The AI Flywheel Paradox: OpenAI's Bet on More Compute Amid Overcapacity Fears

OpenAI’s compute argument surfaced again this week. Their official channels put it plainly: “Compute is what made our first image generation launch possible, and in the three weeks since, weekly active users have grown 32%. There’s more coming… and we need more compute.”

Analysts have been warning about GPU overcapacity and excessive capex for months. Both claims can be true simultaneously, which is what makes this moment genuinely hard to read.

The Flywheel Logic

More compute leads to better models. Better models drive more users. More users generate more revenue. More revenue funds more compute.

Amazon proved this structural dynamic with e-commerce infrastructure decades ago. The same logic is now playing out in AI, at a pace and capital intensity that has no historical precedent. Whether the timing holds together is a different question. Every technology cycle has seen investment outrun near-term demand. The cycles that survived were the ones where the underlying utility was real enough to eventually catch up.

Why Overcapacity Fears Are Partially Justified

The fundamental risk isn’t that compute becomes worthless. It’s about how much of the future you can pull into the present before the weight of that bet breaks you. Human appetite for future value has historically outpaced the technology’s actual delivery schedule, and that gap is exactly where bubbles form. I don’t think we’re past that risk here.

The Real Bottleneck

Models are improving fast. Training cycles are shorter. Benchmark scores keep climbing.

And yet prompting has become more important, not less. AI performance benchmarks are measured using expert-level queries. Most real user questions fall well short of that level. Capability soars; actual utilization flatlines. You can pour billions into compute and push model performance to extraordinary heights, but if users cannot effectively communicate what they need, all that power goes underutilized. This is the paradox at the center of the flywheel argument.

Vibe coding showed what it looks like when that gap narrows. When the interface between human intent and AI capability becomes seamless, adoption accelerates. The same principle applies to presentations, content creation, and data analysis, where AI can theoretically help but practically frustrates.

Closing the Gap

The compute arms race only sustains itself if someone wins the user experience battle. The companies that invest in closing the distance between what AI can do and what users actually get from it are the ones that will justify the flywheel’s next turn. More compute is necessary. It is not sufficient.