Hyperscaler earnings show the AI infrastructure boom is still compounding

If you want less keynote poetry and more cash math on the AI boom, read the hyperscaler prints. Microsoft's fiscal 2026 fourth quarter, ended June 30 and reported around July 29, put commercial remaining performance obligation at $678 billion, up 84% year over year. Azure and other cloud services grew 43% in the quarter. Azure fiscal-year revenue crossed $100 billion for the first time. Amazon's AWS second quarter of 2026 delivered about $42.2 billion in revenue, up 37% year over year, which CRN flagged as the fastest AWS growth rate in 18 quarters. Google Cloud's second quarter landed at $24.8 billion, up 82% year over year on the same CRN face-off. This is not a vibes chart. It is contracted cloud demand and GPU-era run rates showing up in cloud earnings.

What is commercial RPO and why $678B matters?

Remaining performance obligation is contracted revenue not yet recognized. Microsoft's commercial RPO at $678 billion, up 84%, is the order book for Microsoft Cloud consumption that customers have already committed to. It is not the same thing as this quarter's cash. It is the multi-year demand signal investors treat as the AI capacity backlog. When that number jumps this hard, you are looking at enterprises locking in spend before they finish arguing about which copilots will stick.

Pair that with Azure and other cloud services at plus 43% for the quarter and Azure fiscal revenue above $100 billion. Microsoft Cloud itself printed $59.3 billion for the quarter, up 27%, in the same release. CNBC's wrap hammered the same spine of the story. The narrative Microsoft wants you to hear is simple. Customers are pre-committing for AI-era workloads even while capacity remains tight.

Read RPO next to capex debates elsewhere in tech. Some model builders and enterprise buyers spend first and monetize later. Hyperscalers selling the capacity get to show the sell-side of that trade in cloud growth and backlog. That asymmetry is why cloud earnings have become the least dishonest public scoreboard for whether AI infrastructure demand is still compounding.

RPO can also mislead if you treat it as guaranteed GPU happiness. Contracts slip. Usage ramps miss. Credits get negotiated. Still, an 84% jump in commercial RPO is not a rounding error from a single logo deal. It is a stack of commitments large enough that Microsoft's capacity planning and your allocation politics both get noisier.

AWS and Google Cloud join the acceleration tape

AWS at $42.2 billion and plus 37% is the scale leader still accelerating. CRN's Q2 2026 face-off called out the 18-quarter growth high. That is the base load of enterprise IT plus the new AI training and inference layer fighting for the same racks. AWS does not need to win the percentage race to matter. When the largest cloud still posts mid-thirties growth, the AI spend is not a side project living only in Azure marketing decks.

Google Cloud at $24.8 billion and plus 82% is the growth-rate headline. CRN tied the jump to GCP enterprise AI solutions, AI infrastructure, and core platform services. Absolute dollars still trail AWS and Microsoft Cloud. The slope is the strategic point. A smaller base can print eye-watering percentages and still leave sales teams hunting for capacity and committed-use discounts that match the pitch.

None of these prints prove every AI pilot will become a durable software product. They prove buyers are still paying for GPUs, networking, storage, and managed platforms at rates that rearrange hyperscaler mix. Training clusters, inference endpoints, vector stores, and the boring databases next to them all show up somewhere in that revenue, even when the press release only says AI.

I also read these numbers as a warning about vendor concentration. If your architecture assumes you can casually multi-cloud your way out of allocation pain, the earnings tape says the scarce thing is still accelerated capacity, not PowerPoint diagrams of three providers. Growth this loud usually means queues, not surplus.

The CRN face-off framing is useful precisely because it refuses to let one vendor's narrative dominate. Microsoft wants you staring at RPO and Copilot seats. Amazon wants you staring at absolute AWS dollars. Google wants you staring at percentage growth. Reading all three in one sitting is how you keep from being captured by whichever IR deck arrived in your inbox first.

Three clocks inside one growth percentage

Separate three clocks. Training cluster buildouts. Inference traffic that follows successful products. Ordinary cloud migration that still pays the bills when demos fail. Earnings releases mash those clocks into one growth percentage. Operators should not. A training spike can look like product-market fit from the outside while the customer still has no durable application revenue.

If you buy cloud, the practical read is capacity and pricing power. Demand exceeding supply shows up as allocation politics and committed-use negotiations. Reserved instances and spend commitments become less about spreadsheet optimization and more about whether your team gets GPUs when the model team swears the deadline is real.

If you sell software on top of these clouds, the practical read is that your customers' AI budgets are still landing somewhere real. That does not mean every feature flag you ship will get funded. It means the infra line item is not imaginary, which is the floor under a lot of agent and analytics startups that need someone else's metal.

Investors will keep arguing about depreciation schedules and whether capex is overbuild. Fair fight. The late-July and early-August prints still say the sell-through of capacity is not stalling in the public numbers that matter most for cloud operators.

One more cut I use in notes. Ask whether growth is coming from a few mega-deals that concentrate risk, or from a broad base of mid-market and ISV workloads that keep renewing when a single logo stalls. Earnings releases rarely hand you that cut cleanly. Customer anecdotes and capacity language in Q&A sometimes do. When management talks about long-lived AI commitments rather than one-off training bursts, RPO becomes more believable as a durable backlog and less like a temporary reservation spike.

How to read the boom without drowning in run rates

Ignore vanity ARR screenshots from demos that never hit production. Anchor on hyperscaler growth, backlog, and the qualitative capacity language in prepared remarks. When Microsoft talks $678B commercial RPO and Azure over $100B for the year, when AWS posts a multi-year growth high, when Google Cloud prints 82%, you have a coherent industry signal even if your favorite chatbot wrapper is struggling.

Also ignore the opposite cope, the idea that cloud growth somehow proves every enterprise AI program is rational. Plenty of that spend is experimentation, prestige, and fear of missing a board slide. The earnings still clear a lower bar. Money is moving into infrastructure at scale, and the hyperscalers are capturing it.

For practitioners, translate the tape into procurement behavior. Expect harder conversations about committed spend floors. Expect regional capacity maps to matter as much as list prices. Expect your finance team to ask why inference unit economics still look soft while the cloud bill climbs. Those questions are healthier than pretending the boom is either fake or infinitely rational.

I am filing late-July and early-August earnings as confirmation that the AI infrastructure boom remains a hyperscaler revenue event, not only a conference theme. Watch the next quarter for whether RPO growth cools while usage stays hot, or the reverse. Either pattern would tell a clearer story than another keynote about agents. More cloud coverage continues with Younes Bekrar.

Related reading on Skarvonix: our cloud category, the authors directory, and more from Younes Bekrar.

Primary sources and further reading: Microsoft FY26 Q4 earnings release, CRN AWS vs Microsoft vs Google Cloud Q2 2026, CNBC Microsoft FY26 Q4 coverage.