the internet. It’s obvious that humanity is still in the very early phases of AI adoption and that in the coming decades, we will see explosive growth in daily token consumption. The big question for investors is whether Anthropic can capture this opportunity?

The narrative on social media, and the mainstream media, is that the model layer in the AI stack will commoditize, with open-source models successfully fragmenting the market. While overall token consumption is without any doubt strongly fragmenting with a large variety of models handling user requests over API, the market for frontier AI in terms of revenue generation, on the other hand, remains heavily consolidated. Currently, this is a two horse race between Anthropic and OpenAI, with both companies racing towards $100 billion in gross ARR. Competitors, on the contrary, are lucky if can bring in a few billion in ARR.

We believe there is a credible bull case why frontier AI—as measured by models that are generating high gross margins and thus revenues—will remain a consolidated market. The reason is that we see two flywheels which should give the largest players a competitive advantage.

First, there is a data flywheel from customer workloads. Anthropic can see on which tasks Claude is being deployed, where it does a good job, and where it struggles. Anthropic can then analyze workloads where Claude did a poor job to then create new data sets—either synthetically or by acquiring them—to train Claude on how to solve these types of tasks and do a better job next time.

This is very similar to the data flywheel Google has in Search. Google can see on which websites users are clicking and spending their time, and which websites they quickly click away from. This gives Google a strong competitive advantage to have the best possible search ranking in the world when you type in a query.

This data flywheel is also how Anthropic identified the opportunity in coding. Cursor—the AI coding editor—used Claude as its backend AI model to write code. Anthropic subsequently saw that token demand in coding was exploding and that customers were willing to pay top dollar for these tokens. This led Anthropic to heavily train Claude in the area of coding, which then led to the company’s revenue explosion at the start of 2026.

Claude can write out code now in a single day what would have taken us probably one year or so to manually type in. So, a $100 or $200 monthly subscription for us feels like a steal. We’re not interested to use cheaper models, as even saving a few hours of additional debugging in a single month already justifies Claude’s premium pricing. This is why engineers just pay for the top model—it pays itself back in time saved and in building superior solutions.

As a result, the market value of writing out 50,000 lines of code has dropped with 99.99% or so. While the macroeconomic consensus is that we will remain in a higher inflationary environment in the coming 5-10 years due to high fiscal deficits and aging populations around the globe, it’s clear that AI can actually be a massive deflationary force as well as a massive productivity booster. AI gives strong potential for global GDP growth to accelerate and at low inflation rates.

For example, while the value of lines of code has dropped with 99.99%, skilled engineers remain in high demand as you want to cross-check with the AI how it’s writing out the code base and how it’s designing the backend in the cloud. So, the value of engineers that understand how to cleverly design and scale systems on cloud platforms such as AWS, Azure, and GCP has gone up, as there is now much more demand to build and deploy apps in the cloud.

Anthropic explains the flywheel based on customer data and feedback in their prospectus.

“Our enterprise strategy is built on the belief that frontier AI is most valuable in real-world settings, as deployment across enterprises is where economic impact is greatest and where the improvement loops from customer feedback are richest. Accordingly, we have organized our research priorities, our product roadmap, and our go-to-market and Applied AI teams around enterprise use cases, which range from routine knowledge work to the most complex problems. The customer feedback we gather directs our research focus and makes Claude more capable over time. This broadens where Claude can be deployed and further expands enterprise adoption, which helps ensure that the benefits of our AI reach people and organizations at scale.”

The key problem for competition is that their models will perform much worse on tasks they haven’t been trained on. On the other hand, the models with the widest variety of customer requests can get continuously retrained in areas where they struggle.

A second competitive advantage for the large players that generate substantial revenue is that scaling laws are still alive and well. In pre-training, models continue to grow in size. For example, Alibaba mentioned that their next model, Qwen 5, will have 10 trillion parameters. There won’t be many labs around the world which can afford such expensive training runs.

However, also in post-training, the largest labs will be able to generate superior performance due to their much larger budgets and data sets for reinforcement learning. For example, the largest labs could let their AI solve an almost infinite number of math, coding, and engineering problems with reinforcement learning to create the world’s best scientist-engineer-computer scientist-mathematician. This can both be done with synthetically generated and purchased data sets, e.g. by hiring junior investment bankers to teach the AI how to do financial modeling in spreadsheets. It’s only the largest and most cash-rich labs that will be able to hire large numbers of highly skilled professionals to construct advanced data sets in their field.

The Information reported that Anthropic discussed spending over $1B a year on RL environments. Clearly, not many labs will have the capabilities or resources to do multi-billion dollar training runs in order to compete at the frontier in AI. We suspect that, over time, only a handful of labs globally will be able to do so, which includes 2 or 3 Chinese labs. Smaller labs which don’t generate sufficient revenue, and for which funding will dry up at some stage, will have to drop out of the race. This is similar to what happened in leading edge semi manufacturing, where only three fabs currently are still competing (TSMC, Samsung and Intel).

Despite all the hype on social media and the mainstream media that Chinese open-source models are ‘just behind’ the Western frontier labs, we haven’t been very impressed with these models when we’ve tried them out on real-world, fairly complicated workloads. The founder of Vals AI—an AI benchmark for real-world tasks—finds similar results. This is from the Joe Lonsdale podcast:

“Right now we see a pretty considerable gap between Chinese open-source models and domestic closed-source models in multi-turn environments. A lot of the benchmarks in coding today are just single-turn. Chinese models do very well in those, but when you translate them to a back-and-forth conversation with a human, they really fall apart. When the human says, “Oh man, this sucks. I hate this output”? They really don’t know what to do; they’ve never seen that before. But a lot of the US labs have seen that. Claude Code was created, Codex was created, and so in their desire to build better products, they’ve built models that can sustain longer user interaction sessions. And so they have evaluations and benchmarks that allow them to hill-climb on the axes of quality of responses and communication ability.”

Vals’ benchmarks cover Excel modeling, legal and tax research, full-stack web applications and code migrations, and some runs last as long as 5 days.

The big problem is that China is obviously heavily short on AI compute. While in the West there is a shortage in AI capacity as well, this is nothing compared to the shortage that China is suffering from. US sanctions have obviously been effective. Vals AI has a similar conclusion why China is considerably behind:

“One is the idea of chips being a considerable asset that we have. I think it surprises most people, but we’re drowning in compute compared to what China has. And so as long as that’s the case, we’ll continue to train more capable systems, but also we’ll be able to diffuse the models much more effectively because we can host and run inference.”

Interestingly enough, Vals AI doesn’t find a considerable pro-China tilt in Chinese models:

“What we’re trying to measure is: does a model show more geopolitical tilt towards a US view of the world or a Chinese view of the world, as laid out by the PRC? It’s actually a much more nuanced picture than I thought. For instance, what we’re finding is that GLM-5.3, this Chinese open-source model, is actually exhibiting very little pro-China bias. We actually presented this to some labs, and there was suspicion that maybe it’s because it’s just entirely distilled from US models.It’s possible. I think there’s also maybe another view in which those models are actually hosted on systems with guardrails in China. And so once you put them around guardrails, they’re going to follow that PRC view of the world more consistently. But China also wants these models to be adopted by the BRICS, and so that means they may have to skew to a more balanced worldview. So anyway, regardless of the reason for it, I think there’s actually a big disconnect in what people think about the geopolitical tilt and what actually we see in practice.”

Distillation is obviously one key risk factor, where Chinese competitors via bot farms send tons of queries to Claude to then save Claude’s reply. This data can then be used as a basis to train Chinese models to replicate Claude’s answers. However, two points to make here.

First, Chinese competitors don’t have access to all the intermediary steps Claude made in order to construct that solution. As chain-of-thought reasoning continues to expand, it becomes harder and harder to figure out how Claude came with a solution for complex problems. While you can ask Claude to write out how it came up with a solution and all the steps, Anthropic can limit this functionality for suspect accounts. Second, frontier labs are well aware of this problem and can write, with the help of Claude, advanced detection algorithms to limit access for bot farms which are trying to distil the model’s solutions.

Additionally, what makes it hard to close the gap for competition is that smarter AI models are already building even more advanced AI models. The current Claude helps build the next Claude etc., as Anthropic discusses:

“Each version of a Claude model assists in building the next, helping to compress development timelines and produce compounding improvements in the speed at which we innovate. We refer to this recursive dynamic as “closing the loop,” and we believe it is one of the most significant structural advantages in frontier AI.”

Scale advantages, a faster innovation cadence, and more dollar resources all enable the dominant AI labs to also capture new opportunities and new markets where AI can expand. Coding is just the beginning, and over time, AI will also become state-of-the-art at wider engineering tasks, science, and eventually even manual labor with physical AI (i.e. robotics). This is Anthropic discussing the large future opportunity for AI to expand in new areas:

“Frontier AI’s capabilities may extend beyond current markets to open entirely new ones, reaching domains such as robotics, advanced manufacturing, and logistics, and accelerating progress in scientific research, including drug discovery and materials science. For example, in the life sciences sector, where annual global pharmaceutical spend is expected to exceed $2 trillion in 2026, Claude is used in drug discovery and development by our customers, our biotechnology partners, and our own researchers. We are also extending Claude into robotics, where our Model Hardware Standard enables AI agents to operate lab and manufacturing equipment such as microscopes, liquid handlers, and off-the-shelf robotic arms. Ultimately, we believe frontier AI will help drive global economic growth. As frontier AI makes workers and organizations more productive, becomes embedded in the products and services they deliver, and unlocks new business models, it creates economic output that did not previously exist. We believe our opportunity is to capture a share not only of today’s economic activity but also of that newly created output as the global economy expands.We believe the diffusion of AI through the global economy represents one of the largest opportunities in history to improve how humans live and work, and it is just beginning. In our view, the world has only begun to use the capabilities of today’s frontier models, and the gap between what Claude can do and how broadly it is used represents one of our largest growth opportunities, alongside continued advances at the frontier. Claude’s capabilities have evolved significantly in recent years, and we expect they will continue advancing rapidly. This will lead not only to an expanding addressable market, but also to an expanding scope of impact on the world. We aim to ensure that impact is as positive as possible. Our core advantages, the speed and scale of enterprise adoption of Claude, and our unprecedented financial model enable us to continue investing aggressively to build an AI-enabled future that benefits humanity broadly.”

The long term opportunity in AI is vast. Currently, the global economic value of knowledge workers is around $40 trillion. If companies would spend 10% of this amount on intelligent AI systems in the future, we’re already looking at a $4 trillion TAM for frontier AI. Obviously, enterprise spend on frontier and non-frontier tokens could be even much higher.

Our view is that this TAM will likely compound over time. Frontier AI labs are already generating $150 billion in ARR currently, and we suspect that the overall AI TAM will continue to see strong growth in the coming decades. It’s a virtuous circle really—AI will increase the size of the global economy, which again allows more money to be spent on AI.

This is similar to what happened following the industrial revolution. The more the economy grew post the industrial revolution, the more people could be employed in the economy, driving further growth, which allowed again more people to be employed. As a result, the global economy has been compounding at much superior growth rates over the last 150 years compared to historical growth rates. We suspect that a similar scenario will play out with AI, with AI further accelerating global GDP growth rates.

This is Anthropic on the economic opportunity:

“We believe the work that knowledge workers perform represented more than $40 trillion in economic value globally in 2025, and with frontier AI’s potential to make them meaningfully more productive, our near-term opportunity is to expand this value and accelerate economic growth. While it is still early to forecast frontier AI’s full impact across the economy, International Data Corporation (“IDC”) estimates that investment in AI solutions and services will yield a cumulative global economic impact of $22.5 trillion by 2031. We believe the impact could be even greater.”

AI enterprise adoption is forming a clear hockey stick curve. Basically, the number of customers who are spending more than $100k per annum with Anthropic has been going up 7-10x per year:

In addition, spend per customer is massively increasing. Anthropic reported Net Revenue Retention of 900% in September—meaning that for every $1 of revenue generated from customers in year one, that exact same group of customers spent $9 in year two, i.e. a ninefold increase. Anthropic credits this partly to lowering customer costs by making models more efficient:

“Diffusion also accelerates as our models become more capable and efficient, allowing us to lower the cost per task for customers. For example, Claude Opus 5.5 reduced customer costs per task by approximately 40% compared to Claude Opus 5. Lower costs per task drive greater usage, which in turn drives broader diffusion of Claude. Our compute efficiency allows us to deliver these reductions while preserving our margins.”

We’re heavy users of Anthropic’s API, and API costs have remained broadly the same over the last year. Although, the company did recently lower its price for the key Claude Opus model from $25 for one million output tokens to $20. However, the biggest cost savings for customers came from making Claude models more efficient—i.e., Claude has to burn less tokens to find the solution.

While bears such as Steve Eisman—from The Big Short—point out that dropping pricing is a huge negative, this is really nothing new in tech. Clouds have massively lowered pricing during the previous decades, which led to a boom in customer demand and strong revenue growth. Same happened in semis—the dollar cost per transistor fell off a cliff driven by Moore’s Law, which led to an explosion in compute demand and an explosion in fab revenues. We suspect that the same will happen in AI—a falling cost per token will lead to an even greater boom in demand, and even greater revenues for the clouds and labs that provide these tokens.

Confirming the above thesis, we can see that Anthropic’s revenues are exploding with tremendous operating leverage. As we’ll see in the financials, Anthropic is on its way to becoming a cash machine.

If we purely look at the cost of compute that’s used to generate revenue (users that pay for inference), gross margins are 85%. However, more than 70% of compute isn’t used to generate revenue—it’s used for R&D (training better models) and for Sales & Marketing, i.e. giving free users access to Sonnet (which is the funnel for users to subscribe to the more powerful Opus model or purchase API tokens). Including these, margins after all compute are around 45% currently.

If we also include channel fees—when the customer purchases tokens through the AWS, GCP or Azure cloud platforms—compute margins are around 27.5%. Overall, the economics of selling intelligent tokens are attractive.

It’s clear that Anthropic doesn’t want to be reliant on a single hardware vendor. The company’s strategy has been to multi-source various hardware since AI took off, with Anthropic being the first lab to use Amazon’s Trainium accelerators alongside Nvidia. More recently, the company also closed compute deals with Google for their TPUs and with AMD as a second source for GPUs.

“We manage compute as a single, fungible pool from multiple strategic partners, and we dynamically direct available compute between workloads based on where the marginal unit is expected to be most productive at any given time.We have achieved this through our research breakthroughs and our multi-chip compute architecture across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, enabling us to orchestrate workloads to optimize for price and performance.”

Anthropic is also cloud agnostic—Claude is available on every major cloud platform in order to maximize enterprise adoption:

“Claude is accessible on all three major cloud platforms—Amazon Web Services (“AWS”), Google Cloud Platform (“GCP”), and Microsoft Azure (“Azure”). We have a joint go-to-market approach with these CSPs through which we establish direct contractual relationships with our customers while paying a fee to each CSP. This structure allows us to foster stronger commercial alignment with these CSPs, avoid channel conflict, set pricing, and realize the financial benefits of ongoing compute efficiency improvements.”

There are also obvious risks. Bad actors could successfully mislead Claude to help them launch cyberattacks, write malware, create bioweapons, bombs, etc. If Claude helps engineer a new deathly virus in a biolab somewhere, would Anthropic be on the hook for some of the economic damage? While big tech has regularly had to pay heavy fines over the last decade, this hasn’t disrupted the strong growth story. We see a similar scenario playing out in frontier AI.

However, these same risks can also lead to further consolidation. If open-source models are used for cyberattacks, malware, bioweapons, etc., we see regulators in the US, China, the EU, etc. stepping in to limit access to open-source models. We suspect that bad things will almost certainly happen at some stage when it comes to one of these risks—which will then likely lead to more regulation that will favor AI labs that have demonstrated that they can build safe, closed-source AI systems.

Recently, we’ve seen announcements from various governments around the world to make sure people have enough food and water for 3 days or more at home, which we suspect is because they think that bad actors can take out the power grid at some stage with the help of AI. For example, Russian hackers taking out the UK power grid, or potentially Iranian or other hackers taking out the US power grid. Also Anthropic discusses societal risks:

“While the potential benefits of AI are extraordinary, its rapid advancement will impact the economy and society in ways that could create significant disruption, particularly if the technology is not developed and deployed responsibly. For example, bad actors have already conducted high-profile cyberattacks using AI’s capabilities to discover vulnerabilities, generate malware, and evade detection tools; other risks associated with misuse, including in the domain of biology, could result in even greater harm. AI deployed at scale also has the potential to concentrate power and wealth both within and between countries in ways that could erode trust, harm society, and destabilize the geopolitical order. Misaligned AI systems could corrupt institutional decision-making or cause large-scale harm through autonomous actions that are not visible or understandable to their users and operators. We believe that frontier AI companies must invest in safety and alignment to help ensure the responsible diffusion of this technology.”

Next, we’ll go through Anthropic’s financials, our estimates for the future P&L, the risks, and at which valuation we’d be interested to take a position in this company.