Is AI a feature or a company?

In 2026 most investors treat AI as a feature, not a company. A general AI capability on its own is not defensible, because foundation models commoditize fast. The fundable companies use AI as one component inside something hard to copy: proprietary data, a deep workflow, a regulated process, or real-world hardware.

Do VCs still invest in AI wrappers in 2026?

Rarely, and reluctantly. A thin wrapper over someone else’s model has weak margins and no moat, because the model vendor can ship your feature overnight. VCs now ask, ‘Why can’t OpenAI, Anthropic or Google just build this?’ If you cannot answer that convincingly, you are hard to fund.

What is the difference between an AI wrapper and a deep-tech startup?

An AI wrapper is mostly a user interface over a foundation-model API, with little proprietary data or integration. A deep-tech startup solves a hard scientific or engineering problem and may use AI as one input among several, such as data, hardware, or domain depth. The wrapper is easy to copy. The deep-tech company is not.

What makes an AI startup defensible?

Not the model, but the things around it. The durable moats are proprietary data that compounds, deep integration into a customer’s workflow or system of record, regulatory and compliance depth, a real-world or hardware component, and distribution. Stack a few of these and a competitor cannot rebuild you in 18 months.

Is an AI startup considered deep tech?

Only sometimes. Foundational AI research counts as deep tech, but an app built on top of GPT or Claude usually does not. Investors like BDC Capital scope deep tech to fundamental breakthroughs such as foundational AI, quantum, robotics, and advanced materials, where there is real technical risk rather than just an API call.

Why can’t OpenAI or Anthropic just build my AI product?

That is the exact question investors ask now, so you need a real answer. If your product is a thin layer over a model’s API, the vendor can ship it as a feature overnight. You are defensible only if you own something they will not build: proprietary data that compounds, deep integration into a customer’s workflow, regulatory or compliance depth, or a real-world or hardware component. If you cannot name that moat in one sentence, you are hard to fund.

AI Is a Feature, Not a Company: What VCs Actually Fund in 2026

The short answer: In 2026, most VCs treat AI as a feature, not a company. A general AI capability is not a moat, because foundation models commoditize too fast and the model vendor can ship your feature overnight. The startups getting funded use AI as one component inside something hard to copy: proprietary data, a deep workflow, a regulated process, or real-world hardware. The question every investor now asks is blunt: “Why can’t OpenAI just build this?”

What I heard at Edmonton Unlimited’s Deep Tech Showcase

Earlier this month I sat in on Edmonton Unlimited’s first Deep Tech Showcase, a room full of founders and investors, with an investor panel on what actually makes a deep-tech company fundable. The lineup included Patrick Lor of Panache Ventures, Aditya Aggarwal of BDC Capital, and Neha Khera of Innovobot, moderated by Arden Tse of Yaletown Partners. “Deep tech” was framed broadly: AI, health, energy, space, advanced manufacturing, and quantum.

I went in expecting the usual “AI is eating everything” enthusiasm. That is not quite what I heard.

The thread that ran through the panel, and the thing that stuck with me, was a gentle correction to the hype. The investors were not chasing “AI companies.” They were interested in deep tech where AI is one ingredient, not the whole dish. Panache, for instance, has backed quantum-computing and AI-infrastructure companies, businesses where the hard part is not a prompt, it is the science underneath. The message, paraphrasing the room: show me something genuinely hard to build, and tell me where AI fits into it. Do not show me AI and hope it is a business.

That reframing is worth unpacking, because it is now the dominant view across venture capital, and it changes how you should pitch.

“AI is a feature, not a company”

The phrase has been around since people said the same thing about Dropbox being “a feature, not a product.” In 2026 it has become the default lens for AI startups, and the data explains why.

AI is absolutely where the money is going. By most counts, more than half of all 2025 venture dollars went to AI, and in Q1 2026 it was closer to 80% of global venture funding. But look at where inside AI the money lands. The capital is wildly concentrated at the two ends of the stack. OpenAI and Anthropic alone accounted for roughly 14% of all venture investment in 2025, and a handful of foundation-model mega-rounds made up the majority of AI dollars. (Crunchbase)

The middle, the thin application layer that just calls those models, is exactly where investors have pulled back. Khosla Ventures partner Samir Kaul, an early OpenAI backer, put it plainly to CNBC: “The question I ask every time one of them presents is, why can’t OpenAI, Anthropic or Google do this? For most of them, the answer is, ‘They can.’” (CNBC, June 2026)

That is the whole thesis in one sentence. If the foundation-model company can absorb your product into its next release, you do not have a company. You have a feature waiting to be “Sherlocked.”

AI wrapper vs. deep tech: what VCs see

Thin AI wrapper	Deep tech with AI as a component
What it is	A UI over a foundation-model API	A hard science or engineering product that uses AI as one input
Moat	Almost none, easy to copy	Proprietary data, workflow lock-in, regulation, or hardware
Gross margins	About 40%, since you pay token costs per query	Healthy, because the value is not the model
Biggest risk	The model vendor ships your feature	Execution and technical risk, but defensible if you win
The VC test	“OpenAI can build this”	“OpenAI cannot inherit our data, integration, or hardware”

As Allen Lau of Toronto’s Two Small Fish Ventures wrote earlier this year: “Shallow tech… including a lot of AI wrappers, are generally not investable from a VC perspective because they are so easy to build, copy, and replace. The moat is not in the software. The moat is the unique technology embedded in the software, plus the data and the system it connects to.” His co-founder Eva Lau adds the other half: “Brilliant science without ruthless productization is just an expensive science project.”

It is the same point Daniel Wigdor of the AI venture studio AXL makes when he warns that people “confuse enabling technologies for applications,” and that betting on the AI infrastructure itself is “a race to zero.” And it is what Golden Ventures’ Nick Chen meant: “AI can enhance an application’s value proposition and product-market fit, but it doesn’t broadly create it.”

So is AI dead as a pitch? No, the opposite

Here is the nuance that trips founders up. The lesson is not “don’t build with AI.” Building with AI is now table stakes. As Inovia Capital’s Magaly Charbonneau puts it, investors treat “deep AI integration as the price of admission: not a differentiator, but an operational necessity.”

So you have to use AI. You just cannot be only AI. The differentiation has to live in the parts that are not the model. That is why capital is visibly shifting “from bits to atoms”: robotics startups alone raised a record 40.7 billion dollars in 2025, and money is pouring into chips, energy, defense, and real-world data, places where AI is embedded in something physical and hard to replicate.

This is also how BDC Capital scopes its Deep Tech Venture Fund: foundational AI, quantum, robotics, advanced materials, all with real technical risk. As BDC’s Thomas Park said, “We look for innovations in the algorithm and not in the data set.” In other words, fund the breakthrough, not the API call.

What this means for founders: building a real moat

If you are raising in 2026, position your company as deep tech where AI is one component of a defensible system, and lead with the defensibility, not the AI. One proven shape of this is the vertical AI agent, software that owns a narrow industry workflow so completely that the model inside it is almost beside the point. Investors are looking for some stack of these:

Proprietary data that compounds. Data your product generates that no general model has, and that gets better the more you are used.
Deep workflow or system-of-record integration. When OpenAI ships a smarter model, it does not inherit your Epic integration, your court-filing workflow, or your risk dataset. That asymmetry is what makes you fundable.
Regulatory and compliance depth. Auditable, certified handling such as HIPAA, SOC 2, or GDPR that a wrapper cannot just ship.
A real-world or hardware component. Atoms are hard to copy. Bits are not anymore.
Distribution and switching costs. Owning the customer relationship and making yourself painful to rip out.

A useful gut-check I took from all this is the “5 million dollars and 18 months” test. If a well-funded competitor could rebuild your product with 5 million dollars and 18 months, then your moat is the AI, which means you do not have one. If they could not, because of your data, your integrations, or the physics of what you have built, that is the company.

A concrete example: AI inside Taxformify

Let me make this concrete with my own product, because it is exactly how I think about AI as a feature.

I built Taxformify, a tool that helps tax accountants and their clients get through tax season faster. The most genuinely useful, time-saving AI feature in it is automated expense categorization for both US and Canada based individuals and corporations. Feed it the transactions, and it sorts them into the correct tax categories in seconds, instead of the hours an accountant normally spends doing it by hand. For a firm handling a hundred clients, that is a huge amount of time saved, and time is the entire constraint in a tax practice.

Here is the important part. The AI categorization is the feature, not the business. Plenty of people can ask an LLM to categorize a handful of transactions. What makes Taxformify a product, rather than a wrapper, is everything around that feature: the accountant and client workflow, secure document handling, the fact that it understands both US and Canadian tax categories, and the way it fits into how a firm actually preps its returns. The model is swappable. The workflow, the dual-country tax logic, and the accountant relationship are not.

That is the thesis in miniature. The AI is what saves the time, but the defensibility lives in the boring, hard, specific parts around it. If I pitched “an AI that categorizes expenses,” any investor would rightly ask why OpenAI cannot do that. If I pitch “the tax-prep workflow that US and Canadian accountants run their season on, with AI doing the categorization grunt work,” that is a very different conversation. (I wrote about actually building Taxformify with AI tools in Vibe Coding a SaaS App.)

How to find the right investors (using the directory)

This is exactly the kind of question I built my Startup Funding Directory to answer. If you are building deep tech with an AI component, you can filter the directory to the investors who actually fund it:

Filter by sector AI/ML or Hardware/Deeptech to surface the relevant funds.
The Canadian deep-tech and applied-AI names from this article each have a profile: Panache Ventures, BDC Capital, Radical Ventures, Two Small Fish Ventures, and AXL.
Match your stage, whether pre-seed, seed, or growth, so you are not pitching a Series B fund at idea stage.

The directory has 450+ investors and programs across 10 regions worldwide, from North America and Europe to Africa, the Gulf and Asia, including VCs, accelerators, incubators, angel networks, grants, and halal/Sharia-aligned funds, and it is free to search.

The takeaway

The AI gold rush is real, but the easy money for thin wrappers is over. What I heard in that Edmonton room is now the industry consensus: AI is a feature, not a company. The founders who get funded in 2026 are not the ones with the slickest GPT demo. They are the ones who can explain why their data, their workflow, or their hardware makes them impossible to copy, with AI quietly doing its part inside. My own app is a small example of the same idea: the expense-categorization AI saves accountants real time, but the business is the tax workflow it lives in.

Build the moat first. Let AI be the feature.

---

Sources & further reading

---