The business of AI in 2026

Some of the most critical trends driving AI this year

In our end-of-year post, we noted that 2025 marked a shift from leaps at the foundation-model layer to steadier, incremental gains. What surprised us was that many of the things we predicted correctly stopped being interesting. Benchmarks are the clearest example. In 2024 they dominated the discourse, but last year, we almost completely stopped paying attention to ARC AGI and LMArena Elo.

With how quickly the core technology and its applications evolving, broad themes matter much more than point predictions. Humans are notoriously bad at exponentials, so our belief is that what matters is whether you’re headed in the right direction or not – the probability distribution around how far down the road you get is distinctly less important.

There are four key trends we’re watching coming into 2026: (1) data centers & GPU buildouts; (2) acquisitions and acqui-hires; (3) where frontier labs push next; and (4) a shift toward harder application-layer problems.

Data centers

The incredible investment in building new data centers dominated the latter half of 2025. You’ve all seen the predictions, ranging from this being the second coming of a railroad or broadband bubble to this investment foreshadowing American dominance in the age of AI.

Regardless of the macro view, our belief is that the investment will ultimately be justified by demand. We’re confident that we’re only just scratching the surface with how AI is applied in most domains – across both consumer and enterprise applications. The limitation still is the ability to invest in thoughtfully building more things – and perhaps the number of high-quality product thinkers. We regularly find more AI applications that we can integrate into our daily workflows, and there’s plenty more that we’d love to see.

From a finance perspective, the question is how quickly demand ramps up. As we saw with the broadband buildout around the dot-com bubble, it took over a decade for that infrastructure to become truly valuable. Our hunch is that this will be different – as these data centers come online (much of which won’t happen til well past this year) and models continue to improve incrementally, we should generally see inference costs drop which makes more applications economical, which will in turn drive more usage.

While the specific financial predictions are hard to comment on, we’re generally bullish on data center & GPU investment.

Acquisitions vs. acqui-hires

The 2025 trend of acqui-hires that decapitated a company continued right to the end with Nvidia’s “acquisition” of Groq. Acquisitions are a natural, healthy part of the startup ecosystem. As we’ve mentioned before, the acqui-hires are more concerning and very likely are a negative for the startup ecosystem. The economics of these acqui-hires are opaque, and they create anxiety amongst startup employees who are wondering whether they will be left working at a company that’s suddenly lost its leadership. Continued acqui-hires might very well start to spook employees considering joining hot startups. Unless the federal government signals a dramatically different policy in anti-trust enforcement, this trend is likely not going anywhere.

Independent of the mechanics, big companies will continue to eat startups. In fact, we believe the general acquisitiveness of larger incumbents is going to increase. We saw a surge of acquisitions at the end of last year (e.g., Qualified → Salesforce, Veza → ServiceNow). We predicted this trend a year ago – unlike with past tech trends, the large incumbents are very plugged into AI and are very anxious about how they’re going to get their foothold. They’re not going to sit back.

As we see more and more AI applications continue to grow at unprecedented rates, the number of acquisitions will probably also correspondingly increase this year. While these aren’t specific predictions, it’s hard to imagine that companies that are growing quickly but not valued in the tens of billions yet (e.g., Legora, Granola, or Serval) aren’t extremely attractive acquisition targets for the incumbents in their spaces.

Acquisitions and acqui-hires are both here to stay.

Frontier labs

We started to notice at the end of 2025 that each of the three main frontier model labs seemed to be carving out a lane in the application layer – OpenAI in consumer apps, Anthropic in coding, and Google in enterprise productivity. While these investments will continue, it’s hard to imagine the lanes staying fixed. It’s likely to be more of a 3-man weave for the next couple of years.

That said, it’s very clear that the frontier models are going to continue to encroach (or acquire) into the biggest application markets. It’s become clear that the inference business is low-margin and high-volume, especially if model improvements continue to be incremental, so it’s natural for the frontier labs to use their gobs of cash to move into the higher-margin application layer. This was likely inevitable.

The open question is which applications can differentiate themselves sufficiently and which ones will be Sherlocked by OpenAI. As model capabilities improve, the most significant value add is going to come in orchestration, composition, and workflow management – using the right models & tools at the right time. In the B2B world, there are plenty of examples of complex applications that will be safe. For consumer apps, things like Wisprflow that build separate workflows outside of the frontier labs’ walled gardens are likely safe as well.

On the other hand, more generic applications like both consumer and enterprise search seem less defensible to us. Starting with o3, ChatGPT’s web search basically made it our go-to choice, and with every LLM provider integrating in enterprise productivity tools, enterprise search tools will need to differentiate on workflows rather than data access.

Frontier labs are going to continue to push into the application layer, and the real differentiation for startups will lie in workflow complexity.

Easy problems and hard problems

We firmly believe that if all LLM innovation were paused today, there’s 5-10 years of application-layer innovation left to be done with the existing models. What we’re starting to see is that the distribution of where that effort will be spent might not be so even. All of the “obvious” application areas already have large, established startups that have popped up overnight – legal (Harvey, EvenUp), support (Sierra, Decagon), and marketing (Writer, Jasper) to name a few. That isn’t to say that other startups won’t pop up in these markets, but these larger players now seem more like pseudo-incumbents.

That means interest and adoption is going to start moving to harder problems – the ones that can be solved with AI but probably aren’t quite as immediately obvious as the areas listed above. Examples of this include healthcare, security, and SRE (where we’re focused at RunLLM). Adoption here likely isn’t going to move at the same breakneck pace as we’ve seen in the applications listed above. Harder problems means buyers that are more skeptical and evaluation cycles that are more complex. None of this is bad – it’s just a different adoption model than what we’ve seen in the last few years.

The elephant in the room is, of course, AI coding agents. Coding is not an easy problem, but it’s rapidly being solved for a few reasons. First, it’s easy to adopt: Everyone has their codebase downloaded locally, and most engineers pick their own IDEs. Ease of adoption has created a data flywheel that’s enabled tools like Cursor and Claude to improve rapidly. The added attention has come from the fact that every researcher at a frontier lab is a software engineer and wants to solve this problem. Despite those tailwinds, what’s telling is that it is being solved by an incredible scale of investment, both within the frontier model labs and by startups like Cursor, who are rapidly innovating to keep ahead of the labs. Other hard problems likely won’t receive this same level of attention, but these areas will require similar levels of investment in technical solutions over time.

AI application attention will move up the complexity curve, from the most obvious problems to harder ones.

---

A year ago, we thought 2025 would be the year “rubber meets the road.” That was premature: the hype didn’t end; it shifted from model releases to the application layer — and that shift was a necessary step toward durable value.

In 2026, we expect things to trend in the same direction. As capacity comes online and inference gets cheaper, more applications become economically viable – especially in complex domains. Frontier labs will keep moving up-stack where it’s strategic, but their focus is inherently limited — leaving meaningful space for startups to build, differentiate, and increasingly get acquired.

The winners won’t be defined by who has the best model weights; they’ll be defined by who turns models into reliable systems inside real workflows. If we’re right, the most important progress by the end of 2026 will look less like benchmark jumps and more like adoption in harder domains.