Anthropic CEO Dario Amodei called forecasting computing needs the company’s “deepest business dilemma” during a November 13 livestream with JPMorgan CEO Jamie Dimon, according to Business Insider. His answer was a procurement strategy: buy relatively little capacity far in advance, then use the company’s liquidity to purchase additional compute on the spot market if products take off.
The stakes are already substantial. Anthropic has disclosed roughly $900 billion in infrastructure commitments, with about half expected to come due by the end of 2030. At September 30, it held more than $95 billion in cash and equivalents, plus $36 billion in short-term investments, Business Insider reported.
Those figures describe two different things: obligations accumulated over many years and liquid resources available to manage nearer-term decisions. Understanding the gap between them is more useful than treating Amodei’s comments as another prediction about an AI bubble. Anthropic is trying to decide how much future demand to lock into contracts before that demand becomes visible.
The Cost of Getting Compute Demand Wrong
Amodei described two ways the forecast can fail. If Anthropic buys too much computing capacity, costs can run ahead of revenue. If it buys too little, it can struggle to serve customers and lack the resources needed to develop its next models, Business Insider reported.
That distinction matters because compute supports both current sales and future products. Serving Claude requests consumes capacity today. Research and model development require resources before the resulting products generate revenue. A purchasing decision therefore has to account for customer usage and a development schedule that may have a different timeline.
Rapid growth alone cannot make a bad forecast harmless. A company can expand quickly and still fall short of the revenue assumed when it signed its infrastructure contracts. Conversely, preserving cash by buying less can become expensive if scarce capacity prevents it from serving paying customers.
My view is that this is the most useful part of Amodei’s explanation: he identifies a business risk that survives even if AI demand keeps growing. The question is whether demand arrives in the right quantities, at the right prices, and soon enough to support the capacity already ordered.
What the $900 Billion Actually Represents
The commitments are spread across years. Fortune’s November 4 analysis put the total at $895.4 billion, extending through 2045. CNBC’s November 11 report rounded that to approximately $900 billion and provided nearer-term payment expectations.
| Disclosed measure | Reported amount | What it tells readers |
|---|---|---|
| Future infrastructure commitments | About $900 billion | The scale of obligations across multiple years |
| Commitments expected by the end of 2030 | Roughly half the total | A substantial portion falls within the next several years |
| Commitments expected over the next year | About $70 billion | A nearer-term measure of the purchasing burden |
| Cash and cash equivalents at September 30 | More than $95 billion | One part of the liquidity buffer |
| Short-term investments at September 30 | $36 billion | Additional liquid financial resources |
The commitment figures come from company disclosures reported by CNBC and Fortune; the liquidity figures come from disclosures reported by Business Insider. The cited reporting does not independently audit those amounts.
Adding the two liquidity categories produces more than $131 billion. Using the rounded figures, that is about 1.9 times the approximately $70 billion in commitments expected over the next year.
That calculation is a scale comparison, not a cash-runway estimate. It excludes other spending, revenue collections, financing, investment maturities and payment timing. It also cannot show how much of the balance Anthropic can comfortably devote to opportunistic compute purchases.
Still, it helps explain why Amodei presents liquidity as a buffer. Anthropic has substantial resources alongside its obligations, giving it some ability to wait for clearer demand signals before making additional purchasing decisions.
“Relatively Little” Needs a Denominator
Amodei’s description of limited advance buying sits alongside nearly $900 billion in commitments. The two statements can coexist, but the available report leaves an important detail unresolved: relatively little compared with what?
Business Insider does not specify the share of expected future compute demand Anthropic intends to secure in advance, or the exact purchasing horizon covered by the strategy.
The cautious interpretation is that Amodei is describing how Anthropic intends to manage additional capacity decisions. His comments do not show that existing commitments have disappeared, become optional or can be canceled without consequences.
For investors, that missing denominator matters. A procurement policy becomes easier to assess when a company explains how much anticipated demand is already covered and how much remains exposed to future prices.
Cash Buys Flexibility, but Capacity Still Has to Exist
Amodei’s proposed sequence is straightforward: forecast demand using product momentum, commit cautiously, and purchase more capacity at shorter notice when usage justifies it. Business Insider identified Claude Code as one product informing Anthropic’s forecasts.
In this context, spot purchasing means obtaining additional computing resources on short-term, immediate terms instead of securing all anticipated needs years ahead. The reporting does not identify a particular cloud product, purchasing mechanism or price schedule, so readers should not assume a specific provider’s interruptible-instance rules apply.
The economic benefit is flexibility. If a product grows more slowly than expected, Anthropic avoids some additional advance commitments. If growth accelerates, it can deploy cash against demand that is already visible.
The tradeoff is exposure to the market when that extra capacity is needed. Availability and pricing remain uncertain. If many AI companies seek capacity simultaneously, cash gives Anthropic purchasing power, but it cannot guarantee that suitable resources will be available on acceptable terms.
This shifts part of the risk. Advance buying exposes Anthropic to paying for capacity before demand materializes. Buying later exposes it to the cost and availability of capacity after demand has materialized.
I’d assess this as a flexibility strategy with a price attached. Its success depends partly on whether avoiding unused capacity saves more than the company pays to obtain additional resources at short notice.
Gross Margin Provides a Second Buffer
Amodei also pointed to gross margins as protection, praising the revenue Anthropic can generate from each unit of compute, according to Business Insider. The report did not provide a quantified per-unit improvement supporting that assessment.
The underlying mechanism is sound: when customer revenue leaves more room after the direct cost of serving usage, a company has more capacity to absorb higher input costs. But that room depends on product pricing, workload costs and utilization.
CNBC reported that Anthropic’s IPO roadshow projected computing and infrastructure expenses would eventually represent 45% to 50% of revenue, supporting a long-term adjusted operating margin of roughly 35%. Those are company projections, not independently measured future outcomes. They also concern a broader financial model than the gross-margin buffer Amodei described.
Supplier Financing Helps Preserve the Cash Buffer
Frequently Asked Questions
3 questions
1What Is Anthropic’s Compute Dilemma?
Anthropic must buy enough computing capacity to serve customers and develop models without committing to costs that grow faster than revenue.
