Microsoft owns Azure. Azure is the world's number two cloud. And in June 2026, Microsoft started paying Amazon, its number one rival, to keep GitHub online.
That is not a typo. GitHub commits jumped from 1 billion in all of 2025 to a pace of 14 billion in 2026. AI agent pull requests went from 4 million in September 2025 to 17 million by March 2026. GitHub Actions compute hit 2.1 billion minutes in a single week, up from 500 million in 2023.
The platform buckled. Nine service incidents in May alone dragged availability below the 99.9% uptime GitHub promises enterprises, after it had already breached that bar in February and March. So Microsoft did the thing nobody expected: it rented capacity from the enemy. Here is why that single decision tells you the cloud wars are over, and what comes next.
The Capacity Pooling Principle
Here is the framework. When demand grows faster than any single player can build, rivals stop competing on supply and start pooling it. I call this the Capacity Pooling Principle.
How fast agentic development outran a hyperscaler's own forecast
The old cloud model was a fortress. Each hyperscaler locked customers in and treated buying from a rival as surrender. That worked when demand grew in straight lines you could forecast and provision against.
AI broke the straight line. GitHub planned a 10x capacity increase in October 2025. By early 2026 it realized it needed 30x. When your demand triples your own forecast in three months, you cannot pour concrete fast enough. You borrow.
The principle in one sentence: when growth outruns build time, the fastest capacity wins, even if it wears a competitor's logo.
Why a Systems Thinker Sees This Coming a Mile Away
Look at GitHub as a system, not a brand. A system has inputs, throughput, and constraints. The constraint here is not money. Microsoft projected $190 billion in capex for calendar 2026. The constraint is time.
Data centers take 18 to 36 months to build. AI agent traffic quadrupled in six months. That is a throughput mismatch, and throughput mismatches do not care about your strategy deck. The bottleneck moves to whoever has slack capacity right now, and right now that is AWS.
This is the classic theory-of-constraints trap. You can have infinite budget and still get strangled by one binding limit. GitHub's binding limit was usable Azure capacity at the exact moment agentic development exploded. So the system rerouted around the constraint. That is what healthy systems do under load.
Now apply the second lever any scaling operator knows: never optimize a part at the expense of the whole. Microsoft could have protected Azure's pride by rate-limiting agents or throttling Actions. That protects one subsystem and damages the whole. GitHub hosts over 100 million developers, and a degraded GitHub threatens the entire Copilot and Visual Studio flywheel.
So the right move was to sacrifice the local optimum, Azure purity, to save the global optimum, an online GitHub. That is systems thinking in action. You feed the constraint with whatever capacity exists, then fix the architecture behind the scenes.
There is a tell that proves Microsoft thinks this way. The same company walked away from a reported $3 billion Oracle cloud deal over security and compliance concerns. That shows selective tolerance, not desperation. Microsoft will pool capacity, but only where the security model holds. The system optimizes for survival and trust, not pride.
I think the people calling this a humiliation are reading the wrong layer. A founder who refuses to borrow capacity to protect the customer experience is not principled, they are stubborn. Build the system, then let it route load to wherever it flows fastest.
Whether this stays temporary is unclear. Microsoft frames the AWS arrangement as a stop-gap while the Azure migration finishes by 2027. The risk is platform creep: once GitHub stands up datastores and queues on AWS, repatriation gets harder. That is real technical debt, and it is the kind a systems thinker watches like a hawk.
2031
Three signals inside the same shift
The binding constraint is build time, not budget.
Microsoft projected $190 billion in capex, yet data centers take 18 to 36 months to build. AI agent traffic quadrupled in six months, so the system rerouted to whoever had slack capacity right now. That was AWS.
Supplier, competitor, and customer are merging.
Google agreed to pay SpaceX roughly $920 million per month, near $29 billion over 32 months, and sells compute to Anthropic, an OpenAI rival. Each hyperscaler manufactures demand and overflows into its rivals.
Pooling only works if spikes hit at different times.
By 2031 capacity pooling could be a published market, but the safety valve seals shut if every cloud booms at once. Whether AI demand is staggered or correlated decides whether pooling scales or shatters.
Pull back five years. The question is not "who wins the cloud war." It is "what does cloud even mean when the three giants are each other's backup generators."
By 2031, I expect capacity pooling to be a published market, not a secret. We already see the shape forming. Google agreed to pay SpaceX roughly $920 million per month for AI compute from October 2026 to June 2029, near $29 billion over 32 months. Google Cloud also agreed to sell compute to Anthropic, an OpenAI rival, in April 2026. Supplier, competitor, and customer are collapsing into one role.
This is the flywheel nobody mapped. The more aggressively each hyperscaler sells AI tools, the more demand they manufacture, and the more they overflow into rivals. They are both the arsonist and the fire brigade.
The asymmetric risk here is correlation. Capacity pooling only works if demand spikes hit providers at different times. If every cloud booms at once, the safety valve seals shut and everyone runs hot together. The data is mixed on whether AI demand is correlated or staggered, and that single variable decides whether pooling scales or shatters.
My read: the durable advantage in 2031 will not be raw capacity. It will be the architecture that routes across clouds cleanly, with security intact. Whoever owns that routing layer owns the toll booth, the same way Costco never made money on hot dogs but on the membership around them.
What to Build This Weekend
You will not build a hyperscaler this weekend. But you can build the mindset that wins in a pooled-capacity world: design for portability, not loyalty.
First, take any small project and split it across two providers. Run your app logic on one, your storage on another. The goal is not cost savings. It is learning what breaks when you stop assuming one vendor holds everything.
Second, get reps with a cloud coding agent so you understand the demand surge from the inside. Try OpenAI Codex with its new persistent cloud workspaces from the Ona acquisition. A persistent workspace just means the agent keeps your project state between sessions instead of starting fresh.
Third, run something local so you are not fully dependent on any cloud at all. The AMD Lemonade SDK is a local AI runtime that started on AMD hardware and now runs on NVIDIA GPUs through CUDA. If you want a quantized model to test, grab NVIDIA's DiffusionGemma 26B in NVFP4 from Hugging Face and see how it performs on your own machine.
If you do not know which tool fits your project, browse There's An AI For That, a directory with search, ratings, and real-time trends. Pick one. Ship one tiny thing.
Things will break. Your cross-cloud setup will throw weird latency errors, and your agent will open a pull request you did not ask for. That is normal. Test aggressively, fix one constraint at a time, and you will understand the GitHub story better than most engineers reading the headlines. Get your reps in.
Design for portability, not loyalty.
- Split one project across two providers. Run your app logic on one cloud and your storage on another. The goal is not cost savings, it is learning what breaks when you stop assuming a single vendor holds everything.
- Get reps with a cloud coding agent. Try OpenAI Codex with its persistent cloud workspaces from the Ona acquisition so you feel the demand surge from the inside, where the agent keeps project state between sessions.
- Run something fully local. Use the AMD Lemonade SDK, a local runtime that now also runs on NVIDIA GPUs through CUDA, and test NVIDIA's DiffusionGemma 26B in NVFP4 from Hugging Face on your own machine.
The durable advantage in 2031 is not raw capacity. It is the routing layer.
Microsoft owning Azure and still paying AWS to keep GitHub online is not a humiliation, it is a healthy system feeding its constraint. The fortress model died when AI broke the straight-line demand curve and forced rivals to pool supply. The real moat is now the architecture that routes load across clouds cleanly with security intact, the same way Costco earns on membership, not hot dogs. The open question is correlation: if every cloud booms at once, the safety valve seals and everyone runs hot together. Build for portability now and you will read the next outage better than the engineers chasing the headlines.