Goldman’s AI Token-Price Warning: What It Means for Infrastructure Investors

AI inference is becoming dramatically cheaper. That can broaden access to models and stimulate new workloads, but it also complicates the investment case for the chips, cloud capacity and data centers being built to serve them.

Goldman Sachs’ warning is essentially about a mismatch: infrastructure investment is made years ahead of demand, while the price charged per unit of model usage can fall quickly. Silicon Data’s LLM Token Expenditure Index recently reached $0.97 per million tokens after declining 29% in August, according to reporting on the Goldman note. A separate frontier-model pricing index put prices 84% below its March 2023 baseline. 245

If usage growth does not offset those declines, the result could be too much costly compute capacity competing for lower-revenue inference workloads. That would matter across the AI infrastructure stack.

The economic question: volume must outrun price deflation

A falling token price is not, by itself, bearish. Lower prices can make AI practical for more applications, encourage experimentation and expand total consumption. The challenge for providers is whether growth in paid token volume—and improvements in their own operating efficiency—can exceed the decline in revenue per token.

Goldman’s concern, as reported, is that competitive model pricing may make per-token monetization less durable and ultimately leave the industry with excess infrastructure if computing consumption does not rise fast enough. 12

That distinction is important:

• Cheaper AI is good for users when it enables useful applications at lower cost.
• Cheaper AI can be difficult for infrastructure owners when capacity has been financed on assumptions of scarce, high-priced compute.
• The outcome depends on demand elasticity: whether a price reduction creates enough additional usage to preserve or increase total revenue and compute demand.

Why the capital-spending stakes are so high

The potential mismatch is occurring alongside a very large projected build-out. Goldman Sachs’ baseline estimate totals about $7.6 trillion of AI capital investment across compute, data centers and power from 2026 through 2031. Its model implies annual spending rises from roughly $765 billion in 2026 to $1.6 trillion in 2031; Goldman explicitly notes that these are assumption-sensitive estimates, not certainties. 9

This is why token-price trends matter beyond model providers. A large share of the expected spending supports long-lived assets—accelerators, servers, networking, facilities and power infrastructure. Those investments require sustained workloads and reasonable returns over time.

What it could mean for chipmakers

For semiconductor suppliers, a price-driven slowdown would not necessarily show up first as an immediate collapse in demand. Customers may continue building capacity while they pursue new AI products or attempt to secure supply.

The risk emerges if deployed capacity begins to outrun profitable demand. In that scenario, customers could slow new accelerator and server purchases, become more price-sensitive, or seek greater performance per dollar. Businesses with differentiated hardware, strong software ecosystems or efficient inference offerings may be better insulated than suppliers dependent on broadly similar capacity.

The practical question is whether orders are translating into sustained utilization and revenue-generating workloads—not simply whether equipment is being delivered.

Why data-center operators face a utilization test

Data-center projects involve major, long-lived commitments to land, power and facilities. If model inference becomes cheaper or more efficient, customers may need less infrastructure for a given level of application output. That could pressure occupancy, lease economics and returns on projects that assumed persistent capacity scarcity.

This does not mean data-center demand disappears. Power availability and the technical requirements of AI deployments can still constrain supply. But a capacity shortage in one location or power market can coexist with weaker economics elsewhere. Investors should separate physical scarcity from evidence that customers are consuming capacity profitably and renewing or expanding commitments.

Cloud providers and model hosts must replace price with volume and value

Cloud platforms and model hosts can benefit from lower inference prices because affordability may bring more developers and enterprise workloads onto their platforms. Yet lower prices also reduce revenue per unit of usage.

To protect gross profit, providers need some combination of:

1. Much higher token volume from recurring production use;
2. Efficiency gains that reduce their cost faster than prices decline; and
3. Higher-value services around models, such as deployment, security, data integration and workflow tools.

Competitive pricing in proprietary models is a central part of the risk described in the Goldman-related reporting. 12 The key is not the nominal token price alone, but the relationship between customer price, provider cost and total paid usage.

Agentic AI is the potential offset

The bullish counterweight is agentic AI: systems that perform multi-step tasks and may consume substantially more tokens than simple chatbot interactions. Goldman Sachs Research forecasts that consumer and enterprise adoption could increase monthly token consumption 24-fold between 2026 and 2030, reaching about 120 quadrillion tokens per month. 7

That forecast makes clear why falling token prices do not automatically signal falling compute demand. If agents move from demonstrations to frequent, valuable tasks, total usage could grow enough to support the infrastructure build-out.

But it is a forecast, not a guarantee. The infrastructure case relies on agents creating recurring workloads that users or businesses will pay for—not merely on a large number of low-cost experimental queries.

Metrics that matter more than headline token prices

Investors assessing the compute-oversupply thesis can focus on a few connected indicators:

• AI revenue growth versus token-price declines: Is total monetization expanding even as unit prices fall?
• Accelerator and server orders versus deployments: Are customers buying capacity that is actually being put to work?
• Data-center utilization, contracted capacity and renewals: Do demand commitments remain firm after facilities open?
• Cloud gross margins: Are efficiency gains and higher-value services offsetting lower inference pricing?
• Agent adoption in production: Are agents driving repeatable, business-critical workflows rather than isolated trials?
• Power availability: Is it a durable constraint on useful capacity, or simply a bottleneck ahead of demand?

The bottom line

Goldman’s warning is not that lower LLM token prices are inherently bad. Cheaper inference can be the mechanism that unlocks wider AI adoption. The risk is that price deflation and model efficiency reduce revenue and required compute faster than agent-driven usage grows.

For AI-infrastructure investors, that turns the story from a simple capacity build-out into a utilization and monetization test. The most resilient businesses may be those able to convert cheaper tokens into durable, high-value workloads while maintaining an advantage in cost, performance or service depth. 17