AI Is Outrunning Systems Built for Human Scale

Dr. Ryan Ries here. This week I want to discuss two Anthropic stories showing that AI is running faster than the pipelines and contracts designed to hold it.

We tend to measure AI progress by the model: how smart it is, how quickly it responds, and what new tasks it can perform. But two developments this week point to a different question. What happens when the systems around the model were designed for human scale, and AI starts operating far beyond it?

That tension is now showing up in two very different places: software engineering and AI infrastructure. In both cases, capability is accelerating faster than the surrounding process was designed to absorb.

One quick note – if you’re thinking about AI token costs and how to manage them, make sure you register for our live session tomorrow. We’ll break down tokenomics, answer your questions, and share practical ways to get costs under control.

And if coding with AI is driving a significant share of your token usage, we can also help you explore whether AI-DLC makes sense for your team and how to use AI models more efficiently throughout the development process.

When Claude writes 90% of Claude Code, who reads it?

Anthropic published a candid account of how they review software now that Claude produces roughly 90% of the product’s own codebase. Anthropic says changes landing in its monorepo climbed from roughly 2,500 per day in May to more than 30,000 in September.

At that scale, the interesting problem is whether the rest of the software-development system can safely keep up.

Anthropic says merged pull requests to the Claude Code codebase more than doubled from June through September while the share that had to be reverted fell by more than 40%. The team did not eliminate review. It moved much of the reading to machines, made review continuous rather than a gate at the end, and concentrated human attention on higher-level questions and scrutiny.

Now for my skeptic hat. Every reviewer in that pipeline is Claude. One outside developer called it “Claude grading Claude’s homework”, and reported that in his own logs, a second model from another vendor regularly caught serious bugs Claude had already approved. One engineer’s logs aren’t a controlled study. Still, monoculture risk in automated review deserves far more airtime than it gets.

A shifting bottleneck

When code generation becomes abundant, the scarce resources become validation, architecture, security, testing, governance, human judgment, and the ability to recover quickly when something goes wrong.

This is why we preach at Mission Cloud that the conversation needs to move beyond “AI coding assistants” and toward an AI Development Lifecycle. The development lifecycle itself has to adapt to agents that can plan, generate, test and revise work at machine speed. The goal is not to remove people. It is to put human judgment where it has the most leverage, while making repeatable controls part of the workflow.

What we should all be paying close attention to

• Risk-based review rather than identical review for every change
• Automated validation that runs continuously, not only at the end
• Fast rollback and reversibility as core engineering capabilities
• Whether source control and CI/CD pipelines can absorb agent scale throughput

Source: Anthropic, “How the Claude Code team does code review”

Can AI Growth Outrun AI Infrastructure?

The second story I want to chat about is financial rather than technical, but I think it describes the same underlying phenomenon: AI is pushing another human-scale system beyond its familiar boundaries.

Anthropic filed its S-1 on October 26, and it puts extraordinary growth next to extraordinary infrastructure commitments. Reporting on the filing puts first-half 2026 revenue at approximately $16.3 billion and operating loss at roughly $693 million. The same disclosures describe about $265.6 billion in expected future fixed or minimum guaranteed compute commitments, with terms ranging from one to 20 years. Separate categories of leases and commercial agreements can produce different headline totals, so those figures should not be casually combined.

The headline is not as simple as “AI infrastructure is expensive.” We already know that. The more interesting question is what happens when future capacity has to be secured years before the revenue needed to justify it is certain.

This is where the second derivative matters.

Revenue doesn’t have to fall for these economics to pinch. Growth only has to decelerate faster than the contracts assumed. A company can keep expanding and still feel squeezed if capacity booked in 2026 was priced against a curve that flattens in 2028.

That turns AI infrastructure into more than a GPU story. The stack now reaches through cloud capacity, data centers, networking, power, cooling, silicon supply chains, and the cost of capital. Compute demand may remain enormous while financing, capacity, or energy becomes the limiting factor.

For enterprise leaders, that suggests a different way to think about AI architecture: not “cloud or local” as a religious choice, but workload placement, economics, and optionality. The ability to use cloud, private infrastructure, edge systems, and increasingly capable local models can become an architectural lever rather than a compromise.

What I’m watching

• Compute commitments versus realized AI revenue
• Power and data-center capacity as constraints on AI growth
• Whether more capable local and edge models change workload economics
• How pricing moves as providers race to fill capacity they’ve already paid for

Supporting source: Data Center Dynamics on Anthropic’s compute commitments

Some Final Thoughts

These stories look unrelated until you zoom out. One is about code review. The other is about hundreds of billions of dollars of compute. Both are examples of the same transition: AI is generating demand, output and change at a scale that challenges systems built around human rates of production and decision-making.

I’ve said it once and I’ll say it again. The bottom line, and what I hope you take away from this today is that the durable enterprise advantage will come from the architecture around AI's capability: how we validate change, govern risk, place workloads, secure infrastructure and preserve the human attention needed for the decisions that actually matter.

I’ve been having a lot of conversations with customers lately about AI-DLC. If you’re feeling a little uneasy about the architecture surrounding your AI stack, let’s chat. Reach out to us here.

Until next time,
Ryan

Now time for this week’s AI-generated image and the prompt I used to create it.

A fuzzy puppet lighthouse keeper stands on a tiny stone lighthouse at dusk, holding up a single brass lantern as an endless fleet of glowing paper boats pours in from the horizon, far faster than he can count them. Behind him, construction cranes assemble a second, enormous lighthouse out of stacked server racks, its scaffolding strung with crackling power lines. Painterly, spooky cinematic, warm amber light against a deep blue sea.