GitHub’s platform uptime over the last 90 days is 86.98%. That is not a typo. Marek Šuppa, who built an independent tracker because GitHub stopped publishing aggregate uptime numbers on its own status page, has reconstructed the figure from GitHub’s incident feeds.
IncidentHub counted 257 separate incidents between May 2025 and April 2026, 48 of them major, with February being the worst month on record with 37 incidents. GitHub Actions, the CI/CD pipeline service that most development workflows depend on, had 57 outages in the period.
Three years ago, an Actions outage was a quarterly event. Now they’re happening roughly once a week.
By comparison, GitHub’s official status page reports 99.83% uptime. The gap between that number and the lived experience of anyone who uses the platform daily is the kind of gap that erodes trust slowly, and then all at once.
The obvious explanation is the Azure migration. GitHub began moving its core platform from its Northern Virginia data centres to Azure in October 2025, and a half-migrated system is inherently fragile. GitHub CTO Vlad Fedorov confirmed in March that 12.5% of all GitHub traffic was being served from Azure Central US, with a target of 50% by July. But the incident trend started years before the migration began. Actions went from 2 incidents in 2020, to 59 in 2023. The Azure migration is a compounding stressor. It is not the cause.
The cause is AI.
Fedorov said so himself in April. GitHub had planned for a ten-times capacity growth in October 2025: “By February 2026, it was clear that we needed to design for a future that requires thirty times today’s scale,” driven by “a rapid change in how software is being built” since the second half of December 2025.
GitHub COO Kyle Daigle confirmed the numbers on Twitter: there were 1 billion commits on GitHub in all of 2025. But the platform is now processing 275 million commits per week, on pace for 14 billion this year. GitHub Actions went from 500 million minutes per week in 2023, to 1 billion in 2025, up to 2.1 billion in a single week in early April 2026. AI-opened pull requests grew from roughly 4 million in September 2025, to over 17 million in March 2026.
That’s the same number of developers, running agents that each behave like ten to a hundred of the developers who are instructing them.
A human developer clones a repository, reads it for twenty minutes, opens a few pull requests, and calls it a day. An agent hits the API continuously, runs parallel investigations across repositories, retries aggressively when CI fails, and never logs out. The load profile is not just bigger, it is qualitatively different. And most agent traffic flows through free accounts and open source repositories, meaning the infrastructure cost per “user” has changed by orders of magnitude, while the revenue per user has not.
The most striking sentence in Fedorov’s April post is not about load. It is about architecture: “While we were already in progress of migrating out of our smaller custom data centers into public cloud, we started working on path to multi cloud.” Here’s a Microsoft-owned subsidiary, publicly admitting it cannot get the reliability and elasticity it needs from Azure alone. GitHub has not named a second cloud. The admission is enough.
Mitchell Hashimoto, the founder of HashiCorp, publicly quit GitHub in April. “This is no longer a place for serious work if it just blocks you out for hours per day, every day.” The Zig language project migrated to Codeberg in December 2025, citing a “rotted” engineering culture. These are coal-mine canaries, not a flock. Network effects are extraordinarily sticky, and most teams will tolerate outages because the cost of migration is higher than the cost of the outages. For now.
There is an obvious feedback loop going on here, no less vicious for being obvious. GitHub built Copilot. Copilot drove agent adoption. Agent adoption broke GitHub.
GitHub Copilot has over 20 million cumulative users, 4.7 million paid subscribers, and is deployed at 90% of Fortune 100 companies. At GitHub Universe in October 2025, the company launched Agent HQ, explicitly designed to bring coding agents from Anthropic, OpenAI, Google, and others into the GitHub workflow as part of the paid Copilot subscription. GitHub is literally, and aggressively, shipping the accelerant for the fire it is fighting.
The clearest signal that GitHub’s own leadership sees the architecture as mismatched? In August 2025, CEO Thomas Dohmke stepped down. Rather than being replaced, GitHub was folded into Microsoft’s CoreAI organisation. Dohmke then went on to launch Entire in February, with a $60 million seed round at a $300 million valuation. That’s the largest seed round in developer tools history. His stated rationale: “Our manual system of software production — from issues, to git repositories, to pull requests, to deployment — was never designed for the era of AI in the first place.“
The man who ran GitHub for four years just concluded that GitHub itself is the wrong architecture for what comes next.
GitLab Act 2
GitLab is the interesting counterpoint, because it is the first major platform to publicly admit the architectural problem and restructure around it rather than serially patching capacity.
Earlier in the month, CEO Bill Staples published “GitLab Act 2,” announcing layoffs amounting to roughly 7% of the workforce and a restructuring that flattened management from eight layers to five, alongside an R&D reorganisation into roughly 60 autonomous teams, alongside an exit from up to 30% of the countries GitLab currently operates in. The market reacted badly. GTLB shares fell 8.2% in after-hours trading, with Raymond James downgrading the stock to Market Perform, and Bank of America to Neutral.
But the layoffs are the least interesting part of the announcement. The interesting part is the architectural thesis underneath it.
Staples’ letter lays out five bets. Git itself is being rebuilt for machine-scale workloads. The monolith is giving way to API-first composable services. Agent-specific APIs are being built so agents are first-class users of the platform rather than consumers of human-shaped interfaces. CI/CD is being reimagined as an orchestration runtime that coordinates agents, validates their work, and drives changes to production at machine rate. The full lifecycle data model – years of planning, code review, security scanning, and deployment data – is being exposed as an API-accessible context store.
Staples was careful to frame the restructure as reinvestment rather than cost-cutting: “This restructure process is not like others you may be seeing in the news. We intend to reinvest the vast majority of savings back into the business to accelerate our unique opportunity in the agentic era.” Whether you believe that framing depends on whether you believe the five-bet architecture will ship. GitLab has committed to rebuilding Git itself, along with a connected data model service, and an orchestration runtime. These are massive engineering efforts. The letter acknowledges as much.
Staples summarises, “Bolting AI onto platforms not built for agents is the biggest mistake of this era.“
That is the GitLab CEO agreeing with GitHub’s strategy without calling it out. You have two CEOs of the two biggest developer platforms publicly agreeing that human-rate infrastructure is breaking under agent load. They just disagree about whether to retrofit or rebuild.
What GitLab conspicuously has not done is to weaponise GitHub’s outage record directly. But back in February, just nine days after GitHub’s worst cascade, GitLab launched a 99.9% availability SLA backed by service credits for Ultimate customers. The post does not mention GitHub, choosing architecture-narrative over reliability-narrative, which is either a sign that enterprise gravitational lock-in to GitHub is so strong that direct attacks won’t move accounts, or that GitLab’s leadership is betting the structural moat will be more durable than a few months of GitHub instability.
Broken pricing models
In April, GitHub announced that at the start of June all paid individual Copilot plans will move to usage-based billing with base credits matched to subscription price, a variable flex allotment on top, and a new Max plan at $100 per month for heavy agent use. This is GitHub conceding that per-seat pricing cannot survive when one seat can host an autonomous agent burning through model calls and Actions minutes at machine-speed.
GitLab is already further along the same path. The Duo Agent Platform, which became generally available in January, uses group-level credits rather than per-seat allocations. Premium customers get 12 credits per user per month. Ultimate customers get 24. Overages cost $1 per credit. Different models burn credits at different rates. Staples has said explicitly that the medium-to-long-term answer is a shift from pure seat-based pricing to hybrid seat- plus usage-based pricing.
Per-seat SaaS pricing was built on an assumption that each seat represents something that looks like equivalent resource consumption. A human developer commits a few times a day, runs a handful of CI pipelines, and occasionally browses the issue tracker. When that seat is occupied by an agent that opens a few thousand pull requests a month, runs parallel investigations across hundreds more repositories, and triggers pipeline after pipeline in retry loops, the assumption collapses. The marginal cost of a “user” has gone from near-zero to highly variable, and frequently negative on a fixed-price plan.
This is not just a GitHub problem. Atlassian’s CEO Mike Cannon-Brookes announced 1,600 jobs cuts in March, surrounding the statement with language about self-funding AI investment. Jira’s 30-day uptime dropped to roughly 94% after a major multi-product outage earlier in the month, which they attributed to their cloud provider. The pattern here is the same: platforms that don’t own their cloud and that carry large legacy state are the most exposed when the load curve goes vertical.
Where do we go from here?
GitHub does not own its infrastructure. It sits on Azure. It is in the middle of migrating from owned colocation to Microsoft’s cloud, but it is now publicly admitting that even its parent’s cloud is not enough. When the load curve goes vertical, any link in that chain that is not vertically integrated becomes a bottleneck, and it is the software layer that suffers the visible outages, even though the root constraint is physical.
GitLab does not own its infrastructure either, but it is at least redesigning the software to handle the load shape rather than just throwing capacity at the old architecture. Whether that bet pays off depends on execution; their June earnings call will be the first real signal.
The broader pattern is Jevons paradox at play in the infrastructure layer. AI agents make a unit of software production radically cheaper. Total production volume explodes. And the platforms charging by user count cannot capture the value of the explosion. The platforms that own infrastructure, compute, networking, and identity can charge for what is actually being consumed. The platforms that own only software get the demand curve, but not the supply curve.
The infrastructure was not built for this. GitHub’s CTO knows it. GitLab’s CEO knows it. The man who used to run GitHub knows it well enough that he quit to build something else. The question is not whether the architecture needs to change. It is whether the incumbents can change it fast enough, or whether the Thomas Dohmkes of the world will build it from scratch.