EuroSys has always been a good place to see where real systems pressure is building.

The 2026 edition is especially revealing. The accepted-paper list shows a community that is no longer just building generic distributed systems abstractions. It is increasingly shaped by AI-scale workloads, accelerators, network bottlenecks, cloud efficiency, and production-grade reliability constraints.

This report synthesizes the EuroSys 2026 accepted papers into a high-level map of the field:

  1. the key areas covered by the conference,
  2. the most popular areas,
  3. the major trends visible across the program,
  4. and the follow-up deep dives worth turning into a full post series.

Methodology and scope

This report is grounded in the EuroSys 2026 accepted papers list and the linked proceedings entry:

A note on scope: at the time of analysis, the ACM proceedings pages were challenge-gated for lightweight automated fetching, so this synthesis is primarily a titles-led analysis of the full accepted-paper set rather than a full abstract-by-abstract review. That means the report is strongest on conference-wide structure, topic density, and directional trends, and intentionally more cautious on paper-specific claims that would require full abstract access.

The corpus contains 138 accepted papers. To estimate topic popularity, I grouped papers into primary themes using their titles and system focus. These buckets are heuristic rather than official conference tracks, but they are still very useful for seeing the shape of the program.

1. The key areas covered in EuroSys 2026

1) AI/LLM systems has become a central systems topic

The most visible change is that AI infrastructure is no longer a side theme. It is one of the main bodies of systems work.

This shows up across multiple layers:

  • LLM inference serving
  • MoE training and memory management
  • speculative decoding and token scheduling
  • parameter-efficient tuning and LoRA execution
  • end-cloud and on-device deployment
  • GPU-aware scheduling and utilization
  • power, cost, and efficiency for AI workloads

Representative papers include:

  • AdaServe: Accelerating Multi-SLO LLM Serving with SLO-Customized Speculative Decoding
  • TokenFlow: Responsive LLM Text Streaming Serving under Request Burst via Preemptive Scheduling
  • MegaScale-MoE: Large-Scale Communication-Efficient Training of Mixture-of-Experts Models in Production
  • SkyWalker: A Locality-Aware Cross-Region Load Balancer for LLM Inference
  • KUNSERVE: Parameter-centric Memory Management for Efficient Memory Overloading Handling in LLM Serving
  • TZ-LLM: Protecting On-Device Large Language Models with Arm TrustZone

The notable thing is not just the number of AI papers. It is that the papers are deeply systems-flavored: they care about memory hierarchy, scheduling, reliability, network behavior, resource fragmentation, and production economics.

2) Networking remains foundational, but is now shaped by AI traffic patterns

Networking is still one of the deepest veins in EuroSys, but the flavor has shifted.

Classic themes remain present:

  • congestion control
  • routing and packet spraying
  • traffic pattern detection
  • programmable networks
  • RDMA and collective communication
  • load balancing and cross-region transport

But several of these papers are clearly motivated by accelerator clusters, GPU clouds, and AI-era communication patterns.

Representative papers include:

  • Learn-to-Probe: Achieving Signal Distinguishability in Learning-based Congestion Control
  • REPS: Recycled Entropy Packet Spraying for Adaptive Load Balancing and Failure Mitigation
  • Multipath Collective Communication Beyond Scale-up Networks in GPU Clouds
  • Rearchitecting Programmable Networks For In-Network Computing: From Hardware To Language
  • EMVOD: Elastic Multi-Path QUIC Scheduling for CDN Video-on-Demand Service

The message is clear: networks are increasingly being redesigned around modern large-scale model training and serving workloads, not just traditional web traffic.

3) Cloud runtime and serverless systems are maturing into fine-grained infrastructure work

Serverless is no longer treated as a novelty. The 2026 papers suggest it has entered a more mature phase focused on hard operational details:

  • resource pool management
  • workflow routing
  • data passing efficiency
  • VM memory reclamation
  • cost transparency
  • scaling strategies in production platforms
  • multi-cloud replication

Representative papers include:

  • iRoute: Local Routing Table-based Workflow Management in Serverless Computing
  • DROPS: Managing Serverless Resource Pools in Microsoft Azure Functions
  • Squeezy: Rapid VM Memory Reclamation for Serverless Functions
  • Efficient Data Passing for Serverless Inference Workflows: A GPU-Centric Approach
  • Demystifying Serverless Costs on Public Platforms: Bridging Billing, Architecture, and OS Scheduling

This is a signal that cloud abstraction is still advancing, but the research frontier has moved closer to the metal.

4) Storage, memory, and data-path efficiency still matter a lot

EuroSys 2026 also has a strong substrate layer: memory systems, filesystems, flash, persistent memory, metadata services, and storage scheduling.

Representative concerns include:

  • total cost of ownership for memory tiers
  • block storage scheduling
  • persistent memory transactions
  • distributed filesystem metadata updates
  • flash lifetime and coding efficiency
  • compression for memory and data movement

Representative papers include:

  • TierScape: Harnessing Multiple Compressed Tiers to Tame Server Memory TCO
  • Scheduling Cloud Block Storage Proactively and Reactively with Omar
  • FUR: Fast and Unlimited Reads on Persistent Memory Transactions
  • SwitchFS: Asynchronous Metadata Updates for Distributed Filesystems with In-Network Coordination
  • ColdCode: Cold Data Encoding for Enhanced Reliability and Lifetime in 3D NAND Flash

These papers matter because AI infrastructure does not replace classical systems bottlenecks. It amplifies them.

5) Security, isolation, and verification are being pulled closer to systems design

A striking pattern in this year’s program is how often security and reliability appear as first-class systems concerns rather than separate specialist topics.

We see work on:

  • TrustZone-protected on-device LLMs
  • secure multi-tenancy in Kubernetes and containers
  • fuzzing embedded OSes and nested virtualization
  • privacy control for IoT systems
  • blockchain consensus robustness
  • failure detection and fault propagation analysis
  • integrity and attestation in multi-cloud AI settings

Representative papers include:

  • Pyramid: A Secure, Resource-Efficient, and Pluggable Kubernetes for Multi-Tenancy
  • SKernel: An Elastic and Efficient Secure Container System at Scale with a Split-Kernel Architecture
  • Effective On-Hardware Fuzzing of Embedded Operating Systems
  • NecoFuzz: Effective Fuzzing of Nested Virtualization via Fuzz-Harness Virtual Machines
  • Turnstile: Hybrid Information Flow Control Framework for Managing Privacy in Internet-of-Things Applications
  • TrustWeave: Integrity Measurement and Attestation For Multi-Cloud LLMs

The center of gravity has shifted from “add security later” to “design for secure operation from the start.”

6) Heterogeneous hardware and systems architecture are becoming normal assumptions

Another strong theme is the normalization of heterogeneity.

This includes:

  • GPUs and SmartNICs
  • DPU and FPGA architectures
  • Arm TrustZone and RISC-V
  • chiplet-aware runtime mapping
  • binary translation and ISA heterogeneity
  • OS specialization for specific workload classes

Representative papers include:

  • Chimera: Transparent and High-Performance ISAX Heterogeneous Computing via Binary Rewriting
  • LightDSA: Enabling Efficient DSA Through Hardware-Aware Transparent Optimization
  • CHARM: Chiplet Heterogeneity-Aware Runtime Mapping System
  • Practical and Efficient x86-64 Emulation on RISC-V
  • NutCracker: A Compilation Framework for Hybrid DPU Architectures
  • Wayfinder: Automated Operating System Specialization

The old assumption of a mostly homogeneous server fleet keeps fading.

Using a primary-topic grouping over the 138 accepted papers, the program roughly breaks down like this:

  • AI/LLM systems: 41 papers (~30%)
  • Networking, transport, and traffic systems: 28 papers (~20%)
  • Security, isolation, and verification: 15 papers (~11%)
  • Systems architecture, OS, and compilers: 14 papers (~10%)
  • Cloud and serverless runtime: 12 papers (~9%)
  • Storage, memory, and filesystems: 12 papers (~9%)
  • Data, consensus, and specialized domains: 5 papers (~4%)
  • remaining papers are cross-cutting or hard to place cleanly in one bucket

This is the headline.

Roughly a third of the accepted papers are directly tied to AI or LLM infrastructure when grouped by primary theme. That does not mean EuroSys has become an AI conference. It means AI has become one of the strongest demand generators for systems innovation.

This is an important distinction.

The dominant pattern is not “AI applications” but systems for AI:

  • serving under latency SLOs,
  • training at large scale,
  • managing fragmented clusters,
  • handling communication bottlenecks,
  • reducing GPU memory pressure,
  • supporting edge and device deployment,
  • securing inference and multi-cloud execution.

Networking remains the second major pillar. That is not surprising historically, but the context is changing. Increasingly, the networking papers intersect with:

  • GPU cloud communication,
  • AI training collectives,
  • multi-path scheduling,
  • traffic adaptation under high-throughput workloads,
  • and reliability in large-scale distributed environments.

In other words, AI is not replacing traditional systems topics. It is reshaping them.

The middle of the program is broad, not narrow

After the top two areas, EuroSys 2026 remains nicely diversified.

The program still includes substantial work in:

  • cloud platforms,
  • storage and memory,
  • operating systems,
  • compiler/runtime co-design,
  • virtualization,
  • verification,
  • consensus and blockchain,
  • specialized scientific and edge workloads.

That breadth is healthy. It suggests the field is evolving, not collapsing into a single trend.

Trend 1: AI is now a systems stress test, not just a workload category

The strongest conference-wide signal is that modern AI workloads are acting like a forcing function across the stack.

You can see this in papers about:

  • token scheduling,
  • speculative decoding,
  • MoE communication,
  • GPU power and utilization,
  • memory offloading,
  • cluster scheduling,
  • cross-region inference,
  • serverless inference workflows,
  • training fault tolerance.

The implication is simple: if a systems idea does not survive AI-scale pressure, it may no longer be enough.

Trend 2: GPU-era systems research is becoming end-to-end

A few years ago, a lot of accelerator work was localized to one layer: kernels, frameworks, or cluster scheduling.

In EuroSys 2026, the interesting work is increasingly end-to-end:

  • model architecture interacts with systems scheduling,
  • network topology interacts with training strategy,
  • memory layout interacts with serving latency,
  • compiler/runtime techniques interact with cloud resource fragmentation,
  • power and TCO are treated as first-class constraints.

That is a sign of a maturing field. The easy single-layer optimizations are being exhausted.

Trend 3: Efficiency is no longer just about speed; it is about economics

A lot of titles point toward TCO, power, cost visibility, compressed tiers, scheduling efficiency, and resource overloading.

That matters.

Systems research in 2026 is not just asking, “Can we make it faster?” It is also asking:

  • Can we make it cheaper?
  • Can we make it denser?
  • Can we make it power-aware?
  • Can we make multi-tenant infrastructure economically viable?
  • Can we keep accelerator-heavy systems from becoming cost black holes?

This feels like one of the clearest post-hype signals in the conference.

Trend 4: Reliability is moving from reactive fault handling to built-in resilience

Papers on failover for distributed AI training, fault propagation, change-risk detection, consensus roles, and production cloud reliability point in the same direction.

The shift is from:

  • detect failure after the fact

to:

  • build systems that continue to function under likely failure modes

That is especially important for AI infrastructure, where retrying is expensive and recovering large in-flight workloads is painful.

Trend 5: Security is becoming infrastructure-native

The papers on secure containers, secure Kubernetes, TrustZone-protected on-device LLMs, attestation for multi-cloud LLMs, and virtualization hardening all point to the same story:

security is getting embedded deeper into the systems substrate.

That is probably a necessary response to three things happening at once:

  • more sensitive data flows into AI systems,
  • more inference happens across edge, cloud, and multi-cloud boundaries,
  • and more tenants share expensive accelerator infrastructure.

Trend 6: Heterogeneity is winning

The architecture papers make it hard to believe in a “one machine model” future.

Modern systems have to handle combinations of:

  • CPUs,
  • GPUs,
  • SmartNICs,
  • DPUs,
  • FPGAs,
  • chiplets,
  • diverse ISAs,
  • and cloud environments with uneven hardware pools.

A lot of systems work is therefore becoming translation work: how to preserve portability, performance, isolation, and correctness across messy hardware realities.

Trend 7: Serverless is entering its realism phase

The serverless papers are less about idealized abstraction and more about operational mechanics:

  • billing behavior,
  • scaling strategies,
  • memory reclamation,
  • workflow routing,
  • resource pools,
  • and efficient data paths for inference.

That is exactly what you would expect from a platform model that has survived first contact with production.

4. What EuroSys 2026 suggests about the field overall

If I had to summarize the conference in one sentence, it would be this:

systems research is being reorganized around the practical consequences of AI scale, heterogeneous infrastructure, and production economics.

That does not mean everything is suddenly about LLMs.

It means the entire systems stack is being re-pressurized by new workload realities:

  • AI training and serving are expensive,
  • accelerators are hard to utilize well,
  • networks become bottlenecks sooner,
  • memory hierarchies matter more,
  • cross-region and multi-cloud placement gets trickier,
  • security and isolation get harder,
  • and reliability failures become costlier.

EuroSys 2026 looks like a conference where the community is trying to answer one big question:

How do we make the next generation of infrastructure not just possible, but operable?

5. The follow-up deep-dive plan

This report should really be the opening map, not the last word. The accepted-paper list points to a strong follow-up series.

Here is the post plan I would recommend.

Deep Dive 1 — Why AI infrastructure dominated EuroSys 2026

Focus:

  • why AI/LLM systems became the largest theme,
  • what kinds of systems problems AI is creating,
  • why systems researchers are now tackling inference and training directly.

Candidate paper cluster:

  • AdaServe
  • TokenFlow
  • MegaScale-MoE
  • SkyWalker
  • KUNSERVE
  • MFS

Deep Dive 2 — The new bottleneck is not compute, it is coordination

Focus:

  • communication, collective ops, routing, multipath, packet spraying,
  • why distributed AI shifts the bottleneck toward transport and topology,
  • what network papers reveal about future accelerator clusters.

Candidate paper cluster:

  • REPS
  • Multipath Collective Communication Beyond Scale-up Networks in GPU Clouds
  • Learn-to-Probe
  • EMVOD
  • Rearchitecting Programmable Networks For In-Network Computing

Deep Dive 3 — Serving LLMs is becoming a full-stack discipline

Focus:

  • latency SLOs,
  • fragmented clusters,
  • expert offloading,
  • memory overloading,
  • cross-region balancing,
  • cost and power-aware serving.

Candidate paper cluster:

  • AdaServe
  • FlexPipe
  • Taming Latency-Memory Trade-Off in MoE-Based LLM Serving
  • KUNSERVE
  • SkyWalker
  • Untangling GPU Power Consumption

Deep Dive 4 — Serverless after the hype

Focus:

  • what mature serverless research now cares about,
  • why routing, memory reclamation, resource pools, and billing transparency matter,
  • how serverless is adapting to AI inference workflows.

Candidate paper cluster:

  • iRoute
  • DROPS
  • Squeezy
  • Efficient Data Passing for Serverless Inference Workflows
  • Demystifying Serverless Costs on Public Platforms

Deep Dive 5 — Reliability engineering for AI-scale distributed systems

Focus:

  • failover in AI training,
  • cascading failure analysis,
  • proactive change-risk detection,
  • counterfactual evaluation and state-machine resilience,
  • what “reliable AI infrastructure” actually requires.

Candidate paper cluster:

  • Handling Network Faults in Distributed AI Training: Failover is Now an Option
  • CSnake
  • Proactive Change Risk Detection in Production Cloud Systems
  • Avicenna
  • Lessons Learned from Incorporating Formal Methods in Huawei Cloud Reliability

Deep Dive 6 — Secure-by-design infrastructure is no longer optional

Focus:

  • containers, Kubernetes, TrustZone, fuzzing, attestation, privacy,
  • why AI and multi-tenancy are pushing security deeper into systems architecture.

Candidate paper cluster:

  • Pyramid
  • SKernel
  • TZ-LLM
  • TrustWeave
  • NecoFuzz
  • Effective On-Hardware Fuzzing of Embedded Operating Systems

Deep Dive 7 — Heterogeneous computing is the default future

Focus:

  • chiplets, DPUs, FPGAs, mixed ISAs, binary rewriting, specialized runtimes,
  • how systems software is being rebuilt for messy hardware fleets.

Candidate paper cluster:

  • Chimera
  • CHARM
  • NutCracker
  • Practical and Efficient x86-64 Emulation on RISC-V
  • Proteus
  • Wayfinder

Deep Dive 8 — The economics of systems research: TCO, energy, and density

Focus:

  • compressed memory tiers,
  • power-adaptive storage,
  • cost-aware serverless,
  • GPU power analysis,
  • carbon-aware learning,
  • the shift from peak performance to operational efficiency.

Candidate paper cluster:

  • TierScape
  • PASS
  • Untangling GPU Power Consumption
  • Carbon-Aware Continuous Learning for Sustainable Real-Time Machine Learning Analytics
  • Demystifying Serverless Costs on Public Platforms

6. Bottom line

EuroSys 2026 does not look like a conference that has abandoned core systems topics.

It looks like a conference where core systems topics are being reinterpreted through the pressures of:

  • AI-scale inference and training,
  • accelerator-rich and heterogeneous hardware,
  • cloud runtime realism,
  • cost and power constraints,
  • and infrastructure-native security and reliability.

The hottest area is clearly AI/LLM systems, but the deeper story is broader than that.

The program suggests that the next chapter of systems research will be defined by the ability to integrate:

  • performance,
  • economics,
  • security,
  • operability,
  • and heterogeneity

into one coherent infrastructure story.

That is why EuroSys 2026 is worth paying attention to.

It is not just showing what systems researchers are building.

It is showing what modern infrastructure is demanding.

References