Google thinks SpaceX’s Starship has to launch 1,600 times before space data centers get off the ground

Google launched its first orbital-compute prototype on a SpaceX rocket to test whether a Tensor Processing Unit can operate in space, initially running in 15-minute bursts due to power and thermal constraints. Project Suncatcher ultimately envisions 81 closely coordinated satellites processing AI workloads in parallel, with a two-satellite laser-link demonstration targeted for next year. Google estimates launch costs could approach $200/kg by 2035, but this would require Starship to deliver roughly 370,000 tons to orbit through about 1,800 launches over 10 years. Radiation testing supports five-year inference use, with an estimated logic error rate of roughly one in one million, though reliability remains insufficient for mega-scale training runs.

Analysis

This is strategically relevant to GOOG’s long-duration AI infrastructure optionality but immaterial to its 12-month earnings: orbital inference does not relieve the company’s terrestrial power, networking, or depreciation constraints until launch cost and on-orbit serviceability improve by orders of magnitude. The nearer strategic implication is that Google is preserving a proprietary accelerator path rather than conceding all frontier-compute economics to NVDA; however, radiation-tolerant inference is not evidence that TPUs can support high-availability training clusters. NVDA’s near-term moat remains defined by software, interconnect, memory bandwidth, and installed developer workflows—not chip survivability in a niche deployment environment.

PL has the clearest public-equity read-through because successful qualification could create a higher-value satellite-bus and payload-integration category, shifting its narrative beyond Earth-imagery data monetization. That optionality should not be capitalized without disclosed contract value, gross-margin structure, or repeat-order economics: bespoke compute spacecraft may initially consume engineering capacity and dilute margins. A successful multi-satellite laser-link demonstration over the next 12-18 months would matter more than the current test because it addresses distributed-compute networking, the real technical gating item after power and thermal management.

The contrarian conclusion is that the binding constraint is launch cadence and capital intensity, not demand for compute in sunlight. If reusable heavy-launch economics fail to reach a materially lower cost curve, this remains an R&D validation exercise and favors terrestrial power, cooling, and grid-infrastructure beneficiaries instead of satellite operators. Falsify the “no near-term impact” view if GOOG discloses committed orbital-compute capex, PL announces a scaled production award with defined unit economics, or launch providers demonstrate sustained high-frequency heavy-lift operations rather than isolated test flights.

Key Decisions for Investors

• No directional GOOG trade on this item; retain existing AI-infrastructure exposure but treat orbital compute as zero in 2026-27 earnings models. Reassess only on disclosed capex commitments or a commercial constellation contract.
• Place PL on an event-driven long watchlist rather than initiate immediately: buy only after a follow-on contract specifies satellite count, revenue, and gross-margin profile. A 12-18 month position would target multiple expansion from platform validation; exit if integration spending rises without backlog conversion.
• Maintain NVDA exposure independently of this development; do not interpret TPU space qualification as a competitive demand signal. The relevant falsifier is evidence that Google shifts material internal training workloads—not inference experiments—from Nvidia systems to TPUs.
• For a structural infrastructure expression, prefer terrestrial data-center power/cooling beneficiaries over orbital-compute proxies for the next 1-3 years; revisit only if heavy-lift launch cadence and delivered-cost data show a credible commercial inflection.