In 2023, there were a few hundred AI startups. By early 2025, there are thousands. Most of them sell thin wrappers around the same foundation models, differentiated by industry jargon and slide decks rather than genuine capability. The consolidation is coming. The question is whether your AI investments are on the right side of it.
What You Need to Know
- The AI vendor market is unsustainably fragmented. Thousands of startups, many with overlapping capability, funded by venture capital that expects returns. Not all of them survive.
- Wrapper vendors are most at risk. If a vendor's primary value is a UI on top of an OpenAI or Anthropic API, they're vulnerable to the underlying provider adding that UI themselves.
- Platform consolidation will accelerate. Enterprise buyers want fewer vendors, not more. The vendors who build genuine platform capability (not just model access) will absorb the rest.
- Your architecture determines your exposure. If you've built on a vendor's proprietary platform, their failure is your problem. If you've built on open standards with abstraction layers, it's a vendor switch, not a rebuild.
3,400+
AI startups funded in 2024, up from ~1,200 in 2022
Source: CB Insights, State of AI Report, Q4 2024
The Three Layers of AI Vendors
Not all AI vendors carry the same consolidation risk. Understanding where a vendor sits in the stack helps assess their durability.
Layer 1: Foundation Model Providers
OpenAI, Anthropic, Google, Meta, Mistral. These companies build the underlying models. They have the deepest technical moats: research teams, compute infrastructure, training data. Consolidation risk here is low. Some will grow faster than others, but the major players are durable.
Layer 2: Platform and Infrastructure
Companies that provide the orchestration, deployment, monitoring, and governance layers for enterprise AI. This includes AI platforms, MLOps tools, vector databases, and integration middleware. This layer has moderate consolidation risk. The best platforms add genuine value. But there are too many players doing similar things.
Layer 3: Application and Wrapper
Companies that build end-user applications on top of foundation models. This ranges from genuine vertical applications with deep domain expertise to thin wrappers that are essentially a chat interface with a custom system prompt. This layer has the highest consolidation risk.
AI Startup Growth (funded startups)
Source: CB Insights, State of AI Report, Q4 2024
The risk increases when the vendor's differentiation is the model itself rather than the domain expertise, data pipeline, or workflow integration. If Claude or GPT gets better (and it will), the wrapper's value proposition evaporates.
How to Assess Vendor Durability
We use five criteria when evaluating AI vendors for enterprise clients:
1. What's their moat? Is their differentiation the model (no moat, model providers will match it), the data (moderate moat, if the data is proprietary), or the domain expertise and workflow integration (strongest moat)?
2. What happens if the underlying model changes? Can the vendor switch models, or are they locked to a single provider? Vendors locked to one model share that model's risks.
3. What's the revenue model? Per-seat, per-transaction, or consumption-based? Consumption-based models aligned with model API costs are vulnerable to price compression. Vendors need margin between what they pay for inference and what they charge.
4. Can you export your data? If the vendor fails, can you extract your configuration, your training data, your custom models, and your operational data? If not, you're building on rented land.
5. What's the integration story? Does the vendor integrate with your existing systems, or does it require you to adapt to theirs? Vendors that play well with existing enterprise architecture are stickier and more valuable.
The Wrapper Test
Ask your AI vendor: "If I had direct API access to the underlying model, what value does your platform add?" If the answer is essentially "we've made it easier to use the API," that's a wrapper. Wrappers are fine for getting started, but they're not durable enterprise investments.
Positioning for the Shakeout
Invest in architecture, not just vendors. Build abstraction layers that let you switch vendors without rebuilding. This isn't hypothetical risk management. We've already helped clients migrate between vendors when pricing changed, features were deprecated, or quality degraded.
Prefer platforms over point solutions. A platform vendor that handles orchestration, monitoring, and governance across multiple AI capabilities is stickier than a point solution that does one thing. The platform vendor has more surface area for value creation.
Own your data and your prompts. Your domain-specific data, your engineered prompts, and your evaluation datasets are your intellectual property. Ensure your vendor agreements allow you to retain and export all of them.
Build internal capability alongside vendor relationships. If your entire AI capability depends on a single vendor's platform, you have a single point of failure. Internal understanding of your AI systems (even if a vendor builds them) protects you.
The NZ/AU Context
New Zealand's AI vendor market is smaller but faces the same dynamics. Local consultancies and system integrators are adding "AI capability" by wrapping foundation models. Some are building genuine expertise. Others are reselling API access with a PowerPoint.
For NZ enterprises, the consolidation wave means being deliberate about vendor selection now. The vendor you choose in 2025 needs to be durable enough to support you through 2027 and beyond. That means assessing technical depth, financial sustainability, and alignment with your long-term AI strategy.
The AI vendor market in 2025 reminds me of the SaaS market in 2012. The lesson from SaaS consolidation is clear: invest in platforms that own a genuine layer of value, not wrappers that can be replicated by the platform below them.
Isaac Rolfe
Managing Director