Navigating financial anxiety around paying the bill for industrial AI

Editor’s note: This is the first of a series from our ARC Advisory Group colleague, Colin Masson, adapted from an article he wrote for ARC. Colin is a valued industry SME who will occasionally help Smart Industry interpret what is happening in the fast-changing manufacturing technology landscape and conversations in the community.

A notable wave of financial anxiety has filtered through the industrial manufacturing and process C-suite regarding the long-term costs of AI.

As autonomous agents begin performing actual, long-horizon operational work, executives face a highly variable, unpredictable, and potentially volatile cost loop.

Some technology commentators have gone so far as to declare a double-headed “SaaS apocalypse” and a “token apocalypse,” pointing out that the traditional software billing model is breaking down and that the sheer volume of processing cycles will overwhelm IT budgets.

See also: Siemens, HighByte team to offer industrial AI software platforms

To understand these structural realities, we must look at the macro capitalization of the AI market itself. The early phases of corporate AI adoption were heavily subsidized by massive venture capital inflows and hyperscaler market-share plays, allowing enterprises to run frontier models at an artificial loss.

We are now transitioning out of that venture-subsidized honeymoon. Hyperscalers are increasingly passing the astronomical capital costs of gigawatt-scale data center buildouts, advanced silicon procurement, and soaring utility loads to the end user.

What some interpret as a market failure is actually the sound of industrial users finally paying the true, raw bill for the underlying infrastructure footprint.

Breaking down tokens and their meanings

To navigate this landscape without hitting an operational wall, leadership teams must first resolve a common vocabulary confusion running rampant through corporate boardrooms. In the modern industrial technology stack, the word “token” carries two separate, non-competing meanings:

• API compute tokens (the hyperscaler model): The fractional units of data, characters, or text sequences processed by a large foundational neural network, typically billed as a variable operational expenditure (OpEx) based on continuous utilization.
• Software licensing tokens (the value-based model): A highly flexible, shared currency pool utilized within the OT and engineering technology domains to dynamically check specialized application entitlements in and out on demand.

The narrative of an unmitigated cost crisis occurs when executives conflate these two concepts, assuming that running a software-defined factory means exposing their entire bottom line to a variable, cloud-metered transaction fee.

By understanding how these two token models interrelate, industrials can successfully insulate their margins while supercharging their computational scale.

The hyperscaler OpEx token trap

The anxiety surrounding AI processing costs is rooted in a real structural friction point: "Tokenomics is the new headcount."

As organizations transition routine task execution from human personnel to autonomous software agents, managers must shift from measuring operational capacity by human full-time equivalents to calculating the runtime compute cost of model inference cycles.

See also: Siemens: You’ve got an (AI) co-worker in me

The challenge for the industrial sector is that if you attempt to apply a standard, carpeted enterprise IT cloud framework to an uncarpeted plant floor or process network, you walk directly into the hyperscaler OpEx token trap.

Frontier reasoning models are mathematically capable of remarkable contextual planning, but they are computationally expensive to run continuously.

If your data science division deploys always-on, high-frequency autonomous execution agents to monitor telemetry across thousands of real-time factory floor tags (running queries continuously to optimize a complex process or check visual quality on a high-speed assembly line), a cloud-tethered billing architecture is unsustainable.

Because every character read, every instruction generated, and every tool called by the agent incurs a public cloud transaction fee, your monthly software billing lines are suddenly tied directly to your active data velocity and plant throughput.

See also: With new software team-up, AVEVA and IFS think they’ve closed the gap between AI insight and action

Furthermore, it introduces acute infrastructure vulnerabilities. If an industrial enterprise relies entirely on external, centralized public cloud APIs for real-time edge reasoning, they are fully exposed to vendor pricing volatility, algorithmic token inflation, and API gating.

If a public cloud hyperscaler alters its API billing structures or caps token throughput during a peak global demand cycle, your physical operations face immediate, unmitigated constraints.

Calibrating token economics via prepackaged skills

To counter this vulnerability, forward-thinking platforms are deploying structural mitigation strategies centered around pre-coded, modular “skills." In this context, a skill is a highly efficient, composable block of encapsulated domain logic, localized data mappings, and pre-configured business rules.

Rather than letting a large language model burn expensive compute tokens trying to open-endedly figure out data sources, operational boundaries, or process constraints, the platform invokes a specific, well-defined “skill” to anchor the scope.

A premier real-world example of this architecture in action is Aera Technology and their Decision Cloud platform, which utilizes a structured URAL (Understand, Recommend, Act, Learn) framework to re-engineer token economics.

Instead of making a probabilistic LLM the primary engine of heavy computation, the architecture treats the AI agent as a user of a highly deterministic platform. When an agent needs to evaluate an operational shift, simulate an inventory balance, or solve a complex supply chain constraint, it calls upon prepackaged, deterministic skills.

See also: L2L study repeats what many industrials already know: ‘Data chaos’ is stalling their digital transformations

By keeping language models focused strictly on reasoning over outcomes rather than processing raw data queries, this calibrated approach delivers an astounding 90% reduction in token consumption.

When structured correctly, the explosive expansion of computational scale becomes an indicator of operational success rather than an uncontrollable cost center. For example, Cognite recently reported an extraordinary 900% year-on-year growth in token consumption, driven entirely by the massive customer rollout of Atlas AI, their low-code industrial AI agent workbench.

This surge in token velocity represents a highly positive milestone: because it is backed by an industrial DataOps core that handles heavy contextual data preparation, every token consumed translates directly to scaled, real-world deployment and accelerated user adoption rather than wasteful, unguided cloud retries.

The core CapEx escape route: shared value pooling

This operational bottleneck explains why the push for heavy local edge computing is fundamentally a financial strategy designed to dismantle the OpEx token loop.

By investing up front in high-density local edge silicon (CapEx), industrial companies can permanently break their dependency on metered public cloud APIs. This infrastructure layer allows organizations to pull highly compressed, specialized open-weight reasoning models entirely out of the cloud to execute them natively on the plant floor.

While the hardware ecosystem provides the physical means to run local inference efficiently, traditional OT providers are deploying their own counterweight to the token economy through value-based licensing (VBL). This is the layer that completely refutes the narrative of a software billing apocalypse.

Long before Silicon Valley began debating token metrics, industrial software vendors recognized that forcing manufacturing engineers to purchase rigid, permanent concurrent seats for specialized software resulted in massive “shelfware waste”—expensive software assets sitting unused for 90% of the lifecycle.

Under a portfoliowide VBL system, the OT market is rapidly standardizing on a pooling currency model (such as Siemens making Simcenter SimSolid available via VBL directly within the Simcenter and Designcenter X ecosystem, mirroring the maturity of the AVEVA Flex Credits subscription platform or the long-established Altair Units mechanism).

Instead of buying restrictive individual seats for every specialized CAD tool, multi-physics simulator, or operations optimization module, an industrial enterprise pre-purchases a capped pool of shared corporate licensing tokens managed via a centralized cloud registry. This pooled currency behaves as a fluid asset across the multidisciplinary team:

• In the morning, an engineering technology (ET) designer can draw down a block of tokens to execute a complex geometric deep learning surrogate model. Once that upfront simulation task closes, those exact same tokens automatically return to the shared corporate registry.
• In the afternoon, a data science or reliability team can check out those same tokens to run an advanced process customization app inside an MES suite or tweak an automated work instruction.

VBL completely insulates the procurement cycle from the volatility of individual seat licensing, providing the predictable, capped cost parameters that procurement leaders demand while maximizing user accessibility across the entire value chain.

See also: Why industrial AI requires a data ops foundation to scale

Here are some actionable takeaways for industrial operators:

• Apply the right AI/ML tool for the job: Preserve high-cost cloud reasoning models strictly for upfront, low-frequency exploration, semantic data discovery, and high-level knowledge mapping. For continuous plant optimization, lean heavily on highly specialized, lower-cost model classes—such as causal AI, neuro-symbolic logic, and localized reinforcement learning policies—that deliver absolute mathematical determinism at a fraction of the computational overhead.
• Localize the inference loop: Shield your corporate balance sheet from variable public cloud invoices by fine-tuning and executing specialized open-weight models directly at the machine face using unmetered local edge iron drawing under 30W.
• Negotiate shared portfolio pools: Mandate that all core automation and operational software agreements transition to portfoliowide value-based licensing systems, allowing human engineers and automated agents to share a single, capped corporate software currency without triggering unexpected licensing violations.

Read the expanded version of this article at ARC Advisory Group, or continue the conversation with the author on his LinkedIn.