Why Seattle’s AI ambitions started with a hypervisor migration

3 key takeaways

• Prioritize workload-specific placement over “all-cloud” mandates to ensure higher uptime, lower costs, and total data sovereignty.
• Maintain a “recoverable position” by using extensible platforms that allow you to reclaim and redeploy data across different providers.
• Clean up legacy “bloat” first. A solid hybrid foundation is necessary to scale AI pilots into meaningful operational value.

Seattle, the largest city in the Pacific Northwest, was facing a $250 million budget gap.

Like many large organizations, the city suspected inefficiencies were hiding in plain sight — including in its legacy IT infrastructure. CTO Rob Lloyd and his team set out on a hypervisor migration project, partnering with Nutanix to consolidate thousands of legacy virtual machines, reduce server sprawl and save the city millions.

Still, the modernization effort wasn’t simply about moving everything to the cloud. While as-a-service environments can offer flexibility and resiliency, Lloyd said cloud isn’t the right answer for every workload.

“There is a lot of flexibility we want to have, because it’s hard to predict — we’ve seen these technology waves before,” he said. “Our philosophy has been very acutely cloud-smart, not cloud-first.”

From budget crisis to IT reset

Seattle is home to roughly 816,000 residents and more than 50,000 businesses, and is what Lloyd calls an $8 billion organization. The city’s electric, water and wastewater, sanitation, and parks and recreation departments are among the largest in the country. Its engineering department oversees billions of dollars’ worth of construction, and it runs massive police and fire departments.

When facing a large 2025-2026 budget deficit, Lloyd and his team began taking stock of the city’s IT infrastructure. Their key questions: What were their needs moving forward? What cost options did they have? What flexibility did they need? How could they build something to support future AI projects? And, importantly, what were different departments asking for?

The IT team performed a seven-month analysis of different environments (from full cloud to hybrid), analyzed a half-dozen platforms, projected total-cost-of-ownership (TOC), evaluated feature parity, and mapped out every risk.

Ultimately, they settled on Nutanix; Lloyd cited the company’s ability to quickly answer their key questions, collaborate, strategize on AI ambitions, and offer an extensible environment for numerous departments and use cases.

Within a year, the city successfully migrated 2,500 legacy VMs to the Nutanix Cloud Platform, all while keeping services online. They quickly saw benefits in speed, uptime, and costs.

From a cybersecurity perspective, Lloyd said that Nutanix baked encryption and microsegmentation directly into the hypervisor, and provided native support for federal security standards and automated containerization.

Ultimately, the city is saving between $1.6 and $2 million a year with Nutanix; this is not only due to the reduction of systems and servers, but lower licensing costs and “efficiency plays and optimization,” Lloyd said.

“One of the objectives in the project is, how can we actually see bloat over the years, subtract that and yield that savings back to the environment?,” he said. Now, they have visibility into network performance and can optimize as needed.

What enterprises can take away from Seattle’s philosophy

While the Nutanix migration was an integral component of Seattle’s IT modernization, the city has maintained its “cloud-smart, not cloud-first” strategy.

“The hyperconverged environment is our philosophy of saying that the future isn’t entirely cloud,” Lloyd said. “We’ve never believed in entirely-cloud.”

In some cases, the city’s on-premises environments have provided higher uptime and reliability than its cloud environments, he said. In others, though, when service-level agreements (SLAs) aren’t hitting the mark, “we’ll up our game” to as-a-service. Cloud environments are optimal when it comes to security and other “compliance regimes,” he said.

The capability to choose based on use case is critical. “There is a lot of rationale in making sure that you can use every flexibility that cloud offers,” said Lloyd.

Flexibility is also important in vendor relationships. Vendor lock-in is always a risk, so organizations should maintain control of their data and establish guidelines for vendor performance.

“We want the ability to reclaim and redeploy to another provider if need be,” LLoyd explained. While vendors may be able to launch, execute, and hit high service levels, it’s equally important that they have strong defensible and recovery positions.

“If it’s crown jewels, if there’s a digital Dark Age, we know that we’re okay,” said Lloyd. “If a cloud partner has some real struggles, we have a recoverable position.”

This is something a lot of organizations don’t necessarily think through, he said. However, it’s critical in government environments. “The bet is that governments will outlast most companies,” said Lloyd. “We want the ability to deploy, reclaim, redeploy, and optimize our resources, compute, and application environments, to keep the city sound.”

Ultimately, infrastructure is changing for everyone, and there are “good foundations on which to build,” said Lloyd. However, organizations must think long-term— even as short-term thinking is tempting in the rapid-pace technology world.

“The long arc is being very flexible and giving yourself lots of options and a good foundation to return to.”

Defining Seattle’s AI ambitions

Like every other organization, Seattle is also exploring its AI options, having rolled out an AI policy and 2025-2026 AI plan. The city is piloting OpenAI chatbots and has identified about 50 proof-of-concepts.

Notably, Seattle has rolled out Jasper and Smartcat to improve communication with residents, and has also used AI to:

• Identify the most dangerous intersections to make engineering improvements;
• Inspect sewer and water pipes needing repair and improvement (which is critical in a high-rainfall city like Seattle).

However, Lloyd and his team were realistic when piloting. They anticipated that the first 25 to 30 applications wouldn’t stick — which proved correctly, as about 80% didn’t pan out in a meaningful way.

But that doesn’t mean they weren’t valuable, Lloyd said; they’ve served as learning exercises. New technologies are often on a “matrix of learn and impact.”

“The next batch, we want to be much more effective,” he noted. “Are they going to produce, with high certainty, or much higher certainty, the value we’re looking for?”

Because “AI isn’t the answer for everything, large language models (LLMs) aren’t the answer for everything,” he said. Sometimes small language models, or a string of models could be the best bet.

“We’re very, very careful with AI,” Lloyd said. “Because it’s fire: You can do a lot of good with it. You can do a lot of damage with it.”

This has required a pragmatic approach to training and using a relevant vernacular. The first step is fundamental: What is AI? What is it not? Then it’s about establishing AI in context: How will it impact workers? How should they consider AI in program design, or in change management?

Lloyd made a creative correlation: Improper AI training (or training with any new technology, for that matter) is like randomly putting soccer balls in front of 3 or 5-year-olds. “If you don’t teach the sport or the goal, they’ll probably kick each other more than they’ll ever get the ball in the goal.”