Code is free. Running the system is the bill.

The code war is over

Guillermo Rauch, the CEO of Vercel, told the Sourcery podcast that coding agents now originate 60% of deployments on the platform, up from 3% at the start of the year. He summed up the shift: "I'm a yes code company because I'm now saying coding agents basically make the code free." The numbers describe a year in which writing software collapsed from the bottleneck to a near-free input, at least on the platforms where agents already do most of the deploying. m:…4210

The same split showed up at Jane Street, where John Crepezzi told Signals & Threads that roughly 45% of released features now originate from an asynchronous flow involving a large language model. Generating the code, in other words, is table stakes. Jad Naous, the CEO of Grepr, argued on the a16z Show that operating distributed systems remains fundamentally harder than generating code, because unpredictable workloads and vast state spaces make production a different problem from authoring. m:…6908

Where the bill lands: telemetry and tokens

The first place that difficulty shows up as money is observability. Naous said Grepr starts by cutting customers’ telemetry spend — “you put us in the data stream, plug and play, pass the data through, reduce your cost by 90%” — and told the a16z audience that buyers prioritize shrinking bills, Datadog’s among them, over the underlying AI technology. This suggests the observability line item, inflated by years of log-everything defaults, is where engineering leaders are quietly finding budget for their AI build-outs. m:…3865

The second bill is compute itself, and it is getting harder to forecast. Carmen Li, the CEO of Silico (Silicon Data), told the Pitch the PM podcast that AI companies are shifting from premium models to cheaper ones in what she described as the post-token-maxing era, and that reliance on volatile on-demand hyperscaler GPU rates is producing negative margins. Her firm’s data shows the GPU forward curve flipped from downward-sloping last year to contango this year — a market structure that punishes anyone buying compute on the spot. m:…5319

The infrastructure answer: sandboxes and agent-native services

The response taking shape is infrastructure built for agents rather than for developers typing at keyboards. Rauch said: "Vercel is also building the sandboxes and the capabilities so that companies can operate agents, deploy agents, and really trust them." He described agents already buying services on the Vercel marketplace end to end: "I'm enabling that transaction end-to-end without any human intervention." The deploy target is no longer a human reading a dashboard; it is an agent with a budget. m:…4210

That shift has a physical cost. Agentic use is exponentially increasing compute, energy, and network resource demands, Rauch said — the very demand curve that makes Li’s contango observation bite. This suggests the platforms winning the agent era are the ones that meter, sandbox, and price the run, not the ones that generate the most code. m:…4210

What changes for engineering leaders

The management consequence is a new job description. Naous said strong engineers are already using agents to compress work that took months into days, and Rauch framed the engineer’s emerging role as a capital allocator, directing AI tokens toward the tasks where they pay off. The scarce skill is no longer writing the code — it is knowing which work deserves the tokens. m:…3865 m:…4210

For funders and founders, Naous offered the sharpest warning: VC-backed companies trying to solve operational problems are having a harder time than those focused on code generation — a reminder that the “free code” narrative prices in the easy half of the stack. The value now sits in everything that happens after the code is written: observability, cost control, sandboxes, and the metering that makes agents trustworthy tenants. m:…3865