Not Memory. Not Energy. This Is the Biggest Bottleneck in Artificial Intelligence, and These Are the Companies That Can Benefit.

There's a lot that goes into building new AI data centers. Companies need to secure a site, ensure it can supply it with enough power and water to run thousands of servers and keep them cool, and build the structure itself. Only then can it start installing racks full of AI chips and networking equipment.

As the industry looks to build as fast as possible to support the growing demand for AI computing, it's faced several challenges. Ensuring there's enough power to run each data center is a challenge. Meanwhile, wafer shortages have impacted the entire semiconductor industry. Memory chips have been in particularly short supply, as the three biggest memory chipmakers hesitated to expand production capacity, fearing another big cyclical downturn.

Despite the supply chain challenges for building new data centers, hyperscalers and other AI data center companies could face an even bigger bottleneck, according to the CEOs of Broadcom (NASDAQ: AVGO) and Intel: capital. Long lead times and massive price tags for data centers means capital constraints could be holding back some companies from building as fast as they want.

Here are the companies that can benefit the most.

Image source: Getty Images.

The biggest companies are starting to feel the pinch of building

The biggest hyperscalers are facing negative free cash flow as they continue to increase their AI spending.

Alphabet (NASDAQ: GOOG) (NASDAQ: GOOGL), for example, reported negative free cash flow for the first time as a public company last quarter. That free cash flow challenge is expected to get much worse in 2027, and it's raised capital in the debt and equity markets to fund its future build-out.

Amazon (NASDAQ: AMZN) saw its trailing-12-month free cash flow dip into negative territory last quarter. Management cited a $66.1 billion increase in capital expenditures as the reason why.

Meta Platforms, Oracle, and Microsoft are all in similar boats, although Microsoft intends to remain cash flow positive.

Negative free cash flow combined with rising interest rates means companies could face challenges raising enough capital to keep building. The problem is much more pronounced for neoclouds and the leading AI labs, which don't have other sources of revenue or the creditworthiness to support the massive loans needed to continue their build-outs.

Both Nvidia and Broadcom have taken the approach of securing financing for their customers to ensure they can keep buying. That comes with additional risks, but it also creates loyal customers and helps solve the biggest bottleneck for some. Still, the best way to overcome the financing challenge is to reduce the cost of building new data centers in the first place.

The biggest winners

There aren't many ways to reduce the cost of standing up data centers and installing new hardware. The biggest cost is the chips themselves, though, and there's a clear shift going on there. The biggest hyperscalers along with OpenAI and Anthropic are using custom silicon for a growing share of their computing.

Custom chips, or XPUs, come with much lower upfront costs and can deliver better price-performance on specific tasks. That said, GPUs are more flexible and offer superior processing capabilities for many tasks, so a certain number are necessary for a complete cloud computing platform offering.

Using XPUs can save cloud platforms a lot of cash. During Amazon's first-quarter earnings call, CEO Andy Jassy explained the potential savings from its custom Trainium chips. "At scale, we expect Trainium will save us tens of billions of dollars of capex each year and provide several hundred basis points of operating margin advantage versus relying on other chips for inference," he said.

The head of Google Cloud, Thomas Kurian, said the payback period on its custom TPU servers is half that of its average server. While the chips may be more power efficient, the bulk of that accelerated payback likely comes from lower upfront capital needs.

As such, I expect XPU makers to be some of the biggest beneficiaries as companies look to keep capital spending in check. Broadcom is the market leader in XPU chips, but smaller competitors like Marvell Technology (NASDAQ: MRVL) could also see rapid growth as XPU sales outpace those of GPUs and other chips. While Marvell is expected to grow slightly faster, Broadcom sits in a very strong position with leading AI labs and hyperscalers, and its stock trades for half the price on a forward P/E basis.

Adam Levy has positions in Alphabet, Amazon, Broadcom, Meta Platforms, and Microsoft. The Motley Fool has positions in and recommends Alphabet, Amazon, Broadcom, Intel, Marvell Technology, Meta Platforms, Microsoft, Nvidia, and Oracle. The Motley Fool has a disclosure policy.