Key Takeaways

  • AI infrastructure development should prioritize an "output maxing" philosophy, focusing on getting maximum value and optimal outcomes from existing resources rather than simply adding more compute. This is a deliberate shift from a simple resource-acquisition mindset.
  • Anjney Midha, CEO of Amp, warns against abandoning established engineering principles in the rush of AI progress, stating that AI scaling actually puts a premium on common sense and sound infrastructure due to higher costs of wastage.
  • Founders should demand high utilization rates for their AI compute, similar to mature industries. Midha specifically cited targets like 95% node utilization and 60-70% MFU (Memory Free Units, an efficiency metric) as achievable goals.
  • The old startup mantra of “move fast, break things” needs to evolve into “move fast with responsible infrastructure” when it comes to AI, reflecting a necessary maturity in how compute resources are deployed and managed.

The New Playbook: Output Maxing Your AI Stack

Many ambitious founders chase the biggest GPUs, the largest clusters, and the most compute power they can get their hands on. But according to Anjney Midha, CEO of Amp, that's precisely the wrong way to think about building AI infrastructure. Midha argues we're missing the forest for the trees, focusing on brute-force scale rather than smart optimization. He calls for a philosophy of "output maxing" – extracting optimal outcomes from the compute you already have.

“From an engineering perspective it's basically it's output maxing. You know, it's the it's the it's the department of output maxing. of what we have,” Midha explained. This isn't just a philosophy; it’s a direct challenge to the often-wasteful, growth-at-all-costs mindset. He pointed to Mark Zuckerberg's famous shift from "move fast, break things" to “move faster, stable infrastructure” at Facebook. Midha believes AI demands an even further evolution: “I think now we need to move fast with like responsible infrastructure.”

This isn't just about being frugal; it's about engineering discipline. Midha argued that the cost of error and the potential for wastage are so much higher in AI scaling that common sense and sound infrastructure are no longer optional. They're table stakes.

Lessons From the Old Guard: Why Common Sense Still Wins

Midha isn't inventing new engineering principles here; he's reminding us of old ones. He drew direct parallels to the semiconductor and DSA (design for manufacturability/assembly) industries, fields known for their rigorous, iterative approaches to infrastructure. “I think we know the answer, which is just do iterative bring-ups,” Midha said. “If you spend time with people who've been in the semiconductor industry or the DSA industry for a long time, this is not new. And I don't think AI should be an excuse.”

These are industries where every percentage point of utilization matters. Midha didn't just speak generally; he offered concrete targets: "95% node utilization, 60-70% MFU." These aren't aspirational numbers for some distant future. They're benchmarks for today's AI infrastructure, suggesting that if your team isn't hitting these, you're leaving performance and money on the table. The frantic pace of AI development has made some forget these basic truths. Midha's message is simple: don't let the hype lead you to bad engineering. “In fact, if anything, AI scaling should be putting a premium on the value of common sense and infrastructure because the the margin of error now is so much lower and the cost of wastage are so much higher.”

What to Do With This

This week, challenge your engineering and operations teams to conduct a granular audit of your existing compute infrastructure. Don't just look at total cluster uptime; specifically dig into node utilization and MFU for your active AI workloads. Before approving any new compute purchases, demand a clear plan for how your team will hit utilization targets like Midha's 95% node and 60-70% MFU on your current resources. Make them prove they're "output maxing" before they ask for more.