A new exchange-traded fund (ETF) launched today signals a major shift in how Wall Street views artificial intelligence investment. Rather than chasing the headline-grabbing race to build bigger training models, Defiance ETFs introduced the Defiance Inference AI Chip ETF (ticker: AINF) on August 18, 2026, focusing instead on the chips that actually run AI models after they're trained. The fund tracks companies designing and manufacturing specialized processors for inference, the computational work that happens every time someone uses an AI application.

What Exactly Is AI Inference, and Why Does It Matter?

If you've ever asked a chatbot a question or used an AI agent to complete a task, you've triggered an inference workload. Training an AI model is a one-time, massive computational effort that happens in a lab or data center. Inference is what comes after: the ongoing, repeated process of running that trained model to generate outputs for users. As AI applications reach more people and more devices, inference workloads compound exponentially.

The distinction matters because it reshapes where computing power is needed. For years, AI investment focused almost entirely on training infrastructure, the enormous server farms required to build models like GPT-4 or Gemini. But once a model is trained, it needs to run somewhere, and that "somewhere" spans everything from hyperscale data centers serving millions of users to edge devices like smartphones and factory equipment running AI locally.

"The first phase of the AI buildout was about training models. The next phase is about running them. Every time someone uses an AI application, that is an inference workload, and we believe those workloads will keep compounding as AI reaches more users, more businesses, and more devices," said Sylvia Jablonski, Chief Investment Officer of Defiance ETFs.

Sylvia Jablonski, Chief Investment Officer, Defiance ETFs

What Types of Chips Does AINF Track?

The AINF fund targets companies across six distinct segments of the inference chip ecosystem. Each segment solves a different inference problem, from massive data center deployments to ultra-low-power edge devices.

• Inference GPUs: Graphics processing units originally designed for rendering video but repurposed for the massive parallel mathematical operations neural networks require.
• Custom AI ASICs: Application-specific integrated circuits engineered from scratch for maximum efficiency on particular inference tasks, rather than adapted from general-purpose designs.
• FPGA-based accelerators: Field-programmable gate arrays that can be reconfigured after manufacturing, allowing hardware acceleration to adapt as AI models evolve.
• AI-optimized CPUs: General-purpose processors enhanced with dedicated AI acceleration features like specialized instructions or neural processing engines built into the chip.
• Accelerator modules: Integrated hardware systems combining multiple AI chips, memory, and high-speed interconnects into single units for massive-scale inference in data centers.
• Neuromorphic chips: Brain-inspired processors using spiking neural networks to achieve ultra-low power consumption, ideal for inference on edge devices.

To be included in the fund, companies must derive at least 50 percent of their total revenue from, or demonstrate material involvement in, at least one of these segments. They must also meet minimum size and liquidity thresholds: a market capitalization of at least $100 million, free-float ownership of at least 10 percent, and average daily trading volume of at least $1 million.

How to Understand the Fund's Investment Strategy

The AINF fund uses a weighted approach to balance exposure across the inference chip market. Here's how the fund structures its holdings:

• Weighting method: Holdings are weighted by free-float market capitalization, meaning larger companies get proportionally larger positions in the fund.
• Concentration limits: No single company can exceed 20 percent of the fund's total weight, and the cumulative weight of all companies above 4.5 percent is capped at 40 percent, preventing over-reliance on a handful of mega-cap stocks.
• Rebalancing schedule: The fund rebalances and reconstitutes quarterly, after the close of business on the third Friday of March, June, September, and December, with provisions to add newly public companies or firms pivoting into the inference theme between scheduled rebalances.
• Geographic diversity: As of June 19, 2026, the underlying index included 26 companies, with 12 listed on non-US exchanges and significant exposure to companies domiciled in the United States and Taiwan.

Why Is Wall Street Betting on Inference Now?

The timing of AINF's launch reflects a fundamental recognition in the investment community: the AI infrastructure buildout is entering a new phase. Training models grabbed headlines and venture capital for years, but inference represents a larger, longer-lasting market opportunity. Every deployed AI model generates inference workloads continuously, and as AI adoption spreads across industries, those workloads multiply.

The fund's creation also signals confidence that specialized inference chips will become increasingly important. Rather than relying solely on general-purpose processors, companies are designing chips optimized for inference tasks, whether that means maximizing throughput in a data center or minimizing power consumption on a mobile device. This specialization creates a distinct investment category separate from broader semiconductor or AI infrastructure plays.

The underlying index was established on July 29, 2026, and is owned, calculated, and administered by BITA GmbH, a Germany-based index provider. The fund itself is managed by Defiance ETFs, with Tidal Investments LLC serving as sub-adviser.

For investors watching the AI sector, AINF represents a bet that the next wave of AI computing demand will come not from building bigger models, but from running existing ones faster, cheaper, and closer to where they're needed. Whether that thesis plays out will depend on how quickly AI adoption spreads and whether specialized inference chips prove more valuable than general-purpose alternatives.