The market has grown cynical. Every pitch deck now features ‘AI-powered’ somewhere in the first three slides. Every SaaS company has bolted on a chatbot. Every legacy software vendor has rebranded their analytics dashboard as ‘machine learning-driven insights.’
Acquirers see through it.
The valuation gap between genuine AI infrastructure and cosmetic AI features has widened dramatically. Companies with real machine learning moats are commanding 8–12x revenue multiples. Those with ChatGPT wrappers are lucky to get 4x. The difference isn’t subtle. It’s existential.
The Commoditization Trap
Most businesses claiming AI transformation have simply purchased commoditized tools. They’ve integrated OpenAI’s API. They’ve added a recommendation engine from a third-party vendor. They’ve implemented off-the-shelf sentiment analysis.
But none of this creates defensible value.
Buyers perform technical due diligence. They bring in ML teams. They ask to see the training pipelines, the data moats, and the proprietary architectures. When they discover the ‘AI’ consists of API calls to models anyone can access, the valuation conversation shifts immediately. The business becomes a feature — not a platform. And features get acqui-hired while platforms get premiums.
The math is brutal. If your AI capabilities can be replicated by a competitor in three-to-six months using publicly available tools, you have no AI moat. You have a temporary feature advantage that’s already eroding.
What Actually Commands Premium Multiples
Buyers pay premiums for three types of AI infrastructure. Everything else is noise.
Proprietary data flywheels
These are businesses that generate unique data as a byproduct of operations. This data continuously improves model performance. The improvement drives user adoption and adoption generates more data. The flywheel compounds. Competitors without the initial data can’t enter the cycle. This is defensible.
Medical imaging companies that have processed millions of annotated scans command premiums because the data can’t be purchased or replicated quickly. Fraud detection systems that have seen billions of transaction patterns in specific industries possess knowledge that new entrants can’t match. While the models themselves may be standard architectures, the training data is irreplaceable.
Domain-specific architectures that create structural advantages
Generic large language models are available to everyone. Specialized models tuned for specific tasks within regulated industries are not. A legal AI that understands jurisdiction-specific precedent hierarchies and can navigate conflicting circuit court decisions isn’t just ‘fine-tuned’ — it embodies years of domain expertise translated into model architecture.
These systems can’t be replicated by hiring ML engineers and buying compute. They require deep domain knowledge that takes years to encode. In such cases, the technical debt is actually an asset and the complexity is the moat.
Inference infrastructure that operates at dramatically different cost structures
If your AI delivers similar outputs to competitors but at 10% of the compute cost, you’ve built something valuable. This usually requires custom chips, novel compression techniques, or architectural innovations that emerge from years of production optimization.
Edge deployment that keeps data local. Inference that runs on hardware three generations old. Models distilled to a hundredth the size without performance degradation. These create lasting competitive advantages because they require sustained engineering investment and operational knowledge that only accumulates over time.
Due Diligence Questions That Reveal Reality
Buyers now send technical teams before term sheets. They ask questions that expose the difference between real and cosmetic AI.
‘Show us the training pipeline.’ If your answer is ‘we use Claude/GPT-4,’ the conversation is over.
‘What happens if OpenAI raises prices 10x?’ If the answer is ‘our margins collapse,’ you don’t have AI infrastructure — you have vendor dependency.
‘Walk us through how you curated your training data.’ If you purchased a public dataset and ran standard augmentation, you have no data moat.
‘How do you handle model drift in production?’ If monitoring and retraining aren’t automated and instrumented, your AI is a prototype dressed up as infrastructure.
‘What’s your marginal cost per inference as you scale to 10x current volume?’ If costs scale linearly or worse, you haven’t solved the hard engineering problems that create value.
These questions reveal where genuine technical work happened. They separate companies that built AI from companies that integrated it.
The Integration Problem
Many businesses mistake integration complexity for AI value. They’ve stitched together multiple ML services. They’ve built elaborate pipelines connecting data warehouses to inference endpoints. They’ve created dashboards showing model performance metrics.
But buyers don’t want Rube Goldberg machines. They want clean, maintainable systems or genuine technical innovations. Complex integrations that merely connect commodity services represent future technical debt, not assets. The acquirer will rip out the plumbing and replace it with their own infrastructure. Your integration work becomes worthless.
There is one exception: if the integration itself creates proprietary value. Some businesses have built sophisticated systems that orchestrate multiple models in ways that are genuinely difficult to replicate. The orchestration logic encodes domain expertise. The failure handling represents years of production learning. This is defensible, but most integration isn’t. It’s just API calls with error handling.
Positioning for Maximum Valuation Impact
If you’re building or operating a business with ML components, the path to premium valuations requires honest assessment.
Map your AI capabilities into three categories: commodity (available to anyone), competitive (temporary advantages), and structural (genuine moats).
Be ruthless. Most capabilities fall into the first bucket.
For commodity capabilities, stop marketing them as differentiators. Buyers will discover the truth in due diligence. Leading with commodity features erodes trust and frames the entire business as imitative rather than innovative.
For competitive capabilities, document the lead time. If your ML model took six months to train and a competitor could replicate it in three months with the right team, you have a temporary advantage. Price your expectations accordingly. Don’t expect strategic premiums.
For structural capabilities, build the documentation and metrics that prove defensibility. Track data accumulation rates. Measure how model performance improves with scale. Demonstrate cost advantages. Quantify the time and expertise required to replicate your approach. Make the moat visible and measurable.
Acquirers look for businesses where AI creates compounding advantages that grow stronger over time. They avoid businesses where AI is a feature that competitors will match by the time the acquisition closes.
The Model Ownership Question
Here’s one of the sharpest dividing lines in valuation: do you control your models, or do your models control you?
Businesses that own their model weights, training infrastructure, and improvement processes command premiums. They have strategic flexibility. They can optimize for their specific use cases. They capture the full value of improvements.
Businesses dependent on third-party APIs are at the mercy of providers who can change pricing, access, or terms at will. Buyers see this as vendor risk masquerading as AI capability. Some businesses have been built entirely on top of OpenAI or Anthropic. When those providers inevitably move upmarket or change their API economics, the dependent businesses may face existential pressure.
The defense: ‘we can switch providers’ doesn’t hold up. Switching costs are often very high. Model behaviors differ, prompts require rewriting, integration logic breaks and customers notice quality changes. In practice, API dependency is stickier than businesses admit, and buyers know it.
Own your models or accept commodity valuations.
The Talent Signal
Acquirers assess the technical team as part of any AI valuation. The presence or absence of specific expertise sends clear signals.
A team of ML researchers who have published papers in top-tier conferences signals genuine technical capability. These teams can build proprietary approaches and solve novel problems.
A team of application developers who call APIs signals integration work, not innovation. These teams can ship features quickly but don’t create defensible moats.
The talent composition reveals the truth about AI capabilities faster than any pitch deck. Buyers look at LinkedIn profiles, research publications, GitHub contributions, and conference talks. They assess whether the team has the expertise to build what the business claims to have built.
You can’t fake this. The expertise gap between research-level ML engineering and API integration is visible to anyone who understands the technology.
Performance Metrics That Matter
Remember, traditional SaaS metrics — retention, LTV/CAC, gross margin — still do matter. AI businesses that neglect these fundamentals in favor of ML-only metrics confuse innovation with business value.
The AI-specific metrics that drive valuation are different from what most companies track.
Data accumulation rate relative to market growth. If your proprietary dataset is growing faster than the market, you’re strengthening your position. If it’s growing slower, competitors will eventually catch up.
Model improvement curves. Plot model performance against data volume and training compute. Steep curves that haven’t flattened indicate you haven’t hit diminishing returns. Flat curves suggest you’ve extracted most available value from your approach.
Inference cost trends. The unit cost per prediction should decrease over time as you optimize infrastructure and architectures. If costs are flat or rising, you’re not solving the hard engineering problems that create scalable value.
Retraining frequency and cost. Models that require constant retraining at high cost are liabilities. Stable models that maintain performance with minimal intervention are assets.
These metrics tell buyers whether your AI gets better or worse as the business scales. This determines whether AI represents a growth accelerant or a drag on margins.
The Deployment Reality Check
Where you run your AI matters more than most businesses acknowledge. Cloud-based inference looks clean on architecture diagrams, but it can create vendor lock-in and margin pressure in practice.
Businesses that have solved edge deployment command premiums because they’ve solved hard engineering problems. Running sophisticated models on resource-constrained devices requires compression, quantization, and optimization work that takes years. Here, the technical debt is valuable.
Edge deployment also changes business model options. One-time licensing becomes possible, so subscription fatigue stops being a growth limiter. Customers in regulated industries or with data sovereignty requirements become accessible.
Cloud-only inference means you’re perpetually paying rent to AWS or Google. Your margins are their revenue and acquirers discount future cash flows accordingly.
The Technical Debt Paradox
In traditional software, technical debt is seen as a liability. In AI systems, the right kind of technical debt is an asset.
Generic ML models with clean, simple implementations are easy to replace. Sophisticated systems with years of domain-specific optimization are not. The complexity represents accumulated knowledge. The truth is that the ‘debt’ is really embedded expertise that can’t be easily reproduced.
Of course, this doesn’t excuse sloppy engineering. The distinction matters: Is the complexity necessary to solve hard problems, or is it accidental complexity from poor architecture decisions?
Necessary complexity in model architectures, feature engineering, and production optimization creates defensible moats. Accidental complexity in integration code, deployment pipelines, and monitoring systems creates liabilities.
Buyers can tell the difference. They bring in technical teams to audit the codebase and assess whether the complexity represents genuine innovation or accumulated bloat.
Preparing for Technical Due Diligence
Most businesses are unprepared for the depth of technical scrutiny that AI acquisitions now face. The gap between marketing claims and technical reality collapses during due diligence.
Build the documentation before you need it: architectural decision records that explain why you chose specific approaches, benchmark comparisons showing how your models perform against alternatives and cost breakdowns that demonstrate your infrastructure efficiency.
Prepare to explain your data sources and licensing. Many AI businesses have built on questionable data foundations — scraped content, ambiguous copyright, questionable consent. This creates legal risk that tanks valuations when discovered.
Document your model development process. Show the experiments you ran, the hypotheses you tested, the dead ends you explored. This demonstrates genuine R&D rather than taking the first approach that worked.
Make your model performance metrics auditable. Buyers will want to reproduce your results on held-out test sets. If your claimed accuracy was overfit to your evaluation data, this becomes immediately apparent.
The businesses that command premium valuations treat technical due diligence as an opportunity to prove their moat, not just a risk that needs to be managed. They’ve built real capabilities and are eager to demonstrate them.
The Reality Most Won’t Accept
Here’s a hard truth: most businesses adding AI capabilities aren’t building anything defensible. They’re integrating commodity tools and calling it transformation.
This isn’t necessarily a failure. Commodity tools can improve operations and user experience, and so they do generate value. However, that value just doesn’t command strategic premiums in acquisition scenarios.
The market has bifurcated. Genuine AI infrastructure companies with strong moats are getting extraordinary valuations. Everyone else is getting grouped with traditional software businesses, and their AI features are largely ignored in valuation discussions.
What’s more, the gap will widen. As foundation models improve, more AI capabilities will become commoditized. The bar for ‘defensible AI’ rises constantly. What required custom ML engineering two years ago can now be done with API calls. What requires custom work today will be commoditized tomorrow.
The businesses that maintain premium valuations will be those that continuously push into genuinely difficult technical territory. They’ll own their models, control their data flywheels, and solve problems that don’t have off-the-shelf solutions.
Everyone else should adjust their expectations and stop pretending their ChatGPT integration is worth a strategic premium.
By Ivan Ruzic, Ph.D. · Originally published on Medium