Open Weights and Production Tokens: Scarcity Is Moving

Scarcity has moved from the arm to coordination

Open-weight production tokens are growing, the model layer is edging towards commodity, and value is migrating to the layer that remains scarce.

That layer is coordination. The robotic arm has stopped being the hard part of flexible manufacturing.

The hard part is deciding who picks up the part, who yields the cell, who waits for a peer. That is the conclusion coming out of the sector's first reproducible benchmark.

On 13 September 2026 Fouad Bahrpeyma, David Heik and Dirk Reichelt published AssemblyGrid v1^[1], a benchmark for repeated multi-robot manufacturing. It includes three workload families, each across three scenario levels. The yardstick is a privileged centralized reference, that is, a controller that sees everything.

Decentralized policies learn productive behaviours starting from local observations. This is a regime change: scarcity migrates from hardware to coalition software.

Why the consensus is watching the wrong metric

The market measures the arm: price per unit, payload, repeatability, downtime hours.

These are real metrics, and for years they have been sliding down the classic industrial curve. They are also metrics that describe the present and say little about the pace of change.

The metric that predicts is a different one: how many parts come out of a cell when several robots share space, material and geometric constraints. The authors say it explicitly: every decision changes the feasibility of the others. Process progression, material routing, resource allocation, temporary cooperation, simultaneous execution.

Here the interesting curve has three control points, ranked and measured with the same protocol: the privileged centralized reference, the structured decentralized controllers, the policies trained with IPPO, MAPPO and QMIX.

Ninety per cent of analysts are right about the present. They are wrong about the pace.

Open weights are taking share of production tokens

The model layer is becoming a commodity, and the signal is coming from the token side.

The share served by open-weight models grows quarter after quarter, for a pricing reason: whoever serves volume pays the marginal cost, and downloadable weights crush it. This is common knowledge among those running inference at scale.

The practical consequence bears on robotics more than it seems. When the general-purpose brain is cheap, margin shifts to the layer that stays scarce: the policy that decides who does what, inside a real cell, with partial information.

The underlying rule of this desk holds: the bottleneck migrates, and every cycle makes something different scarce. From training silicon to inference, from fab to packaging, from memory to judgement. Now from the arm to the coalition.

What AssemblyGrid v1 measures

The point that matters for capital is a single one: the success measures stay independent of the reward and of the solution method.

Task success is defined outside the reward function. So a learning method and a classical solver tackle the same manufacturing problem and generate comparable numbers.

The three families separate three ways of failing. Flow looks at material flow along the process. Coalition looks at temporary coalitions, that is, two or more robots joining for one step and then dissolving. Concurrency looks at productive simultaneous execution, with workspace compatibility as the constraint.

The authors add executable compliance checks, mechanism studies and algorithmic experiments, along with eleven figures and twenty-three tables. It is the structure of a benchmark, instead of the structure of a demo.

From the demo to the curve: how capital changes

The causal mechanism sits in three steps.

First: as long as a domain lacks a shared benchmark, every vendor shows its best video and the buyer pays for the narrative. Second: as soon as a reproducible protocol exists, two systems become comparable on the same number.

Third: capital changes its object. It stops funding demonstrations and starts funding points on a curve, because the curve now exists and progress can be verified.

The pattern has been seen elsewhere. ImageNet did this to computer vision, the language suites to text models, the autonomous driving benchmarks to on-road perception. In every case the money moved from proclamations to leaderboards within a few years.

AssemblyGrid v1 is the industrial version of that moment.

Three categories that change shape by 2030

Three categories exit the stage in their current form, and the profiles are specific.

• Robotic cell integrators selling wiring, commissioning and reprogramming hours
• Vendors of centralized schedulers (MES and APS) built on the assumption of complete information
• Low-cost arm manufacturers competing on price per axis

The first group sells engineering hours for every line reconfiguration. A policy that reassigns tasks on its own compresses those hours, and with them the revenue per project.

The second group has the deeper problem. A centralized scheduler assumes complete information and low latency, two assumptions a flexible factory contradicts on every shift.

The third group faces hardware prices falling while value climbs back towards the layer above. Arm is bringing the theme of scaling physical AI to RoboBusiness (The Robot Report^[2]), a sign that on-machine compute is being treated as a platform, instead of an accessory.

What changes for whoever is signing now

Whoever reads this desk signs decisions, so let's get to the point.

• CTOs and heads of innovation: reassess the control stack, before decentralized coalition becomes obvious
• Venture capital: the unpopular bet is the coordination layer, barely visible and hard to tear down
• Heads of strategy: a three-year plan with hand-reprogrammed cells describes a world on its way out
• Technology procurement: demand results on a public benchmark before signing

The concrete risk sits in multi-year contracts. A five-year deal on a centralized scheduler locks in an architectural assumption that the next research cycle is dismantling.

Ask the vendor for one thing only: results on a public benchmark, with a reproducible protocol and a comparison against a centralized reference. Whoever has the numbers shows them; whoever has the narrative offers another video.

For growth capital the signal is symmetrical. A policy that ingests factory data accumulates a defensible advantage, because that data stays where the process runs.

The position, the counterargument and what falsifies it

My position: in flexible robotic manufacturing the dominant constraint is now coalition software, and market value will follow that layer.

The serious counterargument exists, and it deserves to be stated in full. AssemblyGrid remains a grid environment, with simplified geometry and discrete time. A real cell adds friction, dirty sensors, tolerances and functional safety.

Whoever objects that simulation overstates transfer has history on their side: robotics has already paid for the gap between environment and shop floor.

The thesis holds for a reason of ranking. A benchmark ranks the methods, then engineering closes the gap. The ranking of methods changes far more slowly than the physical details.

Forecast

By 31 December 2027 at least three industrial robotics vendors will publish results on a public, reproducible multi-robot benchmark, with decentralized policies trained using MARL methods.

Confidence: 65 out of 100. Horizon: 469 days. Kill signal: as of 31 December 2027, results on reproducible multi-robot benchmarks remain confined to academic groups, with zero publications signed by industrial vendors.