Two Orders of Physical AI: How the System Works and How Liquidity Rotates

The thesis

Physical AI has two sequences that are easy to confuse: the order in which the system works, and the order in which capital discovers the system.

A machine must perceive its environment, reason over that state, choose an action, execute it, and assess the result. That is the operating loop. Compute, data, and trusted execution support several stages at once.

Markets tend to focus first on an obvious constraint with liquid assets that offer exposure to it. Compute often fits that description. Once the market has priced in much of the compute opportunity, capital starts looking for the next layer that turns capacity into useful work. The process continues from one underpriced dependency to the next.

1. One stack, two orders

The operating sequence describes what a physical system needs to do to complete a task. The liquidity sequence tracks which parts of that system investors understand and can gain exposure to first.

“First” and “last” therefore mean different things in each row. Perception comes first in the machine because it needs information about its surroundings before it can act. Compute can lead in the market because investors are already familiar with GPU scarcity, deployment costs, and the price of capacity. Inference is already part of today’s systems. Its position at the end of the liquidity sequence reflects the point at which repeated use provides economic evidence for the investment in the preceding layers.

2. What actually flows

Liquidity rotation does not require one investor to sell AKT and immediately buy FET, then sell FET for GEOD. What moves is the market’s marginal capital: the next dollar looking for exposure, the profits recycled from an earlier leader, and the balance-sheet capacity that market makers and funds allocate to a new group of assets.

The flow usually has three sources. Profit rotation moves gains from an already re-rated layer into smaller adjacent names. New thematic inflows enter through the recognized leader and then broaden as investors learn the surrounding stack. Fundamental spillover begins when customers, integrations, or capacity constraints in one layer produce measurable demand for another.

Fundamental spillover helps explain why a rotation might last. Profit rotation and fresh inflows can lift prices; demand from one layer for another gives investors an operating reason to keep allocating capital. A compute rally broadens into a Physical AI rotation as expanding capacity supports more applications and brings demand for physical data, private workloads, and inference into view.

3. The rotation and its causes

The case for each hand-off rests on a change in the question the market is trying to answer.

The arrows track shifts in market leadership as new dependencies come into focus. Compute can remain strong while execution begins outperforming. Perception can become the new high-beta group while agent networks keep their earlier gains. A healthy expansion can bring more of the stack into the rally.

4. How a hand-off develops

A rotation normally begins with concentration. One or two liquid names carry the whole idea because they can absorb size and their story needs little explanation. Strong performance attracts attention, deeper markets, and new entrants. Eventually, exposure to that bottleneck becomes expensive relative to opportunities elsewhere in the stack.

Differences in valuation and liquidity help explain the hand-off. The leader may have the deepest liquidity, but it may also be the most fully valued. Assets in the next layer are often smaller, less understood, and more sensitive to a modest increase in demand. A credible link to the expanding layer gives investors a reason to accept their lower liquidity.

A broad rally on a shared headline tells us less. Without shifts in relative trading volume, integrations, paid usage, or market depth, there is little evidence that investors are identifying a new bottleneck.

5. One rollout, seen twice

Consider a delivery company building an autonomous fleet. Inside the machines, the task begins with perception: precise location, road conditions, obstacles, and local geometry. The system reasons over that state, chooses a route, sends commands to the vehicle, and measures the result. This follows the operating sequence in the first row of the diagram.

The market sees the same rollout differently. The earliest visible costs may be simulation, deployment, and GPU time. That could put compute proxies such as AKT or RENDER in the lead. Once that capacity is available, attention can turn to the agents and execution infrastructure that generate workloads for it, bringing FET, OLAS, LINK, or NEAR into focus.

When the pilot leaves a controlled warehouse and reaches public roads, the operating bottleneck changes. Precise positioning and fresh spatial data matter more, directing attention toward GEOD, HONEY, ROVR, and OVR. These workloads involve routes, customer locations, proprietary models, and signing keys. Protecting them creates a case for confidential execution through PHA, RLC, ROSE, or a confidential NEAR environment.

Attention can then shift to decision quality: which models choose better routes, handle changing conditions, and reduce costly errors? Comparing those results makes the reasoning layer easier to assess. As the fleet moves into regular operation, the next question is how often that intelligence is used and what users pay for it.

Investors can examine Bittensor’s subnet output and the value validators assign to it. NEAR AI offers a different view through confidential inference and hosted-agent usage, while paid compute credits provide a way to assess demand on Akash. Recurring inference spending feeds back into capacity demand, closing the liquidity loop. The machine’s operating sequence still begins with perception.

6. How the loop closes

To connect operating demand to asset value, we need to trace how each service uses its token. The mechanisms include AKT burned to create ACT compute credits, HONEY redeemed for map access, GEOD used within the positioning economy, TAO emissions directed toward useful subnet output, and staked NEAR producing credits for AI services.

Recurring inference sends spending back toward compute, data updates, coordination, and trusted execution. This makes the liquidity map circular. Compute can lead the initial phase, recede from attention as investors price the rest of the stack, and return to focus as usage drives another round of demand.

By that stage, investors have a different basis for assessing compute. The initial thesis rested on anticipated scarcity; they can now examine consumption from the applications built during the expansion. Compute assets return to focus with clearer evidence of recurring demand.

Conclusion

Physical AI works from state to action and back to state. Liquidity usually moves from the easiest visible constraint toward the next underpriced dependency: compute, execution, perception, trust, reasoning, inference, and eventually back to compute.

We look for leadership to move toward the layers facing new demand as the system develops. When profit rotation and fresh inflows follow those same dependencies, there is a clearer link between market performance and the growth of the underlying system. That is how a broad Physical AI theme can develop into a rotation through the stack, one bottleneck at a time.