Sunny Madra, whose X profile describes him as a VP at Nvidia, wrote at 23:34 UTC on 3 November of SpaceX’s planned AI build: “Building the equivalent of 17 years of capacity in 18 months.” He added: “All on @nvidia gear!” SpaceX’s IPO prospectus puts a price on part of that gear: Anthropic agreed to pay $1.25 billion a month for compute that “includes approximately 325,000 NVIDIA GPUs”.
That is about $3,846 per GPU a month, or about $5.27 per GPU-hour, by our arithmetic. The fee also covers the CPUs, storage and networking the prospectus says back those GPUs, so it is an all-in price per GPU, not a chip-only one.
Business Pill · Usage-Based Pricing
An evergreen Business Engineer explainer on usage-based pricing. It teaches the general idea only and does not discuss SpaceX, Anthropic or Nvidia.
The key insight: SpaceX’s prospectus prices part of the Nvidia gear Madra describes: about 325,000 Nvidia GPUs rented to Anthropic for $1.25 billion a month, about $5.27 per GPU-hour all-in by our arithmetic, in data centers it lists in Memphis and Southaven.
SpaceX’s final IPO prospectus, filed on 12 June, says that in May 2026 it entered into Cloud Services Agreements with Anthropic for compute across COLOSSUS and COLOSSUS II. It says the capacity is “backed by hyperscale-class CPUs, exabyte-scale storage and high-speed networking and interconnects purpose-built for AI workloads”.
The customer “has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee”. The prospectus adds: “After the initial three-month period, the agreements may be terminated by either party upon 90 days’ notice.” At the full fee, that is $15 billion a year, by our arithmetic.
SpaceX describes the deal as “monetization of unused compute capacity”. It says the structure lets it “monetize a portion of the compute capacity in our infrastructure, while still permitting reallocation of that capacity for our own internal initiatives if needed in the future”.
From the fee in SpaceX’s 12 June 2026 prospectus to a price per GPU: Anthropic agreed to pay $1.25 billion a month for compute that includes approximately 325,000 NVIDIA GPUs, which is about $3,846 per GPU a month and about $5.27 per GPU-hour over a 730-hour month, by our arithmetic. The fee also covers the CPUs, storage and networking backing the GPUs; ramp-period fees were lower.
What Was Posted on X
Building the equivalent of 17 years of capacity in 18 months.
Sunny Madra (@sundeep), whose X profile describes him as a VP at Nvidia, quoting Trung Phan’s post on SpaceX’s 13GW build through 2027: “Building the equivalent of 17 years of capacity in 18 months. All on @nvidia gear!” The 13 GW matches what Musk said on SpaceX’s 20 October webcast; the 17-year comparison and the Nvidia claim are Madra’s post, not SpaceX disclosures.
Data centers in space looking pretty appealing right now
Elon Musk’s post of 06:12Z on 4 November on data centers in space, the subject of this morning’s first piece. It gives no reason and does not mention Nvidia, Anthropic or Texas.
What Musk Said About the Gear
On SpaceX’s 20 October webcast, Elon Musk said: “We’re tracking to be able to bring roughly 200,000 GB300s online per month for every month for the remainder of this year, and continuing that pace next year, with next year being primarily Rubin, Vera Rubin compute”, according to SpaceX’s edited transcript.
By our arithmetic, the roughly 325,000 GPUs in the Anthropic deal equal about 1.6 months of the pace Musk described. He also said: “I would be surprised if we brought on less than 13 GW of compute by the end of next year”.
SpaceX’s 10-Q reports 1.7 GW of nameplate compute at 30 September, against 0.6 GW a year earlier. It defines nameplate compute as “the number of GPUs installed in our data centers at the end of the period multiplied by their respective all-in power”, and reports it in gigawatts. The prospectus says SpaceX believes COLOSSUS II was among the first data centers to deploy GB200s and GB300s “at significant scale”.
Where the Gear Sits
The prospectus lists its data centers under one heading, Memphis, Tennessee and Southaven, Mississippi: “We operate a cluster of high-density data centers in the Greater Memphis Area extending into northern Mississippi along the state border”. COLOSSUS is on Paul R. Lowry Road in Memphis; COLOSSUS II is on Tulane Road in Memphis and on Stateline Road in Southaven.
Its Texas entries in that June list are Starbase, McGregor and Bastrop, none of them described as a data center. For Bastrop, the Starlink factory, it says SpaceX expects to add “the production of Starlink gateway antennas, solar cells and AI compute satellites”. Musk’s 06:12Z post on space data centers came 1 hour 55 minutes after AP called the Texas governor race, as we reported; the post gives no reason.
The Nvidia Side
Anthropic’s amended S-1 of 2 November says its agreements with SpaceX’s AI unit are for access to compute powered by Nvidia GPUs, as we reported. It expects to pay about $3.1 billion a month once all capacity is delivered, expected in December 2026. That is about 2.5 times the prospectus fee, by our arithmetic. It does not give a GPU count for that fee, so we do not apply the per-GPU figure to it.
Nvidia does not name its customers. Its latest 10-Q, for the quarter to 26 July 2026, says one direct customer represented 16% of total revenue, primarily in its Compute & Networking segment. We make no link between that customer and SpaceX. SpaceX’s prospectus says its “direct chip suppliers are dependent on a concentrated group of advanced semiconductor fabrication facilities”, and describes a Terafab initiative with Tesla aimed in part “to alleviate potential future chip shortages at SpaceX”.
As we read the prospectus, SpaceX sells the same installed GPUs two ways: as training capacity for its own models and as rented capacity, a structure it says still permits reallocation of that capacity for its own internal initiatives.
As we read the filings together, the 10-Q counts installed capacity in gigawatts while the Anthropic deal counts it in GPUs and dollars; the per-GPU-hour figure, by our arithmetic, is the one unit that links the two to a price.
SpaceX IPO prospectus, 12 June 2026
“Compute capacity provided includes approximately 325,000 NVIDIA GPUs, backed by hyperscale-class CPUs, exabyte-scale storage and high-speed networking and interconnects purpose-built for AI workloads.”
Three Implications
FOR ANTHROPIC The prospectus says either party may end the agreements on 90 days’ notice after the initial three-month period; Anthropic’s amended S-1 puts the expected fee at about $3.1 billion a month once all capacity is delivered.
FOR NVIDIA Nvidia’s latest 10-Q says one unnamed direct customer represented 16% of total revenue; SpaceX’s prospectus says its direct chip suppliers depend on a concentrated group of fabrication facilities.
FOR THE TEXAS QUESTION SpaceX’s June property list puts its data centers in Memphis and Southaven; for its Bastrop, Texas factory it says it expects to add the production of AI compute satellites.
The framework puts it this way: “The May 2026 Map of AI argued that the seven-layer AI stack is now governed from the bottom — energy, silicon, compute capacity.”
As we read it, the Anthropic deal puts a price on the silicon and compute-capacity layers at once: Nvidia GPUs, with the CPUs, storage and networking around them, rented by the month. The framework does not discuss SpaceX, Anthropic or this agreement.
The prospectus does not say which Nvidia GPU models make up the approximately 325,000, and the $5.27 is an average across the whole package at the full monthly fee; ramp-period fees were lower. Anthropic’s filing says the agreements were amended through October, and we have no per-GPU price for the amended terms.
The 17-year comparison is Madra’s wording on a third-party estimate, and the 13 GW is Musk’s statement, not a filed figure. The data center list is the prospectus’s of 12 June. We make no forecast on prices, capacity or the agreements.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Sunny Madra, whose X profile describes him as a VP at Nvidia, says SpaceX’s build is “All on @nvidia gear!” SpaceX’s own prospectus prices part of it: about 325,000 Nvidia GPUs rented to Anthropic for $1.25 billion a month, about $5.27 per GPU-hour all-in by our arithmetic, in data centers it lists in Memphis and Southaven. Its 10-Q reports 1.7 GW of that kind of capacity installed at 30 September.
A note on sourcing. The Anthropic deal terms, data center locations and supplier lines come from SpaceX’s 12 June 2026 prospectus on SEC EDGAR; nameplate compute from its 10-Q for the quarter to 30 September 2026; Musk’s remarks from SpaceX’s edited transcript of its 20 October webcast; Nvidia’s customer concentration from its 10-Q for the quarter to 26 July 2026. Madra’s post was read on X. Anthropic’s S-1/A figures are from our earlier reporting. Ratios are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
