You can always buy more GPUs. But eventually, you have to squeeze more out of the ones you've got.
OpenAI just hired Brendan Gregg, the performance engineer behind flame graphs and two books on systems optimization, away from Intel. In a blog post explaining the move, Gregg frames his new job in surprisingly stark terms: the goal isn't just cost reduction. It's environmental sustainability. AI datacenters are expensive and power-hungry, and Gregg believes performance engineering at scale is essential to making them viable long-term.
He joins as a Member of Technical Staff, working remotely from Sydney. First project? Developing what he describes as a "multi-organizational strategy" for performance improvement and cost reduction across ChatGPT's infrastructure.
The Guy Who Literally Wrote the Book
Gregg isn't just a senior hire. He's the person who literally wrote the book on making systems faster. Two books, actually. His work at Netflix on cloud performance became industry standard. His visualization techniques, flame graphs and heat maps, are now default debugging tools across the industry.
If you've ever optimized a slow service, you've probably used his methodology.
That OpenAI is bringing in someone of this caliber signals they're serious about the software side of their cost equation. And it fits a pattern worth noting: as we covered last week, OpenAI announced a 40% time-per-token speedup for GPT-5.2, achieved through better systems engineering rather than model changes. Gregg's hire suggests that wasn't a one-off improvement. It's the beginning of a sustained infrastructure push.
Gregg conducted 26 interviews across multiple AI companies before choosing OpenAI. His reason for picking them? They had "the largest number of talented engineers" he already knew.
That's a notable data point on OpenAI's engineering bench strength.
There are two ways to handle AI's compute demands: build more datacenters, or make the existing ones work harder. The industry has mostly chosen door number one. OpenAI is betting they need both. Gregg specifically calls out the tooling he'll bring: eBPF, Ftrace, performance monitoring counters. These are deep systems observability tools, the kind that let you find the 5% inefficiencies that compound into real savings at scale.
What Convinced a Skeptic
Here's a detail that says more than any benchmark: Gregg admits he was initially skeptical about AI's real-world adoption. What changed his mind?
His hairstylist enthusiastically telling him "I use ChatGPT all the time!"
It's a small anecdote, but it captures the adoption curve driving OpenAI's infrastructure crunch. When mainstream users are hammering your product daily, every percentage point of efficiency matters.
Our read: This hire is OpenAI acknowledging that the GPU arms race isn't sufficient. Software optimization is becoming as strategic as hardware acquisition. Gregg's track record suggests he can find meaningful efficiency gains that translate to real cost savings. The question is whether those savings can keep pace with the scaling demands of whatever comes after GPT-5.