The Master Orchestrator: What Orchestration Means Once It Is Defined, and What the Defined Thing Could Be Worth
Abstract Orchestration is one of the most used words in AI infrastructure software and one of the least defined. It names at least eleven different jobs, from the container manager inside one cluster to the software that splits a single inference request across different kinds of chips, and the figures attached to it describe different things: a research firm counts under $12 billion a year of agent and workflow software sold under the name, while an investor on a compute podcast valued one company doing the right version at hundreds of billions of dollars. This paper proposes a definition. Orchestration is the authority to decide where, when, and on what each piece of work runs, within rules the work’s owner sets, together with three things the decision depends on: the reach to choose among capacity, the ability to carry the decision out, and a record of what ran and what it produced. Measured against that definition, the products sold as orchestration supply the decision inside one estate, one silicon family, or one provider’s cloud. What sets the decision’s value is reach across owners and silicon, execution below the provider’s interface, and a shared record, and the decision layer holds none of them in full. The paper then takes the investor’s figure and asks, as theory, how it could be accurate and under what scenarios. At an illustrative license near $3,700 per GPU and revenue multiples of 15 to 35 times, the bottom of the investor’s range, 200 billion dollars, is reached when between one in 17 and one in 40 of the GPUs installed in 2028 are licensed. Five hyperscalers held about 70 percent of global AI computing power at the end of 2025 and build much of their own orchestration software, so the share open to an independent vendor is unknown. From a 40 percent baseline, ten points of effective utilization is worth about 4,400 dollars per GPU a year to the buyer, little more than the license, so the price is defensible only if the layer delivers performance and commercial benefit beyond utilization and meets the buyer’s other demands. The transaction record mostly points one way: NVIDIA gave away the scheduler it had bought, owners of capacity bought the software above them, and companies that began as renters started building capacity of their own. The exception is a model router that owns no hardware and was sold for a reported $7.5 billion. A master orchestrator worth hundreds of billions would have to be built at the point where hardware is allocated and work is placed, which is where this series has located the unbuilt layer. Last, the paper sets the definition against what CIOs say is slowing them down and reads three demands in it: no further layers, a record of what ran, and more work from what they already own. The paper argues for no valuation, every figure it derives is theoretical, and it closes with six outcomes that would prove it wrong.
Authors
- JR Poeta
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24205763
- Primary Topic
- Cloud Computing and Resource Management
- Type
- article
- Field-Weighted Citation Impact
- 0.00