SpaceX is reportedly lining up $40 billion of new debt to buy Nvidia chips, split into roughly $30 billion of bonds and $10 billion of bank loans; Apollo Global Management is said to be leading the process, with PIMCO evaluating allocations. That magnitude would reshape SpaceX’s capital structure and the AI-capex race in 2026. [1][2]

Management is pointing to a fast-growing Memphis/Southaven “Colossus” data-center business—backed by compute leases reportedly totaling $41.1 billion annually and a fresh $1.11 billion per month hosting deal starting December 1, 2026—to argue cash flow can service the leverage. [1][3][4]

If SpaceX deploys $40 billion into Nvidia systems at 2025 pricing, it could be buying about 1.0–1.1 million GPUs—enough to operate as a top-tier compute lessor—assuming it secures ≥1.0 GW of power, sustains high utilization, and avoids being boxed into 6–7% coupons for a decade. [4][5][6]

Back‑of‑envelope: interest burden

• Assume $30B of bonds priced near the Moody’s Baa composite yield of ~6.3%–6.5% and $10B of bank loans at ~7.0% all‑in (illustrative). Annual interest ≈ ($30B × 6.4%) + ($10B × 7.0%) ≈ $1.92B + $0.70B ≈ $2.62B per year. Even with a tighter bond print at 6.0%, total interest likely lands around $2.3–$2.7B. [6][5]
• Compare to AI‑compute ARR: CFO remarks include $1.11B/month (~$13.3B/year) starting December 2026, layered atop earlier deals; DataCenterDynamics pegs total compute leases at $41.1B annually post‑award. Gross coverage looks ample; net coverage depends on power, networking, cooling, and depreciation. [3][1]

Back‑of‑envelope: how much hardware is $40B?

• A typical 8×H100 system listed in 2024–2025 reports at ~$280k–$300k implies ~$35k–$37.5k per GPU equivalent; Reuters has cited H100 pricing from ~$16k to ~$100k depending on config and volume. Using $37.5k/GPU as a conservative blended figure, $40B buys ≈ $40,000,000,000 / $37,500 ≈ ~1.07 million GPUs. [6][7]
• Fortune reported SpaceX’s Colossus plan contemplates ~1 million GPUs, aligning with the capex math; the capital would fill Colossus 1–2 and seed a next wave if power and cooling arrive on time in Shelby and DeSoto Counties. [4]

2×2 framework: where this can go right—or wrong

• Axis 1: Demand visibility (contracted vs. speculative). Axis 2: Financing structure (on‑balance‑sheet vs. ring‑fenced/SPV).
  • Contracted + SPV: “Utility model.” If Apollo syndicates asset‑backed structures against take‑or‑pay compute, bonds price safer and cheaper; SpaceX preserves parent flexibility. [2]
  • Contracted + On‑balance: “Heavy core.” Execution is simpler, but consolidated leverage and refinancing risk rise at the corporate level.
  • Speculative + SPV: “Option with recourse‑lite.” Works only if GPUs retain resale value and there’s a deep secondary market.
  • Speculative + On‑balance: “Overhang.” Worst quadrant: elevated net leverage, front‑loaded capex, and utilization risk in fast chip cycles.

Named‑stakeholder breakdown

• Apollo Global Management: Wins by templating AI‑infra ABS—chips, racks, and contracted workloads packaged for pensions and insurers—earning fees and a marquee mandate. [2]
• PIMCO and other IG buyers: They want duration with coverage. If compute ARR is durable and cancellability is limited, they’ll consider BBB‑area paper at mid‑100s bps over Treasuries; Bloomberg Línea noted SpaceX credit spreads widening as risk repriced in 2026. [2][8]
• Nvidia: A 1‑million‑GPU order anchors Blackwell‑generation volumes and tightens ecosystem lock‑in across CUDA and NVLink in 2026–2027. [7]
• Anthropic/Google and other tenants: Near‑term compute certainty, potentially at premium pricing; mid‑term risk if contracts carry 90‑day outs and market prices normalize by late 2027. [4]
• TVA/MLGW (power): Memphis‑area power interconnection and on‑site generation are gating factors; delays or rate disputes would hit revenue timing more than 50–100 bps of bond spread would. [4][10]

Contrarian read

• Consensus: “$40B of debt is aggressive, but contracted AI ARR makes the math work.”
• Counter: The bond math only holds if compute scarcity persists and tenants stay locked. Fortune flagged 90‑day cancellation mechanics on marquee contracts; if hyperscalers overbuild by 2027, SpaceX could become a price taker facing ~$2.5B+ in annual interest with rapidly depreciating accelerators. In that world, ring‑fencing and take‑or‑pay terms become existential, not optional. [4][5][6]

Back‑of‑envelope calculation for 2026–2029 cash flows

• Total value if it runs full term: $150 million × 42 months (Jul 2026–Dec 2029) ≈ $6.3 billion. That’s $900 million for 2H26 and $1.8 billion per full year thereafter. [1], [4]
• Capacity lens: If Colossus 1 was ~220,000 Nvidia GPUs across >300 MW for Anthropic, Reflection’s tranche likely targets Colossus 2’s newer GB300 inventory. GB300 NVL72 packs 72 Blackwell Ultra GPUs per rack with an in‑rack 130 TB/s NVLink domain; selling time slices of such tightly coupled racks commands premium pricing because many training runs don’t decompose across disjoint clusters without heavy efficiency penalties. [3], [2]

A 2×2 to decode the 2026–2029 market

• Axis A: Model strategy
  • Open models (Reflection, select academia/defense pilots)
  • Closed models (OpenAI, Anthropic, Google)
• Axis B: Compute sourcing
  • Asset‑light buyers (rent compute): Reflection today; many Series B–D labs
  • Asset‑heavy builders (own DCs): Microsoft, Google; portions of OpenAI
• Where this deal sits: Open × Asset‑light. Advantages: speed to train, procurement optionality, and political palatability for U.S. government buyers who want source‑inspectable systems. Risks: termination rights (90‑day clause after the initial quarter) and renewal pricing exposure if GB300 supply tightens further. [1], [2], [4]

Named‑stakeholder breakdown (2026–2029)

• SpaceX: Proves Colossus is not a vanity project. It’s monetizable, modular, and now diversified across Anthropic (Colossus 1) and Reflection (Colossus 2). Post‑IPO, it becomes a credible third pillar beside Starlink and launch, with utility‑like revenue visibility. [3], [6]
• Reflection: Gains frontier‑class compute without a decade of data‑center capex and permitting. That turns its ~$25B valuation ambition from story into schedule: models out sooner, pilots with DOE and defense in a posture consistent with open procurement. [9], [1]
• Nvidia: Sells the picks and shovels, then benefits twice as labs rent time on GB300 NVL72 racks that entrench Nvidia’s full stack (NVLink, Quantum‑X, libraries). Every GB300 domain increases switching costs away from Nvidia. [2]
• Anthropic: Counter‑intuitively benefits from SpaceX scaling as a neutral lessor; its own deal locked up Colossus 1, and a bigger, healthier lessor reduces counterparty risk—until queues collide. [3]
• Utilities and regulators (TVA, MLGW; Mississippi Southaven build): Must keep adding firm power, water, and interconnects to maintain SLAs tied to Colossus near Memphis and the new Mississippi site. Delays would hit SpaceX’s compute P&L as contracted racks sit idle. [3], [5]

Contrarian read in 2026

• Consensus: “SpaceX is becoming a cloud provider.”
• My take: SpaceX is becoming an AI utility, not a cloud. Clouds multiplex VMs; Colossus monetizes whole‑rack, high‑bandwidth NVLink islands engineered for tightly coupled training and reasoning. The product isn’t elastic compute; it’s guaranteed access to a specific fabric topology with deterministic latency and power—closer to capacity offtake in energy markets than AWS‑style instances, and the contract form (fixed monthly, cancelable after a lock‑in) looks more like a power purchase agreement. [2], [1], [4]