In March 2026, Intuit CEO Sasan Goodarzi stated directly that large language models are commodities. He was not being provocative. He was describing a market reality that the pricing data had been pointing toward for months. GPT-5 Nano is available at $0.05 per million tokens. DeepSeek V3.2 processes frontier-class work at $0.27 per million tokens. The price gap between the cheapest and most expensive frontier model in 2026 is 35 times. In 2024 it was 100 times. In 2023 it was effectively infinite because competitive open-weight models did not exist. The compression is not a temporary pricing war. It is the structural consequence of multiple well-resourced labs producing comparable general intelligence capability, with open-weight models from DeepSeek, Meta’s Llama, and others providing a zero-marginal-cost floor that commercial models must price against.
The enterprise planning question that follows from commoditisation is not which model to buy. It is what happens to the infrastructure investment thesis when the model itself stops being a source of competitive advantage. The answer reshapes nearly every aspect of how operators evaluate, design, and run AI infrastructure. The AI infrastructure market has not yet caught up to that change.
The Competitive Advantage Has Shifted to the Infrastructure Layer
When frontier AI models become genuinely commoditised, competitive advantage in enterprise AI shifts from model access to the infrastructure and operational layers that determine how effectively organisations deploy those models. Softkraft’s May 2026 enterprise AI trends analysis identifies the build-versus-buy reassessment driven by model commoditisation as the most strategically significant enterprise AI shift underway. Enterprises that spent 2023 and 2024 trying to secure the best model are now asking a different question: given that five models can perform comparably on most tasks, what determines which enterprise uses AI most effectively?
The answer is not the model. The competitive advantage comes from the data pipeline that feeds the model with proprietary context, the fine-tuning and retrieval infrastructure that adapts commodity capability to specific workflows, the orchestration layer that manages multi-agent systems without generating runaway cost loops, and the security and compliance infrastructure that allows models to access enterprise data without violating governance requirements. All of those are infrastructure problems, not model problems. And they are the problems that create genuine competitive differentiation in an environment where the model itself is no longer differentiated.
The implication for AI infrastructure investment is direct. Enterprises that built their AI strategy around securing the best model are building on a competitive advantage that is depreciating as the model landscape commoditises. Enterprises building proprietary data infrastructure, fine-tuning pipelines, governance frameworks, and operational AI capabilities are building advantages that compound over time — harder to replicate than model access, and improving with each additional enterprise-specific data point and workflow integration. The infrastructure investment the commoditisation trend makes most valuable is not GPU compute. It is the data and workflow infrastructure that gives commodity compute proprietary context to work with.
What Model Routing Changes When Every Model Is Good Enough
The output gap between premium and budget model tiers has narrowed significantly in 2026. The cost gap has not. That asymmetry is the most operationally important fact about the current model landscape for enterprise infrastructure teams. An enterprise routing every query to a frontier model at $15 per million tokens when a capable model at $0.27 would produce equivalent output for 80% of workloads is spending 55 times more on inference than task quality requires. Model routing capability — directing queries to the appropriate tier based on complexity, quality requirements, and cost — is not a nice-to-have optimisation. It is the operational discipline that separates enterprises whose AI infrastructure economics are sustainable from those whose inference bills are growing faster than the value they generate.
The FinOps Foundation’s 2026 State of FinOps Report found that organisations reporting AI as an active FinOps concern jumped from 31% in 2024 to 63% in 2025. That jump reflects the arrival of commoditisation economics in enterprise budget conversations. When models were scarce and expensive, enterprises paid whatever frontier access cost because the capability was irreplaceable. When GPT-4-class performance costs $0.05 per million tokens, paying $15 for the same capability on tasks that do not require frontier performance is a financial governance failure, not a technology investment. The model routing infrastructure that prevents that failure is where real enterprise AI infrastructure value is being built in 2026.
The Open-Weight Question That Is Reshaping Self-Hosted Infrastructure
The open-weight dimension of commoditisation introduces a specific infrastructure question that enterprise AI teams are actively evaluating: when capable models become freely available for self-hosting, what total cost of ownership does self-hosted inference carry relative to cloud API access, and when do enterprises gain enough competitive advantage from self-hosting to justify the added operational complexity?
Open-weight models including GLM-5, Kimi K2.5, and Meta’s Llama variants are free to self-host for enterprises with GPU infrastructure, making zero-marginal-cost inference achievable for organisations whose data sovereignty, compliance, or latency requirements preclude cloud-based access. Regulated industries with data that cannot leave on-premise environments, enterprises with latency requirements that cloud API round-trip times cannot meet, and organisations with volume requirements at which self-hosted inference is substantially cheaper than API pricing are all evaluating this path as newly viable. The Google-Blackstone TPU venture documented that the alternative silicon story is being enabled partly by model commoditisation — when the model runs on anything, the silicon matters less than it did when CUDA optimisation was a frontier capability requirement.
The model commoditisation trend will accelerate before it stabilises. The labs producing frontier models will not stop improving capabilities, but competitors are matching those improvements faster than they matched previous generations. The enterprise that builds its AI strategy around model access as a competitive differentiator is building on a foundation the market is actively undermining. The enterprise building around proprietary data, fine-tuned specialisation, operational AI discipline, and infrastructure that leverages commodity model capability at maximum efficiency is building on a foundation the commoditisation trend strengthens rather than erodes.
