The Hidden Tax of AI Scaling: Why Output Maxing is the New Frontier

In this conversation, Anjney Midha explains that the AI bottleneck is not just a shortage of GPUs, but a failure in how we manage them. While frontier labs chase capital expenditure, they often accept Model FLOPs Utilization (MFU) rates below 10 percent. At Google, this would be considered a major outage. The consequence of this inefficiency is not just economic waste, but a compounding alignment tax that separates capital from operational reality. For leaders and engineers, the competitive advantage no longer lies in simply buying more compute, but in mastering output maxing. This is a disciplined approach to infrastructure that prioritizes stability and utilization over the move fast and break things ethos. Those who treat infrastructure as a reliable, long-term system rather than a disposable commodity will be the ones to scale when others hit the wall of operational complexity.

The Hidden Cost of Fast Solutions

Most teams approach AI infrastructure with a hustler mentality, prioritizing speed at the expense of stability. Midha argues this is a miscalculation. When organizations scale too quickly without iterative bring-ups, they introduce a radian effect: small misalignments at the start of a project grow into massive operational failures as the system expands.

I think 95% at Google, which is where my co-founder, Seb, came from, he built the Borg scheduler at Google, and there I think 95% was considered an outage, so 96% node utilization should be standard. And most single-tenant clusters are not running at that.

-- Anjney Midha

This reveals a deeper systems problem: the distance between the capital providers and the cluster managers. When these groups are misaligned, the system loses its ability to self-correct. The downstream effect is a brittle architecture that suffers from stranded compute, which are pools of resources that cannot be used because the scheduling and economic layers were not built for fungibility.

The 18-Month Payoff: Why Hardship is a Feature

Conventional wisdom suggests that abundant capital is the ultimate advantage for an AI lab. Midha flips this, arguing that the hardship Anthropic faced in its early days was a competitive moat. Because they lacked the resources to do everything, they were forced to define a P0, which was coding, and ignore the noise. This enforced focus allowed them to achieve takeoff while better-funded competitors suffered from culture rot caused by too much money and no clear constraints.

The hardest part is in the earliest days when you don’t have a group of people who are going through difficulty, stress, crisis together, then your culture doesn’t get defined sharply enough, and that’s what I’m worried about right now, is there’s so much money going to these labs. There’s no hardship.

-- Anjney Midha

When a team has unlimited resources, they avoid the confrontational trade-offs required to build a great company. They end up with a fragile culture that collapses the moment they face a real technical hurdle. By contrast, teams that survive early-stage scarcity develop a prepared mind, a state of readiness that allows them to capitalize on breakthroughs when they arrive.

How the System Routes Around Your Solution

Midha’s vision for AMP, an independent system operator for compute, is a response to the market failure of research hoarding. When frontier labs treat their compute and research as proprietary silos, the entire industry suffers from negative externalities.

The systemic shift Midha proposes is to treat compute like the electrical grid. By pooling demand and supply, organizations can move from full-stack integration to a more resilient, horizontal model. This requires output maxing, a discipline that treats infrastructure as a public utility. The advantage here is delayed. While competitors are busy fighting fires in their proprietary, inefficient stacks, those who build on open protocols and stable, community-aligned infrastructure will be able to scale their research far more efficiently. The backlash against data centers is another systemic constraint. By giving back to local communities, companies can secure the social license to operate that their move fast competitors will eventually lose.

Key Action Items

• Audit your MFU (Model FLOPs Utilization): If your utilization is significantly below 60 to 70 percent, treat it as an operational failure, not a technical limitation. Immediate action.
• Adopt an Output Maxing framework: Shift your engineering culture from moving fast to moving fast with stable infrastructure. Prioritize reliability over feature velocity for the next quarter.
• Define your P0: Identify the single capability that, if cracked, accelerates all other work. Force your team to say no to everything else until that P0 is achieved. 12 to 18 month investment.
• Build community alignment: If you are procuring data center capacity, explore ways to provide tangible value back to the local community, such as energy cost reduction or direct investment. This creates a moat against regulatory and community backlash. Over the next 6 to 12 months.
• Prioritize Star Athlete Researchers: When hiring, look for researchers who have already contributed to SOTA and are willing to be confrontational about their convictions. This is the foundation of a high-performance culture. Ongoing.
• Piggyback on open standards: When building new hardware or infra, use existing reference architectures, like NVIDIA’s, rather than fighting to build a new standard from scratch. Focus innovation on co-design, not data center footprint. Immediate.