A new BCG analysis says enterprises would have to carry most of the cost of data centers, chips, and model development
Artificial intelligence has a bill coming due, and someone has to pay it. According to a BCG Institute analysis, the AI ecosystem will need about $2.6 trillion in annual end-market spending by 2030.
That figure is meant to cover data centers, chips, model development, and returns for investors. It is the amount customers would need to spend each year for the whole buildout to make economic sense.
The report is titled “Who Will Pay the $2.6 Trillion AI Infrastructure Bill?” BCG’s answer: mostly businesses.
Breaking down the $2.6 trillion
Enterprises carry the largest share by far. BCG estimates they would account for $1.8 trillion of the annual total.
Consumers come in a distant second at $366 billion. That covers individuals paying for AI products directly, from chatbot subscriptions to AI features built into devices and apps.
Hyperscalers, the giant cloud operators that build and run most of this infrastructure, would account for $204 billion through their own internal use.
The final slice is $200 billion for emerging technologies, with physical AI named as an example. Physical AI refers to systems that act in the real world, such as robots and autonomous machines, rather than living purely in software.
Two important qualifiers sit behind all of these figures. The estimates reflect global spending excluding China, and the underlying infrastructure costs are concentrated mainly in the US, where most new capacity is being built.
How BCG built the model
The model starts with the cost of operating data centers. That includes power, chips, and depreciation, which is the accounting cost of hardware losing value as it ages.
From there it moves to AI labs, the companies training and running frontier models. Their costs include salaries and data licensing, the fees paid to use content for training.
On top of that, BCG layers a modeled 10% operating margin on external revenue.
BCG argues that spending at this scale is plausible when measured against current IT budgets. The case rests on two trends moving together: demand for intelligent technology intensifying, and costs becoming more manageable for enterprises.
The report also lists nine key indicators to watch as the decade unfolds. Among them are chip utilization rates and the shift from training AI models to monetizing inference.
Inference is the stage where a trained model actually answers questions and does work for users. Training is the expensive school years; inference is the paycheck that is supposed to justify the tuition.
Why the number matters
For enterprises, the takeaway is pointed. If businesses are expected to supply $1.8 trillion a year, AI has to move from pilot projects and experiments into core operations that justify sustained budget lines.
The $366 billion consumer estimate is notable for its relative size. Despite the attention consumer chatbots attract, BCG’s model suggests individual users are a supporting player rather than the main source of revenue.
The hyperscaler figure highlights a quieter dynamic. Large cloud companies are both landlords and tenants in the AI economy, and $204 billion of internal use means a chunk of demand comes from these firms deploying AI across their own products and operations.
The physical AI category is the most speculative piece. Robotics and autonomous systems are earlier in their commercial life, so the $200 billion slice depends heavily on those technologies maturing over the next several years.
What to watch from here
For investors tracking chipmakers, data center operators, and AI software providers, BCG’s indicators offer a practical scorecard. Chip utilization is a straightforward test: idle hardware is a sign that capacity is running ahead of demand.
The training-to-inference transition may be the more telling signal. Training spending is largely a cost center, while inference is where usage, and therefore revenue, shows up.
The geographic concentration also carries weight. With most new capacity being built in the US, American power grids, local permitting, and domestic supply chains become central to whether the buildout proceeds on schedule.
Power appears directly in BCG’s cost model alongside chips and depreciation. That means energy prices and electricity availability flow straight into the size of the bill that customers ultimately have to cover.
The exclusion of China is a reminder that the figure does not represent the entire global AI market. China’s AI ecosystem operates on its own track, so the true worldwide spending picture would be larger than BCG’s estimate.
