Plexo Capital's Lo Toney: AI 'inference economics' important to watch in 2026

Key Concepts

• Training (AI): The initial, one-time cost of building an AI model.
• Inference (AI): The ongoing operational cost of running an AI model in real-world applications.
• Inference Economics: The study of who pays for AI each time it’s used (inference) and whether revenue from that use scales faster than the cost.
• Vertical Integration: A company owning and controlling multiple stages of its supply chain (e.g., Google’s TPU chips and data centers).
• Monetization of Attention: A business model (like Meta’s) that relies on generating revenue from user attention.

AI Trade & Software Performance: Outlook for 2026

The discussion centers around the performance of tech stocks, particularly in relation to the AI “trade” and expectations for 2026. 2025 saw strong performance from hardware companies (Western Digital, Micron, Seagate) while software companies (Adobe, Workday, Datadog) underperformed, leading to investor skepticism about AI monetization within the software sector. The core question is whether software can “catch up” in 2026.

Apple vs. Meta: A Fundamental Difference

Leo, Founding Managing Partner at Plexo Capital, highlights a crucial distinction between Apple and Meta regarding AI. He argues that Apple’s strong underlying economics protect it even if its AI execution isn’t perfect. “What’s important about Apple is that even if their AI execution stumbles slightly, their economics still protect them. And that’s very different from a company like Meta.” This difference stems from how each company handles the costs associated with AI.

Training vs. Inference: Shifting Costs

A key framework presented is the distinction between AI training and AI inference. Training is a one-time build cost, while inference represents the ongoing operational cost of running AI. The conversation emphasizes a shift in focus from training to inference. “Training creates capability, inference determines profitability.” The critical question is whether revenue generated from inference will outpace the associated costs – this is termed “inference economics.”

Infrastructure Bottlenecks & Workflow Challenges

John raises a significant point about potential bottlenecks beyond infrastructure costs. He suggests that existing company workflows and data organization could hinder the scaling of AI use, even with sufficient infrastructure. “There are certain ways that people are used to getting things done, and there are certain ways that people have organized their data centers and their data. I’m not sure that enough companies are in position to scale their AI use…to capture the full expectation of the market.”

Company-Specific Analysis: Exposure to Inference Costs

The discussion then breaks down how different tech giants handle inference costs:

• Apple: Minimally exposed to inference costs because AI functionality largely runs through third-party apps. Apple benefits from AI without directly bearing the marginal inference costs.
• Microsoft: Inference costs are largely absorbed into existing seat licenses and contracts, minimizing direct exposure.
• Google: Leverages vertical integration with its TPU (Tensor Processing Unit) chips and centralized data centers to manage both training and inference. Google doesn’t subsidize inference; it “taxes” it, meaning users effectively pay for the computational resources used.
• Meta: Highly exposed to inference costs because its business model relies on monetizing user attention. Rising inference costs could outpace monetization if pricing power doesn’t improve.

Meta’s Acquisition of Manus: Addressing Financial Concerns

Meta’s acquisition of Manus is presented as a strategic move to address these financial concerns. Manus specializes in turning AI into practical tasks and workflows. This acquisition is intended to improve Meta’s Profit and Loss (P&L) statement. The acquisition is framed as fixing Meta’s financials in two ways: first, by improving the balance sheet through a previous transaction with Blue Owl (removing $30 billion from the balance sheet), and second, by incrementally improving the P&L through workflow optimization.

Logical Connections & Overall Perspective

The conversation logically progresses from a broad overview of market performance to a detailed analysis of the economic factors driving success or failure in the AI space. The emphasis on inference economics provides a framework for understanding the differing vulnerabilities of various tech companies. The discussion highlights that simply having AI capabilities isn’t enough; the ability to profitably run those capabilities is paramount.

Data & Statistics

• The performance of NASDAQ 100 companies in 2025 was cited: Hardware companies (Western Digital, Micron, Seagate) outperformed, while software companies (Adobe, Workday, Datadog) underperformed.
• Meta’s balance sheet was improved by a $30 billion reduction through a transaction with Blue Owl.

Conclusion

The key takeaway is that the success of the AI trade in 2026 will depend heavily on “inference economics” – the ability to generate revenue from AI usage that exceeds the ongoing operational costs. Companies with strong underlying economics (like Apple) or vertically integrated infrastructure (like Google) are better positioned to navigate this challenge. Meta, reliant on monetizing attention, faces greater risk and is attempting to mitigate this through strategic acquisitions like Manus, aiming to improve workflow efficiency and profitability. The conversation suggests that the market’s expectations for software performance in 2026 may be overly optimistic if companies cannot effectively manage inference costs and adapt their workflows.