
At Citi's fifth annual TMT Leadership Summit (technology, media and telecom) in Pebble Beach, California, few speakers questioned demand for AI. Most of the talk was about what stands in the way of meeting it. The three-day program ran 21 sessions for almost 116 guests from outside Citi, representing nearly 103 companies worth about $12 trillion combined. The week also brought a sharp market selloff and a U.S. election.
The financing numbers were large. Pankaj Goel, Citi's global co-head of technology investment banking, said the tech industry raised nearly $1 trillion this year and needs $5 trillion to $6 trillion more over the rest of the decade. A senior fixed-income investor said that by his figures capital spending by the largest cloud providers, known as hyperscalers, has gone from $180 billion in 2024 to about $1.7 trillion expected next year. He projected that AI-related debt will exceed the high-yield market by the end of 2027 and expects equity issuance to return as equity and debt costs converge.
The same investor read the week's selloff as crowded positioning, not a bubble: "Everybody's long AI and short interest rates. I've never seen an unwind of that in such spectacular fashion." He noted that markets are narrow, with the 10 largest stocks accounting for 40% of the S&P 500's market cap and net income. He added that capital spending does not respond to rate moves, so higher rates weigh on housing and small business without slowing AI investment.
On where the risk sits, the clearest answer came from the chip side. A semiconductor executive said the risk in circular financing, where suppliers help fund their own customers, depends on the counterparty: AI labs will be around for decades, while some middle-layer players, such as newer AI-focused cloud providers known as neoclouds, carry much higher risk. A chip-design CEO noted that chipmakers now extend credit to their customers.
The shortage will outlast 2026, but speakers disagreed on by how much. A memory-chip executive said 2027 and 2028 look tighter still. The chip-design CEO called it at least a five-year problem: "The problem and headache we have in 2026 is a 2023 problem," he said, because nobody planned for this scale. An enterprise-software CEO put the range at five to seven years, while a venture investor said it could last "probably decades." The bottlenecks are physical, speakers said: chipmaking capacity, packaging, power, cooling, transformers and turbines.
Even with capital and chips, permission is no longer assured. The enterprise-software CEO said 30 of 31 data center moratoriums passed this week; published tallies of the Nov. 3 local ballot measures vary but show a similarly lopsided result. He said the industry "did a very poor job at community engagement," and contrasted the U.S. backlash with the Middle East, where data centers can be permitted in 14 days. The chip-design CEO warned that if the U.S. hesitates, the rest of the world is "absolutely ready to fill that void," and the venture investor said jobs and capacity could go to Canada or the Middle East.
Several speakers said the industry brought part of the backlash on itself. The venture investor said "our big labs keep telling us it's going to kill us all and take all the jobs." A former senior U.S. official said, as he described it, that industry leaders had warned of half of jobs disappearing and a 10% to 20% chance that civilization ends, and that the public had rejected that message. An AI-lab finance chief said the industry came in too strong on job losses.
For fixes, speakers pointed to evidence and early engagement. The AI-lab finance chief called for more concrete stories of benefit, and a health care AI founder put it this way: "The real conversation we need to have is not AI versus humans, but AI versus nobody." The former official urged the industry to ask for regulation before it is imposed: "It's coming. Now, it can either be shaped, or you're going to get it done to you."
Speakers also made the case for benefit with usage data. One said about 90% of the weekly users of a leading AI assistant are outside the U.S., in markets such as India, Brazil, Indonesia and Vietnam, and that about 300 million people ask it health questions each week and about 200 million use it for personal finance.
Speakers who build AI said the technology is ahead of the organizations trying to use it, and that people are the main obstacle. A data-platform CEO argued that artificial general intelligence (AGI) arrived in the fourth quarter of last year but expects the transformation of the economy to take five to 10 years. He listed four obstacles (context, control, cost and choice) and a fifth that he called the biggest: people who will not change how they work. A cybersecurity CEO predicted that 75% of enterprise AI work done in the last 12 to 18 months will be thrown away, because companies are designing AI to be controlled by humans.
The shift also changes how organizations are built and how software is sold. An AI-lab product executive said the gap between frontier users and the median employee is widening. An enterprise-software co-founder and the product executive described smaller teams and wider spans of control, and the co-founder said every role is becoming a builder. An AI-coding CEO expects pricing to move from per-seat charges to usage-based charges and eventually to payment for outcomes.
Security drew the sharpest warnings, and speakers said controls are not keeping up with the technology. The AI-lab finance chief said, "Cyber, I think, is probably the single thing keeping me up at night right now," and thinks CEOs are glossing over the risk. The data-platform CEO expects cyber to be the biggest headache of the next 12 to 18 months, with agents inside companies becoming a new insider threat.
The cybersecurity CEO offered the specifics. He said AI can now break in and steal data in about 18 minutes, while the average North American company takes four days to find an intruder. He called today's so-called agents "minions": real agents would need explainability, verifiability and auditability, plus a supervisory AI that watches other AI, and none of that exists yet. When he asked how many in the room used coding agents, many hands went up. When he asked how many knew those tools were deployed securely, one stayed up. His company sells security products, so these are a vendor's assessments.
Looking ahead, speakers tied the next wave to national interests. An autonomy-software CEO said sovereignty matters more in physical AI (AI that runs machines such as vehicles and industrial equipment) than anywhere else, because no country wants an adversary's autonomy running inside its borders. The cybersecurity CEO expects the bigger change over the next five years to come in consumer applications, bringing intelligence to 6 billion people. He described the present as "the flip phone era of AI, which is going to be amazing."