Executive Summary

This article presents a central argument. The AI cycle is being rewritten, and the shift is not driven by technical breakthroughs. It is being shaped by the rise of an energy-based language and a new form of governance.

Microsoft CEO Satya Nadella framed AI efficiency as the number of useful tokens produced per gigawatt. This reframes performance from a discussion about GPU cost to one about energy. NVIDIA reinforced this shift through the concept of offtake, which turns demand from a market behavior into a structurally guaranteed outcome. These signals may appear separate, yet together they form the starting point of a new phase of market reflexivity.

Although the industry still lacks a killer application that can support large-scale compute demand, this gap has been filled by narrative language and governance structures. They allow the cycle to continue and delay the timing of a potential reversal.

Introduction

Two recent signals in the AI world may appear unrelated at first, yet they point toward the same direction.

The first comes from Microsoft. Satya Nadella described the core design principle of the Fairwater data center as producing the highest number of useful tokens per gigawatt of power. This shifts the discussion of AI efficiency from chip specifications to energy allocation. The language itself introduces a new frame of comparison.

The second comes from NVIDIA. In its latest earnings call, Jensen Huang emphasized three major platform transitions, five hundred billion dollars of visibility, the rise of sovereign AI, and the demand certainty created through offtake agreements. These signals seem rooted in supply chain planning and corporate strategy, yet together they reveal an emerging governance structure.

Interpreted separately, each development looks like a routine update within its own domain. Placed on the same timeline, however, they form a clear pattern. The constraints shaping AI are shifting from technology and cost to energy and governance. The industry narrative is also moving from performance comparisons to power allocation and institutional guarantees.

This shift is beginning to reshape the AI capital cycle. It is no longer driven only by technical progress. It is being accelerated by the interaction of narrative and governance. This is why the AI bubble does not appear close to a reversal. Instead, it has entered a phase where the turning point is pushed further out. The following sections explain how this structure is taking shape.

Tokens per GW Has Become a Language of Price Governance

Microsoft’s adoption of tokens per gigawatt appears to emphasize energy efficiency, but its meaning goes much deeper. It shifts the unit of comparison in AI from GPU pricing to power, and introduces a new language for procurement and capital allocation.

Over the past year, Jensen Huang has repeatedly mentioned the idea of the “free GPU.” The point is not that GPUs come without cost. It means that under a fixed power budget, the price of hardware becomes less important as long as it produces more tokens.

Once performance is measured through energy, the cost of the GPU naturally becomes secondary. Within a fixed gigawatt budget, the architecture that generates the most model output is effectively the cheapest one.

This is why the adoption of tokens per gigawatt represents a shift at the level of narrative.

The Shift From Cost Metrics to Energy Metrics

When companies evaluate AI investments, procurement teams and CFOs often face the same challenges:

• Highly volatile GPU prices
• Rapid and overlapping hardware generations
• Complex and opaque TCO calculations
• Unpredictable AI development costs

In such an environment, companies hesitate to anchor long-term investment decisions to specific hardware models.

This is why Satya Nadella’s comment matters. It anchors AI cost not in hardware, but in energy.

Once energy becomes the unit of comparison, everything becomes simpler. Companies no longer ask which GPU is cheaper. They ask which architecture produces the most revenue under a fixed amount of power.

This shift in language effectively rewrites the pricing system of the market.

Anchoring AI Economics to Power Changes the Logic of the Industry

When companies adopt tokens per gigawatt:

• High GPU prices become less of an issue as long as energy-to-token efficiency is superior
• Investment cycles naturally extend because energy infrastructure has long planning horizons
• Larger data centers become reasonable and necessary
• The market begins to see AI not as a product, but as an energy conversion system

This language is not only about efficiency. It is also a form of governance. It reshapes how companies plan AI, how cloud providers design data centers, and how capital markets understand growth potential.

Once energy becomes the center of the narrative, AI is no longer a competition of technology but a competition of energy allocation. This shift in narrative is the first reason why the AI bubble may take longer to reach a turning point.

Offtake Is Not a Demand Guarantee. It Is a System for Manufacturing Demand

NVIDIA referenced the idea of offtake in its earnings call. The term is unfamiliar to many in the tech industry. It is commonly used in energy, mining, petrochemicals, and large-scale infrastructure. It does not mean guaranteed shipping. It means that future output is guaranteed to be absorbed. In other words, it is not market demand. It is institutionalized demand.

This is why offtake rarely appears in the context of traditional technology companies. NVIDIA’s decision to introduce it into the AI sector is one of the most important signals for understanding this phase of the AI cycle.

When NVIDIA uses offtake to describe its customer relationships, it is expressing something deeper. NVIDIA is building a market that can absorb future supply. It is not only assuring that GPUs will be used. It is shaping an environment where the entire industry believes that once you invest in GPUs, there will be users to consume that compute in the future.

How Offtake Reshapes Investment Behavior Across the AI Supply Chain

In the energy and mining sectors, the purpose of offtake agreements is to help:

• Investors commit capital
• Supply chains expand production
• National-level projects move forward

When NVIDIA brings this model into AI, the effects become clear:

• GPU cloud becomes investable
• CSP expansion gains legitimacy
• Frontier models gain real operating space
• Supply chains extend planning horizons and expand capacity
• The market believes that once compute is built, someone will use it

Within this framework, NVIDIA is no longer only a supplier. It becomes the engine that generates global demand for AI compute.

It is not waiting for demand to appear. It is creating demand. It is not observing where the market will go. It is setting the path the market is likely to follow. This is a key reason why the AI bubble is unlikely to unwind quickly.

Satya and NVIDIA Are Redirecting AI’s Constraints Into the Physical World

In earlier phases of AI expansion, the main constraints were financial and operational. Companies slowed down because of GPU costs, development expenses, TCO pressures, and uncertainty about model performance. The recent statements from Satya Nadella and NVIDIA signal a shift away from these limits.

Satya moves the constraint from cost to energy. When companies adopt tokens per gigawatt, the question is no longer which GPU is more cost effective. It becomes how to generate the greatest amount of model output within a fixed amount of power. This shift turns hardware prices into secondary considerations and places energy at the center of cost anchoring.

NVIDIA moves the constraint from product to governance. Through offtake agreements, the market no longer worries about uncertain demand for GPUs. It begins to believe that once compute capacity is built, someone will use it. Demand becomes an institutional certainty rather than a market variable.

These two developments represent different sides of the same movement.

Once Cost and Product Fade as Limits, Physical Boundaries Take Over

The combined effect of Satya’s language and NVIDIA’s governance model is clear. AI is becoming less constrained by financial factors and more constrained by the physical world.

Physical constraints appear later than financial constraints and are more difficult to overcome. They cannot be adjusted as quickly as budgets and cannot be redesigned as easily as model architectures. Physical constraints involve:

• Building power plants
• Expanding data centers
• Strengthening the grid
• Scaling liquid cooling
• Constructing new campuses

These shifts occur over years, often longer.

As AI’s constraints shift toward physical conditions, the cycle evolves in four ways:

• The bubble naturally lasts longer
• Declines in efficiency do not cause immediate reversals
• Investment accelerates as narrative strength grows
• Companies plan AI based on energy scale rather than expense scale

This structure is rare in the technology sector, yet in this AI cycle it is emerging rapidly.

Satya’s energy language and NVIDIA’s governance architecture confirm this shift almost simultaneously.

The Three Platform Transitions Reflect a Single Underlying Shift in Energy

In NVIDIA’s earnings call, the three platform transitions are often interpreted as the natural progression of a technical roadmap.

• From CPU to GPU
• From ML to GenAI
• From GenAI to Agentic AI

On the surface, these transitions represent higher performance, stronger architectures, and more complex models. Viewed together, however, they reveal a deeper common axis.

Each transition raises AI’s energy requirements to a new level.

• The move from CPU to GPU shifts workloads from general-purpose computing to highly parallel processing, which sharply increases power density.
• The move from ML to GenAI transforms statistical optimization into large-scale model inference, leading to exponential growth in energy demand.
• The move from GenAI to Agentic AI extends generation into longer sequences, deeper reasoning, and continuous decision making, which pushes energy consumption even higher.

These shifts follow the same underlying curve. Every new phase of technology moves toward more energy, deeper computation, and longer sequences.

Platform transitions are therefore not only technical events. They are energy events. They form an energy staircase that forces AI demand to expand in step with available power.

This staircase is also what makes the language of tokens per gigawatt meaningful. When the primary difference between platforms is the amount of power needed to complete the work, power naturally becomes the unit of comparison for AI.

Technical differences become energy differences, and energy differences become the new units of narrative and governance.

This is a central reason why the AI cycle is being rewritten.

Treating Energy as the Unit of Value Turns AI Into Infrastructure

The lifecycle of technology products is usually short and compressed. They innovate quickly and reverse quickly. Demand can swing with prices, product generations, or market sentiment. But when the basic unit of comparison for AI moves from FLOPS to gigawatts, the entire rhythm of the industry begins to change.

Energy systems and infrastructure operate under a very different logic. They share several characteristics:

• Long-term investment
• Long-term returns
• High fixed costs
• High visibility
• National or cross-border capital

Once energy becomes the unit of value for AI, the supporting institutions and cost structures begin to resemble those of the energy sector rather than the tech sector. In other words, AI stops behaving like a product and begins to behave like infrastructure.

When AI is treated as infrastructure, two effects emerge.

First, the cycle becomes significantly longer. Companies plan data centers and energy facilities on three to seven year horizons rather than seasonal or product timelines. This makes AI investment less prone to sudden contraction.

Second, the speed of a bubble’s reversal decreases. Infrastructure assets are not cut aggressively in a short period. They are delayed, phased, or adjusted in pace. Even if model progress is faster than expected or revenue lags in the near term, capital does not immediately retreat.

This is also why NVIDIA frames AI as a sovereign-level deployment. Once AI enters national budgets, its cycle begins to follow the logic of energy and grid investment rather than tech valuations.

The constraints on AI therefore shift from commercial constraints to physical constraints. Physical limits appear later than financial limits and are more difficult to reverse.

Reflexivity Accelerates When Narrative and Governance Interlock

Reflexivity tends to emerge when two conditions appear at the same time.

First, the narrative must provide a clear justification for investment.

Second, the institutional structure must provide certainty about future demand.

In this phase of the AI cycle, Satya Nadella and NVIDIA each supply one of these conditions.

Satya provides a new narrative language. If power is fixed, the architecture that produces more useful tokens has a stronger economic rationale. This gives companies a stable reference point even as AI technologies evolve rapidly.

NVIDIA provides institutional certainty. Once you invest in GPUs, NVIDIA ensures that your compute capacity can be absorbed. It turns demand from a market behavior into a structured guarantee, which removes the anxiety of whether your compute will be used.

When narrative language and governance structures connect, a classic reflexive loop begins to operate.

• Higher GPU prices imply the potential for more tokens
• More tokens expand the imagined revenue space
• Larger imagined revenue leads to more aggressive investment
• Faster investment increases energy needs
• Rising energy needs make tokens per gigawatt more central
• The more central this metric becomes, the stronger NVIDIA’s governance position grows

This is a self-reinforcing cycle, not a linear growth path.

Within this cycle, the AI bubble does not move toward an early end. It enters a phase where narrative and institutional support extend the timing of any reversal.

Conclusion: The Real Turning Point Will Form on the Energy Side

Placing Satya Nadella’s language and NVIDIA’s institutional approach in the same frame reveals a clear structural shift. The constraint on AI is moving from technological limits to energy limits. Energy constraints appear later than financial constraints and are far more difficult to adjust quickly.

This means we are not at the end of a bubble. We are in the phase where market reflexivity accelerates.

The greatest uncertainty today comes from the demand side. OpenAI has not yet generated usage at a scale that can support global GPU deployment. Enterprise adoption still takes years. Large models, agentic AI, robotics, and world models are all in early stages without a clear killer application. Judging by demand alone, AI investment appears too fast and not fully grounded.

Yet this is exactly where reflexivity begins to operate. When demand is uncertain, narrative and governance temporarily substitute for demand itself.

• Energy language, expressed through tokens per gigawatt, provides a rationale for continued investment.
• Offtake agreements provide certainty about future absorption of compute.
• Sovereign AI moves capital spending from the corporate level to national budgets, which lengthens the cycle and slows reversals.
• Supply chain construction requires three to seven years, which means investment does not contract quickly even when demand lags.

A true turning point will emerge only when several conditions appear at the same time:

• Power expansion fails to keep pace with investment plans
• Cooling and data center capacity reach real limits
• GPU and HBM supply begins to exceed demand
• Token growth significantly outpaces revenue growth

Some early signs have already surfaced. Data centers in the United States, Europe, and Asia face delays in power access. Cooling and facility upgrades lag behind the rapid rise in GPU power density. Large AI companies are beginning to show token growth outstripping revenue growth. These are early indicators that a future reversal is possible.

For now, these constraints are not strong enough to slow the overall expansion. Narrative language creates a new basis of comparison, and governance structures provide demand assurance. Companies plan AI according to energy scale rather than cost scale. Supply chains believe that investment will find a market. As a result, even with emerging signals, the force of reflexive acceleration remains stronger.

Until then, AI investment will continue through a phase propelled by both narrative and governance. The path will not be smooth or even, but it will continue to extend.

We are living within this cycle.

Note: AI tools were used both to refine clarity and flow in writing, and as part of the research methodology (semantic analysis). All interpretations and perspectives expressed are entirely my own.