AI Startups: What's Overhyped and What's Real

The AI Funding Explosion

In 2025, artificial intelligence startups raised over $110 billion in venture capital globally. That figure represents roughly 38% of all venture dollars deployed worldwide, an unprecedented concentration of capital in a single technology category. To put this in perspective, the entire global venture market in 2019 was about $250 billion. AI alone now absorbs nearly half of what the entire venture ecosystem invested just six years ago.

The numbers are staggering at every level. Foundation model companies have collectively raised tens of billions. AI infrastructure companies are closing $100M+ rounds. Even application-layer AI startups are commanding seed valuations that would have been Series A prices two years ago. Our AI Valuations tracker shows that AI startups at Series A are raising at 2-3x the valuations of comparable non-AI SaaS companies with similar revenue. The premium is real, persistent, and, in many cases, disconnected from the underlying economics.

This explosion of capital has created a gold rush mentality that veteran investors will recognize from previous technology cycles. Not every company riding the AI wave will survive. In fact, most will not. The critical question for founders, investors, and operators is the same one that defined the internet era, the mobile era, and the cloud era: which innovations represent genuine, durable value creation, and which are hype masquerading as substance? After spending years investing in and studying this space, I have a strong point of view.

What Is Genuinely Transformative

Not everything in AI is hype. There are categories within the AI ecosystem that represent real, defensible, and transformative businesses. These are the areas where the technology is creating genuine value that did not exist before, not just automating workflows that were already possible with traditional software.

Vertical AI Applications

The most exciting category in AI right now is vertical AI: companies that apply AI to specific industries with deep domain expertise, proprietary data, and integrated workflows. Think AI for drug discovery, AI for legal contract analysis, AI for construction project management, AI for insurance underwriting. These companies win not because they have a better model, but because they have a better understanding of the problem domain, access to unique data, and distribution channels that horizontal AI tools cannot replicate.

Vertical AI companies tend to have several characteristics that make them genuinely defensible. They accumulate proprietary training data from their customers over time, creating a compounding data moat. They embed deeply into existing workflows, making them difficult to rip out. They often operate in regulated industries where compliance requirements create natural barriers to entry. And they can charge enterprise-grade prices because they are replacing expensive human labor, not competing with free consumer tools.

Our AI Landscape map tracks hundreds of vertical AI companies across industries. The ones that are winning share a pattern: they chose a specific wedge, became indispensable within it, and are now expanding from that beachhead.

AI Infrastructure

Every gold rush needs pickaxe sellers, and AI infrastructure companies are the pickaxes of this era. This category includes companies building compute orchestration platforms, data labeling and curation tools, model evaluation and monitoring systems, vector databases, inference optimization engines, and fine-tuning infrastructure. These companies benefit from a powerful tailwind: as more organizations build with AI, they all need infrastructure, regardless of their specific use case.

The best AI infrastructure companies have several advantages. They are model-agnostic, meaning they benefit regardless of which foundation model wins. They often sit in the critical path of production AI deployments, making them sticky and essential. And they tend to have usage-based business models that grow with their customers. The AI Spending dashboard shows that enterprise AI infrastructure budgets are growing at 80-100% year-over-year, and that growth is flowing directly to these companies.

AI Agents

AI agents, autonomous systems that can perform multi-step tasks with minimal human oversight, represent the next frontier. Unlike chatbots that respond to single prompts, agents can plan, execute, iterate, and complete complex workflows end-to-end. We are seeing early-stage agent companies in software development, customer support, sales outreach, financial analysis, and research. The most compelling agent companies are those replacing entire job functions rather than just augmenting individual tasks.

The agent category is still early and carries real technical risk, reliability, hallucination, and trust are genuine unsolved challenges, but the economic potential is enormous. If an AI agent can reliably perform the work of a $150K per year employee at a fraction of the cost, the addressable market is essentially the global payroll for knowledge work. That is not hype. That is a structural shift in how companies operate. The question is execution timing: the companies that solve reliability and trust first will capture disproportionate value.

What Is Overhyped

For every genuinely transformative AI company, there are a dozen riding the hype wave with fundamentally fragile business models. Identifying what is overhyped is not about being bearish on AI. It is about being honest about which specific applications of AI have durable economics and which are living on borrowed time.

Thin Wrappers on Foundation Models

The most obvious category of overhype is the thin wrapper: a startup that takes an existing foundation model (GPT-4, Claude, Gemini), adds a basic prompt template and a user interface, and calls it a product. In 2023 and early 2024, these companies raised significant capital because investors were eager for AI exposure and the demos were impressive. By 2026, the reality has set in.

Thin wrappers suffer from a fatal structural problem: they have no moat. When your entire product can be replicated by a competitor in a weekend, you are not building a business, you are building a feature. The foundation model providers themselves are the biggest threat. Every capability that a wrapper company adds, from web search to file analysis to code generation, gets absorbed into the foundation model's native feature set within months. OpenAI, Google, and Anthropic are not going to leave money on the table for thin-wrapper startups to capture.

The graveyard of thin wrappers is already growing. Dozens of AI writing assistants, AI meeting summarizers, and AI research tools that raised $5-20M in 2023-2024 have either shut down, been acqui-hired, or are slowly bleeding users to native features in the platforms they depend on. If your startup's differentiation can be described as "it is like ChatGPT but for X," you should be very concerned about your defensibility.

Undifferentiated Chatbots

Closely related to thin wrappers are the undifferentiated chatbot companies. These startups build conversational AI interfaces for customer support, sales, or internal knowledge management without meaningful differentiation from the dozens of other companies doing the exact same thing. The chatbot market has become commoditized to a degree that would shock anyone who saw the enthusiastic seed rounds of 2023.

The challenge is not that chatbots do not work. They do. The challenge is that building a chatbot is no longer a technical moat. Any competent engineering team can deploy a functional AI chatbot in days using existing APIs and open-source frameworks. The companies that will win in conversational AI are those with deep integrations into specific enterprise systems, proprietary training data from millions of customer interactions, or unique distribution channels that competitors cannot replicate. Generic "AI assistant for your business" is not a venture-scale business.

AI Companies with Terrible Unit Economics

Perhaps the most insidious form of AI hype is companies that have real revenue and real customers but terrible underlying economics. AI inference costs, while declining, are still substantial. Many AI startups are selling their products at prices that do not cover their compute costs, betting that inference prices will drop fast enough to make the economics work eventually. Some will be right. Many will not.

I have seen AI companies with $10M in ARR that are burning $15M per year on compute alone, before salaries, rent, or marketing. Their gross margins are negative. They are essentially subsidizing usage to grow, hoping to figure out the economics later. This is a dangerous game. If inference costs do not decline as fast as projected, or if competitive pressure prevents price increases, these companies will hit a wall. Real AI businesses have 60-80% gross margins. If your AI company has 20% gross margins and no clear path to 60%+, you have a cost problem masquerading as a growth story.

Defensibility in AI: What Actually Protects You

The question of defensibility is more critical in AI than in any previous technology cycle. When the underlying technology, foundation models, is available to everyone via API, what can a startup actually own? This is the question I ask every AI founder I meet, and the ones with strong answers tend to build the most durable companies.

Data Moats

The strongest form of AI defensibility is proprietary data. If your product generates unique data through customer usage, and that data makes your product better over time, you have a compounding advantage that competitors cannot shortcut. This is why vertical AI companies are so compelling: a legal AI company that has processed 10 million contracts has training data that a new entrant simply does not have. The more customers use the product, the better it gets, and the harder it is for competitors to catch up. Data moats are the AI equivalent of network effects.

Workflow Integration

The second strongest form of defensibility is deep integration into existing workflows and systems. If your AI product sits inside a customer's ERP, CRM, or core operating system, the switching cost is enormous regardless of whether a competitor has a marginally better model. Enterprise customers do not switch tools because a competitor is 10% more accurate. They switch when the cost of staying is higher than the cost of switching, and deep integration dramatically increases that switching cost.

Distribution and Brand

In a world where technology is commoditizing, distribution becomes the differentiator. Companies that build strong brands, large user bases, and efficient go-to-market engines can defend their position even as the underlying technology evolves. This is why some of the most successful AI companies are not the most technically sophisticated. They are the ones that figured out distribution first and layered AI capabilities on top of an existing user relationship. Consider what traditional SaaS companies are doing with AI and how incumbents with distribution advantages are leveraging them.

Where the Next Wave Is Heading

The AI landscape is evolving rapidly, and the opportunities of 2027 will look different from the opportunities of 2024. Based on the patterns I am seeing across the portfolio, in our AI Landscape tracker, and in conversations with hundreds of founders and investors, here is where I believe the next wave of durable AI companies will emerge.

AI-Native Services Companies

One of the most underappreciated trends is the emergence of AI-native services companies: businesses that combine AI technology with human expertise to deliver outcomes rather than software. These companies price on value delivered, not seats or usage. Think AI-powered accounting firms, AI-powered law practices, AI-powered recruiting agencies. They look more like professional services firms than SaaS companies, but their margins are dramatically better because AI handles 80% of the work and humans handle the remaining 20% that requires judgment and accountability.

AI for Physical Industries

The first wave of AI startups focused on knowledge work: writing, coding, analyzing, communicating. The next wave will focus on physical industries: manufacturing, agriculture, logistics, construction, energy. These industries have massive datasets, complex optimization problems, and enormous economic value at stake. They are also harder to disrupt because the AI needs to interact with the physical world, which creates natural barriers to entry. The companies that figure out how to deploy AI reliably in physical environments will build some of the most valuable businesses of the next decade.

AI Safety and Governance

As AI deployment scales, the need for safety, governance, and compliance tooling is growing exponentially. Every enterprise deploying AI needs tools to detect hallucinations, ensure regulatory compliance, monitor for bias, manage access controls, and audit AI decision-making. This category barely existed two years ago and is now a billion-dollar market. Regulatory requirements in the EU, and increasingly in the US, are making AI governance not optional but mandatory. The companies building the picks and shovels for responsible AI deployment have a massive tailwind.

Multi-Agent Systems

Beyond individual agents, the next frontier is systems of agents that collaborate, coordinate, and check each other's work. Imagine a team of AI agents that can handle an entire business process: one agent does research, another drafts a proposal, a third reviews it for accuracy, and a fourth negotiates terms. Multi-agent systems are still early and face significant coordination challenges, but the potential is transformative. The companies that solve agent coordination at scale will define the next era of enterprise software.

The Bottom Line

AI is not overhyped as a technology. It is the most significant platform shift since the internet, and it will reshape every industry over the next decade. What is overhyped are specific applications of AI that lack defensibility, have unsustainable economics, or are building features rather than products. The distinction matters enormously for founders, investors, and anyone allocating capital or career time to this space.

As I wrote in our State of VC Funding in 2026 overview, roughly 40% of all venture capital is now flowing into AI. That concentration creates both opportunity and danger. The opportunity is that transformative AI companies will be generationally valuable. The danger is that the abundance of capital is funding companies that should not exist, inflating valuations that will eventually compress, and creating a class of AI startups that look successful on paper but are structurally fragile.

If you are building an AI startup, be honest with yourself about which category you fall into. Do you have a genuine data moat, deep workflow integration, and sustainable unit economics? Or are you a thin wrapper hoping the wave carries you to an exit before the tide goes out? The market will eventually answer that question for you. Better to answer it yourself first.