Every AI output runs a meter. As token volume multiplies, the companies that match each task to the right model, and own what that model learns, will control their costs.
How the inference bill grows.
Inference works like a utility bill. Each AI result triggers measured calls to one or more models. More users, longer prompts, bigger context windows, agents, tools and retries all raise the bill.
Google's numbers show how fast this grows. At Google I/O in May 2026, CEO Sundar Pichai said Google processed 9.7 trillion tokens a month two years earlier and about 480 trillion a year earlier, almost 50 times more. It now processes more than 3.2 quadrillion tokens a month, another 7x in twelve months.
At that pace, inference is on track to become the largest market software has ever had, and Orchestra sees making it efficient as the largest opportunity capitalism has produced.
Match the model to the job.
The useful question is not which model is best. It is which model fits this task at this quality bar, latency, cost, security posture and level of control.
Closed frontier models suit genuinely novel problems that need the strongest reasoning, and they are convenient. They are still the best way to learn what AI can do, and we use them every day. Routine work does not need that much capability, in the same way a mailroom does not need a PhD.
For real workloads, the strongest choice is a post-trained open-weight model. It can cost less, run faster and operate inside your own environment. It can also be shaped by your private evals, your traces and the harness you already built.
Keep what the model learns.
With a frontier API, what the model learns from your work either disappears when the session ends or flows into training for the model your competitor also uses. Microsoft CEO Satya Nadella put the alternative plainly in his July 2026 essay, The Reverse Information Paradox: "In the cloud era, enterprises accumulated data. In the AI era, they accumulate learning."
That loop is the real IP. Private data, evals, workflow traces, domain feedback and reinforcement environments improve the model, which improves the product, which produces more data for the next model.
Why few companies do this yet.
It takes too many people, too much time and too much money. Orchestra co-founder and CEO Luis Manrique lived through cloud migrations at Google and Instacart, where teams looked for vendors to take on the pain and wished they had started earlier. The AI shift is bigger, and so is the infrastructure work.
The timing has turned. Open-weight models are now good enough for enterprises to treat them as real alternatives to closed APIs, and enterprises are following AI-native companies for capability, cost and control.
What Orchestra does.
Orchestra hides the messy infrastructure behind one surface. Record what your current model does, define what success looks like, train or tune a lower-cost model for the repeatable work, and run it wherever you like. Ongoing feedback keeps it improving.
One design partner more than halved its monthly frontier spend and now owns a model that is still getting better.