Opinion
The AI post-hype era will be won below the model layer
The winners of this era will be the organizations that look below the flashy model layer and focus on the plumbing.
The enterprise artificial intelligence market is not running out of momentum, it is running out of patience. Over the past three years, organizations rushed to launch copilots, proof-of-concept tests, and autonomous agents. The early, permissive question was simple: Can this technology do something impressive? Today, CFOs and boardrooms are asking a far harsher question: Does it solve a high-value business problem, operate reliably, and generate a measurable return?
This shift is exposing a widespread market gap between AI ambition and enterprise production readiness. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, while 60% of projects will be abandoned due to poor data readiness. The lesson is not that enterprises should reduce their AI ambitions, but that too many initiatives have started at the most visible layer- the model, chatbot, or copilot interface- while neglecting the unsexy technical and operational foundations beneath it.
This failure rate is exacerbated by how the cloud giants are approaching the enterprise. Rather than offering a single blueprint, AWS, Microsoft, and Google are entering from fundamentally different angles, leaving enterprise buyers with disjointed architectures.
Microsoft and Google are leveraging productivity as their main front door, embedding tools like Copilot and Gemini directly into daily employee workflows and workspace applications. Yet, giving employees access to a productivity interface does not automatically generate ROI. If enterprise data remains fragmented, access permissions are mismanaged, and core business applications aren't integrated, productivity tools merely generate usage metrics rather than operational value.
AWS, conversely, approaches the market from its traditional foundation: infrastructure, developer choice, and AI economics. By enabling enterprises to choose between managed frontier models, specialized open-source models, and custom accelerators, AWS targets architectural flexibility. But this choice introduces its own complexity. Deciding which workloads belong on expensive reasoning models versus smaller, cheaper open-source models on Kubernetes requires sophisticated engineering that few non-tech enterprises can execute alone.
This strategic divide creates both a bottleneck for enterprises and a major opportunity for cloud services firms. Buying access to cloud tools is no longer the hard part; the hard part is assembling them into a production-grade system that moves a business metric.
To bridge this gap, modern IT and cloud services partners cannot simply add AI products to a generic service catalog or sell open-ended implementation hours. The real enterprise value lies in seamlessly connecting five previously siloed layers: defining a clear business hypothesis up front, redesigning workflows around productivity tools, establishing a governed, AI-ready data foundation, engineering optimized multi-model cloud infrastructure, and managing ongoing production operations, including latency, cost control, and security. An infrastructure project without a business workload becomes an expensive asset searching for a purpose. Conversely, a productivity tool without data readiness generates hype without return.
This shift fundamentally changes the economics of tech consulting. As Sequoia Capital highlighted in its thesis on "services as the new software," the future belongs to firms that sell completed outcomes rather than raw labor. For cloud services providers, scaling in the post-hype era requires moving away from traditional, headcount-driven consulting. Generic migration and testing work is being rapidly automated. To remain defensible, delivery models must become productized: proprietary assessment methodologies, reusable architectural blueprints, standardized open-source deployment patterns, and recurring operational capabilities.
The customer should no longer be purchasing a bundle of billable engineering hours to conduct open-ended experiments. They should be purchasing a proven, repeatable way to deploy and run a production system.
The next phase of enterprise AI will be defined by ruthless operational discipline. The model layer itself will remain hyper-volatile; today’s leading LLM will inevitably be commoditized, repriced, or superseded tomorrow. Durable enterprise capability cannot be built on the model alone. Whether an initiative originates at Microsoft or Google’s productivity layer or deep within AWS’s infrastructure stack, all strategies ultimately converge in the same place: the production environment. The winners of this era will be the organizations that look below the flashy model layer and focus on the plumbing - connecting business strategy, clean data, scalable infrastructure, and productized delivery to produce sustainable economic value.
Gil Ron is the Chief Commercial Officer at Sela, a cloud and AI service provider.
