There is a huge gap between what AI can do and what AI does do inside most companies today, because raw intelligence needs a rich context layer to point it in the right direction. This is the opportunity for founders building enterprise applications on top of foundational models.

• The commoditization of the intelligence is a massive tailwind for founders building on top of models.

• SaaS incumbents look strong for now, but their moats are drying up.

• Founders with a deep understanding of their customers can build incredibly valuable and novel applications for AI.

• As these applications scale, customer data and feedback become an asset that allows them to direct raw intelligence from the models in increasingly useful ways for their customers.

(1) The Utility of Intelligence

Frontier labs are in a historic competition to push the boundaries of AI and are evolving into something that looks like digital infrastructure.

Anthropic was first to market this year with its frontier model, Fable. One month later, OpenAI released Sol at half the cost per token. Another two weeks later, Chinese lab Moonshot released Kimi K3 at roughly half the cost of Sol. Then in August, another Chinese lab Z.ai released Ox Alpha at a fraction of the cost per token of Kimi K3.

This competition is an incredible tailwind for everyone using the models and building on top of them:

1. model capabilities are going through the roof while costs are dropping through the floor, and

2. frontier labs have to focus on their core business of pushing the frontier of AI.

Concerns that Anthropic and company might devour the entire digital economy now look overblown, as frontier labs can’t build deeply contextualized applications for every use case while also fighting an existential battle to release the best models. Their focus is on making the highest quality intelligence at the lowest cost. To their credit, AI leaders have been saying this all year:

“We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter and use it for whatever they want to use it for.”

-Sam Altman at Blackrock Infrastructure Summit, March 2026

“AI should be treated as essential infrastructure like electricity or roads.”

-Jensen Huang at Davos, January 2026

So if frontier labs are building intelligence as a deliverable commodity, who gets to deliver the package?

(2) SaaSpocalypse Now Postponed

The first beneficiaries of labs’ evolution towards infrastructure have been public SaaS companies and their shareholders. In the short term, it’s an opportunity, but in the long-term, it’s a threat as AI begins to erode their moats.

In our last update, we wrote that the SaaSpocalypse is not the death of SaaS but the beginning of a new, more competitive era, in which models are unlikely to take the cake but legacy SaaS incumbents will face more competition.

Three months later, it’s official: the SaaSpocalypse is postponed. Salesforce announced Claudforce with Anthropic last week and their stock is up more than 50% from its low. CrowdStrike, Palo Alto Networks, Snowflake, and other cybersecurity and infrastructure players are in high demand as the proliferation of agents online multiplies cybersecurity threats and data storage needs.

But this is not a happily ever after.

Take Salesforce as a bellwether for the alphabet soup of public and private SaaS systems of record (CRMs, ERPs, DMSs, PMSs, EHRs, PLMs, SCMs, etc). Claudeforce will let users interact with Salesforce through Claude, turning Salesforce into a headless CRM. Now Salesforce is positioned as a critical layer in the emerging AI stack, but it is also admitting that one of the incumbents’ historical moats is evaporating.

For a long time, Salesforce has been the default CRM, in part because every sales rep in America knows how to use it. In ceding the user interface to Claude, Salesforce loses that moat. If sales reps no longer need to use Salesforce’s UI, Salesforce will have to compete on capability alone against every other headless CRM.

For all incumbent systems of record, this is a threat and an opportunity. The systems of record that build systems of action around the data and context they control will offer a much more powerful experience to their users. Salesforce, to their credit, sees the writing on the wall and is running towards this future. Others will not be so quick to adapt, and even for Salesforce it will be a challenge to keep up with cutting-edge AI while defending their profitable SaaS business.

Incumbents typically aren’t the best innovators, for the same reason that Google invented the architecture for AI but their frontier models are way behind OpenAI and Anthropic. It’s very hard to be the dominant player in any market while also building the technology that will eat that market.

This is the opportunity for AI-native startups.

(3) Who Takes the Cake?

Jensen described the AI economy as a five-layer cake. The base is energy, followed by chips and data centers, then models. The frosting on top is the application layer. Each layer of the cake has seen very roughly an order of magnitude less investment to date.

Jensen Huang’s “AI is a Five-Layer Cake” (Our edits in red)

As we move beyond chatbots and coding assistants, the work of bringing AI to end users in *useful* ways is increasingly falling to the application layer. So why has the application layer seen relatively so little investment?

First off, each layer of the cake is built on the layer beneath. We’ve also had fear, uncertainty, and doubt that the models might eat the entire software stack (see: SaaSpocalypse). Maybe we also have a negative association with applications from the last major tech paradigm shift to mobile, where every mobile app had to through Apple to reach their customers and Apple captured the lion’s share of the value.

This time around, it’s different. The applications will own the customer relationship. Models perform invisible magic inside of applications and are often interchangeable. In that sense, AI-native applications are positioned less like mobile apps and more like the tech giants of today, who built their businesses around the massive infrastructure investment of the ‘90s that made internet access ubiquitous.

The time to invest in the application layer is now, and for at least the next five years. The models are incredible, SaaS incumbents’ moats are eroding, and we’ve barely begun to realize AI’s potential.

We have presumptive winners at the application layer in software engineering, but coders make up about 1% of the U.S. workforce. Who is going to build effective AI for the other 99%?

Most people’s exposure to AI today is characterized by AI slop on every homework assignment and content feed (even the WSJ). McKinsey reports that 88% of enterprises are using AI, but less than 40% are seeing an impact on their bottom line.

To realize the vision of intelligence as a utility, more people need to find utility in intelligence.

We need systems of intelligence that understand how people work, their incentives, the weird ways they’ve actually operated for the past 30 years, and critically, what they want to automate with AI vs leave for people to decide.

(4) Customers as Context

AI needs context to do well, and that context lives in the users. Every new user brings with them context from their industry that lives in people’s brains, systems of record, regulatory environments, data warehouses, and company cultures.

For AI-native startups, understanding the user’s unique context starts a flywheel where more customers lead to more context to tune their agents and models, which creates a better product and in turn, more customers.

AI-native startups have already proven they can scale up quickly. To data from our own portfolio, our pre-seed investments from the past three years added $40 million revenue collectively in 2025, and they’ve already added more than that already in 2026. The question we often hear is how do we know this revenue is defensible? Or in other words, are they just repackaging tokens or are they building something novel and valuable?

For AI applications, customer data becomes an asset that the models do not have. Not just the raw data, but the way that customers use the product, what they ask for, what delights them or what seemed like a good idea but users just don’t care.

The library of data these companies build over time allows them to point models and agents at the right problems. The models themselves are like a brilliant, bright-eyed graduate who has never worked a day in their life. With no context they are clueless, but with some context and training, they can be incredible. That context is the asset that application-layer startups are building.

Take Cursor for example, which sold to SpaceX this month for $60 billion. Cursor started as a great user interface for agentic coding. They drove adoption by building a product that is deeply clued in to how coders work. As they scaled, they collected a tremendous amount of data on how their users are prompting AI, and ultimately launched their own coding model. Cursor’s context flywheel generated the biggest private venture exit of all time.

(5) How Analog Invests in the Application Layer

To find the teams that will define new categories in application-layer AI, we look for both technical capability and deep customer empathy.

In the last two years, we’ve backed

• a leading expert in grinding mills to build AI for mines

• an AI hedging business co-founded by the youngest person to
  run a hedging book at Barclays

• an AI-native govtech startup founded by a Navy pilot who
  spent two years on Capitol Hill

• an AI asset monitor for heavy industry founded by a woman
  who spent her 20s sleeping in trucks in oil fields.

What makes these founders uniquely suited to this moment is that they have a deep understanding of their users. They close their first customers by speaking their language and building an AI harness that fits their systems and workflows. Every new customer they add accelerates the flywheel of customer context.

i3 Specialty’s installation inside of a high-security clearance gold mine in Canada

In a year when the public narrative around AI is increasingly skeptical and hostile, this feels like important work. AI is working very well for Silicon Valley, and not so well yet for everyone else.

To deliver on the promise of AI for the rest of America, we need founders who have walked a mile in their shoes, and making AI a useful tool in the hands of everyone is enough work for an army of founders over the next decade.