Surging server sales surely shouldn't stop

Welcome to Runtime! Today: IDC's most recent data shows both hyperscalers and server huggers are on a spending spree, the Trump administration might authorize U.S. security professionals to go after nation-state hackers, and the latest funding rounds in enterprise tech.

(Please forward this email to a friend or colleague! And if it was forwarded to you, sign up here to get Runtime each week.)

Serve and volley

There are many ways to quantify the excesses of the generative AI boom; startup funding, offers made by Meta to AI researchers, or rent in San Francisco's Hayes Valley neighborhood all work in a pinch. But the most direct way to understand the magnitude of the frenzy is to look at sales of the raw ingredient for both traditional and AI applications.

Worldwide server sales reached an all-time high during the third quarter of 2025, according to new research from IDC. Companies spent $112.4 billion on servers during the period, up 61% from the same period last year, and it's not just Nvidia's GPUs driving the boom.

• Sales of servers with x86 processors (made by Intel and AMD) increased by 32% to $76.3 billion, while sales of servers using other types of chips — in most cases, probably Arm chips like AWS's Graviton — increased 192.7% to $36.2 billion.
• But sales of servers with an embedded GPU increased 49.4% compared to last year, which IDC said was a key part of the growth over the past year and accounted for "more than half" of the overall market.
• It's been a little hard to tell how much of the enterprise market was interested in Nvidia's GPU/CPU combinations, as opposed to companies buying its GPUs separately and plugging them into their own servers, but the data suggests Nvidia is becoming an even bigger part of the data center mix thanks to those products.

Most of those servers are going into the hyperscalers (AWS, Microsoft Azure, Google Cloud, Meta, and Oracle to some extent), however, rather than fueling a big cloud repatriation movement that has been talked about at length over the five years. The big companies don't buy servers off the shelf; they spec them out with custom hardware and software to improve performance and reliability.

• Sales of servers from original design manufacturers (ODMs) increased much faster than the overall market, up 112.2% to $66.8 billion and taking a significant amount of market share away from the original equipment manufacturers (OEMs).
• Those custom shops now control more than half the market for servers, with 59.4% market share, up from 45.1% at this time last year.
• Dell led the OEMs with $9.3 billion in revenue during the quarter, up 37.2% from last year, but actually lost market share overall.
• And sales of external storage systems for on-premises data centers grew just 2.1% during the quarter according to IDC, suggesting that companies are modernizing their data strategies on the cloud.

But all anybody wants to talk about as the year winds down is whether this pace of investment will continue into 2026, given that enterprises are still struggling to see a return on their own investments into the AI services run by all those servers. IDC's prognosticators think the overall market for IT spending will increase by 10% next year, down from this year's pace "but still representing one of the strongest years for the industry since the 1990s," it said in a recent press release.

• The answer might hinge on how long companies are able to use the servers they're snapping up in 2025; in the first half of the decade, hyperscalers extended depreciation schedules for their servers from three years to six years after realizing they could still make money on older technology.
• But nobody is really sure how long AI servers will be useful, given how new they are to the enterprise IT stack and how quickly the market is evolving.
• In a world where AI scaling laws hold up and hyperscalers and frontier model companies still believe throwing obscene amounts of computing power at the training process is the only way to improve performance, they'll keep spending money to replace aging server fleets.
• In a world where businesses start to realize the benefits of smaller models and new techniques like neurosymbolic AI grow popular, all the servers that were snapped up in 2025 could remain useful for a long time.

Hacking as a service

Amid growing threats to governments and businesses around the world from cyberattacks, the Trump administration is considering allowing private businesses in the U.S. to strike back. According to Bloomberg, the government wants "to enlist private companies in more aggressive efforts to go after criminal and state-sponsored hackers in a new national cyber strategy, a draft of which has been viewed by industry officials and experts."

The trial balloon, if formalized, would involve an executive order offering legal protection to companies that undertake offensive operations against known threat actors, which range from get-rich-quick schemers to quasi-military operations according to Microsoft's research. The U.S. government doesn't spend a lot of time talking about its offensive cyberattack capabilities, but it has unleashed government agencies against foreign targets before, perhaps most notably through the development of the Stuxnet worm that targeted Iran's nuclear infrastructure in 2010.

There are probably several American cybersecurity companies that would jump at the opportunity to take down targets that have unleashed ransomware and stolen sensitive data for years, but such a policy would open a whole new can of worms. "There is currently no legal basis for private firms to conduct their own offensive cyber operations," as Bloomberg put it, and escalation seems very likely.

Enterprise funding

Databricks raised $4 billion (!) in Series L (!!) funding, which values the company at $134 billion, nearly double the market cap of rival Snowflake as of the close of Tuesday's stock market.

Harness scored $240 million in Series E funding, which values the software-development pipeline company at $5.5 billion.

Serval landed $75 million in Series B funding for its IT service management product, which will help it take on ServiceNow and Freshworks.

Outset raised $30 million in Series B funding as it builds out a customer-research platform using AI.

Prime Security landed $20 million in Series A funding for its AI agents, which the company says help companies instill security into the software-development process.

Kilo Code scored $8 million in seed funding for its open-source coding agent, which can work in editors like Visual Studio Code and Jetbrains or through the command line.

The Runtime roundup

Three U.S. Senators sent letters to the hyperscalers Tuesday asking for more information about the energy consumption of their data centers, another sign that it could get much harder for AI companies to build U.S. data centers as local communities start to push back.

Nvidia acquired SchedMD, which develops an open-source workload scheduler that is popular with high-performance computing organizations as well as neoclouds like CoreWeave.

Thanks for reading — see you Thursday!