To Serve or Not to Serve Hyperscalers

In my recent discussions with three companies operating within the high-tech value chain, I am starting to see a pivotal question emerging: Should these companies prioritize serving the AI needs of hyperscalers or focus on enterprise customers? This dilemma impacts various industry players, including data center co-location providers, server OEMs/ODMs, and companies supplying software stacks for cloud providers.

Many server OEMs, for instance, report razor-thin margins on AI servers sold to hyperscalers—often in the low single digits—compared to the 10-15% margins typically achieved on non-AI servers for enterprise customers. This margin disparity is not unique and reflects a broader industry challenge.

Historical precedents provide valuable insights. The hard drive industry, for example, has experienced long-term margin compression as its customer base became concentrated among hyperscalers and cloud providers—entities often 10-100 times larger than their suppliers, leaving little room for pricing power. Their bigger challenge is centered on the fact that hyperscalers are the only customer base left.

Drawing from Geoffrey Moore’s "Crossing the Chasm," the go-to-market (GTM) and sales strategies for innovators and early adopters differ significantly from those needed to address the mainstream market (early and late majority, i.e., enterprise customers). Legacy technology companies (laggards) are unlikely to be the target market for these advanced AI solutions.

While much has been written about DeepSeek, skepticism remains regarding claims such as its $6 million model training cost. I personally believe unit volumes will rise 100-fold with a 10-fold price reduction, in line with Jevons Paradox. This is reflected in the significant increases of CAPEX of AWS, Google, Microsoft, and Meta from $220 billion in 2024 to more than $300 billion in 2025. I personally believe while this may continue to increase, the next wave of CAPEX install will come from enterprises and companies in the high-tech value chains need to position themselves to profitably serve this customer segment in the near future.

While serving hyperscalers may not be immediately profitable, it is crucial for building technical expertise that can later be leveraged to serve enterprise AI data centers expected in 2026-2027. Two compelling examples highlight this advantage: learnings that these companies may have gained while building X.AI’s data center, completed in 122 days, and Grok’s inference cluster in Saudi Arabia, built in 51 days. These timelines are unprecedented—even for hyperscalers, which typically take 2-3 years to construct a new data center.

Companies collaborating with hyperscalers and cutting-edge data center firms (e.g., X.AI, CoreWeave, Vultr, Lambda Labs) may initially operate on low margins. However, they gain essential technical capabilities, cost efficiencies, and accelerated deployment expertise—critical learnings that will serve them well when enterprise AI data centers scale in the coming years.

In conclusion, while the decision to serve hyperscalers may present immediate financial challenges, the long-term benefits in terms of technical expertise, cost efficiencies, and market positioning are invaluable. Companies in the high-tech value chain must carefully weigh these factors as they navigate their strategic priorities in an evolving industry landscape.

?