Microsoft is betting that its custom chips can lower costs, but the AI equation is more complicated than that

Microsoft is expanding efforts with its Maia chip business.

Microsoft wants to make artificial intelligence cheaper to run. For investors trying to judge the return on its enormous infrastructure spending, that sounds encouraging.

But a lower computing bill for each task leaves an important question unanswered: How much of the savings will Microsoft (MSFT) actually keep?

Some of the money saved could become pure profit. Some could reach customers through lower prices. Some could finance more ambitious AI services that consume the extra capacity. Microsoft's new chip puts all three possibilities in play.

Earlier this week, Microsoft announced plans to expand its Maia systems to multigigawatt deployments starting in 2027, making them its primary infrastructure for inference across its products and services. Inference is what happens when a trained AI model answers a question or works through a task.

Diversifying beyond expensive chips made by Nvidia (NVDA) and doing more with its own custom chips could be a real opportunity for Microsoft to improve its AI economics. The investment case depends on turning that engineering progress into profitable customer demand while controlling the capital required to serve it.

Don't miss: Don't ditch tech stocks just yet. Make these moves to take advantage of the latest AI reset.

Maia 300, scheduled to begin production deployments in 2027, is central to that effort. Microsoft expects it to deliver 40% more tokens - pieces of information that AI models process and generate - per megawatt than leading systems based on graphics processing units.

That is a company projection, and the announcement does not identify the competing systems behind the comparison. Investors should also resist reading it as a promise of 40% lower costs. Power-efficiency measures only part of the expense of delivering AI. Chips, memory, networking, buildings, software and maintenance all have to be paid for.

Even the power arithmetic needs care. Producing 40% more output with the same power would imply roughly 29% less energy per token under comparable operating conditions. Neither figure establishes the savings on Microsoft's total bill.

Nevertheless, the business logic is compelling. Microsoft sells applications and cloud services, giving it opportunities to design infrastructure around work its customers already perform. It says Maia 200 is already serving production workloads, including Microsoft Copilot, the company's AI assistant. The successor therefore builds on an operating program, although its promised performance remains ahead.

The financial stakes are substantial. In the June quarter, Microsoft Cloud revenue reached $59.3 billion, up 27% from a year earlier. Across the company, cash additions to property and equipment totaled $35.8 billion, compared with $17.1 billion a year earlier. Those figures cover more than AI, but they show the scale of the business and investment that efficiency improvements must support.

The clearest route to shareholder gains would be lower costs for serving existing customers. If Microsoft can deliver the same useful service for less while maintaining prices, it retains more revenue after delivery expenses. Repeated across millions of interactions, modest savings could become meaningful.

Consider this example: A service brings in $100 and costs $60 to deliver, leaving $40 before other expenses. Reduce delivery costs to $50, and that margin rises to $50. But if competitive pressure also reduces revenue to $90, the margin stays at $40. Technical progress has benefited the customer without increasing the provider's profit dollars on that sale.

That second outcome can still be attractive. Lower prices could bring in customers who previously found the service too expensive. Higher volume could then generate more total profit. The question is whether the additional business compensates for the lower price and any extra infrastructure required.

A third possibility is that Microsoft keeps prices steady while giving customers more AI work for their money. An assistant might examine more documents, check its answers more thoroughly or complete longer assignments. The customer receives a better product, but heavier usage can absorb the savings from cheaper computing.

That makes cost per useful completed task more revealing than cost per token alone. A system that produces inexpensive answers but requires repeated corrections can be costly to use. A more demanding system might be worth its higher computing bill if it reliably finishes work that customers value.

For Nvidia investors, Maia raises a broader question: How much pricing power can a supplier retain when its biggest customers develop credible alternatives? Microsoft does not need to replace every Nvidia chip to strengthen its negotiating position. It needs an alternative that performs enough of its work reliably and economically to make shifting orders a realistic option.

That could affect Nvidia well before any outright decline in sales. Microsoft could direct more of its next round of spending toward its own chips, or use that option to negotiate better terms on GPUs. Nvidia's revenue could still grow in an expanding market while it captures a smaller share of a customer's budget or accepts lower margins to secure business. Investors should distinguish those outcomes: Continued sales growth would not, by itself, settle the question of bargaining power.

The shift is not automatic. Microsoft must demonstrate that Maia's savings survive the costs of development, deployment and adapting software to run on it. A chip that performs well on selected tasks is useful, but its negotiating value depends on how much spending Microsoft can realistically redirect. Nvidia can defend its position by making its systems more valuable overall, including through performance, software support and reliability.

The question, then, is how much of AI's future growth Nvidia can capture, and on what terms. Successful custom chips could let Microsoft retain more of the money it earns from AI, while putting pressure on the prices Nvidia can charge. Both companies could benefit from rising demand, but shareholders should watch how the profits are divided.

See also: Microsoft hits a big milestone that solidifies its revival

For Microsoft shareholders, three measures will help show whether those potential savings are becoming financial gains.

First, cloud margins alongside growth as deployment progresses. Improving margins with sustained demand would support the case that Microsoft is retaining some efficiency gains. Aggregate figures cannot isolate Maia's contribution, so management's explanation of infrastructure costs and product mix will matter.

Second, customer spending and retention. Lower prices or more generous usage allowances need to translate into durable business. Investors should look for customers who continue paying and renew their contracts after the initial enthusiasm fades.

Third, cash generation relative to investment. A more efficient chip can improve operating economics while a larger deployment consumes additional cash. Investors should compare operating cash flow, capital spending and financing commitments over several quarters, allowing for the timing of construction and equipment purchases.

Microsoft's announcement strengthens its case for controlling the cost of AI. It leaves open how the resulting benefits will be shared among customers, suppliers and shareholders. The strongest evidence will come when more useful AI services produce durable revenue and cash returns that justify the infrastructure behind them. That is the standard by which investors should judge Maia.

-Jurica Dujmovic