SpaceX is Planning to Spend Up to $500 Billion on Data Centers. Here's Why Elon Musk is Pushing All-In

SpaceX’s AI infrastructure expansion from 1.4 GW to 6–10 GW is underwritten by extremely strong inference economics: Semi Analysis estimates ~$100B of inferencing value per GW (8.3x payoff). SpaceX has signed deals paying it $1.25B/month from Anthropic for ~300MW and $920M/month from Google for ~110,000 GPUs (2026–mid-2029), implying ~$31B and ~$48B annual value per GW, respectively. The article highlights risks of a “bullwhip” if spot compute prices roll over or if TSMC’s advanced-node capacity constraints (U.S.$265B Arizona commitments; $60–$64B 2026 capex; full-year revenue guide +40% YoY) eventually shift supply beyond demand.

Analysis

The key market mechanism is not "AI demand" but who controls the scarce inputs. TSM sits at the choke point for advanced-node supply, so it should capture the first-order economics while hyperscalers absorb the depreciation, financing, and execution risk that comes with racing ahead of monetization. That makes TSM the cleaner duration trade: it benefits whether the end customers are MSFT, GOOGL, or another platform, while the platform names are exposed to a future margin reset if incremental inference revenue does not compound as fast as capex.

Near term, the market usually rewards capex acceleration because it signals share defense, but that tends to become a multiple problem once investors start capitalizing lower free cash flow and higher maintenance intensity. MSFT is the highest-quality participant, yet its AI upside is partly offset by the risk that growth is being purchased rather than organically harvested; GOOGL has more to prove because catch-up spending can depress ROIC before product momentum shows up in the numbers. AMZN is the stealth relative loser if the arms race forces AWS into a spend cycle where pricing power is less visible and the return on new dollars is harder to see.

The contrarian risk is that consensus is extrapolating today’s inference economics too linearly. If model efficiency improves, spot compute prices soften, or advanced-node capacity finally broadens, the current scarcity rents can flip into an overbuild cycle and compress the entire AI infrastructure complex over 6-18 months. The main falsifiers are a slowdown in hyperscaler capex guides, weakening compute rental rates, or any sign that TSMC’s capacity expansion closes the gap faster than expected.

Key Decisions for Investors

• Long TSM vs short GOOGL for a 3-6 month horizon: TSM monetizes the scarcity bottleneck, while GOOGL is forced to spend defensively to keep up. Risk/reward is better than shorting the broader AI basket because the pair isolates capex intensity versus supply-chain control.
• Use MSFT as a quality long only on pullbacks, not as a chase buy: the stock can work over 6-12 months if AI demand stays tight, but the near-term overhang is lower FCF conversion and higher scrutiny on incremental returns. Prefer calls or a smaller cash equity position rather than outright size.
• Avoid chasing any illiquid private-market exposure to the buildout; the better public proxy is TSM. If TSM starts to lag while hyperscaler capex keeps rising, that would be an early warning that supply is catching up and the trade is overcrowded.
• Set an alert on GOOGL and MSFT capex commentary next quarter: if spending stays elevated while cloud growth decelerates, rotate out of platform longs and toward suppliers. If capex growth inflects down before revenue re-accelerates, the bull case is breaking.
• Watch for a relative short in AMZN if AWS begins matching the spending cadence without an equal re-rating in growth expectations. That is a months-long catalyst, not a days-long trade, and it only works if management tone implies the capex arms race is becoming mandatory rather than optional.