Anthropic's Prospectus Turns Its $15B SpaceX Compute Contract Into an SPCX Test

The prospectus gives the customer side of a contract SpaceX already disclosed#

Anthropic's newly reported IPO financials create a more direct listed-options question in SpaceX than in the usual cloud-stock proxies. Reuters reported from a prospectus it reviewed that Anthropic generated about $4.59 billion of revenue in 2025, spent $7.33 billion on compute and infrastructure, posted an $8.06 billion operating loss, and ended the year with $20.28 billion of cash, cash equivalents and short-term investments. Those figures matter to SPCX because SpaceX has already disclosed a direct commercial relationship that is unusually large and unusually specific. Under compute-services agreements signed in May 2026, Anthropic agreed to pay SpaceX $1.25 billion per month for access to COLOSSUS and COLOSSUS II capacity through May 2029, after an initial ramp period. The agreement covers roughly 325,000 NVIDIA GPUs and can be terminated by either side with 90 days' notice after the initial three-month period. The useful options question is therefore not whether Anthropic's IPO is good or bad for SpaceX. It is whether new information about a major SPCX customer's economics changes how the market prices SpaceX's AI-infrastructure exposure.

Anthropic confidentially submitted its draft S-1 on June 1. The detailed financial picture now reported from the prospectus adds something that SpaceX's own filing could not provide: a view of the customer's historical revenue scale, cost intensity, losses, liquidity and revenue concentration.

The contrast is striking.

Anthropic's 2025 compute and infrastructure spending of $7.33 billion was about 160% of its roughly $4.59 billion of revenue. Its operating loss of $8.06 billion was about 176% of revenue. Reuters also reported that two customers represented nearly one quarter of Anthropic's revenue and that many large customers were not locked into long-term contracts.

Those figures do not show that Anthropic cannot support its current infrastructure commitments. They do show why counterparty quality is now a more measurable part of the SpaceX compute story.

The comparison with the SpaceX contract needs careful timing. At $1.25 billion per month, the agreement has a $15 billion annualized headline run rate once fully ramped. That is a little more than twice Anthropic's entire 2025 compute and infrastructure spend.

But the figures belong to different periods. The $7.33 billion number describes actual 2025 spending across Anthropic's infrastructure stack. The $15 billion figure annualizes one contract entered in May 2026. It should not be presented as evidence that Anthropic suddenly doubled total compute spending or that every dollar of the contract will be recognized by SpaceX in one calendar year.

The mismatch is the research value. A historical cost base can now be compared with a current contractual commitment without pretending they are the same metric.

SpaceX is a cleaner listed exposure than the cloud proxies for this question#

Anthropic has deep relationships with Amazon, Google and Microsoft. That makes AMZN, GOOGL and MSFT relevant to Anthropic's growth, but it also makes attribution difficult.

AWS is Anthropic's primary cloud and training partner, distributes Claude through Bedrock and has secured more than $100 billion of future Anthropic technology commitments over ten years. Google provides TPU capacity and cloud distribution. Microsoft has a separate Azure capacity and investment relationship.

Those companies are enormous diversified businesses. Anthropic compute demand reaches each through several channels at once.

SPCX is different for this particular research question because the economic bridge is documented in a single customer contract with a stated monthly payment, duration, hardware capacity and termination provision. The link is not merely thematic.

That does not make SPCX a pure Anthropic proxy. SpaceX has Space, Connectivity and AI businesses, operates its own AI workloads, sells compute to other parties and has many other drivers of equity volatility. It does make the Anthropic relationship more directly observable than a generic statement that higher Claude usage should help a cloud provider.

The $42B net loss headline should not replace the operating picture#

Reuters also reported a 2025 net loss of nearly $42 billion. That number is much larger than the $8.06 billion operating loss because it included roughly $34 billion of accounting expense related largely to financing instruments whose estimated value increased.

For the SpaceX counterparty question, the operating loss and cash position are therefore more useful starting points than the net-loss headline alone.

The difference matters because a financing-related accounting charge is not the same thing as a cash payment for GPUs, power or cloud capacity. A counterparty analysis should separate operating economics from financing-accounting effects before connecting the prospectus to an infrastructure supplier.

Even the $20.28 billion year-end cash figure has limits. It is a December 31, 2025 balance-sheet snapshot. Anthropic raised substantial additional capital in 2026, so that number cannot be treated as the cash available today to meet the SpaceX agreement.

The prospectus improves the baseline. It does not freeze the company in its year-end 2025 financial condition.

Customer concentration and supplier concentration are different risks#

The disclosure that Anthropic's two largest direct customers produced roughly one quarter of 2025 revenue is relevant, but not because those customers can automatically be mapped to the same companies that provide Anthropic's compute.

Customer concentration describes where Anthropic earns revenue. Infrastructure concentration describes where Anthropic spends money and obtains capacity.

Those are different sides of the business.

Anthropic's compute stack spans several providers and architectures. It has large AWS commitments, Google TPU capacity, NVIDIA-linked infrastructure and the SpaceX compute agreement. A reduction in one revenue customer's spending does not mechanically identify which infrastructure contract would be changed.

Likewise, a large compute commitment does not show which Anthropic end customer is economically supporting it.

For SPCX options, the observable question is narrower: does new evidence about Anthropic's funding, operating trajectory or compute commitments change the perceived distribution of outcomes around SpaceX's third-party AI-infrastructure revenue?

That question can be tested without guessing who Anthropic's two largest customers are.

September 28 provides a useful pre-disclosure options baseline#

The timing of the Reuters report creates a clean event boundary.

Reuters published its prospectus report after the U.S. listed-options session had already ended on September 28. That means the day's SPCX option quotes could not have incorporated the newly reported Anthropic financial details.

A delayed Public snapshot for the October 2 expiration later displayed SPCX around $146.31. At the $150 strike, the call ask was $2.01 and the put ask was $4.95, for a combined offer-side quote of $6.96.

That $6.96 figure is not an expected move. It is the sum of two asks from a delayed retail feed, not a synchronized midpoint calculation. It also contains every other source of SPCX uncertainty through the October 2 expiration.

Its value is as a timestamped pre-report reference.

The first stronger observation comes on September 29, when the same contracts can be measured after the market has had a normal session to process the Anthropic disclosures. The comparison should use the same vendor, the same quote convention and similar moneyness.

Short-dated options cannot answer the full contract question#

The October 2 contract can show whether the prospectus report creates an immediate change in short-horizon uncertainty. It cannot measure the economics of a compute agreement that runs through May 2029.

The termination clause also changes the relevant horizon. After the initial three-month period, either party can terminate with 90 days' notice. That means the economically relevant risk is not simply a fixed stream of $1.25 billion monthly payments through 2029.

Longer-dated SPCX expirations can contain more of the contract horizon, but even the currently listed January 2029 contract expires before the stated May 2029 endpoint. No single listed expiration maps perfectly onto the full agreement.

A useful term-structure study would therefore compare several clocks rather than searching for one "Anthropic option."

Near-term expirations can test reaction to the prospectus disclosure. November and December contracts can span more of the anticipated IPO process. Longer maturities can reflect broader beliefs about SpaceX's third-party AI-infrastructure business, but they also contain far more unrelated SpaceX events.

The farther the expiry, the harder causal attribution becomes.

The next test is whether SPCX reprices differently from the wider AI complex#

A September 29 change in SPCX options would still not prove that Anthropic caused it.

The strongest competing explanation is broad technology volatility. SpaceX, NVIDIA, Amazon, Alphabet and other AI-linked companies can reprice together because of rates, semiconductor news, market-wide risk or another infrastructure headline.

The cleaner observation is relative.

If SPCX's near-term implied volatility or same-moneyness option premium changes materially after the prospectus report while broader technology controls remain comparatively stable, the Anthropic contract becomes a more plausible hypothesis to investigate. If the entire AI complex moves together, a common factor becomes more credible.

The same logic applies to longer expirations. A persistent change in SPCX's term structure would be more interesting if it aligned with new information about Anthropic's compute commitments, financing or the status of the SpaceX agreement rather than with a broad market repricing.

Options can reveal that the distribution changed. They cannot reveal the motive of every participant.

What would change this interpretation?#

Three future observations matter most.

First, a synchronized September 29 SPCX option surface can establish whether the prospectus report produced any incremental repricing relative to the September 28 baseline.

Second, Anthropic's eventual public registration materials can clarify the current infrastructure-obligation schedule, liquidity position and concentration risks more completely than the financial excerpts reported so far.

Third, future SpaceX filings can show how material third-party compute revenue becomes inside the AI segment and whether the Anthropic agreement remains in force on the disclosed terms.

The important new information is not simply that Anthropic grew quickly while spending heavily. It is that public investors can now begin connecting the financial profile of a large private AI customer to a specific, disclosed SPCX revenue relationship.

That turns the Anthropic IPO process into a counterparty-research problem for SpaceX options rather than another generic AI headline.