$1 trillion AI capex by 2028: Justified or inflated?

Episode Summary

Executive Summary: The episode debates whether AI capex is heading toward a $1T peak by 2028. Tom O’Malley argues demand from AI labs, hyperscalers, enterprises, and sovereigns—driven by agentic AI, inference growth, and recursive self-improvement—supports even higher spending than consensus, though power and infrastructure are real constraints. The host remains skeptical, questioning ROI, funding feasibility, and the risk of overbuild.

Main Topics: AI capex outlook and the $1T 2028 peak (Priority: 5/5): The central thesis is that AI-related capital spending can peak around 2028 at roughly $1 trillion, well above Street expectations, because compute demand is still early and expanding rapidly. AI labs as demand signals (Priority: 5/5): OpenAI, Anthropic, and other labs are treated as leading indicators of real compute demand, with forecast revisions and spending plans feeding hyperscaler investment decisions. Recursive self-improvement and agentic AI (Priority: 4/5): Tom argues that model improvement loops and agentic workloads will drive sustained training and inference demand, justifying continued capex into 2028. Funding constraints and ROI skepticism (Priority: 5/5): The host challenges whether hyperscalers can fund such enormous spending and whether AI ROI will arrive fast enough to justify it, especially if demand disappoints. Inference, custom silicon, and chip mix shifts (Priority: 4/5): The discussion covers how inference growth, custom chips, and future NVIDIA generations could change capex needs, with Tom arguing these factors likely push spending higher, not lower. Power, permitting, and infrastructure bottlenecks (Priority: 5/5): A major concern is whether the physical buildout is feasible given power availability, data center timelines, permitting, transmission, and labor constraints. Valuation and market positioning (Priority: 3/5): Tom argues that large-cap semis, especially NVIDIA, already reflect some of the skepticism, while fundamentals and revenue estimates continue to improve.

Key Arguments: Street estimates have consistently undercounted AI capex since late 2022, so current consensus may again be too low. Using AI labs' OpEx and translating it into hyperscaler CapEx is a better forecasting framework than relying only on hyperscaler plans or bottom-up GPU utilization assumptions. Agentic AI and recursive self-improvement could trigger a sustained compute arms race, increasing both training and inference spend. Hyperscaler spending is backed by operating cash flow, and demand is broadening beyond hyperscalers to sovereigns and enterprises. The biggest real-world constraint is power, but current and expected supply additions suggest the $1T capex scenario is still feasible. Custom silicon can reduce spend in some cases, but flexibility and workload uncertainty limit how much it can displace high-end GPUs. Efficiency breakthroughs like DeepSeek may increase overall usage via Jevons paradox rather than reduce total compute demand. The market’s skepticism is partially reflected in valuations, so the risk is not necessarily unpriced exuberance. The biggest uncertainty remains whether AI ROI and adoption can materialize quickly enough to justify the pace of investment. Large projects may encounter delays, but those are framed as normal execution hiccups rather than signs that the buildout is broken.

Data Points: AI capex peak estimate: ~$1 trillion in 2028 - Tom and Ross Sandler's framework for AI-related capital spending peak Capex vs Street in 2027: ~$230 billion above Street - Projected AI capex upside versus consensus for calendar 2027 Capex vs Street in 2028: close to $300 billion above estimates - Projected AI capex upside versus consensus for calendar 2028 Capex trend into 2029: ~5% year-over-year decline - Tom characterizes post-2028 as stabilization, not a collapse Hyperscaler capex as % of OCF: 85% to 90% - Model assumption for hyperscaler operating cash flow consumed by capex from 2026 to 2028 Demand outside hyperscalers: 40% - NVIDIA GTC disclosure cited for demand from sovereigns and enterprises OpenAI forecast revision window: September 25 to February 26 - OpenAI forecasts were revised sharply higher over a five-month period U.S. data center development time: about 2 years - Timeline to develop a data center in the U.S. New large gas power plant sourcing/commissioning time: 5+ years - Power infrastructure bottleneck discussion Transmission line permitting/development time: over 10 years - Long lead times for grid expansion New U.S. data center supply in 2027: 19 gigawatts - Estimated added supply from ISOs in calendar 2027, excluding behind-the-meter and global additions Estimated industry capacity need in 2027: 13 gigawatts - Ross/Tom estimate for capacity required in 2027 Estimated industry capacity need in 2028: 21 gigawatts - Ross/Tom estimate for capacity required in 2028 NVIDIA valuation: <14x calendar 2027 earnings - Used to argue the stock already reflects some AI skepticism NVIDIA five-year median valuation: 30x - Comparison point for current valuation OpenAI-NVIDIA MOU: 10 gigawatts - Walked back agreement cited as a sign of project turbulence

Pivotal Quotes: "The Street has been consistently underestimating the CapEx forecast since the AI wave started in late 2022, and today is no different." — Tom O’Malley: Tom defends the bullish capex forecast against consensus skepticism "Cliffs are scary words, and I'm not saying this is a collapse. It's more of a stabilization, with capex down about 5% year-over-year in calendar year 2029." — Tom O’Malley: Clarifying that the post-2028 outlook is a plateau rather than a crash "We are in early stages of an adoption curve and model breakthroughs may help instead of hurt." — Tom O’Malley: Argument that efficiency gains can still expand total compute demand

Implications: If Tom is right, AI infrastructure spending could remain elevated for years, benefiting semis, data centers, power, and networking. The main risk is not demand alone but whether power, construction, and ROI can keep pace with the investment cycle.