Recently, Reuters obtained the confidential prospectus Anthropic prepared for its initial public offering.
The document contains some extraordinary numbers.
A potential valuation of more than $2 trillion. A twelvefold increase in revenue in 2025. And roughly $42 billion in net losses.
That last figure deserves an explanation. Around $34 billion represents an accounting charge related to financing that could eventually convert into shares. So Anthropic did not spend $42 billion running Claude. Its operating loss was still substantial, at more than $8 billion.
The valuation, meanwhile, clearly requires a great deal of optimism about what AI could become.
But the number that interests me most when building an investment thesis appears elsewhere in the document.
$518 billion.
That is how much Anthropic expects to spend on cloud services, computing capacity, and infrastructure over roughly ten years.
A single company, behind a product you may already use, plans to commit more than half a trillion dollars to running and developing its artificial intelligence.
And a large share of those commitments would remain payable even if Anthropic used less capacity than expected.
For Anthropic, that represents an enormous obligation.
For the companies supplying that infrastructure, it represents an enormous potential market.
That is what I want to explore today: where this money is expected to go, which companies could capture the largest profits, and how we can position ourselves at valuations that still leave room for attractive returns.
Understand the thesis. Follow how we position for it.
As a Macro Notes Premium subscriber, you’ll unlock the complete research behind this thesis: the companies we select, why we select them, their valuations, and the risks that could change our view.
We’ll also build a live investment thesis index that you can follow over time—tracking performance, every recorded trade, and the developments that matter, including Anthropic’s potential IPO.
The index will focus on the listed companies we believe are best positioned to benefit from Anthropic’s infrastructure spending over the short to medium term. Each position will have a clear rationale, and you’ll be able to follow how that rationale evolves.
This follows the global macro, top-down approach I described in a previous edition: start with a structural shift, map where the money is flowing, then select the businesses positioned to capture it.
We’ve applied this approach to themes such as fiber optics, where our thesis recorded a return of more than 500%, and memory and storage, where the thesis covered in Public Markets recorded more than 700%.
Premium brings the research and its ongoing tracking together. You can follow an idea from the initial thesis through the decisions, catalysts, and changes in conviction that shape its development.
New Macro Notes Premium subscriptions now also include Public Markets Premium, our second Substack bestselling newsletter.
Subscribe today to unlock the full Anthropic infrastructure thesis, follow the index once published, and access the research from both newsletters.
Before looking at the companies, I want to explain why this spending could create an investment opportunity—and what would have to happen for that opportunity to translate into shareholder returns.
Start with the money already being spent
According to Reuters’ reporting on the prospectus, Anthropic spent $7.33 billion on computing and infrastructure in 2025, more than three times the previous year’s amount. Its revenue that year was approximately $4.6 billion.
Those figures include investment in developing models and supporting future growth. They do not tell us the cost of serving an individual customer.
But they show that computing is already an enormous expense, giving Anthropic a strong incentive to improve how efficiently it uses that infrastructure.
For investors, that creates two questions.
Which suppliers can provide the performance Anthropic needs? And which can retain attractive margins while helping their customer reduce its costs?
The infrastructure keeps working after the model is trained
Training creates the model. Inference is the process of running it when someone asks a question, submits a document, or requests a piece of code.
An AI agent can take several steps to complete one task: inspect files, generate an answer, use a tool, check the result, and try again.
The economic variable I want to follow is therefore the amount of useful work these systems perform—and the computing resources required to perform it.
Efficiency matters here. Faster chips and better software can reduce the resources needed for each task. Our thesis depends on demand growing sufficiently to support additional capacity after those improvements.
That is something we need to measure through actual deployments and utilization.
There is already a specific industrial agreement behind the story
Broadcom’s April regulatory filing describes an agreement to develop and supply future generations of Google’s custom AI chips, called Tensor Processing Units, alongside networking and other components for Google’s next-generation AI racks through up to 2031.
The same filing describes Anthropic accessing approximately 3.5 gigawatts of TPU-based capacity through Broadcom beginning in 2027. It explicitly makes consumption of that additional capacity dependent on Anthropic’s continued commercial success.
This gives us an identifiable supplier, a deployment timeline, and a condition to monitor.
Anthropic also says it uses Nvidia GPUs, Google TPUs, and Amazon Trainium chips. Its growth can therefore support several hardware ecosystems, with different companies capturing value at different points.
A custom chip also creates demand around the chip
Google’s existing Ironwood architecture provides a useful illustration.
Each Ironwood chip contains eight stacks of specialized high-bandwidth memory. A rack contains 64 chips. That means 512 memory stacks in a single rack, before considering the infrastructure connecting racks into a larger system.
Google describes configurations scaling to 9,216 chips, connected through an optical switching network. These specifications illustrate the architecture; they do not establish the exact configuration Anthropic will deploy in 2027.
The investment implication is that accelerator growth can create demand for memory, advanced packaging, and networking.
However, we still have to establish which suppliers are qualified for each platform, how much they supply, and whether additional volume improves their profits.
A component can be essential to the system and still be a disappointing investment if competition compresses its margins.
We can already see substantial growth in a supplier’s accounts
Broadcom reported $16.7 billion in AI semiconductor revenue in its third fiscal quarter of 2026, up 221% year over year and 54% from the previous quarter.
Across the entire company, it generated $13.7 billion in free cash flow during the quarter, equivalent to roughly 46% of revenue. These are consolidated figures, including its software business; they are not the margins or cash flow of its Anthropic relationship alone.
This is evidence of a large, growing hardware business inside a company already producing substantial cash.
The next step is to determine how much further growth can reach shareholders after manufacturing costs, research spending, financing obligations, and changes in the business mix.
The contract is only part of the investment case
Reuters reports that approximately 80% of Anthropic’s infrastructure commitments are non-cancelable or payable regardless of usage. That strengthens their contractual significance, but payment still depends on the customer’s ability to meet its obligations.
That is why we need to examine financing alongside revenue.
We also need to examine the price we pay for the supplier’s shares.
Here is a simple illustration: if earnings per share rise 30%, but the valuation multiple falls 20%, the share price increases only 4%, before dividends.
The business can deliver impressive growth while the investment delivers a modest return.
This is where the Premium research begins: comparing the profits these deployments could generate with the expectations already embedded in each stock.
In the Premium section, we will examine:
Where the $518 billion is expected to go: the partners, contract structures, deployment schedules, and the distinction between equipment sales, leases, and cloud services.
Broadcom’s exposure in detail: potential revenue growth, margins, customer concentration, financing obligations, and cash conversion.
The alternative ways to express the thesis: comparing Broadcom with TSMC, Amazon, Alphabet, and relevant memory suppliers, based on documented exposure.
What current valuations require: scenarios for earnings and cash flow, the assumptions behind them, and the returns those scenarios could support.
How we would build the index: the companies we would prioritize, the rationale for each position, and the developments that would lead us to add, reduce, or exit.
The objective is to identify where this infrastructure expansion could produce enough additional profit to justify the price we pay—and to follow the evidence as the thesis develops.