The AI Flywheel: The AI Trade Is Becoming the Data Trade

People are no longer using AI only for occasional experiments. They are beginning to reach for it before they search, write, analyse, translate, design or code. The important change is not simply that a model can answer a question. It is that the answer changes the user's next behaviour. One useful result creates trust. Trust creates repetition. Repetition creates dependence. Dependence turns a novelty into infrastructure. The same change is moving through companies. A worker asks for a summary; then the summary becomes a template; the template becomes an automated workflow; the workflow becomes an agent; the agent is connected to company data and given permission to take action. What began as one prompt becomes a permanent, measurable workload. That is why the growth curve can become self-reinforcing. Better models create better outcomes. Better outcomes attract more users. More users create more inference, more stored context and more demand for speed. Providers build more capacity, which improves availability and price-performance, making the next class of applications possible. The wheel turns again.

The next acceleration comes from reusable skills: small packages of instructions, code, tools and domain knowledge that let an AI system perform a task consistently. A person can describe an outcome in natural language; the system can select a skill, call software, read data, write code and complete the job. That makes computing easier to consume, but it does not make computing disappear. It makes it scalable. A successful skill can be reused by one person hundreds of times, then by a team, then by every customer. Each execution can require model inference, CPU work, database reads, object storage, vector retrieval, network calls, identity checks, logs, backups and monitoring. This is the hidden bridge between software intelligence and physical demand. Natural language lowers the barrier to creating workloads. Skills make those workloads repeatable. Agents make them persistent. Persistence creates a need for memory. Memory creates data, and data needs somewhere to go.

The visible centre of the AI trade is accelerated compute: designers such as $NVDA and $AMD, manufacturing capacity such as $TSM, and the specialised systems built around them. But a working AI service is a chain, and the slowest or scarcest link can govern the economics of the whole system. Cloud platforms including $MSFT, $AMZN and $GOOGL assemble that chain into capacity customers can rent. Networking suppliers such as $ANET and $AVGO help data move between processors and clusters. Storage groups such as $WDC and $STX retain training data, model artefacts, retrieval libraries, media, logs and the growing memory of agentic workflows. Power-and-cooling providers such as $VRT help dense racks operate, while generators, grids and utilities have to support the physical load. The UK-listed map is smaller but still relevant. #IOM provides cloud and managed infrastructure; #CCC sits in the enterprise technology supply chain; #VLX supplies power and connectivity products; #SGRO has highlighted data centres within its broader property and power strategy. These are different businesses with different exposures—the list is a map of the system, not a claim that every name will capture the same value. The investment idea is therefore broader than 'buy AI'. It is to ask where data is created, where it travels, where it waits, what keeps it available, and which supplier owns the constraint.

Recent company and energy disclosures show the physical side of the loop. $MSFT said it added 31 data centres across five continents in its June quarter, added another gigawatt of capacity, and increased throughput for some Copilot workloads fourfold since the start of the year. It also said Azure demand continued to exceed available capacity. Those facts connect usage, efficiency and construction in one operating system: squeezing more output from existing infrastructure while adding more infrastructure. Storage is participating too. $WDC reported fiscal fourth-quarter revenue of $3.75 billion, up 44% year on year, and described growing storage demand as global data creation accelerates. The important point is not one quarter or one share price. It is the mechanism: AI produces and reuses data far beyond the moment of inference. The International Energy Agency projects global data-centre electricity consumption to more than double to about 945 TWh by 2030 in its base case, with AI the most important growth driver alongside other digital services. That forecast is not proof that every data-centre project will earn an attractive return. It is proof that the digital experience has an increasingly material physical footprint.

Efficiency is often presented as the argument against the infrastructure build-out. If models need fewer tokens, chips become faster and inference becomes cheaper, surely less equipment will be required. At the level of one task, that can be true. At the level of the system, a lower cost per useful outcome can unlock many more outcomes. Faster responses make voice natural. Cheaper inference makes always-on agents practical. Better coding systems let more people build software. Larger context windows invite richer documents, video and memory. Higher reliability moves AI from drafting into execution. The central question is therefore not only how much compute one prompt consumes. It is whether efficiency falls faster than total demand expands. If the cost of intelligence falls and consumption grows even faster, aggregate demand for compute, data movement and storage can continue rising. The flywheel is powered by usefulness.

A powerful story still needs an invalidation map. First, capital intensity matters. Capacity can arrive before revenue, and depreciation, financing, energy and lease costs can pressure cash generation. Secondly, infrastructure is cyclical: shortages attract supply, supply can overshoot, and pricing power can reverse. Thirdly, value capture will be uneven. A company can sit inside a growing market and still lose economics to competitors, customers or technological substitution. Power connections, planning permission, cooling water, chip supply and network availability can delay deployments. Regulation, privacy and data-residency rules can change where information is processed or retained. Better compression and model architecture can reduce storage or compute per task. Enterprise adoption may also arrive more slowly than consumer excitement suggests. The bullish thesis strengthens when paid AI use, contracted demand, utilisation and revenue rise alongside disciplined capital returns. It weakens when capacity growth outruns durable consumption, when unit economics fail to improve, or when debt and dilution absorb the benefit. The data trade should be measured in cash flows, not adjectives.

To analyse the flywheel, follow one piece of information through the system. 1. Creation — a user, sensor, company or model creates data. 2. Intelligence — accelerators and software turn that data into an output. 3. Movement — switches, fibre and networks move it between machines and regions. 4. Memory — databases, flash and high-capacity storage retain what must be used again. 5. Operation — power, cooling, buildings and security keep the service available. 6. Distribution — cloud platforms and applications place the result in front of more users. 7. Repetition — a useful result becomes a habit, skill or agent, producing the next workload. This is why the AI trade is becoming the data trade. Intelligence is not a finished object stored inside one model. It is an active system that repeatedly creates, moves and remembers information. As AI becomes easier to use, the number of people capable of creating those workloads expands. As the system becomes faster, expectations rise. As expectations rise, more capacity is required. The interface may be a sentence. Beneath it sits an industrial build-out.

Primary company disclosures and the IEA's energy model were used to verify the operating evidence. Company references elsewhere in this edition identify layers of the infrastructure chain; they are not recommendations or a complete sector screen. Source cut-off: 12 September 2026.