External Publication

OpenStack Gazpacho is a dish best served cold for hot cloud networks

Network World [Unofficial] April 1, 2026

OpenStack has been running production cloud infrastructure for 15 years, and its 33rd release keeps that record going.

The OpenStack community today released OpenStack 2026.1, code-named Gazpacho, delivering improvements across compute, bare metal, networking and storage focused on operator experience and workload mobility. Gazpacho is the first release since OpenStack 2025.2 Flamingo, which centered on eliminating technical debt and advancing confidential computing support.

Around 500 contributors from 100 organizations delivered 9,000 code changes during the six-month development cycle. Notably, 40% of contributions came from European contributors, a figure tied to growing digital sovereignty initiatives driving OpenStack adoption across the region. Gazpacho is also a SLURP (Skip Level Upgrade Release Process) release, meaning operators running the previous SLURP release, 2025.1 Epoxy, can upgrade directly to 2026.1 without an intermediate step.

The release arrives as VMware migration continues to drive new OpenStack deployments.

“VMware escapees represent currently a lot of the new deployments that are coming to OpenStack right now,” Thierry Carrez, general manager of the OpenInfra Foundation, told Network World.

Nova closes the gap with parallel live migrations

OpenStack as a platform consists of a series of constituent projects. The Nova project is the primary compute technology and gains a series of enhancements in the Gazpacho update.

The compute improvements in Gazpacho are directly tied to what operators migrating from VMware have been requesting. The headline Nova feature is parallel live migrations. The concept of a live migration in VMware is commonly known as vMotion and is often cited as a critical feature by users in their deployments.

Previously, live migration in OpenStack used a single memory transfer connection to copy VM memory from one host to another. The process works by copying the full memory state, then copying incremental deltas until no delta remains and the cutover can complete. That transfer ran as a single thread.

Gazpacho changes the underlying algorithm. Multiple memory transfer connections now run simultaneously, fragmenting the network transfer across parallel threads. “This process of transferring the memory is happening through multiple threads,” Carrez said. “It’s no longer, let’s transfer the whole thing and then transfer the new delta, and then the new new delta.”

Carrez said the performance improvement brings OpenStack’s live migration behavior closer to what operators experienced in VMware environments. The feature has been in high demand from operators deploying OpenStack as a VMware replacement.

The second major Nova addition is live migration support for instances backed by a virtual Trusted Platform Module (vTPM), which store cryptographic secrets used to protect workload data. Moving a VM that relies on a vTPM has historically required separate handling of the secret material, since the secret stored in one instance’s vTPM cannot automatically transfer to a destination host.

Gazpacho addresses this by persisting the TPM secret in Barbican, OpenStack’s key management service, and transferring to the destination host during migration. “It allows the secret to be restored into the next vTPM, and that’s really enabling the secure movement of sensitive workloads,” Carrez said.

Ironic moves toward smarter defaults (no AI required)

The Ironic bare metal service receives several changes focused on reducing the number of decisions an operator must make explicitly at deployment time.

The autodetect deploy interface removes the requirement to specify a deployment method manually. Ironic now examines image metadata and node configuration to select the appropriate deployment approach.

Trait-based port scheduling extends Ironic’s network port scheduling to incorporate physical network attributes stored as metadata. Previously, port scheduling could group by availability zones or regions but could not factor in physical network characteristics. An operator can now express a requirement such as dual-redundant 10Gb connectivity and the scheduler will assign a bare metal node whose ports match those physical attributes.

Carrez described the overall Ironic direction as moving away from requiring operators to configure everything explicitly. “I wouldn’t say AI, because there’s no AI in it, but it’s like, how about you use more smart defaults,” he said.

Neutron extends OVN driver for large-scale networking

The Neutron project is OpenStack’s networking technology, and it benefits from a series of incremental improvements.

Gazpacho’s networking changes center on the OVN (Open Virtual Network) driver in Neutron, extending work begun in the Flamingo release, which added stateless NAT support.

The OVN driver now includes BGP support. Carrez said the addition addresses routing requirements at scale. “It allows you to manipulate BGP routes from the OVN driver, which is pretty interesting to me for large-scale deployments where you have to actually care about BGP,” he said.

The release also adds north-south routing for external ports, covering both SR-IOV ports and ports exposed through PCI passthrough. SR-IOV ports bypass the hypervisor software switching layer by giving VMs direct access to physical NIC functions. The north-south routing addition means traffic to and from those ports no longer needs CPU involvement for processing.

“You can have direct north-south routing so that you don’t go through the CPU to actually process those things,” Carrez said.

Cyborg gains renewed attention as AI accelerator demand grows

Like every other form of technology in 2026, OpenStack is also leaning into AI.

When Carrez discussed the Flamingo release six months ago, he predicted that AI inference workloads would start generating new requirements at the OpenStack layer. Gazpacho shows early movement in that direction through the Cyborg project.

Cyborg manages accelerator attachment in OpenStack, handling GPUs, FPGAs, NPUs and other hardware. The project had been in maintenance mode for some time, seeing limited new development. Gazpacho reflects a renewed investment, with a new driver configuration guide covering all supported accelerator types including FPGA, GPU, NIC, SSD and PCI passthrough.

The next OpenStack release, 2026.2 Hibiscus, is scheduled for September 2026. Project teams will gather later this month to define priorities.

“I hope I’ll see a renewed interest in Cyborg and translating to new features,” Carrez said. “I hope that we will be able to better meet the needs of the people that are coming out of VMware, that we will be able to cover all of the digital sovereignty requirements that Europe is asking for and other countries are asking for.”

OpenStack Gazpacho

OpenInfra Foundation

Nova closes the gap with parallel live migrations

Ironic moves toward smarter defaults (no AI required)

Neutron extends OVN driver for large-scale networking

Cyborg gains renewed attention as AI accelerator demand grows

Discussion in the ATmosphere