Agentic Workflows and the Impending Compute Infrastructure Scarcity

The market misprices AI by applying a linear perspective, which ignores how compute scaling compounds. This discussion shows we are nearing a non-linear jump in capability, where agentic workflows will trigger a massive, structural rise in compute demand. Investors and operators who ignore this shift will be caught off guard by the coming infrastructure shortage and the rapid obsolescence of traditional business models. The advantage belongs to those who recognize we are in a two-world economy: one where fast adopters are already seeing major efficiency gains, and another where the broader market waits for a revenue inflection that is already happening at the infrastructure level.

The Scaling Law and the Coming Weirdness

The current disconnect stems from a misunderstanding of how compute translates into capability. Stephen Byrd points to a consistent relationship: a 10x increase in training compute yields a 2x boost in model capability. While this may seem like a steady climb, the compounding effect over time is what Byrd calls getting weird.

If you increase the training compute by 10x, the capabilities of the models go up by 2x. Now, as you and I have talked about this a lot; just meditate on that for a moment.

-- Stephen Byrd

This is not just about faster chatbots. It is about models that can perform a much larger share of the economy's work at lower costs. The systemic result is a forced re-evaluation of business models. Investors can no longer rely on broad sector bets. They must distinguish between businesses that are immune to AI, those that will be enabled by it, and those facing total disruption.

The Hidden Economics of the Agentic Shift

Conventional wisdom suggests the current surge in token usage is a temporary result of developer-heavy coding tasks. Byrd argues the opposite: we are moving toward agentic AI, where models perform autonomous, multi-step tasks. This shift is not just additive; it is multiplicative in its demand for compute.

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 an enterprise moves from simple queries to agentic workflows, the compute cost and the potential savings skyrocket. Byrd points to a 55 dollar human-labor saving against a 5 million-token cost. Even with the five grand credit card bill that comes from unconstrained agents, the ROI for the adopter is a home run. This creates a feedback loop: as adopters realize these gains, they increase token usage, which further tightens the already scarce supply of compute and power.

Why the Infrastructure Crunch is a Feature, Not a Bug

The market is currently bearish on hyperscalers due to concerns over high CapEx and free cash flow pressure. However, this view ignores the pricing power inherent in the scarcity of compute.

The system is responding in real-time. Infrastructure providers see a massive uptick in urgency from the AI community. The revenue inflection is not a distant 2028 event; it is a 2026 phenomenon playing out in labs and among early adopters. The disconnect exists because the average enterprise still lags behind the fast adopters who have already stopped writing code entirely. The market will eventually bridge this gap as these early success stories become impossible to ignore.

Key Action Items

• Audit your business model for AI Immunity: Conduct a rigorous assessment of your core workflows. Determine which are vulnerable to disruption, which can be augmented, and which are truly immune. (Immediate)
• Implement Guardrail Economics: If you are experimenting with agentic AI, do not deploy without strict parameter constraints. As Byrd notes, unconstrained agents can lead to unexpected, massive compute costs. (Immediate)
• Track the Fast Adopter Signposts: Monitor early-adopter industries for tangible economic benefits. These will serve as the leading indicators for broader market shifts before they appear in aggregate earnings reports. (Next 3-6 months)
• Shift from Query to Agentic Thinking: Begin mapping your internal knowledge-work processes to agentic workflows. The 10x increase in token demand is inevitable; prepare the budget and infrastructure for this transition now. (Next 6-12 months)
• Invest in Infrastructure Visibility: Prioritize monitoring the supply chain for compute, power, and memory. This is where the pricing power resides, and it is the most durable layer of the AI value chain. (12-18 months)