Exponential Compute Scarcity Drives Long-Term Infrastructure Pricing Power

The market tends to view AI through a linear lens, expecting steady, incremental gains. Stephen Byrd of Morgan Stanley suggests instead that we are in the middle of an exponential shift driven by clear scaling laws. This creates a significant disconnect: while hyperscalers face short-term pressure from capital expenditures, the economic reality for adopters is a massive increase in efficiency. The hidden consequence of this gap is a looming, structural shortage of compute power that will give immense pricing power to those who control the underlying infrastructure. Investors and business leaders who mistake this transition for a standard tech cycle will likely miss the revenue inflection expected in 2026. This analysis is intended for those looking to distinguish between temporary market noise and the durable, long-term competitive advantage being built in the AI infrastructure layer.

The scaling law and the revenue gap

The core of the current AI dynamic is a mathematical reality: historically, a 10x increase in training compute results in a 2x increase in model capability. This is not just a technical metric; it is the engine of economic disruption. Because AI capabilities are advancing faster than the average enterprise can integrate them, a capability gap has emerged.

The capabilities of the models are advancing so fast that the average corporate user is not yet keeping up. There is this gap. But that will happen quickly.

-- Stephen Byrd

This gap creates a temporary illusion of low demand for some observers, but it masks massive, latent demand. Fast adopters are already seeing the benefits, such as software companies ceasing manual code writing, which provides a preview of the economic utility that will eventually force widespread enterprise adoption.

The agentic explosion and compute scarcity

Conventional wisdom suggests that the current surge in token usage is a temporary spike driven by software developers. Byrd argues the opposite: we are moving from simple query-based interactions to agentic workflows, where AI autonomously executes multi-step tasks. This shift is multiplicative.

When you go from a query-based usage of LLMs to an agentic use for any occupation, you see about a 10x increase in token usage per use of those models.

-- Stephen Byrd

When agents are set loose on knowledge work, token demand explodes. This creates a feedback loop: as models become more capable, they are used for more complex, agentic tasks, which in turn consumes more compute. For the hyperscalers and infrastructure providers, this ensures that the scarcity of compute is not a temporary bottleneck, but a structural feature of the next several years.

Why immediate pain creates lasting moats

The market is currently bearish on hyperscalers due to massive capital expenditure. However, this is a case of misaligned timescales. The immediate pain of high capital expenditure is the mechanism that will create a competitive moat. Because the economics for the end user are so favorable, saving $55 in human labor for a mere $5 in token costs, the demand for compute will remain inelastic. Those who secure the power and the hardware now are effectively locking in the pricing power of the future.

Key action items

• Audit your agentic readiness: Over the next quarter, identify high-frequency knowledge work tasks that can be automated via agents. Do not wait for enterprise-wide rollouts; start with small, high-leverage workflows.
• Establish guardrails for cost: As you scale agentic workflows, implement strict parameter controls. As Byrd noted, setting agents loose without constraints can lead to unexpected, exponential costs, such as the $5,000 credit card bill scenario.
• Shift from linear to exponential planning: Reassess your 18-month financial and operational projections. If you are modeling AI adoption as a linear improvement, you are likely underestimating the pace of disruption.
• Monitor infrastructure signposts: Watch the fast adopters in your specific industry. Their early successes are the leading indicators for the 2026 revenue inflection point.
• Prioritize compute-adjacent assets: For investors, focus on the infrastructure value chain, including power, memory, and semi-cap. The scarcity of compute is the primary bottleneck, and those who control the supply will dictate the terms of the market.
• Differentiate disruption vs. enablement: Conduct a business model review to categorize your operations: which parts of your business are immune to AI, which will be disrupted, and which will be enabled? This is a 12-18 month survival exercise.