SpaceX AI: Probability-Weighted Sum-of-the-Parts Valuation

Thesis

A current probability-weighted sum-of-the-parts DCF based on SpaceX's filed segment results, disclosed AI contracts, buildout targets, capital requirements, dilution, institutional market checks, and an explicitly low-probability Elon Case.

Key chart

The argument

The correct question is not only, "Can SpaceX build 10 GW?" It is:

How much capacity becomes commissioned, accepted, billable, and durably profitable—and how much of the value survives refresh capex, financing, and dilution for common shareholders?

FactSpaceX reported $100.0 billion of cash and marketable securities and $39.4 billion of debt and finance leases at June 30, 2026. It also reported $3.5 billion of first-half operating cash flow against $28.5 billion of property-and-equipment purchases. The company has substantial liquidity and a profitable operating engine, but the current investment program is not self-funding.

FactIn Q2 2026, Connectivity produced $1.656 billion of operating income while Space and AI produced combined operating losses of approximately $1.799 billion. Starlink therefore covered roughly 92% of the other reported segments' operating losses. It did not cover the AI buildout's capital spending.

Company GuidanceManagement has described more than 2 GW of installed nameplate compute by year-end 2026 and a year-end 2027 outcome closer to 10 GW than 5 GW. Installed nameplate compute is not the same as generation-ready power, commissioned capacity, customer-accepted capacity, billable capacity, or actual facility demand. The model keeps those concepts separate.

Key findings

• Connectivity is the strongest current source of operating value. In the Base SOTP it contributes about $232.1 billion of enterprise value, even while other segments consume value.
• AI Infrastructure is the central swing factor. Its Base segment value is negative $468.3 billion because expansion and hardware-refresh cash needs remain larger than the modeled present value of operating cash flow.
• AI annualized run-rate revenue is not the same as recognized revenue or free cash flow. A December run rate cannot be treated as if it existed for the entire year.
• Current cloud contracts support very high near-term scarcity pricing, but their termination terms shift remarketing and repricing risk back to SpaceX.
• The Bull case reaches 37 GW, $849.1 billion of 2035 revenue, and $181.5 billion of 2035 FCFF, yet still produces only $65.47/share in the SOTP.
• The Elon Case can clear the current price, but it is assigned only a 5% weight and still has a roughly $177.9 billion external funding requirement in 2028 after a modeled 1.5 billion-share issuance.
• The reverse DCF is more reliable than any single target price. It shows the scale of long-run cash generation already embedded in the market value.

Counterarguments

The strongest optimistic argument is that a conventional DCF understates SpaceX's ability to create markets. SpaceX combines launch, communications, applications, customers, infrastructure, manufacturing, and financing under one organization. Faster construction can capture temporary scarcity pricing, each completed cluster can help finance the next one, and internal applications can capture more value than a conventional GPU lessor.

There is also evidence that the company can execute unusually quickly. Filed materials describe a roughly 130 MW Colossus cluster built in 122 days, an approximately 210 MW GB200 cluster in 91 days, and an approximately 220 MW GB300 cluster in 64 days. Those are genuine build-speed benchmarks, even though they do not prove that multi-gigawatt repetition will be equally fast.

The strongest cautious argument is that the model may still be giving too much credit to scarcity-era economics. SpaceX must simultaneously secure generation, gas, cooling, transformers, networking, millions of advanced GPUs, customer acceptance, and financing. Revenue per MW can fall as competing supply arrives, while the hardware still needs replacement.

The workbook preserves both arguments. Bull gives credit for strong execution and pricing; the Elon Case tests a stars-align outcome; Bear and Base show what happens if capital intensity outlasts the scarcity window.

Risks

Power, construction, and commissioning

Ten gigawatts of reported nameplate compute could require roughly 11–13 GW of gross facility power at a 1.10–1.30 power-usage-effectiveness ratio, before additional reliability reserve. A power plant, a powered data hall, an installed GPU, a commissioned system, and a billable customer workload are different milestones.

The Base case assumes 8 GW installed by year-end 2027. The Bull case assumes 10 GW. Either path would represent a material share of projected total U.S. data-center demand and requires physical evidence beyond a high-level pipeline target.

Customer concentration and contract durability

Those contracts validate premium near-term pricing. They also become terminable on 90 days' notice after their initial windows. SpaceX can try to remarket released GPUs, but replacing several gigawatts at the same price becomes harder as industry supply grows.

Capital spending and hardware refresh

The model uses initial all-in AI capital cost of $50 billion, $40 billion, and $35 billion per GW in Bear, Base, and Bull. It applies hardware replacement by installation vintage rather than charging a new cluster an immediate annual reserve. Modeled IT and network equipment is refreshed after four, five, or six years, with scenario-specific replacement-cost assumptions.

This matters because adjusted EBITDA can look excellent while free cash flow remains deeply negative. EBITDA does not deduct the capital required to build the next cluster or replace aging GPUs.

Funding and dilution

The Base liquidity schedule requires approximately $598.6 billion of cumulative external funding through 2035 if management follows the modeled expansion path without adapting it. Bear requires about $848.8 billion. Bull requires approximately $148.3 billion through 2028 and then becomes self-funding in the modeled schedule.

These are not financing forecasts. A rational company would slow, defer, partner, lease, or cancel uneconomic projects. They are a warning that debt, leases, joint ventures, customer prepayments, and equity cannot be treated as free substitutes for economic capex.

Internal AI double counting

SpaceX can either rent compute to an outside customer or use it internally for Grok, X, Cursor, and other applications. It cannot count a full external cloud fee and full application revenue on the same capacity.

The SOTP uses an illustrative, zero-sum internal compute transfer between AI Infrastructure and AI Applications. Combined AI value is authoritative until workload-level capacity and cost disclosures exist.

Terminal value and orbital optionality

Bear and Base receive no AI Infrastructure terminal value because the segment has not reached a stable, positive post-refresh cash-flow state by 2035. The model requires a longer forecast and three consecutive positive post-refresh FCFF years before allowing that terminal value.

Starmind is kept outside the terrestrial DCF as a probability-weighted option. Vacuum heat rejection, radiation, bandwidth, repair cadence, launch mass, and twelve-month total cost of ownership remain unproven.

Catalysts

Evidence that could justify higher values or higher Bull/Elon probabilities includes:

• at least 8 GW installed and commissioned by year-end 2027 with funded generation and cooling;
• customer acceptance and billable allocation for the delivered capacity;
• third-party AI renewals after 2029 at or above $25 million per MW-year;
• positive post-refresh AI Infrastructure FCFF for three consecutive years;
• Starlink revenue and operating cash flow approaching the modeled 2030 path without a severe ARPU decline;
• capacity-backed financing with disclosed collateral, maturities, covenants, and credible debt-service coverage;
• measured improvements in AI capex per watt from Terafab or other supply-chain integration;
• Starmind test data showing reliable operation and lower twelve-month cost than terrestrial compute; and
• a clearer fully diluted share count after pending awards and transactions.

Scenario analysis

The zero values in Bear and Base are common-equity floors. Their raw DCF equity values remain visible in the workbook as negative $479.2 billion and negative $225.5 billion funding-burden diagnostics.

Buildout and cash-flow checkpoints

The exit-ARR figures are point-in-time run rates. They are not the revenue recognized during 2027, especially when construction is back-loaded.

Sum-of-the-parts valuation

Negative segment values mean the present value of modeled explicit investment and losses exceeds the accepted terminal value. They do not mean a segment can literally be sold for a negative price. They identify where strategy must adapt before capital is committed.

Valuation analysis

Each segment uses a discount rate and terminal economics suited to its risk:

The model does not assume that today's scarcity pricing continues forever. It also normalizes terminal taxes rather than using temporary tax-loss carryforwards as a permanent benefit.

Reverse DCF: what the market price requires

At $133.11 per share and approximately 13.182 billion reported basic shares, the model estimates a basic enterprise value of roughly $1.694 trillion.

Using the Base interim cash flows and the model's terminal assumptions, supporting that enterprise value requires approximately:

• $449.6 billion of 2036 FCFF;
• $526.3 billion of normalized terminal NOPAT; and
• $2.31 trillion of terminal revenue at a 30% EBIT margin and 24% tax rate.

These are conditional hurdles, not forecasts. They show why "SpaceX reaches 10 GW" is only one part of the valuation debate.

Contract and financing anchors

Anthropic and Google

The two disclosed contracts provide the strongest evidence that near-term AI capacity can command unusually high prices. They cover 435,000 GPUs and approximately $26.0 billion of combined steady-state annualized fees.

The model does not assume those terms survive indefinitely. Cancellation windows, customer concentration, replacement pricing, and recontracting downtime are modeled explicitly.

Liquidity and hardware financing

FactSpaceX disclosed approximately $13.3 billion of debt associated with AI-hardware lease arrangements involving Valor Equity Partners. That is stronger evidence of equipment financing than the unconfirmed claim that NVIDIA will provide material vendor financing.

Financing affects when cash is needed and who bears the risk. It does not make the underlying GPUs, power plants, networks, or buildings free in an enterprise DCF.

Elon Case

The Elon Case is not an extrapolated Bull case and it is not management guidance. It is a deterministic test of what the valuation could look like if several of management's longest-shot claims align.

The bridge includes approximately:

• $30 billion of Space revenue in 2030 and $100 billion in 2035;
• $120 billion of Connectivity revenue in 2030 and $500 billion in 2035;
• $701 billion of external AI Infrastructure revenue in 2030 and about $1.08 trillion in 2035;
• $150 billion of AI Applications revenue in 2030 and about $847.5 billion in 2035;
• 9 GW of installed compute by year-end 2027, 30 GW by 2028, and 180 GW by 2035;
• declining AI construction cost per watt, while keeping full growth and vintage-based refresh capex; and
• a $512 billion probability-weighted Starmind option outside the terrestrial cash flows.

The case deliberately excludes separate revenue for Terafab chip sales, robotics, lunar factories, and mass drivers. It does not treat a $100 billion December 2026 run rate as full-year revenue, and it does not use permanent $40–50 million-per-MW scarcity pricing.

Even with those controls, the financing is incomplete. A modeled issuance of 1.5 billion shares at the reference price raises about $199.7 billion, yet the case still needs approximately $177.9 billion of external capital in 2028. The $229.92/share result assumes that gap is met through capacity-backed debt or joint ventures. More common equity would lower the per-share result.

Institutional and market cross-check

Secondary reports attribute targets of roughly $200–$240 to Citi, Bank of America, and JPMorgan, with higher long-shot values from Citi and Morgan Stanley. The underlying reports are paywalled, so their WACCs, site schedules, margins, financing waterfalls, and conflicts cannot be reproduced. The model records them as cross-checks, never as valuation inputs.

Assumption evidence map

The workbook separates filed facts, management claims, analyst assumptions, inferences, and speculation across its Sources Audit and Research Cross-Check tabs.

The strongest evidence supports:

• filed historical financials and segment results;
• cash, debt, backlog, share count, and stock-based compensation;
• the Anthropic and Google contract terms;
• current installed nameplate compute;
• the historical Colossus build-speed benchmarks; and
• issued shares already included in the legal cover count.

The least certain inputs include:

• the full site-by-site 2027 power and GPU schedule;
• normalized revenue per MW after scarcity fades;
• capex by silicon, network, building, generation, and contingency;
• application paid conversion and internal compute economics;
• long-run financing terms;
• Starlink V3 traffic monetization;
• Starmind thermal and maintenance economics; and
• terminal continuation probabilities.

Model checks

Version 2.0.1 passes the workbook's formula, source-link, segment-reconciliation, dilution, funding, and terminal-state checks. Important repairs include:

• replacing an immediate refresh reserve with installation-vintage replacement dates;
• correcting a 1,000× dilution-sensitivity error;
• removing already-issued restricted shares from the incremental dilution waterfall;
• correcting contract source mapping and earliest termination windows;
• adding recurring stock-based compensation as an economic expense;
• correcting 2026 periodization and pending Spectrum-payment timing;
• normalizing terminal taxes;
• suppressing invalid Base AI Infrastructure terminal value; and
• flooring Bear and Base common equity at zero only for probability weighting.

Passing checks confirms that formulas reconcile. It does not prove that the future assumptions will occur.

Disconfirming evidence

The current conclusion would become too cautious if:

• SpaceX commissions at least 8 GW by year-end 2027 with substantially lower capex per GW;
• multi-year renewals preserve pricing above $25 million per MW-year after 2029;
• application revenue scales without requiring a matching increase in internal compute cost;
• Starlink converts V3 bandwidth into revenue and cash flow faster than modeled;
• hardware life or residual value is better than assumed;
• customer prepayments, leases, debt, or joint ventures fund expansion with limited common dilution; or
• a longer forecast establishes stable positive AI Infrastructure cash flow.

The model would become more cautious if:

• power, GPU delivery, commissioning, or customer acceptance slips;
• current customers cancel or reprice sharply;
• all-in capex per GW is higher or refresh occurs faster;
• internal applications fail to monetize their assigned compute;
• debt and lease financing imposes tighter collateral or cash-sweep terms;
• fully diluted shares exceed the scenario waterfalls; or
• orbital testing fails to demonstrate competitive total cost of ownership.

What would change the conclusion

The highest-value new evidence is not another headline capacity target. It is a reconciled site-and-customer schedule showing, for each cluster, funded generation, facility readiness, delivered hardware, commissioning, customer acceptance, billing, contract duration, and replacement economics.

The probability-weighted estimate should be updated when new evidence changes a claim, scenario probability, segment cash flow, financing requirement, dilution denominator, or terminal gate. The public model summary, source archive, research cross-check, version history, and website update record should change together.

Primary and supporting sources

The current SpaceX source archive and model-specific source map include:

• the Q2 2026 Form 10-Q for segment results, cash flow, capex, liquidity, debt, backlog, concentration, awards, and share count;
• the final prospectus and filed offering materials for company structure and Colossus build benchmarks;
• the official Q2 2026 earnings transcript for management's compute, revenue, pricing, Starlink, Starship, and Starmind claims;
• the filed Anthropic and Google cloud-service disclosures;
• filed Cursor and equity-ownership disclosures;
• the official Starmind technical page and NASA thermal-control reference;
• SemiAnalysis as an aggressive secondary technical and commercial thesis;
• public institutional research from JPMorgan, Goldman Sachs, Bank of America, Citi, BlackRock, and Morgan Stanley;
• Synergy Research's neocloud market estimate; and
• secondary summaries of paywalled issuer targets, retained only as non-reproducible cross-checks.

Primary filings and executed agreements control over management aspirations, institutional targets, and secondary research.

Related spreadsheet model

Open the SpaceX AI Probability-Weighted SOTP Model to inspect all four scenarios, quarterly 2026–2028 forecasts, annual 2029–2035 forecasts, segment DCFs, contracts, hardware vintages, funding, dilution, sensitivities, reverse DCF, source audit, model checks, and version history.

Article revision summary

Version 2.0.1 publishes the repaired probability-weighted SOTP as the primary valuation framework. It adds the reader summary, research cross-check, vintage-based refresh schedule, claim-constrained Elon Case, segment discount rates and terminal gates, explicit funding warnings, corrected dilution, recurring stock-based compensation, normalized terminal taxes, and a common-equity floor for probability weighting. The earlier consolidated DCF is retained only as a reconciliation diagnostic.