For investors combing through Anthropic’s financials ahead of its IPO, a key question is what happens to the money the company brings in. The S-1 filing showed Anthropic spent $1.04 for every $1.00 of revenue it booked in the first half of 2026, resulting in an approximately $650 million operating loss on $16.3 billion of revenue. The bulk of Anthropic’s spending goes to compute, most of which flows to a short list of companies: Amazon AMZN , Alphabet’s GOOG Google, Microsoft MSFT , Broadcom AVGO , and a few GPU clouds.
Most of them also own, finance, or resell Anthropic, so a large share of each dollar returns to the companies that fund it. And this compute bill is only going to get bigger in the next two years. In this article, we will break down each significant operating cost, who is collecting on the other side, and how the money comes back in return.
Compute Consumes More than Half of Anthropic’s Revenues The most significant operating cost for any frontier artificial intelligence lab is compute. It is the lifeblood of frontier labs, enabling an AI system to be run across all stages of its lifecycle (pre-training, post-training/reinforcement learning, and most importantly, inference). In the first half of 2026, Anthropic spent $8.9 billion on compute and infrastructure.
Of this, $2.4 billion was revenue-generating, while the other $6.5 billion went toward training runs, research experiments, and free-tier usage. As revenue has scaled over the last year, compute costs have fallen from $2.01 to just $0.48 per revenue dollar, reflecting higher utilization of existing capacity and exponential revenue growth. While compute costs may appear well-controlled today, we expect them to grow materially.
Anthropic’s compute costs can be split into two buckets: training and inference. Training costs are expensed in advance of the revenue the trained models generate. As a result, as Anthropic continues to train larger, more advanced models, we expect the training costs to grow ahead of the revenues these new models stand to bring in.
We also see Anthropic bringing on new capacity for inference, which pushes compute costs higher (though rising inference costs are immediately “paid for,” due to the revenues associated with inference). To sustain future growth, Anthropic has already contracted significant compute capacity to meet rapidly increasing demand. Therefore, we see compute per revenue dollar increasing in the near term as Anthropic’s massive compute commitments come online, with some margin compression likely to follow.
The reason for signing large commitments is pretty obvious. Demand for frontier AI models far exceeds the existing supply of AI compute, with AI labs aiming to bring more capacity online to capture incremental revenues and grow their businesses. This dynamic is also why frontier labs like Anthropic have raced to secure compute agreements with a wide range of suppliers.
However necessary for sustaining growth amid growing demand, these massive future commitments pose a significant risk for Anthropic. CEO Dario Amodei emphasized this on the Dwarkesh Podcast when he explained, “I could buy $1 trillion of compute that starts at the end of 2027. If my revenue is not $1 trillion, if it’s even $800 billion, there’s no force on Earth, there’s no hedge on Earth, that could stop me from going bankrupt.” For this reason, Anthropic aims to be “bold but sane” in its compute commitments, balancing revenue maximization with some downside protection.
According to our estimates, Anthropic has roughly $819 billion in compute commitments, and $278 billion h of those commitments can be exited if demand falls sufficiently short of expectations. While Anthropic has afforded itself some room for error through a few flexible commitments, ultimately, even the existing fixed commitments require a massive ramp-up in revenues for the firm to support them durably. The Cost of the Cloud For many enterprise customers, Anthropic isn’t the vendor.
Cloud platforms such as AWS (Amazon), Azure (Microsoft), and Google Cloud are increasingly popular options for enterprise customers with preexisting relationships and large financial commitments with cloud vendors, who would rather add to budgets than create new ones. Additionally, many enterprises feel more comfortable purchasing AI through cloud vendors, as they believe these providers have better governance and security controls. Nearly half of Anthropic’s revenue now arrives through the same clouds that sell it compute, and each one takes a cut.
These costs are bucketed under the Sales, Marketing, and Partners operating line, and we estimate that they take roughly 15 cents coming through cloud marketplace sales, or roughly 7 cents of total revenue. The Unbearable Weight of Massive Talent Personnel costs amounted to $3.15 billion in the first half of 2026, representing 19 cents per revenue dollar. However, roughly 70% of that was stock-based compensation, as cash salaries and benefits only totaled around $970 million ($0.06/rev dollar) for a staff of about 5,300.
That is a remarkably small payroll for the revenue Anthropic is both currently generating and projecting to generate in the near future. TTM revenue per employee was $6.6 million as of Sept. 30, 2026, up from $1.4 million a year earlier and several times the level of the most efficient large software companies. As revenues continue to trend upward, this number could grow even higher in the near term.
The Money That Comes Back One of the most fascinating dynamics that makes Anthropic unique is that many of their biggest suppliers also fund and invest in the company. This creates a commercial money loop wherein five of the six companies that collect most of Anthropic’s compute and channel fees are also its shareholders or lenders. At first glance, it may seem counterintuitive that hyperscalers such as Microsoft and Google would fund a potential competitor.
However, while several of Anthropic’s investors have competing AI ambitions, they also stand to benefit from helping scale what could become one of the world’s largest consumers of AI infrastructure. By providing capital, suppliers can accelerate Anthropic’s growth, increasing future demand for their clouds, chips, networking equipment, and data center capacity. At the same time, equity ownership gives these companies exposure to Anthropic’s upside, hedging their own in-house AI bets should frontier labs ultimately capture a disproportionate share of the value created by AI.
On balance, we view this loop as healthy while demand is strong, since supplier capital is matched by real end-customer usage. It would become a concern if growth stalled, because the same suppliers would then be exposed on both their equity stakes and their contracted capacity. These factors ultimately work together to create a unique economic ecosystem whereby a portion of the money that goes out to suppliers eventually comes back to Anthropic in one form or another.
Put it all together, compute lays the largest claim to every dollar Anthropic brings in (and will continue to), followed by other spending areas, including personnel and cloud partner fees. The compute dollars flow to a handful of suppliers, many of whom are also Anthropic’s shareholders. With significant contracted compute coming online in 2027 and 2028, it is imperative that Anthropic’s revenues continue to scale to meet the rising compute bill.
The author or authors do not own shares in any securities mentioned in this article. Find out about Morningstar’s editorial policies .
Source: Morningstar
