The most useful artificial intelligence (AI) signal I have is observing how our own product team uses these tools. A year ago, most of our AI usage was interactive. Someone sat in the terminal or a chat window, asked a question, got an answer in seconds, and iterated. Human attention and machine latency were tightly coupled. You wanted an answer as soon as possible, so you can keep working. The unit of work was the answer. That seems to be changing.
Our team is increasingly delegating bounded outcomes to agents that utilize a proprietary harness and reviewing the result later. For example, engineers and quants queue coding tasks that run for hours and return as a pull request. The unit of work is shifting from the answer to the deliverable, and the relevant clock is shifting from seconds to hours or overnight. I think that is an important signal for investors because it changes the economics of AI infrastructure.
From Latency to Utilization
Today's inference stack was largely built around a human waiting on the other end of a prompt. That makes latency valuable: small batches, fast response times and capacity held in reserve for demand spikes. The evolved interaction with AI changes the equation.
When nobody is waiting on the token, the system can optimize for throughput, utilization, and cost. Sail's published work is one example, showing 90.72% accuracy on its BrowseComp-Plus benchmark, at 6–35x lower cost than the comparison providers it tested. Sail's architecture relies heavily on inexpensive open models for the bulk of the work, with more capable models used selectively for harder reasoning and orchestration.1 OpenAI is making the same economic distinction at the API level. Its Batch API processes asynchronous workloads within 24 hours at a 50% discount to synchronous pricing.2 The implication is more important than either example because tokens per task can rise dramatically while the price per token falls.
As inference gets cheaper, more workloads become economically viable: more research, more simulation, more coding, more evaluation, and more agentic workflows. The market can therefore have much higher compute consumption even as inference revenue per token compresses.
The Power Implication
The next bottleneck is electricity. Not all AI loads are flexible. Real-time inference still has a latency requirement. But a growing share of AI work can be scheduled over hours rather than seconds. That creates a new class of load that can be shifted across time or facilities without materially changing the outcome.
This is already moving from theory into practice. A field demonstration at a hyperscale facility in Phoenix, Arizona, used software controls to reduce the power consumption of a 256-graphics processing unit (GPU) AI cluster by 25% for three hours during peak demand while maintaining quality-of-service guarantees. No hardware modifications or energy storage were required.3
The investment implication is straightforward: as more compute becomes tolerant of delay, the valuable megawatt is increasingly the one that is cheap, reliable and available at high utilization, rather than simply the one that can be delivered at the fastest possible response time.
Economics Accrue to Contracts, not Bottlenecks
The question is not which generation technology wins every future data center. It is not even which parts of the stack are bottlenecked on supply, because a bottleneck is a temporary condition that invites competition and capital until it clears. The question is which parts of the infrastructure stack can convert scarcity into contracts, and get paid as compute demand grows regardless of which model or application ultimately wins.
In constrained US markets, natural gas has an important advantage: it can support grid generation, behind-the-meter generation and new power plants where electricity capacity is difficult to secure. And the midstream asset can be contracted for a decade or more.
Energy Transfer's disclosures make the trend concrete. The company has multiple long-term agreements tied to data center and power demand, including approximately 900,000 cubic feet per day (Mcf/d) of natural-gas supply agreements with Oracle's US data centers and a 10-year agreement with Fermi America for an initial approximately 300,000 million British thermal units per day (MMBtu/d) of gas supply and pipeline interconnection. Energy Transfer's disclosures also identify more than 6 billion cubic feet per day (Bcf/d) of contracted pipeline capacity associated with new demand, with an 18-year weighted-average life and more than $25 billion of expected firm-transportation revenue.4
Several of these projects remain subject to customer elections, final investment decisions or other conditions, and long-dated contracts convert model risk into counterparty credit risk rather than eliminating it altogether, so I would treat them as contracted opportunities rather than fully secured cash flow. Nonetheless, this contract structure is the key.
AI models, inference economics and the amount of compute required for a given task will change. But a long-term pipeline contract does not depend on predicting which model wins. It depends on a customer needing large amounts of reliable energy infrastructure.
And the gas market already has other structural demand anchors. US liquefied natural gas (LNG) exports averaged 17.4 Bcf/d in the first half of 2026, up 23% from the same period a year earlier, and the US Energy Information Administration (EIA) expects exports to rise further as new liquefaction capacity comes online.5
The Thesis
I think the market is still underwriting AI primarily as a latency problem: faster chips, faster networking, and faster inference. It is becoming a utilization problem. As more AI work moves from synchronous interaction to asynchronous, throughput-oriented workloads, the cost of useful computation should fall, and the volume of computation should rise. A growing share of that demand can be scheduled around the power system, making cheap and reliable electricity a more important part of the AI cost curve. That pushes value toward the infrastructure that can deliver power at scale.
In the US, gas midstream is one place where that value can become long-duration, fee-based cash flow. That combination is particularly attractive in growth-and-income portfolios: investors can participate in the infrastructure spending required to support a growing economy while also receiving current income and, where contracted volumes and capital investment continue to grow, a path to higher cash distributions over time. The market is focused on who builds intelligence. I think more attention should go to those who get paid to keep it running.
Important Disclosures & Definitions
1 Anand, K & Liang, T. (2026, April 22). Let the tokens flow. Sail powers efficient, reliable deep research. Sailresearch.
2 OpenAI. (2026). Batch API endpoint for asynchronous batch processing. OpenAI FAQ.
3 Colangelo, P., Coskun, A.K., Megrue, J. et al. (2025, December 5). AI data centres as grid-interactive assets. Nat Energy 11. Nature.
4 Energy Transfer. (June 2026). Investor Presentation. Energytransfer.
5 US Energy Information Administration. (2026, September 1). US LNG exports rose 23% in the first half of 2026 because of higher capacity. Eia.
AAI001223