全球人工智能：花旗太空与AI虚拟教学中的AI洞察

Global Artificial Intelligence AI Insights from Citi’s Space & AI Virtual Teach-In CITI'S TAKE Heath TerryAC+1-212-723-4624heath.terry@citi.com Across 10 hours of sessions with public company execs, private companyfounders, and various experts, our inaugural Space & AI Virtual Day framedthe scale of the opportunity at the intersection, the benefits of verticalintegration (like SpaceX’s), and the need for investment across power,semiconductors, compute, models, software, and distribution. The AIopportunity is increasingly an infrastructure and cost-curve challenge, withinitiatives such as Terafab, space solar, and orbital compute positioned aspotential accelerators for AI growth. While demand for compute remainsstrong, as evidenced by recent deal terms, focus is likely to shift towardcapex efficiency, enterprise adoption following the Cursor acquisition, andexecution milestones, like today’s Grok 4.5 release. Michael Rollins, CFAAC+1-212-816-1116michael.rollins@citi.com John GodynAC+1-212-816-8014john.godyn@citi.com Shelby Spencer+1-212-816-0416shelby.spencer@citi.com Vertical integration the playbook.Across sessions, AI was framed as the new stackfor enterprises to internalize, spanning from power all the way to distribution. Thekey debate emerging was the ability of companies like SpaceX to leverage verticalintegration to lower the unit cost of intelligence ahead of the next “event horizon”,the development of Recursive Self Improvement (RSI). From our meetings withindustry experts from Morgan Lewis and Recon Analytics, Starlink will likely needterrestrial solutions to be a full mobile service provider. Within the U.S. market,building its own network was viewed as the likely path forward, rather than buying atelco or obtaining an MVNO. The counter-response from telco and cable incumbentsis likely to be a more aggressive push to offer converged service bundles anchored byfiber/cable broadband. Ashley Kim+1-212-816-6689ashley.kim@citi.com Caitlyn Walsh+1-212-816-6692caitlyn.walsh@citi.com Max Lesnik+1-212-816-2382max.lesnik@citi.com Demand > Supply.Multiple sessions pointed to strong near-term visibility for AIcompute demand, with cloud, software, and adjacent infrastructure providers stillcapacity-constrained over the next 12–24 months, even with previously captiveexcess supply being unlocked by deals like those signed by SpaceXAI last month.Space-based compute was seen as additive, not a replacement for terrestrial datacenters, as latency-sensitive workloads and existing terrestrial ROI remain relevant,and scaled orbital compute still years away, though how many years is in question. Janna Withrow+1-212-723-0439janna.withrow@citi.com Cost-curve levers.Terafab and orbital compute were seen as the clearest cost-curve and supply chain levers for AI, particularly given pressures in EUV, advancedpackaging, memory allocation, and foundry capacity. Terafab was positioned as aneffort to reduce manufacturing cycle times by 10–30%, simplify chip designs, andremove supply-chain margin layers. Similarly, AI satellites require power systemswith radiation resistance, high power-to-weight ratios and scalable manufacturing. Model capability, inference economics and distribution control.Grok 4.5 is atimely proof point that SpaceXAI is pushing forward at the model layer, with today’srelease emphasizing coding, agentic tasks, knowledge work, training across its fleetof GB300s, 80 TPS, and pricing of $2/M input tokens and $6/M output tokens,which compare favorably to frontier lab leaders. Multiple speakers spoke toenterprise customers desiring lower inference costs, open model customization,data control, neutral infrastructure partners, and monetization surfaces. See Appendix A-1 for Analyst Certification, Important Disclosures and Research Analyst Affiliations. Setting the Stage: Frontier Models & Key Themes Across AI - Group VideoConference We hosted Jason Warner, Co-Founder/Co-CEO of Poolside and former CTO ofGitHub, for a wide-ranging discussion on frontier intelligence and AI infrastructure.Warner stated that current open models exceed the capabilities of frontier systemsfrom roughly 18 months ago, and positioned knowledge work as a roughly $29Tcategory where AI may increasingly substitute for or augment human labor. Warnersees model progress moving toward a small number of scaled players, with theinfrastructure and research systems needed to improve models recursively. He alsodescribed Poolside’s approach as industrializing model development, with its10GW data-center project in West Texas dedicated to producing gigawatts ofpower and compute for model training and token economics. A central theme wasthat customers may increasingly seek to own or control more of the stack ratherthan rent intelligence solely through APIs. Warner believes cost discipline shouldbecome more important as enterprise AI matures, and that most enterprise tasksdo not require frontier-level capability as cheaper open models can han