The Physical Frontiers of Compute Expansion: OpenAI’s Billion-Dollar Financing and the Technical Reality Behind Orbital AI Data Centers

Since the explosion of generative AI at the end of 2022, the compute consumption of large language models has climbed at an exponential rate, but as of early 2026, this trend has shown no signs of convergence. OpenAI Chief Financial Officer Sarah Friar recently stated publicly that even though the company has just completed the largest-ever $122 billion private financing round, ChatGPT's weekly active user base of over 900 million continues to overwhelm its compute infrastructure, and the company is preparing for further fundraising. Almost simultaneously, SpaceX's AI arm, SpaceXAI, signed a historic partnership with Anthropic—OpenAI's primary competitor—to provide the latter with the entire ~300-megawatt capacity of its Colossus 1 data center. Furthermore, both parties expressed intentions to collaborate on exploring multi-gigawatt-scale orbital AI data centers. The temporal convergence of these two events is no coincidence; together, they point to the core dilemma currently facing the AI industry: the linear growth of compute supply can no longer match the exponential expansion of model inference demands, and the resulting capital-intensive solutions are dragging the entire industry into orbits that are unsustainable both physically and financially.

The compute shortage confronting current AI enterprises has reached a staggering scale. OpenAI President Greg Brockman disclosed in court testimony that the company's projected spending on AI computing resources in 2026 will reach approximately $50 billion, representing a multi-thousand-fold increase compared to 2017. This figure is still insufficient to cover total demand, as evidenced by the CFO already laying the groundwork for the next round of financing immediately after securing a record-breaking single financing round of $122 billion. During the same period, a structural inversion in AI compute costs is accelerating this trend: industry data shows that inference's share of total AI compute expenditure has climbed from about 50% in 2025 to approximately two-thirds in 2026, with some projections indicating this ratio could ultimately reach 80%. OpenAI's own inference costs corroborate this, rising from approximately $8.4 billion in 2025 to an estimated $14.1 billion in 2026. This means that even if enterprises temporarily contain compute consumption during the model training phase, continuously growing inference demand will act like a rising water level, rapidly submerging any temporarily acquired compute margins.

Turning attention to the immediate terrestrial component of the SpaceX-Anthropic partnership, the deal involves 220,000 Nvidia GPUs at the Colossus 1 data center, with analysts estimating an annual rent of approximately $50 billion. The underlying logic of this transaction warrants scrutiny: the fundamental reason Colossus 1 could be leased out as a single unit is that it is already obsolete. Elon Musk's xAI migrated its primary training workload to the newer Colossus 2 cluster in early 2026, rendering Colossus 1 a sunk cost. While leasing it to Anthropic secures a much-needed injection of compute for the latter, it is inherently a transfer of technical depreciation—Anthropic has secured a massive block of hardware that has already phased out of frontier training utility at a premium annual rent, while the impact of generational hardware degradation on inference quality has yet to be fully assessed technically. The architectural iteration cycle for AI chips is roughly 18 to 24 months; after two to three years of deployment, the performance-per-watt metrics of these 220,000 GPUs will lag significantly behind contemporary new-architecture products. This implies that the long-term financial commitments Anthropic locked in to alleviate short-term compute bottlenecks may ultimately translate into premium-priced access to computing resources with continuously diminishing marginal efficiency.

Looking further ahead into their grander vision of orbital AI data centers, SpaceX has been applying to the FCC since January 2026 to launch millions of satellites to construct an orbital data center constellation. According to the blueprint, these satellites will operate across multi-layered orbits spanning 500 to 2,000 kilometers, networking via inter-satellite laser communications, utilizing solar power, and leveraging cryogenic space radiation for cooling. However, a series of fundamental technical hurdles remain completely unresolved. Foremost among these is thermal management. Nvidia CEO Jensen Huang has publicly admitted that space lacks atmospheric convection, forcing heat dissipation to rely solely on radiation, which requires massive radiation apparatuses, rendering the system complex and costly—a challenge that will take years to solve. The liquid cooling solutions relied upon by terrestrial data centers not only add significant launch weight in microgravity environments, but the reliability of their circulation pumps also faces severe trials during long-term orbital operation. Vacuum environments render thermal conduction and convection near-obsolete, while space radiation continuously threatens chip operations—high-energy protons and particle streams can induce Single Event Upsets (SEUs) and latch-ups, leading to data corruption or hardware damage. The engineering limits of radiation-hardened chips remain unanswered today, and long-term reliability data for commercial GPUs in space environments is virtually nonexistent.

Communication bandwidth bottlenecks constitute another foundational constraint. Inter-satellite data exchange within an orbital data center must reach Terabit-per-second (Tbps) levels to support distributed training or large-scale inference tasks. However, recent systematic analyses from academia point out an order-of-magnitude gap between the Petabit-level data exchange rates inside terrestrial data centers and the merely Gigabit-level capacity of satellite-to-ground links; communication capability is the most fundamental restriction for space data centers. When orbital data centers provide AI inference services externally, data must travel back and forth between satellites and the ground. The resulting uncertainties in latency and bandwidth will directly impact user experience—an AI application delivering real-time services to global users cannot tolerate dozens or even hundreds of milliseconds of satellite-to-ground round-trip latency for every inference request.

An analysis of the cost structure similarly invites skepticism regarding this vision. Even at an equivalent 30-megawatt scale, the total cost of space compute is roughly 10 times that of terrestrial data centers, with launch costs accounting for 30% to 40%, satellite manufacturing representing 20% to 30%, and the remainder allocated to space environmental adaptation modifications, compute chips, and power systems. Optimistic industry estimates suggest that if first-stage rockets achieve 20 reuses, launch costs could drop to approximately 20,000 RMB per kilogram, thereby bringing space-based and ground-based costs toward parity. This assumption relies on rocket recovery and reuse technology maturing and achieving large-scale commercialization as expected. However, SpaceX's own repeated timeline shifts on the Starship project demonstrate that a massive window of uncertainty exists between experimental reuse and normalized operations for reusable launch systems. SpaceX itself conceded in an internal document that its orbital AI compute initiative relies on "unproven technologies" and "may fail to achieve commercial profitability." The phrasing of this statement itself constitutes an internal doubt regarding the project's viability.

From a macro industry perspective, these two events map out a clear trajectory: the expansion of terrestrial compute infrastructure has hit physical bottlenecks, forcing enterprises to seek breakthroughs in two directions simultaneously—one is unlimited capital raising to fight over scarce resources, and the other is reaching toward space, an environment with far harsher physical conditions, to seek incremental volume. Yet, neither direction addresses the core of the issue. The continuous escalation of capital density pushes the financing scales and valuation figures of leading enterprises into the trillions, while the corresponding commercial returns remain insufficient to cover these expenditures within a foreseeable cycle. When a company still needs to raise more funds after completing a $122 billion round, the patience of the capital market itself becomes a scarce resource. Space compute, on the other hand, bears excessive industrial expectations on a far from mature technical foundation—engineering challenges across the four dimensions of cooling, radiation hardening, communication, and cost remain stuck in the verification phase, leaving a technical climbing period of over a decade before commercially scaled orbital AI data centers can become a reality. Once the logic of dual expansion across both earth and space is pushed to its absolute limits, compute efficiency—the variable most easily overlooked amidst the noise—may emerge as the true, unavoidable focus of the next phase.