The AI boom is entering a more complicated phase.
For years, the biggest question was whether companies could build powerful enough models and whether customers would pay for them. Now, the technology is getting cheaper at a remarkable pace, but the infrastructure needed to run it is becoming more expensive.
That creates a problem for investors.
AI prices are falling. AI infrastructure costs are not falling nearly as fast. And the companies buying AI still have a hard time showing exactly what they are getting back.
That gap could become one of the biggest questions for the AI trade over the next few years.
AI intelligence is getting cheaper
Look at what has happened over the past few weeks.
A free model called Ox Alpha appeared online and reportedly performed close to the frontier of what AI models can do. OpenAI has also cut prices on its flagship model several times in a short period.
The broader trend is even clearer when looking at token prices.
The cost of generating AI output is falling rapidly. That is normally a good thing. Technology becomes more useful when it becomes cheaper, because more people can afford to use it.
Think about cloud computing, smartphones or internet data.
Lower prices eventually create much larger markets.
But AI has an unusual problem.
The cost of producing that intelligence is not falling at the same speed.
The expensive part is underneath the AI
The AI model might be getting cheaper for customers, but the machines running those models remain expensive.
Nvidia’s chips are in enormous demand. Memory is becoming increasingly difficult to secure. Advanced chipmaking capacity is tight. Servers containing AI processors are also getting more expensive.
Some Nvidia customers have reportedly been told that server prices could rise by more than 15% for systems shipping early next year.
Samsung has raised prices for some advanced contract chipmaking, while a large portion of its memory capacity is being tied up in multiyear agreements with data-center customers.
SK Group’s chairman has also warned that the memory shortage could become worse in 2027.
So there is a strange situation developing.
The product is getting cheaper while the ingredients needed to make the product are becoming more expensive.
That is where the economics start getting uncomfortable.
But cheaper AI could create much more demand
There is a strong argument on the other side.
If AI becomes dramatically cheaper, companies may simply use much more of it.
That is already happening.
The share of US businesses paying for AI is approaching 60%, while spending has increased significantly across the market. The major hyperscalers are also seeing strong cloud growth, with their combined cloud revenue reaching roughly $106 billion in the latest quarter.
This is the classic technology adoption story.
Lower prices lead to higher usage.
And higher usage can eventually make up for lower prices.
The AI industry does not necessarily need to charge more per token if the number of tokens being consumed keeps exploding.
That is probably the most important argument keeping the AI bull case alive.
The bigger problem: where is the payoff?
This is where things get more difficult.
Investors have heard about AI productivity gains for years.
Companies talk about automation, faster employees, better customer service and new revenue opportunities.
But when researchers look for actual financial results, the evidence is still surprisingly thin.
One study looked at 919 earnings calls from the 60 largest US-listed financial companies over three years.
About four out of five calls mentioned AI.
More than half talked about its cost.
But only one company highlighted a realized dollar return from using AI, and it did so twice, with roughly $19 million in combined benefits.
That does not mean AI is failing.
It means companies are still struggling to translate the AI story into a number that investors can put into a spreadsheet.
Three years into the AI investment cycle, the spending is easy to see. The return is much harder to measure.
Wall Street is starting to notice
This is also showing up in credit markets.
Companies are borrowing enormous amounts of money to finance the next wave of AI infrastructure.
Broadcom, for example, has been in talks to raise more than $60 billion in debt to fund AI chips.
The logic is straightforward.
Build the infrastructure today because demand for compute is expected to be much larger tomorrow.
But debt investors are starting to demand more compensation for taking that risk.
The cost of insuring Broadcom’s debt has risen by around 80 basis points since January.
That does not mean credit markets are predicting an AI crash.
It does show that investors are becoming less willing to assume that every dollar of planned AI investment will automatically produce the expected returns.
This is becoming an economics problem, not a technology problem
The interesting part of the AI story today is that there is less debate about whether the technology works.
It clearly does.
The debate is increasingly about who captures the economic value.
If AI models become commodities, model providers may have less pricing power.
If chips, memory and data-center capacity remain scarce, semiconductor and infrastructure companies could capture a larger share of the economics.
If customers cannot generate meaningful returns from AI, they may eventually slow their spending.
And if infrastructure companies finance too much expansion with debt, investors may start questioning whether future demand will be enough to justify today’s capital spending.
There are a lot of moving pieces.
The Nvidia question is particularly interesting
Nvidia’s latest outlook provides some reassurance for the bulls.
The company expects roughly 70% revenue growth in fiscal 2028, which is well above the roughly 45% growth analysts had expected.
That suggests demand for AI infrastructure remains extremely strong.
But Nvidia has also warned that rising memory costs could put pressure on margins.
That creates an interesting split.
Strong AI demand is good for Nvidia’s revenue. Rising input costs could make that growth less profitable.
For investors, both things can be true at the same time.
What happens next?
There are really two competing versions of the future.
In the bullish version, AI gets cheaper, adoption accelerates and usage explodes.
Companies eventually figure out how to turn AI into measurable productivity gains. Those gains justify continued infrastructure spending, and today’s enormous capital investments begin generating equally enormous returns.
In the bearish version, AI prices fall faster than costs.
Customers push back on pricing, chip and memory suppliers remain expensive, and companies continue spending heavily without being able to demonstrate meaningful returns.
Eventually, investors demand lower valuations or companies reduce capital spending.
The biggest question is not whether AI adoption continues. It probably will.
The question is whether the economic value created by that adoption will be large enough, and arrive quickly enough, to justify the cost of building the infrastructure underneath it.
That is the part of the AI story the market is still trying to figure out.
And for investors, it may be more important than the next model benchmark.
Cheap intelligence is exciting. But someone still has to pay for the machines that produce it.