The Milken Institute Global Conference this week produced a clearer picture of where the current AI boom is straining than most earnings calls will admit: token costs now exceed employee costs at major companies, 80 percent of business leaders privately confess their AI transformation is mostly theater, and 68 percent of workers say they are navigating the largest economic shift of their lives without any guidance from their employer.
The most concrete figure came from Nvidia VP of Applied Deep Learning Bryan Catanzaro, who told Axios that compute costs for his team are now far beyond the costs of the employees. Uber's CTO has already blown through his entire 2026 AI budget on token costs alone, with the year barely half over. Those two data points tell you more about the economics of enterprise AI than any earnings call. Worldwide IT spending is expected to hit $6.31 trillion this year, and AI compute is consuming an outsized share. The narrative that AI reduces costs is real at the margin for specific workflows. The narrative that it reduces costs at the infrastructure layer is not.
BlackRock CEO Larry Fink pushed back on bubble framing but conceded the underlying constraint clearly: supply shortages are real, demand is growing faster than anyone anticipated, and the global opportunity has barely been explored. Fink's framework is actually the bearish one, once you read it clearly. If demand is growing faster than supply and supply is constrained by energy, chips, and hardware, then prices stay elevated, margin compression for compute-heavy startups persists, and the winners are the infrastructure providers, not the application builders. That is the structure of a market where incumbents capture the economics and startups compete for the residual.
The AI-washing data from the Milken-adjacent survey is the most significant finding for founders and investors. When asked confidentially about actual AI progress, 80 percent of business leaders admitted their public posture overstates reality, with C-suite executives more likely to confess than VP-level respondents. The explanation is structural rather than dishonest: markets punish hesitation and reward AI posturing. Companies that do not claim AI leadership face analyst downgrades and talent flight. So they claim it, regardless of implementation depth. That creates a specific distortion in the startup market. Enterprise buyers who are publicly committed to AI investment are privately uncertain which bets will produce returns. Sales cycles are long, pilots proliferate, and actual production deployments remain concentrated in the companies with the data infrastructure and talent to execute.
The talent constraint is more binding than it looks from the outside. The World Economic Forum's Future of Jobs data shows 63 percent of employers identify skill gaps as the primary barrier to AI transformation. Job postings requiring AI proficiency grew 1,800 percent in two years, while the average time-to-fill for AI-adjacent roles in financial services and healthcare runs six to seven months. Nvidia's Jensen Huang said at Milken that AI is creating enormous numbers of jobs, which is technically true and strategically misleading. The jobs being created require skills that take years to develop. The jobs being displaced or restructured are filled today by people who have not been given training plans. Walmart is retraining 2.1 million employees. IBM is tripling entry-level hiring. Most companies are spending on AI systems and not on the humans required to operate them.
For SF founders, the bottleneck map is more useful than the headlines. Compute costs exceeding employee costs is a startup opening in inference optimisation, caching, and retrieval augmented generation that reduces token consumption per workflow. Enterprise AI-washing creates demand for measurement and evaluation tools that prove actual ROI, not just adoption metrics. The talent shortage opens markets for AI-assisted deployment, workflow configuration, and change management software that does not require six-month integration cycles. Energy constraints, flagged by Fink as the primary supply limit, create opportunities in power management, workload scheduling, and data centre efficiency.
The category where AI incumbents are most likely to capture value is the one where compute is the product: foundation model APIs, cloud infrastructure, and hardware. Startups that compete on model capability are fighting with unlimited-budget labs. The startup categories that remain open are precisely the ones that existing infrastructure cannot commoditise: vertical-specific data pipelines, compliance automation for regulated industries, deployment tooling that bridges enterprise systems with model outputs, and measurement platforms that tell buyers whether their AI investment is actually working. Those are not glamorous categories. They are the plumbing of an economy still figuring out whether the promises are real.
Also read: Andreessen Horowitz leads $16 million into Stockholm's Pit, proving US capital is still Europe's AI price-setter • The EU Startup Fund's first bet is a quantum chip company, signalling Europe's deep-tech capital shift is real • China is applying its EV playbook to humanoid robots and the production economics already look familiar
