We are seeing a massive surge in "Neocloud" providers—startups and specialized cloud entities trying to carve out space against the hyperscalers—who are promising astronomical amounts of compute capacity. The problem Altman is pointing out is a fundamental mismatch: these providers are announcing massive, multi-billion dollar capacity expansions without having the actual customer base or long-term contracts to justify the footprint. It’s a classic supply-side gamble that assumes the demand curve will always be steeper than the supply curve, which is a dangerous bet in any industry.
The risk of falling compute costs #
One of the more interesting technical nuances in his warning is the volatility of hardware and energy costs. In the current market, we are all operating under the assumption that compute is a scarce, expensive resource that justifies massive, upfront CAPEX. However, if the cost of training and inference drops faster than expected—due to breakthroughs in model architecture, more efficient quantization, or specialized silicon—the math changes entirely.
If we see a rapid decline in the cost per FLOP, those massive, billion-dollar data center projects being greenlit today could turn into massive stranded assets. Even for a company like OpenAI, which sits at the center of this ecosystem, the risk is real. If you build a massive cluster based on today's cost projections and tomorrow's efficiency gains make that cluster 10x more expensive to operate relative to the new baseline, you've just built a white elephant.
Why this matters for AI workflow deployment #
For those of us focused on practical deployment and building AI agents, this macro-economic tension matters because it dictates the stability of the infrastructure we rely on. If the Neocloud bubble bursts, we could see: Pricing Volatility: Sudden shifts in how much providers charge for H100 or B200 access as they scramble to cover their debt.Infrastructure Consolidation: A move away from niche providers back toward the "Big Three" (AWS, Azure, Google Cloud) as the smaller players fail to meet their capacity promises.Hardware Efficiency Focus: A shift in research toward making models run on much smaller, more efficient footprints rather than just throwing more raw compute at the problem.
It's a weird time to be in the industry. We are simultaneously seeing the most aggressive infrastructure buildout in human history while also being warned that the foundation might be built on speculative air. We need to keep a close eye on whether these capacity announcements actually translate into enterprise adoption or if we're just watching a massive overcorrection in the compute market.
The US government might have just handed OpenAI a massive legal 11h ago
AI is changing religion and religions are trying to change AI 14h ago
AI is killing the startup growth curve before they can even scale 17h ago
The massive push to build more data centers might actually be a 1d ago
How we actually approach AI alignment and security 1d ago Stop struggling with broken OpenAI tunnels for your MCP setups 1d ago
Next How AI-native companies turn simple workflows into core → a library of Claude prompt techniques, with plenty of directly applicable cases.