Episode Overview
The AI landscape has never looked more crowded. Every major tech company — Google, Microsoft, Meta, Amazon — now has at least one flagship large language model. Dozens of well-funded startups are racing to compete. And yet the gap between benchmarks and real-world value creation keeps widening.
Sam Altieri has a front-row seat to this chaos. As the founder of Nexus AI, a platform that helps enterprise teams evaluate and deploy foundation models at scale, he's seen firsthand what separates the companies building something durable from those chasing leaderboard glory.
In this conversation with host Jordan Mercer, Sam unpacks the real dynamics of the model wars — who's winning, who's faking it, and what the enterprise AI market will actually look like in 2027.
The Model Landscape
The first thing Sam pushes back on is the framing that there are "model wars" at all. "The narrative of a zero-sum competition misses the point," he says. "Most enterprises don't pick one model. They're building on three or four simultaneously, routing different tasks to different providers based on cost, latency, and quality."
This multi-model reality has profound implications. The companies building commoditised general-purpose models may be fighting a war they've already lost. The question isn't who has the best model — it's who has the best integration story, the best data flywheel, and the most defensible distribution advantage.
"OpenAI has the consumer brand. Google has the enterprise distribution. Anthropic has the trust narrative. Meta has the open-source moat," Sam says. "They're all playing different games. The mistake is trying to compare them on the same axis."
Building a Durable Moat
What does a real competitive moat look like in the foundation model era? Sam argues there are only three that actually matter: proprietary data, vertical depth, and switching costs.
Proprietary data is the obvious one — if your model gets smarter from being used, and competitors can't replicate your data flywheel, you have a structural advantage. But Sam is skeptical most players are actually achieving this. "The myth of the data moat is that you're collecting proprietary signal. The reality is most companies are collecting noise."
Vertical depth is more interesting. The companies that have gone narrow — deploying AI exclusively in legal, healthcare, finance, or coding — have found that domain specificity creates switching costs that horizontal players can't easily replicate. "A law firm that's integrated an AI deeply into their document workflow doesn't rip that out to save $2 a month per user."
Enterprise Adoption
Sam is characteristically blunt about the state of enterprise AI adoption: "Most enterprise AI deployments are still demos in production. They're not wrong to experiment. But the gap between proof-of-concept and production-at-scale is enormous, and most organisations underestimate it."
The bottleneck, he argues, isn't the models — it's the data infrastructure, the change management, and the evaluation frameworks. "You can't measure ROI on something you haven't instrumented properly. And most companies don't know how to instrument AI workflows."
Nexus AI built its product directly around this problem. The insight was simple: enterprises don't need better models. They need better tooling to understand which model to use, when, at what cost, and with what accuracy guarantees. "We're not in the model business. We're in the model operations business. That turned out to be a much bigger market."
Regulation Ahead
The conversation turns, inevitably, to regulation. The EU AI Act is now in force. The US has a patchwork of executive orders and sector-specific guidance. China has its own rules. Sam doesn't think any of this fundamentally changes the competitive dynamics — yet.
"Regulation always lags the technology. By the time regulators understand what they're regulating, the technology has usually moved on." The bigger risk, he thinks, is the compliance burden falling disproportionately on smaller players. "A startup with 40 people can't afford a regulatory affairs team. Anthropic and OpenAI can. That's a structural advantage that has nothing to do with the quality of your model."
Sam's Predictions for 2027
We end with the predictions segment. Sam makes three:
First, at least one of the current top-five foundation model companies will have pivoted away from the general-purpose LLM market entirely by the end of 2027, either acquired, shut down, or repositioned as a vertical AI player.
Second, open-source models will be "good enough" for 80% of enterprise use cases by late 2026. The cost pressure this creates will compress margins industry-wide. "The democratisation of capability is also the democratisation of commoditisation."
Third — and this is the contrarian one — the biggest AI company by market cap in 2030 won't be building models at all. "It'll be the company that figured out how to make every organisation's data actually useful. The picks-and-shovels play, but for the AI gold rush."
