Supermicro's record year shows AI server demand is still booming
Two readings of the AI buildout landed overnight: the machine that assembles the servers posted a record year and a stock surge, and a new benchmark suggested AI models still can't do original research — only optimize what humans already built.
Supermicro capped fiscal 2026 with record revenue of $39.1 billion, up 78 percent year over year, and fourth-quarter revenue of $11.1 billion, up 93 percent, sending the stock up about 9 percent in extended trading. The headline number for investors was the margin recovery: gross margin hit 17.5 percent in Q4, nearly double the 9.5 percent from a year earlier and above the 15-to-17 percent range management had flagged in July, helped by a richer mix of enterprise customers and broader adoption of its server architecture. Adjusted earnings came in at $1.70 a share versus the $1.59 analysts expected, and the company guided fiscal 2027 revenue as high as $72 billion. AI solutions still contributed roughly 60 percent of Q4 sales, and the company booked over $60 billion in new orders during the quarter — the order book is the more telling signal, since it means demand is still arriving faster than Supermicro can build. The one wrinkle: revenue landed at the low end of guidance because of customer readiness delays, a reminder that the bottleneck in the AI supply chain keeps shifting between chips, power, and now the ability to stand up infrastructure fast enough. For everyone watching the AI capex supercycle, this is the cleanest confirmation yet that enterprise and cloud demand for AI servers is not just holding — it's compounding.
A new benchmark called MLS-Bench, from researchers at the University of Washington, UC and Purdue, put five frontier models through 140 real research tasks — and none of them beat human baselines at discovering genuinely new methods. The tasks span 12 fields including model pretraining, vision generation, robotics, AI-for-science and causal inference, and each requires an agent to modify real code and produce a working implementation, with a full evaluation costing roughly 700 H100-hours per candidate. The finding that matters: models are strong at engineering search — recombining known components, tuning parameters, patching losses — but rarely propose new mechanisms that hold up across conditions. Tellingly, framing the task as "optimize existing methods" outperformed "discover new methods," and scaling up test-time compute improved simple tasks but quickly saturated. It's the kind of rigorous, expensive benchmark the field needs, and the honest read is that current models are excellent research assistants and weak research leaders.
What to watch: whether frontier labs start using MLS-Bench as a training target — and whether the next model generation finally clears the human baseline on method discovery.
If AI agents can run the experiments but can't pick the ideas, how much of research is left for humans to do? Tell us in the comments.
