Our 2026 Outlook

The easy version of an AI investing outlook is a list of sectors.

AI in healthcare. AI in legal. AI in sales. AI in defense. AI in education. Pick your vertical, add "agents," and call it a thesis.

I don't think that's the interesting question anymore. The real question is what happens when software changes again.

That has happened before. The move from licensed software to SaaS did not make software less important. It made software more important. More companies depended on it. More people used it. More work ran through it. But the economic shape of the business changed. The winners changed. The way you built, sold, and valued software changed with it.

That is happening again now. The wrong conclusion is that software is becoming less important.

The opposite is closer to true. We are about to get more software, in more places, built by more kinds of people, against more categories of work than the last generation of company-building could justify. When the cost of creation drops hard enough, more things become worth building. More internal tools. More workflow software. More vertical systems. More weird, specific, valuable products that used to die on the whiteboard because there were not enough engineers, not enough time, or not enough margin.

People notice that building software is getting cheaper and faster, then jump to one of two lazy conclusions. Either software is dying, or nothing fundamental is changing and the old SaaS playbook just gets an AI layer on top.

I think both readings miss the point. The deeper pattern is that great software has always taken capabilities that used to be scarce, expensive, or available only to experts and opened them up to everyone else. Cloud infrastructure did that for building internet applications. Better developer tooling and open source did that for writing and shipping code. QuickBooks did that for bookkeeping by encoding a bunch of expert logic into software normal people could actually use.

That underlying power did not go away. It got stronger.

This cycle is not just making software cheaper to produce. It is making far more capabilities legible, automatable, and available than they were before. That should increase the amount of software built, not decrease it. But it should also change what kinds of software companies matter and how value accrues to them.

Software is not becoming less central to the economy. If anything, it is getting more embedded in everything. What is breaking is the old path for building a large standalone software company. The standard SaaS playbook — build a product, sell seats into the enterprise, compound ARR, expand upmarket, and grow into a giant multiple — is getting weaker right as software itself becomes more abundant. That is the part I think many people still do not want to say out loud.

The reason is not that software stopped mattering. The reason is that more of software is becoming fluid, generated, orchestrated, and bundled into larger systems. The labs will keep shipping more of the obvious surface area. Internal teams will build more for themselves. Buyers will expect more automation and less dashboard tourism. More of the value will move from "access to a tool" toward "reliable execution of a job."

For a long time, it was enough to sell software that helped a person do work better. Increasingly, the more interesting opportunity is software that does the work, or software that lets a very small team build and operate systems that do.

That shift sounds semantic until you sit with it for a minute.

If software can draft, route, reconcile, test, follow up, watch, escalate, and learn, then the boundary between a software company and a business starts to blur. Some companies will still look like software companies in the traditional sense. Many of the best new ones will not. They will look like businesses whose core advantage is that their operating model is software-native from day one.

That changes company shape in practical ways. The next accounting winner may not look like software you log into and operate. It may look like a system that closes the books, flags the exceptions, and gets paid for the work being done. The next security company may not just sell a dashboard to a team. It may continuously test, verify, and help remediate vulnerabilities inside a defined loop. The next great vertical company may charge less for access and more for completed work, earned trust, and outcomes that would previously have required a much larger human operation.

That matters because many investors still talk about "AI companies" as if that is a category. It is not. It is closer to an enabling condition.

The winners in this cycle will not be the teams with the best AI slide. They will be the teams that deeply understand what autonomous or agent-mediated software makes possible, and then use that understanding to build something that could not have existed under the last playbook.

That is also why user expectations changed so fast.

The old product promise was: here is a better tool for doing a hard thing.

The new expectation is increasingly: why am I doing this at all?

I do not want help closing my books. I want my books closed. I do not want a slightly better interface for some repetitive, annoying workflow. I want the workflow handled, monitored, and escalated only when necessary. Once people feel that shift, a lot of last-generation software starts to feel clunky, not because it is broken, but because it still assumes the human should be carrying the burden.

You can already see where value is starting to accrue.

Some of it will accrue to the companies building the new stack: infrastructure, tooling, evaluation, orchestration, observability, permissioning, and all the ugly edge-case machinery required to make agentic software real in production. That layer matters because this new generation of software will not run on vibes. It will run on reliability, monitoring, control, and feedback loops.

Some of it will accrue to founders who use that stack to build entirely new businesses. Not "AI for X" as a feature layer. Not a prettier wrapper around an incumbent workflow. Businesses built around the fact that software can now act, not just assist.

That distinction matters because the business model often changes with the product. A seat-based tool gets paid for access. A system that closes books, handles claims, verifies compliance steps, or runs a recruiting workflow can get paid for work completed, risk reduced, or outcomes delivered. It may still look like software. It may partly look like a service. But the economic center of gravity is different.

Those are not really separate theses. They are different expressions of the same shift. If the way software is built changes, the shape of the businesses built on top of software changes too. The same technical fluency that lets someone build better infrastructure can also let them attack a market that previously required a much larger company.

This is one reason I think 2026 will be a clarifying year.

In 2024 and 2025, it was still possible to tell yourself a lot of comforting stories. Maybe every workflow gets a copilot. Maybe incumbents simply add a model layer and keep moving. Maybe a good interface plus a foundation model is enough to build a venture-scale company. Maybe demand alone will sort out the winners.

Some of that will still work. A lot of it will not. The market is getting less forgiving at the exact moment the tools are getting more powerful.

That combination matters. When the cost of building drops, the penalty for building something undifferentiated rises. When more software becomes economically viable, more people will build it. That makes distribution, workflow position, system-of-record access, and actual permission to take action matter more, not less. When the underlying models improve fast, superficial product advantages decay faster too.

This is why I am increasingly skeptical of startups whose core claim is basically "we applied AI to a known category."

That is not a company description. It is a timestamp.

The stronger question is: why does this team get to exist when the tools are available to everyone? What do they understand about the workflow, buyer, system of record, or distribution path that is not obvious from the outside? Where is the permission to act? Where is the feedback loop? What keeps the system getting better with use? Where is the monetized outcome that makes adoption durable?

Those questions sound basic. They are still where most of the truth lives.

Those questions also frame one of the more important debates inside this cycle: the AI holding-company thesis. If software commoditizes and margins compress, maybe the smarter move is to buy the service business itself, own distribution from day one, and automate it from the inside. I get the appeal, and in some categories it may work. But to me it still feels more like a hedge than a bet. It captures existing value and optimizes it. The more exciting path is when software opens up a capability that used to be locked, expensive, or impossible, and creates a new market around it.

One thing we have seen repeatedly in founder conversations is that AI amplifies people unevenly. The best technical builders get dramatically more leverage. They can ship faster, test more, and cover more surface area than small teams could a few years ago. Middling teams often just produce more slop, faster. The bottleneck does not disappear. It moves.

What used to look mostly like a headcount problem increasingly looks like a taste problem.

That has a second-order consequence that I think matters more than most people appreciate: early team formation is becoming one of the defining bottlenecks in company building.

This follows directly from the expansion in what software can now economically do. If more products, workflows, and businesses are suddenly worth building, then the scarce resource is not generic access to code generation. The scarce resource is people with enough taste and judgment to decide what should exist, what can actually work, and how to assemble a team that can build it without disappearing into noise.

It is a weird moment. The best people can do more on their own than ever. They also have more reasons to start companies themselves. That means a great founder's first five to ten hires are often harder, not easier, to make. You are not just recruiting talent. You are trying to assemble a tiny group with unusually high judgment, unusually high trust, and enough range to build in public while the ground is moving.

That problem shows up earlier than many investors want to admit. It is not a "later stage org design" issue. It shows up right after the first burst of product excitement, when the company has to become real.

Over the last few months, some of the most recurring founder questions we have heard were not about model choice or compute. They were about people. Not "how do I hire twenty engineers." More like: who is actually strong enough to work in this new mode? Who can use these tools without lying to themselves? Who can move from prompt-shaped output to production-grade systems? Who can operate without hiding behind process?

This is another reason the old startup templates are failing. For most of the last generation, founders inherited a fairly legible map. Hire the standard functions. Build the standard org. Sell with the standard motions. Add AI if the market forces you to. That map is less useful now. The best teams are rethinking not only what they build, but how they build, hire, and distribute from day one.

That does not mean every company becomes a two-person science project.

It means the new advantage belongs to teams that are native to this way of working. Teams that understand where human judgment still matters, where software should operate autonomously, and where the handoff between the two creates compounding rather than chaos.

That is also the rough pattern we keep coming back to. The opportunities we find most interesting tend to live where software can actually act, not just advise. They sit inside a workflow with fast feedback loops. They generate information that improves the system over time. They own or move close to a concrete outcome. And they are being built by people who are genuinely fluent in this stack rather than merely adapting to it.

This is also why I think some of the loudest narratives around AI are still off. The shallow version is that AI makes building cheaper. True, but incomplete. The more important version is that AI changes what should be built in the first place. It changes what kind of founder is unusually powerful. It changes what a good early team looks like. It changes which markets are attackable by a small company. It changes whether the right outcome is a software vendor, a vertically integrated operator, or something in between.

In other words, it changes strategy, not just tooling. From an investment perspective, that forces a more opinionated posture.

We are less interested in companies that need the old software market to stay structurally the same. We are more interested in founders who already behave as if the shift has happened. They are not waiting for consensus. They are not adding AI to a static product map. They are building around the idea that software can act, that interfaces become thinner, that more value moves into workflow position and execution, and that small teams can create disproportionate outcomes if they are genuinely native to the stack.

That phrase — native to the stack — is doing a lot of work.

I do not mean people who can talk fluently about model benchmarks. I mean founders with real technical taste about what these systems can and cannot do, and enough judgment to know where that capability becomes a business rather than a demo.

I also mean founders with conviction.

This cycle is producing a lot of intelligent tourists. That is normal. Big platform shifts attract tourists. But the people worth backing are usually not the ones treating the moment like a broad exploration exercise. They have picked a direction. They have a specific view of what is changing. They are willing to build with the risk that they might be wrong in a specific way.

That is the right kind of risk.

Investors are paid to hedge. Founders are paid to be right about something narrow and important before everyone else is.

So if I had to compress our 2026 outlook to one sentence, it would be this:

The next great companies will come from founders who understand that software changed again, and who are building for a world where software is not just used by people but increasingly works on their behalf.

Everything else follows from that.

• It changes what categories matter.
• It changes what moats are real.
• It changes what early teams need to look like.
• It changes how we evaluate founders.

And it should change how investors talk about what is happening, because "AI is big" is not a point of view anymore. It is just ambient reality.

At Antigravity, this is the lens we are using to make sense of 2026. Not because we think every old software company disappears, and not because every agent demo becomes a business. Quite the opposite. We think the amount of software built, deployed, and embedded in the economy is likely to explode. That should widen the opportunity set even as it makes the last cycle's default company template less reliable.

Historically, software created enormous value when it opened up capabilities that used to belong to a small number of experts and made them available to everyone else. I think this cycle will do that again, at a much larger scale. That is why we are more excited by founders trying to build genuinely new capabilities than by strategies that mainly optimize existing ones.

That is what makes this moment interesting.

Not that software is over.

That much more of the economy is about to become buildable in software, and the founders who understand that earliest will build businesses the last cycle could not.