Every major technology transition creates a period where adoption moves faster than understanding. The early internet created businesses that assumed traffic automatically translated into value. Cloud computing created organisations that moved infrastructure externally without fully understanding operational dependency. Mobile platforms created ecosystems where companies built entire businesses on rules controlled by someone else. Artificial intelligence is entering a similar phase.
The current assumption is straightforward: intelligence is becoming abundant, accessible and cheaper. Every generation of models delivers more capability. Token prices decline. New providers emerge. Businesses rapidly integrate AI into customer service, software development, operations, research and decision-making.
The technology is real. The strategic assumptions surrounding it may not be. The most important question facing executives is not whether AI will transform business. It will. The question is whether organisations are building AI capability or simply building dependency on AI providers. That distinction will define many winners and losers in the next decade.
The Illusion of Falling AI Costs
Recently, the proposal of an AI Token Price Index highlighted an emerging issue: unlike traditional software, AI consumption has a measurable underlying cost structure. A token may appear to be a simple unit of text processing, but behind every token sits an industrial ecosystem of advanced semiconductors, data centres, electricity, networking and specialised infrastructure. This creates a fundamental difference between AI and previous software revolutions.
Traditional software benefited from near-zero marginal replication costs. Once Microsoft, Adobe or Oracle created a product, selling another licence added minimal additional cost. Cloud computing changed the location of infrastructure but did not fundamentally change this economic model. Amazon Web Services, Microsoft Azure and Google Cloud still provide predictable compute, storage and networking services where customers understand the relationship between consumption and cost.
AI introduces a more complex dynamic. The customer is not simply renting infrastructure. The customer is renting intelligence. The distinction matters because the supplier controls not only the hardware layer, but also the capability layer. The model architecture, training approach, performance characteristics, pricing model and availability are determined externally. Today, this capability is being offered at prices influenced by strategic competition rather than pure economics.
OpenAI, Anthropic, Google and others are investing tens of billions into AI infrastructure while operating in a market where profitability remains uncertain. Microsoft has committed substantial resources to OpenAI while simultaneously expanding its own AI infrastructure. Amazon and Google are making similar investments to secure their position in the AI ecosystem.
This resembles previous platform strategies where market share is prioritised before profitability. The customer benefits. The dependency accumulates.
The Cloud Comparison Trap
Many executives compare AI APIs to cloud computing. The comparison is understandable but incomplete.Cloud succeeded because it separated ownership from operation. Companies no longer needed to purchase servers, maintain data centres or manage hardware depreciation. They gained flexibility while relying on infrastructure that remained relatively standardised. A company could move workloads between cloud providers. Virtual machines, containers, databases and networking architectures followed relatively well-understood patterns.
AI is different.
A company building on a foundation model is not simply renting servers. It is integrating an external reasoning engine into its processes. If an organisation builds customer service automation, internal knowledge systems, software development workflows or decision-support systems around a particular model, changing providers may require significant redesign. The dependency is therefore deeper. Cloud dependency was often infrastructure dependency. AI dependency can become capability dependency.
The Wrapper Problem
This leads to another strategic challenge: the rise of AI wrappers. The term is sometimes used dismissively, but the reality is more nuanced. Many successful technology companies have historically been built on top of platforms. Value is often created through application design, workflow integration and customer understanding. However, history shows that platform advantage is difficult to defend when the platform owner enters the same market. Apple’s “Sherlocking” of third-party applications is perhaps the most famous example. Features developed by independent developers were later incorporated into the operating system itself. The same pattern occurred in cloud computing. Independent infrastructure companies created valuable capabilities that were eventually absorbed into larger platforms.
AI companies face the same question. If a startup builds a product around access to a foundation model, what prevents the model provider from offering the same capability directly? Jasper provides an early example. The company built a successful AI writing platform and reached significant scale before the arrival of ChatGPT dramatically changed the competitive environment. The lesson was not that Jasper lacked value. The lesson was that access to intelligence alone was not a sustainable advantage. The defensible advantage lies elsewhere: proprietary data, specialised workflows, domain expertise, customer relationships and operational integration.
The Hardware Reality Behind the Software Story
Another misconception is that AI is simply the next software revolution. It is not. AI is software built on an increasingly expensive physical foundation. A large AI training cluster requires thousands of specialised GPUs, massive electricity consumption and sophisticated cooling infrastructure. Nvidia’s accelerated computing business has demonstrated this reality, with demand for H100 and newer Blackwell systems exceeding supply as companies race to secure compute capacity.
A single large-scale AI cluster represents hundreds of millions of dollars in hardware investment before considering energy, facilities and operational costs. Software improvements can reduce the amount of computation required. They cannot eliminate the requirement for computation. This creates an unusual economic tension. The industry is simultaneously experiencing rapid technological efficiency and rapidly increasing infrastructure demand. Both can exist together. The mistake is assuming one automatically eliminates the other.
The Strategic Response: Ownership of Intelligence Assets
The answer is not that every company should build its own AI infrastructure. That would be unrealistic. Only a small number of organisations will ever train frontier models at global scale. The strategic question is different. Which parts of AI capability represent a competitive asset? For some companies, the answer may be proprietary models trained on specialised data. For others, it may be private deployment of open models. For many, it will be a hybrid architecture combining commercial frontier models with internally controlled systems. The important factor is optionality. Companies that understand their dependency points can make deliberate choices. Companies that build without understanding those dependencies may discover that their AI strategy is controlled by decisions made outside their organisation.
The Next Phase of AI Competition
The first phase of AI was a race for capability. The second phase will be a race for economics. The winners will not necessarily be those with access to the largest models. Everyone will have access to powerful models. The winners will be those who understand where intelligence creates strategic differentiation and where it is simply a commodity input. The organisations that treat AI as another software subscription may eventually discover they have outsourced part of their competitive advantage. The organisations that treat AI as strategic infrastructure will make different decisions about data, architecture, suppliers and ownership.
The greatest AI risk is not failing to adopt the technology. It is adopting it successfully while misunderstanding who controls the system on which the business increasingly depends.