Millie Summary:
- AI is entering a new economic phase where the key question is no longer âWhat can AI do?â, but âWhat does AI cost per business outcome?â
- The rise of agentic AI creates a cost paradox: smarter, multi-step AI workflows can become more expensive even as token prices fall.
- The organizations that win will be âcompute-aware enterprisesâ
Written by Jordyn Geiger, Growth Development Analyst
For much of the generative AI era, enterprise leaders have focused on a simple question: What can AI do? That was the right question for the first wave of experimentation. It helped organizations test copilots, automate content generation, summarize knowledge, accelerate software development, and reimagine parts of knowledge work.
The next wave requires a sharper question: What does AI cost per business outcome?
That shift matters because AI is moving from a capability conversation to an operating economics conversation. The unit of this new economy is the token: the fragment of text, code, or data that AI models consume and produce. Every prompt, answer, tool call, retry, search result, document chunk, context window, and agentic reasoning step consumes tokens. In practical terms, intelligence is becoming a metered service, much like cloud compute, bandwidth, or electricity.
For many organizations, the challenge is no longer proving that AI can work. It is proving that AI can work repeatedly, securely, and economically inside the realities of their business. The strategic implication is clear: the winners in enterprise AI will not simply be the organizations with access to the most powerful models. They will be the organizations that know how to orchestrate models, control token consumption, modernize data foundations, redesign workflows, and measure AI value at the task level. For MILL5 clients, this is where AI strategy becomes enterprise transformation.
The market has entered the token economy
AI is becoming cheaper at the unit level and more expensive at the system level. That is the central paradox now shaping the market.
The cost of inference for a GPT-3.5-level model reportedly fell roughly 280-fold in two years, while training compute for frontier models has continued to grow at an extraordinary pace. At the same time, token consumption and enterprise spending are surging. Google was cited as processing 3.2 quadrillion tokens per month by May 2026, and enterprise generative AI spending rose to approximately $37 billion in 2025.
This is not a contradiction. It is the familiar pattern of elastic demand: when a powerful resource becomes cheaper, organizations find more ways to use it. New workflows become viable. More employees experiment. More software products embed AI. More agents run in the background. More tokens are consumed.
The result is a new form of technology demand: falling unit prices, rising total spend, and growing operational complexity. Cloud followed a similar arc. It began as a lower-cost infrastructure option and evolved into a consumption-management challenge. AI is now entering the same phase. Value will not come from simply deploying more AI; it will come from designing, governing, securing, optimizing, and operating AI with discipline.
The agentic cost paradox
The most important shift is the rise of agentic AI. A traditional generative AI interaction often answers a single prompt. An agentic system may plan, reason across multiple steps, call tools, critique its own output, retry failed attempts, re-read context, search external systems, and loop until a task is complete. Each step consumes tokens.
That creates what the source paper calls the agentic cost paradox: even when token prices fall, an AI workflow can become more expensive if the number of tokens required per task rises faster. Some agentic workflows can consume up to 1,000 times more tokens than a standard query. As a result, a task may be technically automatable but economically unattractive.
The implication is not that agentic AI is overhyped. It is that agentic AI must be engineered, governed, and economically instrumented. A poorly designed agent may reason too long, retrieve too much irrelevant context, use expensive models unnecessarily, repeat work that could have been cached, or apply probabilistic reasoning where deterministic software would be cheaper and more reliable.
A well-designed agent does the opposite. It uses the smallest sufficient model, retrieves only the context required, limits unnecessary loops, escalates intelligently, and measures whether the output was accepted, reused, or converted into business value. This is why the next frontier is not model selection alone. It is AI orchestration.
For enterprise leaders, the lesson is straightforward: AI cost control is not a finance afterthought. It is an architecture principle.
Compute is now a strategic constraint
AI may feel digital and weightless to users, but its economics are rooted in chips, data centers, power, cooling, networks, cloud capacity, and capital expenditure. The token economy has a physical footprint.
Data-center electricity consumption was estimated at about 415 TWh in 2024 and is projected to rise to roughly 945 TWh by 2030, according to the International Energy Agency’s Energy and AI report. Hyperscaler capital expenditure is projected to exceed $1.15 trillion from 2025 through 2027. These numbers matter even for companies that will never build a data center.
Compute scarcity influences vendor pricing. Power constraints shape cloud-region availability. GPU capacity affects product roadmaps. Latency and data gravity influence architecture. Model-provider economics affect pricing durability. Energy policy and infrastructure bottlenecks will shape where AI workloads can scale.
For executives, this changes how AI strategy should be governed. AI can no longer be treated as a collection of isolated pilots or departmental productivity tools. It should be treated as a business-critical operating layer that depends on cloud architecture, data readiness, cybersecurity, vendor strategy, and financial discipline.
The board-level questions are changing: Can we afford to scale the AI workflows we are piloting? Do we know our cost per successful outcome? Are we using frontier models where they matter, or everywhere by default? Is our cloud and data architecture designed for AI efficiency? Are we reducing work, shifting work, or simply adding AI cost on top of existing processes? These questions separate AI experimentation from AI transformation.
The future of work is task recomposition
The future of work will not be defined by a simple contest between humans and machines. It will be shaped by how work is decomposed, redesigned, and redistributed.
A job is a bundle of tasks. Some tasks are repetitive and rules-based. Some require judgment, empathy, domain context, or accountability. Some are high-volume and low-risk; others are low-volume but business-critical. AI will affect each category differently.
That is why exposure does not equal job loss. A task can be affected by AI without the job disappearing. In fact, real-world usage appears more balanced than the most aggressive automation narratives suggest, with AI being used for both augmentation and automation. The most valuable enterprise AI programs may not be those that pursue headcount reduction first. They may be the ones that capture expert knowledge, reduce time-to-proficiency, improve decision quality, accelerate throughput, and allow teams to focus on higher-value work.
This is especially important for organizations facing talent shortages, rising customer expectations, technical debt, or operational complexity. AI can become a mechanism for institutional learning: capturing what the best people do, embedding it into workflows, and helping the broader workforce operate at a higher level. But that outcome is not automatic. It requires intentional design.
The MILL5 perspective: build the compute-aware enterprise
At MILL5, we see the token economy as the next phase of enterprise AI maturity. The first phase was experimentation: proofs of concept, copilots, chat interfaces, and model exploration. The second phase is integration: embedding AI into applications, workflows, data platforms, customer experiences, engineering pipelines, and operational processes. The third phase, now emerging, is optimization: managing AI as a production system with measurable economics, resilient architecture, responsible governance, and clear business accountability.
We call this the compute-aware enterprise.
A compute-aware enterprise does not ask, how many AI tools can we deploy? It asks, where can AI create measurable value, and how do we deliver that value with the right model, the right data, the right controls, and the right cost structure?
Five moves for leaders
1. Manage AI at the task level. Start with the work, not the model. Decompose workflows into their component activities and identify where AI can automate, augment, accelerate, or improve quality. For each task, understand the current cost, cycle time, error rate, risk level, required judgment, and value of an accepted AI-assisted outcome.
2. Measure cost per accepted outcome. Vendor spend and license costs are useful, but they are not enough. Leaders need visibility into cost per resolved ticket, accepted answer, approved code change, reviewed claim, generated insight, hour saved, or dollar of revenue influenced. This exposes the difference between AI activity and AI value.
3. Orchestrate models instead of defaulting to the most powerful option. Many enterprise tasks do not require frontier reasoning. Some require classification, extraction, retrieval, deterministic workflow automation, or human review. A compute-aware architecture routes simple tasks to smaller models, reserves premium models for high-value reasoning, caches repeated work, limits agent loops, and uses deterministic code when it is cheaper and more reliable.
4. Modernize cloud and data foundations for AI efficiency. Poor data architecture becomes expensive in the token economy. Fragmented, duplicated, low-quality, or poorly labeled data forces AI systems to consume more context, make more calls, generate lower-confidence outputs, and require more human review. Clean data pipelines, governed access, retrieval-ready knowledge stores, metadata, observability, and cost controls are prerequisites for scalable AI economics. This is the foundation-level work at the heart of MILL5’s cloud and data strategy practice.
5. Redesign workflows and workforce roles together. AI transformation fails when people are treated as downstream recipients of technology decisions. Leaders should define where AI automates low-value work, where it augments expert judgment, where it coaches less-experienced workers, where it escalates to humans, and where it should stay out of the workflow entirely.
What this means for the market
The token economy will reshape enterprise technology markets. AI vendors will increasingly compete not only on model quality, but on cost-to-serve, latency, orchestration, integration, governance, and outcome-based value. Cloud providers will compete on infrastructure capacity, region availability, data services, energy strategy, and optimization tooling. Enterprises will demand more transparency into AI pricing and stronger ways to manage consumption.
Software pricing will also face pressure. Many SaaS vendors are embedding AI into existing products, but customers will increasingly ask whether those features deliver measurable value or simply add an AI premium. As usage scales, buyers will become more sophisticated. They will want to understand how often an AI feature is used, how much work it improves or replaces, what risks it introduces, and whether the economics improve over time.
For consulting, implementation, and managed-services partners, the mandate will shift as well. The market will reward partners who can connect strategy to architecture, architecture to data, data to workflow, workflow to adoption, and adoption to measurable outcomes. That is why AI transformation cannot sit in a single function. It requires strategy, engineering, cloud, data, cybersecurity, operations, finance, HR, and change management to work together.
The leadership question
The central issue is not whether compute will keep rising. It almost certainly will. The more important question is where organizations point it.
AI can be directed toward shallow automation that removes labor without improving the system. It can be directed toward uncontrolled experimentation that increases cost without improving outcomes. It can be directed toward fragmented tools that add complexity to already complex environments.
Or it can be directed toward better work: better customer experiences, better software delivery, better operational resilience, better decision-making, better knowledge transfer, better use of scarce talent, and better economics.
The first era of AI rewarded curiosity. The next era will reward operating maturity. In the token economy, intelligence may become abundant, but profitable intelligence will remain scarce. The organizations that win will manage AI with the same rigor they bring to cloud, data, security, finance, and enterprise operations.
They will know that every token has a cost. They will know that every workflow has an economic threshold. They will know that the most powerful model is not always the best model. They will know that augmentation can be more valuable than substitution. And they will know that AI transformation is not a technology rollout. It is a business redesign.
Ready to make AI measurable?
The next phase of AI will not be won by deploying more tools. It will be won by building AI systems that are secure, scalable, cost-aware, and tied directly to business outcomes.
MILL5 helps organizations move from AI experimentation to enterprise-grade execution by identifying high-value use cases, modernizing the cloud and data foundations behind them, orchestrating the right models for the right tasks, and establishing governance that measures what matters: cost, performance, adoption, risk, and business impact.
AI should not just be powerful. It should be profitable, governed, and built to scale. Connect with the MILL5 team and Jordyn Geiger at jordyng@mill5.com to start building your compute-aware AI strategy today.
Source note
This article draws on Alejandro J. Guipe Salazar’s 2026 working paper, The Token Economy: Compute Power, Automation, and the Future of Work in an Ecosystem of Human Labor, Capital, and Artificial Intelligence [Working paper]. SSRN.
