Read that sentence again. That's not a typo and it's not a hypothetical.
Sometime around June 2026, Microsoft quietly contracted AWS — Amazon's cloud, the thing Azure exists to compete with — for extra capacity to keep GitHub running. Business Insider broke it first, citing two people familiar with the arrangement. Microsoft didn't deny it. A spokesperson confirmed the deal on the record and called it "a multi-cloud strategy." Which is true, technically, the way "we had to borrow the neighbor's generator" is technically a home-energy strategy.
Here's what actually pushed things there.
The number nobody was modeling for
In September 2025, AI coding agents were opening roughly 4 million pull requests a month on GitHub. By March 2026, that number had passed 17 million. GitHub COO Kyle Daigle confirmed, on the record, that the platform was processing 275 million commits a week — a pace that puts 2026 on track for 14 billion commits total, up from about 1 billion for all of 2025.
Say that plainly: commit volume went up roughly 14x in a single year, and almost all of the new volume is agents, not people.
Back in October 2025, GitHub planned for this. The internal target was 10x capacity growth. By early 2026, that number was already wrong — GitHub quietly revised it to 30x. Three times its own projection, inside a few months. That's not a company that got the trend wrong. That's a company that got the trend right and still underestimated the slope.
What breaking actually looked like
GitHub publishes its own monthly availability numbers. They didn't need Business Insider for this part — it's in their own blog:
• January 2026: rough month, repeated incidents.
• April 2026: 78.33% availability. Ten separate incidents, including a 2.7-hour Copilot outage and 8.7 hours of code-search downtime.
• May 2026: recovers to 93.86% — looks like the fix is holding.
• June 2026: slips back to 88.39%. Six more incidents, including a Copilot code-review failure rate that peaked at 93.9% (nineteen out of twenty requests failing isn't degraded service, it's an outage with better PR), and a signup failure that hit 62% of new users trying to join the platform.
May's recovery is the part worth sitting with for a second, because it makes June look less like bad luck and more like a pattern GitHub hasn't actually solved — just intermittently outrun.
The Register's framing on June 12 was blunt: outages persist as AI coding drives traffic surge. Not "outages happened." Persist. Ongoing tense.
Then the AWS story surfaces
This is where it stops being just a reliability story and starts being a genuinely strange one.
Microsoft was already mid-migration, moving GitHub off legacy infrastructure onto its own Azure cloud — targeting roughly half of GitHub's traffic on Azure Central US by July 2026. Running a full platform migration and an unmodeled 14x traffic surge at the same time is already a hard problem. And somewhere in the middle of solving it, Microsoft picked up the phone to Amazon.
Not as a footnote. As capacity. GitHub — the product Microsoft paid $7.5 billion for in 2018, the flagship of its developer ecosystem — needed more room than Azure could give it fast enough, and the fastest room available belonged to its biggest cloud rival.
Microsoft's own statement, on the record: the arrangement responds to "the incredible spike in agentic development" that "tested our infrastructure's limits," and the company is "both accelerating our move to Azure and continuing to explore a multi-cloud strategy."
Read that twice. The company confirmed the strain. It confirmed the AWS deal. It just declined to call it what a lot of engineers reading GitHub's own outage reports would call it: an infrastructure team caught flat-footed by its own product's success.
The part that actually resolves the story
Here's the detail that made this whole thing click for me, and it's the one detail I haven't seen tied back to the outage numbers anywhere else.
On July 8, 2026 — less than a month after the AWS story broke — Thomas Dohmke, GitHub's own former CEO, launched a company called Entire. It's a distributed Git network. The pitch: instead of every AI agent and developer on Earth hammering one central GitHub server, repositories get mirrored across regions, and agents clone from whichever mirror is closest instead of queuing up behind everyone else's traffic.
Dohmke's own words: "In the era of agents, centralized Git hosting has become a fundamental constraint... the strain of billions of agents and developers hammering a central server shows up in the form of rate limits, high latency, or even outages."
He's not guessing at this. He ran the company that's been publishing those exact outage numbers for the last nine months. He left, and the thing he built next is a direct architectural answer to the specific failure mode his old company has been experiencing in public, one availability report at a time.
Worth naming clearly: Dohmke is now selling the alternative to the problem he's describing, so this isn't a neutral bystander's diagnosis — it's also a pitch deck. Both things can be true. His incentive to frame centralized Git hosting as broken doesn't make the underlying strain he's pointing at fictional; GitHub's own numbers already established that part.
Where the two explanations actually disagree
There's a genuinely fair counter-read here, and it's worth stating in full rather than knocking down.
Multi-cloud contracts are completely normal for large platforms — burst capacity from a second provider is standard risk management, not a sign of crisis, and plenty of companies that have never had a reliability problem still keep a second cloud on retainer. And GitHub's outages weren't happening in a clean lab environment — they were happening in the middle of an already-planned Azure migration, which makes it genuinely hard to cleanly separate "AI agents broke this" from "moving infrastructure while under load broke this."
Both of those are reasonable. Neither one explains away GitHub revising its own capacity plan from 10x to 30x within a few months of setting it, and neither one explains a company's former CEO building a literal architectural alternative to the exact bottleneck his old employer was hitting.
My honest read
I don't think Microsoft did anything scandalous here, and I don't think GitHub's engineers are incompetent — 14x growth in commit volume in a single year is not a load profile anyone plans for cleanly, agentic or not. What I think actually happened is simpler and a little more uncomfortable: the industry's collective bet on shipping AI coding agents everywhere, as fast as possible, ran ahead of the infrastructure math for exactly the platform those agents depend on most. GitHub isn't a side character in the AI coding story. It's the thing all of it runs through. And even GitHub, owned by one of the three companies actually building frontier AI, revised its own scaling estimate 3x and still needed help from a competitor to hold the line.
If that's what happened to GitHub — with Microsoft's balance sheet, engineering depth, and advance warning — it's worth asking what the same math does to a mid-sized company's internal CI, code review queue, or self-hosted Git server as agentic coding adoption climbs inside their own teams. Most of those systems were sized for humans. Nobody has published their own version of GitHub's April numbers yet, mostly because nobody's had to yet.
Key things worth holding onto from this:
• GitHub's own commit volume grew roughly 14x year over year — from ~1 billion in 2025 to a 14-billion pace in 2026 — almost entirely driven by AI coding agents.
• GitHub revised its internal capacity plan from 10x to 30x within months, and its own published availability numbers still dropped as low as 78.33% (April) and 88.39% (June).
• Microsoft — GitHub's owner, Azure's operator — contracted AWS, its direct cloud competitor, for extra GitHub capacity, and confirmed the arrangement on the record rather than denying it.
• GitHub's own former CEO, Thomas Dohmke, launched a distributed Git network in July 2026 explicitly designed around the exact strain his old company has been publicly reporting since January.
Sources: GitHub availability reports, January/April/May/June 2026 (github.blog); The Register, "GitHub outages persist as AI coding drives traffic surge" (June 12, 2026); Business Insider (via TechRadar, Yahoo Finance, Memeburn, Neowin), reporting on Microsoft's AWS capacity arrangement for GitHub; Microsoft spokesperson on-record statement re: multi-cloud strategy; GitHub COO Kyle Daigle, public confirmation of commit-volume figures (April 2026); DevOps.com, GeekWire, The Register, Winbuzzer, coverage of Thomas Dohmke's Entire launch (July 8, 2026).
No comments:
Post a Comment