August 2026 Summaries

DigitalOcean has introduced v5 Droplets, a new compute generation powered by 5th Generation AMD EPYC processors that provides up to 30% higher per-core performance than prior Droplets for workloads including AI platforms, machine learning, media transcoding, data pipelines, game servers, and high-traffic web applications. The new offering allows customers to independently configure and pay for vCPUs, memory, and storage, while retaining traditional bundled options; it includes shared s5 Droplets for variable workloads and dedicated general-purpose g5 Droplets with memory ratios from 2x to 8x per vCPU. DigitalOcean states that hardware generations will now provide consistent underlying silicon and performance, and early customers reported up to twice the performance for certain workloads. Existing Droplet plans, prices, bundles, and billing arrangements remain unchanged, while v5 Droplets are initially available in Memphis, Kansas City, Richmond, and Atlanta, including for Kubernetes node pools, with expansion to further regions planned.

DigitalOcean has introduced Managed Agents Runtime Services (M.A.R.S.) in invite-only Private Preview, a managed platform intended to run coding agents and long-running, multi-tool AI workflows independently of developers’ local machines. M.A.R.S. combines Harness Runtime, which provides persistent, scalable agent sessions in isolated Firecracker microVMs, with Action Gateway, which centrally governs agents’ access to external tools and services such as GitHub, Jira, Notion, Linear, and Postgres. The platform supports environment templates that let users retain their preferred agent harnesses, dependencies, and configurations, including Claude Code, Codex CLI, OpenCode, LangGraph, CrewAI, and custom environments. Its features include rapid session startup and resumption, cross-device and team handoffs, parallel agent execution, human approval controls, native GitHub operations, centrally managed authentication, and audit policies, while OAuth credentials remain protected through Secrets Manager rather than being exposed to models or runtime environments.

DigitalOcean describes its responses to two serious cloud infrastructure vulnerabilities in July and August 2026: Januscape, a KVM nested-virtualization flaw that could enable guest-to-host escapes, and an AMD Safe RET issue involving speculative-execution protections that could permit information disclosure across privilege boundaries. For Januscape, the company developed and deployed live patches across most of its hypervisor fleet, while handling several hundred older-kernel hosts through workload evacuation, host upgrades, capacity reallocation, and a separately hardened custom patch, reaching full coverage eight days after disclosure. For the AMD issue, which could not be livepatched and required reboots on roughly 1,600 hypervisors in 12 regions, DigitalOcean used coordinated disclosure time to test kernels, stage updates, migrate workloads, reboot hosts, and complete remediation before the vulnerability became public. The company attributes zero confirmed customer-facing impact to cross-team coordination among kernel, security, capacity, operations, support, and fleet-management teams, along with staged rollouts, continuous monitoring, detection engineering, and reusable automation. It concludes that increasingly rapid exploit development requires even faster fleet-wide remediation capabilities and says it is reviewing both incidents to reduce response times further.

DigitalOcean has added cache-aware routing to its Inference Router, aiming to reduce the cost and latency of multi-turn AI agent workloads by accounting for the value of warm prompt caches when selecting models. The company argues that switching to a nominally cheaper model can be more expensive and slower if it requires reprocessing large amounts of previously cached context, such as system instructions, tools, repository data, and conversation history. Developers can preserve model bindings through an X-Model-Affinity header, rely on automatically inferred session affinity, or set an X-Routing-Max-Switch-Spend-Pct policy that limits the extra cost incurred when a router changes models. The update also expands the Analyze page with cache-efficiency, switching, latency, model, task, and trend information to support routing optimization. DigitalOcean positions cache-aware routing alongside its existing preference-aware routing, custom model pools, and task definitions, emphasizing that effective AI cost management depends on balancing model quality, latency, developer priorities, and cache reuse rather than relying solely on benchmark rankings, usage caps, or per-token prices.