Martin Casado admits tech overprovisions while the public pays
By Imani Sutton · Reporting from Atlanta ·
As Martin Casado and Jad Naous debate automated software on a16z, Silicon Valley’s habit of overprovisioning computing infrastructure shifts the real energy and water costs onto everyday utility bills.
$170 billion of capital into autonomous driving bought the world worse unit economics than hailing an Uber. That was the candid confession from Andreessen Horowitz general partner Martin Casado. He spoke on the "a16z" podcast episode, "Why AI Can Write Code But Can’t Keep It Running (Yet)." Sitting across from him, Grepr CEO Jad Naous explained how coding agents have transformed his company. Naous used Anthropic’s Claude Code and the Opus model on a Windows machine running Microsoft’s Linux subsystem. He completed an architectural project—auto-promotion in a data lake—in days, a task that once demanded months from a solo engineer. Yet both men admitted that while software syntax has been automated, real-world operations remain an untamed wilderness.
The fairy tale of self-healing software
The strongest case for artificial intelligence in systems engineering, as Naous laid out, rests on breaking operational chaos down into manageable steps. Running production systems involves unpredictable workloads, network deadlocks, and an astronomical number of possible states—far more than atoms in the universe. But Naous argued that production failures almost always trace back to a specific change. Instead of forcing an AI model to guess why a metric spiked after a crash, Grepr scans petabytes of daily logs to detect anomalies upstream. When connection limits fail or memory buffers empty, the system feeds that deterministic signal into an AI agent. The agent generates hypotheses, examines code, and runs tests to isolate the root cause. Software firms currently spend between 5 and 20 percent of their infrastructure budget on observability tools like Datadog just to buy peace of mind. Automated troubleshooting offers a path out of the telemetry money pit.
Every efficiency promise ends in overprovisioning
Yet the historical record reveals why this optimism collapses the moment software touches the physical world. Casado acknowledged the grand tradition of computer systems. Theorists design clever algorithms for congestion control and adaptive routing, and then the industry ignores them and simply overprovisions hardware. When computer operating systems ran short of physical memory, swapping data to disks proved fatal, so engineers stopped optimizing and just bought excess RAM.
The same cycle is unfolding with automated code generation. Naous described floodgates of automated pull requests flying across his team, dumping twenty reviews at a time on human engineers. When writing software becomes frictionless, systems sprawl. Small microservices multiply, data lakes swell with semistructured logs, and telemetry data expands faster than actual functional output. Generative models cannot reliably predict availability, access patterns, or correctness under real workloads. Infrastructure operators will respond exactly as history predicts: by massively overprovisioning cloud compute to prevent catastrophic collapse.
Who pays for the digital bloat
Every unoptimized query, runaway agent thread, and overprovisioned server cluster has a physical address. They run on vast data centers operated by companies like Microsoft. They pull continuous baseload megawatts from local power grids and drink millions of gallons of municipal water for cooling. Silicon Valley executives joke about spinning up dozens of agents that burn tokens on basic tasks. Yet the cost of that excess capacity does not stay in Menlo Park. Regulated electric utilities routinely cite surging industrial compute demand to justify building new fossil-fuel peaking plants and substation buildouts. They socialize the capital costs across ordinary residential power bills.
When city councils weigh tax abatements for massive server campuses and state regulators review utility rate hikes, voters must reject the myth of clean, frictionless computing. Demand that cloud operators prove physical water and energy efficiency standards before breaking ground. Insist that public service commissions ring-fence industrial data loads so residential households never subsidize tech’s refusal to manage its own systems.