Variant Investment Memo: Ten Judgments on the AI Era

Authors: Alana Levin, Kayvon Tehranian

Compiled by: TechFlow

TechFlow Editor's Note: Variant just closed its fourth fund of $222 million, shifting its investment theme from "ownership economy" to "autonomy". Partners Alana Levin and Foundation founder Kayvon Tehranian jointly released ten investment hypotheses, covering AI Agents, compute, open source, multi-model systems, and more. This list is essentially a public scorecard—they've laid their bets on the table, waiting for the market to grade them.

Variant invests in technologies that expand autonomy—giving individuals or organizations greater freedom to build, customize, and act in their own way.

Automation helps developers, individuals, and organizations do more with fewer resources. Well-done automation gives users new capabilities while giving them greater control over their tech stack.

This scope is very broad, so in our exploration process, we use a set of hypotheses to anchor our thinking direction.

Below are ten working hypotheses formed from discussions over the past few months. They help us decide where to invest time and attention. Some hypotheses may be incomplete, some may be wrong. But this exercise forces us to state implicit assumptions clearly, giving us a framework to judge: how the world will change, what products are important, where value will accumulate, and who can capture it.

Agents Will Become the Main Source of Internet Traffic

Today's software defaults to having a person clicking buttons and viewing interfaces. But more and more products need to be designed for both humans and Agents. Agents run 24/7, can be infinitely replicated, and have behavior patterns and incentive mechanisms completely different from humans.

Compute Shortages Will Last for Quite Some Time

Scarcity makes infrastructure important. Startups that can reduce compute costs, improve utilization, or open up new supply sources will become increasingly valuable.

Open Source Will Play an Increasingly Heavy Role in the AI Tech Stack

For enterprises, open source provides a viable path to compete with frontier labs. For developers and consumers, it lowers trial-and-error costs and expands access to state-of-the-art models.

The Barrier to Entry for Frontier AI Will Become Higher and Higher

Frontier AI is becoming increasingly "permissioned". Demand for open, permissionless alternatives will grow accordingly, as developers need to retain autonomy.

The Number of Truly Important Models Will Grow by an Order of Magnitude

As models become more specialized, the ability to select and orchestrate models becomes more valuable than building around a single model.

Multi-Model Systems Will Become the Default Choice

The application layer manages the complexity of distributing tasks across multiple models; users don't need to worry about this.

Harness (Orchestration Layer) Is the Core Where Value Accumulates in the AI Tech Stack

Models are becoming easier to replace; lasting relationships shift to products that orchestrate models and own the user experience. The Harness mentioned here refers to the application layer wrapping the model—it is responsible for scheduling, routing, context management, and is the layer directly facing users.

Current Agent Systems Are Mostly Single-Player Mode, the Frontier Is Multi-Player Collaboration

The challenge shifts from "building a single Agent" to "coordinating shared context, permissions, workflows, and collaboration across organizations". Defining and owning the control plane is an important leverage point.

AI Will Create Low-Cost Alternatives for Many Expensive Services in Society

Industries like healthcare, education, and legal services are particularly suitable for new entrants to rethink the economic models of expertise delivery. Some of the most novel solutions may come from directly empowering end users.

AI Will Make Many Previously Impossible Markets Economically Viable

When Agents can automate much of the work required to start and scale a company, entrepreneurship in more fields becomes viable, allowing founders to focus on the hardest problems.

We are releasing this list as a public scorecard for early Q3 2026. We look forward to tracking how well these judgments stand the test of time, and will continue to update as our understanding evolves.

If you are building products around these hypotheses, we want to talk. Especially if you disagree with some of these judgments—tell us where we are wrong.