TL;DR: The hardest problems discussed at EUC World Amplify 2026 sat between platforms, so measurement and process can solve them where another tool purchase cannot. The opening keynote and a session on Entra travel policies both traced configuration drift to habits rather than technology, including migrations treated as finished and temporary exceptions left in place long after they were needed. A session on cloud desktop hardware showed that two identical desktops can run on processors several years apart in age, and a security benchmark showed that an endpoint security stack can more than triple a virtual desktop's CPU use. In each case, a team can diagnose the problem with a simple check before spending money on a new tool.
EUC World Amplify 2026 was the most engineering-heavy conference I have ever attended. The event ran from 28 September to 1 October at the Baird Center in Milwaukee, organized by World of EUC, which grew out of the Citrix User Group Community and other vendor-specific groups, according to StorageReview. Vendors sponsor the event, but practitioners set the tone, and every session I sat in assumed the audience had a production estate in their charge.
Most conference takeaways I bring home are product announcements, and this time they were registry settings, CPU microarchitecture, and disk IOPS. The breakout rooms filled up, and the expo floor stayed quiet for long stretches because people chose sessions over booths. Each afternoon closed with “UnConference” sessions, where attendees propose topics on the spot and vote on which ones get a room, as Daniel Keer explains in his recap. The format rewards whoever has the most practical problem to discuss, and the keynote set up the problem that ran through the rest of the week. This unorthodox approach was, in a nutshell, pretty cool.
Scott Manchester opened the conference by arguing that a cloud desktop migration is an operating model rather than a milestone with a finish line. According to StorageReview, Manchester spent 25 years at Microsoft leading the engineering teams behind Remote Desktop Services, Azure Virtual Desktop, and Windows 365 before joining Nerdio as Chief Product and Technology Officer. That history gave weight to one detail he shared about how closely the two Microsoft platforms are related. As Keer's recap notes, roughly 80 percent of the Windows 365 codebase is Azure Virtual Desktop code, so most lessons learned on AVD carry over directly to Cloud PCs. His central warning was about what happens when teams treat the move to either platform as a project to complete rather than an ecosystem to nurture.
Manchester argued that treating mixed estates of laptops, Cloud PCs, hybrid hypervisors, and GPU tiers as a static lift-and-shift leads straight to configuration drift and administrative fatigue. His remedy was to replace hand-curated golden images with automated image pipelines and to consolidate management where possible. That framing gave the rest of the week a useful lens, because the best sessions kept finding performance problems that only continuous operation can solve for.
Two identical cloud desktops can perform differently because a cloud VM size guarantees a vCPU count and a memory allocation, but not a processor generation. Kees Baggerman, Senior Technical Director at Nutanix, calls this the cloud desktop CPU lottery, and his session was a highlight of my week. Microsoft documents several Intel generations behind some of its most common VM families, and the platform assigns whichever host is available at the same price for every draw. And It works much like reserving a rental car in the "midsize or similar" class, where every customer pays the same rate but one driver gets this year's model and another gets a 12-year-old economy car with 150,000 miles on it.
The part that changed my thinking was how the deployment model decides when a user redraws. On pooled desktops, a user can land on a different host, and a different processor, every time they fully sign out, and the hosts themselves redraw whenever they're deallocated and restarted, for example by autoscale. Dedicated desktops draw once at build and keep that silicon for years.
As Baggerman explains, a user who locks the screen instead of signing out stays on the same host all day, so a help desk that opens a fresh session to reproduce the complaint finds nothing wrong. That detail can explain why so many of these tickets close as "unable to reproduce”.
Baggerman's practical starting point is a logon script that records the processor model behind every session, so the spread becomes visible before anyone argues about remediation. His full post walks through the probability modeling, the deployment-by-deployment behavior, and the remedies for Windows 365, AVD, AWS, and Google Cloud, and I recommend reading it in full rather than relying on my summary.
The second session that stayed with me looked at a different layer of the same desktop.
In independent testing presented by Yury Magalif of Cloud Zebra, a full endpoint security stack more than tripled average CPU use on an identically sized virtual desktop. Average CPU rose from 8.2 percent to 27.9 percent, average memory rose from 38.4 percent to 53.5 percent, and Excel launch time went from 0.88 seconds to 7.33 seconds.
Magalif ran one user for one hour across 14 runs on Omnissa Horizon 2606 over vSphere 8.0.3, using 4 vCPUs and 16 GB of memory on Windows 11 25H2. He left the image unoptimized on purpose so the security tooling's effect would stand out, and he reported averages rather than peaks, noting that peaks ran much higher. The most surprising result was that the effects were not additive, since some combinations of products used less CPU than a single product running alone. A datasheet therefore cannot predict the cost, and the only reliable number comes from measuring the exact stack and image a team runs in production.
The DEX panel on the second morning pointed at the same blind spot from a different angle. Panelists from Citrix, Recast Software, Omnissa, and Liquidware described telemetry turning vague complaints into root causes, including an endpoint security agent consuming half of a client's threads, according to StorageReview's coverage of the panel. The security team picks the tooling, the EUC team sizes the desktops, and the user calls the service desk, so nobody owns the interaction between the three. When users complain about slow desktops, a security stack measurement belongs on the checklist before anyone orders a larger VM.
Daniel Keer's session on Entra Conditional Access for travelers is the most immediately reusable material from the week. As StorageReview summarized the session, standard geo-blocking breaks the moment an employee signs in from an airport or an overseas hotel, and the usual workaround of a temporary manual bypass rule tends to linger long after the trip and leave a security gap behind. Keer showed how Entra Access Packages let users request a time-bound travel window with manager approval and an automatic expiry, paired with compensating Conditional Access rules. He has published his slides and supporting material on a public landing page, which makes the session easy to revisit.
The Empower Panel took on systems engineering burnout, a subject most conferences leave off the agenda entirely. Owen Reynolds covered related ground from the management side, with practical advice on delegating work and letting go of control after moving from individual contributor to IT leader, according to Keer's recap. On the UnConference track, Stephen Wagner and Jaymes Davis compared approaches to running language models on private infrastructure, including how to index runbooks and documentation without that data leaving the building. Taken together with the headline sessions, these conversations pointed at one shared pattern.
The best sessions at Amplify described performance and security problems that live between platforms, where no single purchase can fix them. Manchester traced configuration drift to estates treated as finished migrations. Keer traced it to temporary exceptions that nobody removes. Baggerman showed that customers cannot specify the silicon under a cloud desktop, and Magalif showed that security tooling can degrade the desktop it protects in ways no datasheet predicts. Each of these problems sits in a seam between teams or tools.
Before buying another tool to fix a performance complaint, record the CPU model behind each session, measure the security stack against the production image, and check whether the users who complain share a host generation. Those checks turn a vague complaint into evidence.
I would love to compare notes with anyone working through the same problems, whether that means a CPU spread you just discovered, a security stack you suspect is slowing your desktops, or a session from Amplify you want to dig into further. Grab time on my calendar and bring your questions, because the best conversations from Milwaukee happened in exactly that kind of informal setting.