Global data center capex is projected to grow at an annual rate of 18% and reach $200 billion by 2028. At that point, cloud hyperscalers Amazon Web Services (AWS), Google Cloud, Meta and Microsoft will represent nearly half of global data center spending, according to a new report from Dell’Oro Group.

In 2023, data center providers overwhelmingly shifted their infrastructure investment priorities from general purpose computing and storage to artificial intelligence (AI) and accelerated computing infrastructure, Dell’Oro Sr. Research Director Baron Fung told SDxCentral. As a result, AI workloads are projected to represent 25% of annual data center capex by 2028.

“The shift has been pronounced given the high cost of the equipment and GPUs relative to general purpose computing, and also the sudden shift in AI-related infrastructure spending,” Fung said.

To drive long-term growth, cloud providers will work to streamline general purpose compute costs by adopting next-generation servers and rack-scale architectures. But that plan isn’t without its challenges.

Finding the perfect balance in capacity utilization to prepare for future demand, for example, requires careful planning and sufficient lead time. “Lack of capacity could mean that the cloud services provider would not be able to meet customers’ demand and they could lose share,” plus “one provider cannot add capacity quickly based on immediate demand,” Fung said.

There is also a balance to be struck between replacing and maintaining existing servers. This requires experimentation with “operating servers for a longer time versus replacing servers more frequently to ensure that the performance benefits of each upgrade cycle are captured,” he said.

Analysts also anticipate cloud hyperscalers will ramp up their vertical integration efforts to manage costs and continue to optimize their infrastructure.

Though the major hyperscalers have been building equipment like custom servers and network switches for a while, continuing to cut out the OEM might help reduce supply chain costs. “And they can deploy equipment that is matched for their technical requirements,” Fung said. “Excess components and features could be eliminated, which would reduce cost and increase reliability – fewer components failing.”

Most recently, the cloud hyperscalers have taken vertical integration to the next level by designing their own chips, including CPUs, smartNICs and accelerators. These efforts support workload optimization and control of supply chain costs.

On the other side of the equation, enterprises are set to face headwinds associated with economic uncertainties, leading to greater adoption of hybrid cloud models both for AI and general purpose compute.

Hybrid models allow enterprises to reduce or increase their cloud usage based on variable needs. “If they downsize their business or lay off employees, they can reduce their expense allocation to the cloud. Whereas, if all their infrastructure is hosted in their own private data centers, their data center cost has been sunk and can’t be optimized,” Fung said.