NextFin News - Anthropic's confidential IPO prospectus discloses up to $84.5 billion in computing agreements with Elon Musk's SpaceX, nearly double the roughly $45 billion previously reported, and lays bare a decade-long, $518 billion infrastructure plan in which about 80% of the spending is non-cancelable or must be paid regardless of usage. The filing, reported on September 29, 2026, frames the commitments as a necessity rather than an option: the company told investors that future demand for advanced AI will be
"limited principally by the availability of compute."The xAI portion of that plan is the one flexible piece - most of it can be canceled on 90 days' notice - which makes it the cleanest signal in the filing of how confident Anthropic really is that demand will keep compounding.
The Filing: A Ceiling That Doubled Overnight
The $84.5 billion figure is an upper bound on spending for Nvidia-based computing capacity supplied by SpaceX's xAI unit through 2029, not a firm take-or-pay obligation. That distinction is the difference between a backlog and an option. When SpaceX filed its own IPO paperwork in May 2026, the arrangement was disclosed as $1.25 billion a month through May 2029 - a 36-month run totaling roughly $45 billion, with either party able to terminate on 90 days' notice. The new prospectus nearly doubles the ceiling while leaving the escape clause intact.
The revision sits inside a much larger buildout spanning six partners. Anthropic expects to spend at least $518 billion on cloud, computing, and infrastructure obligations over the next decade. About 80% of that sum is non-cancelable or structured on take-or-pay terms. The company argues the commitments are necessary because access to computing power is becoming the key constraint on AI development.
The non-cancelable core is where the real balance-sheet risk concentrates:
- Google: at least $111.1 billion, April 2026 through July 2033.
- Amazon: at least $110 billion, May 2026 through April 2036.
- Microsoft: $31.4 billion, November 2026 through May 2033, cancelable only in the event of an uncured material breach.
- Broadcom: roughly $161.2 billion in equipment lease obligations, non-cancelable by either party except in default.
Add the four together and the hard core runs to roughly $414 billion. Against that, the xAI arrangement is the flexible outlier. In the filing, the company said of its Google agreement:
"If our actual spend falls short, we must pay Google the difference."It added that similar terms apply to the Amazon agreement. No such language surrounds the SpaceX commitment.
The overall figure is comparable in scale to OpenAI's Stargate project, a $500 billion AI infrastructure plan whose costs are expected to be shared among OpenAI, SoftBank, Oracle, and MGX. Two rival labs, two half-trillion-dollar infrastructure bets, one decade. The AI buildout race has a price tag now, and it is measured in hundreds of billions.
The Financials Behind the Bill
The commitments are easier to absorb when set against the company's actual economics, and the gap is wide. Anthropic's revenue grew roughly twelvefold in 2025 to nearly $4.6 billion. Over the same period it posted a net loss of about $42 billion, though roughly $34 billion of that was an accounting charge reflecting a higher estimated value for financing arrangements that could eventually convert into stock. Excluding writedowns tied mostly to earlier fundraising, the operating loss topped $8 billion.
Total operating expenses reached $12.65 billion, with compute and infrastructure accounting for $7.33 billion - a threefold surge from 2024 and more than half of all spending. The company ended 2025 with $20.28 billion in cash, cash equivalents, and short-term investments. At the current burn profile, that cash balance covers less than three years of 2025-level compute spending alone, before a single dollar of the $518 billion plan comes due.
Two customers generated nearly a quarter of 2025 revenue, and the filing warns that many of the largest clients are not locked into long-term contracts and could reduce or stop spending. That concentration sits uneasily beside a decade-long spending plan that assumes demand keeps compounding. The filing is explicit about the dependency:
"If the compute we have access to from third parties is curtailed, repriced, or terminated ... our business, financial condition, and results of operations could be adversely affected."
Why SpaceX - And Why the Rival Became a Supplier
The arrangement is as much strategic as it is financial, and the politics are awkward. Musk has been a vocal critic of Anthropic, writing in February 2026 that the company "hates Western Civilization" and asking whether there was "a more hypocritical company than Anthropic." Weeks after signing a deal that gives Anthropic access to his data-center capacity, the two companies are structurally linked.
The logic runs both ways. For Anthropic, xAI's Colossus facilities offer dedicated, Nvidia-based capacity outside the three cloud giants that also compete with it in models. For SpaceX, which merged with xAI earlier in 2026, the arrangement monetizes data-center capacity built to train Grok and turns spare training infrastructure into a revenue stream. The launch business does not consume GPUs; a data center can sit idle between rocket cycles.
Analyst estimates put SpaceX's combined AI infrastructure revenue - stacking the Anthropic deal with other disclosed compute contracts - above $28 billion a year once all agreements are billing simultaneously, a scale that would rank compute alongside launch services and Starlink as one of the company's largest revenue lines. The $1.25 billion monthly Anthropic payment alone runs to $15 billion a year, close to SpaceX's entire standalone revenue before the xAI merger.
Market Reaction: A 3% Move That Says More Than a 20% One Would
SpaceX shares added about 3% on the report, trading near $145.47 with a market capitalization of roughly $1.96 trillion. The stock remains below its post-IPO peak of $225.64 on June 16, 2026, when it briefly commanded a market cap near $3 trillion, and above its June 12 offering price of $135. Since the June debut, the shares are down roughly 8%.
The move is modest, and that is the point. Investors are treating the $84.5 billion as potential revenue, not guaranteed revenue - precisely because of the 90-day exit clause. A contract either party can terminate on a quarter's notice is a revenue stream with an asterisk, not a backlog. Had the market read the figure as firm, the reaction would have been larger; the muted response is the market pricing optionality, not obligation.
Cyclical or Structural: The Hybrid Read
The filing's central thesis is structural: compute supply, not model demand, is the binding constraint, and it will remain so for a decade. That is a regime-shift claim, and it carries the burden of proof. The evidence Anthropic offers is its own spending behavior and the concentration of advanced chip capacity in the hands of a few suppliers.
But there is a cyclical leg layered on top. The AI infrastructure boom of 2024-2026 was financed on expectations of compounding model usage that have not all been realized. If utilization rates fall, if model efficiency improves faster than demand grows, or if new chip capacity from Nvidia, AMD, and custom silicon comes online faster than expected, the "compute is scarce" thesis weakens. That is why the 90-day clause is the most important line in the filing: it is the market's way of saying the structural claim is not yet fully trusted.
The right read is a hybrid. The long-dated, non-cancelable Google, Amazon, Microsoft, and Broadcom commitments - roughly $414 billion of the $518 billion - are the structural bet. The xAI piece is the cyclical hedge, an option to scale up or down without a decade-long penalty. Anthropic is effectively telling investors: we are certain enough about the long run to lock in $414 billion, and uncertain enough about the near term to keep $84.5 billion flexible.
Second-Order Effect: The Realignment of the AI Oligopoly
The first-order story is simple: Anthropic needs chips. The second-order story is more interesting. The filing lays bare a market in which Amazon, Google, and Microsoft simultaneously act as Anthropic's investors, customers, cloud providers, distributors, and competitors - roles the company says create incentives that
"may not be fully aligned"with its own.
That conflict is the hidden driver of the xAI deal. Anthropic is diversifying away from a cloud-only model toward dedicated data centers and directly leased chips not just for capacity, but for bargaining power. Every dollar committed to xAI is leverage against the three giants when renewal terms come up. The same logic explains the AMD arrangement disclosed alongside it: the chipmaker agreed to buy up to $5 billion of Anthropic stock and to supply AI computing capacity expected to exceed $20 billion. Equity alignment plus supply is a different relationship than a cloud bill.
The counterparty risk cuts the other way, too. Anthropic's entire model roadmap depends on infrastructure it does not own. The company is moving toward directly leased chips and dedicated facilities precisely because the cloud-only model leaves it exposed to curtailment, repricing, or termination by partners that are also building rival models.
The Counter-Thesis
The strongest case against this read is that the $84.5 billion number is being over-interpreted. It is a cancelable ceiling, not a firm order book, and the market has already priced Anthropic's dependence on third-party compute into the AI trade. SpaceX's shares rose only 3% because the incremental information content is low: everyone knew Anthropic was a major xAI customer; the filing merely revised the ceiling upward.
There is force in that view. But it misses the asymmetry. If the number were a firm $45 billion, the revision to $84.5 billion would be noise. Because it is cancelable, the revision is a signal about expected utilization, not just expected spend. Anthropic is telling investors it may need nearly twice as much xAI capacity as the market assumed. That is a statement about demand confidence - and about how quickly the company expects to burn through dedicated GPU capacity.
The falsifying signal is specific and observable: if Anthropic's utilization of xAI capacity stays flat or declines over the next two quarters while its revenue keeps growing, the "compute is the bottleneck" thesis weakens materially. Conversely, if the company exercises expansion rights or converts flexible spend into firm take-or-pay commitments in the final S-1, the structural read is confirmed.
What to Watch
The IPO itself is the first catalyst. Anthropic confidentially submitted its draft registration statement to the SEC in June, and the public sale is expected after the November US midterm elections, at a valuation above $2 trillion - more than double its own estimated valuation of $965 billion in May. The listing would follow SpaceX's blockbuster June debut, which valued Musk's company at $1.77 trillion at pricing.
Three signals matter most:
- Whether the final S-1 converts any of the cancelable xAI commitment into firm take-or-pay terms.
- Whether the two top customers - nearly a quarter of 2025 revenue - remain concentrated or diversify.
- Whether compute and infrastructure spending, already $7.33 billion in 2025, continues to outpace revenue growth.
A fourth signal sits outside the filing: whether the broader AI capital cycle holds. If OpenAI's Stargate buildout slows, if hyperscaler capex guidance is cut, or if chip lead times normalize, the scarcity premium embedded in these contracts erodes.
Conclusion: Who Benefits, Who Is Exposed
The beneficiaries are clear. Nvidia supplies the underlying GPUs across the xAI arrangement. SpaceX and xAI gain a revenue stream that monetizes data-center capacity without launching a single rocket. The three cloud giants lock in a decade of minimum payments that underwrite their own capital expenditure cycles. Broadcom holds a largely non-cancelable equipment lease book.
The exposed party is Anthropic itself. A $518 billion decade-long plan, 80% of it non-cancelable, against $4.6 billion of annual revenue and an $8 billion operating loss, is a bet that demand compounds without interruption. The xAI arrangement is the pressure-release valve - the one commitment flexible enough to shrink if the AI spending cycle turns.
Short term, the stock reaction says the market sees optionality, not obligation. Medium term, the test is whether Anthropic's revenue can grow fast enough to make $7.33 billion in annual compute spend look like a bargain rather than a burden. Long term, the question is whether compute scarcity is a structural feature of the AI economy or a cyclical shortage that new capacity will cure.
The filing answers one question and poses another. Anthropic has secured the capacity. The harder part - making it profitable - is still ahead.
Explore more exclusive insights at nextfin.ai.
