What we learned about the hyperscalers from earnings season

Recently Amazon.com Inc., Google owner Alphabet Inc. and Microsoft Corp. reported earnings within roughly 24 hours of each other, and this earnings season told us more about long-term artificial intelligence strategy than any keynote or press release this year. Beneath the headlines lie three very different strategies.

Here’s what I took away: Amazon will be the long-term winner because its open, platform-first approach gives customers the broadest choice of models and silicon, and it’s showing that size doesn’t limit scale. Microsoft is a solid second, protected less by strategy than by the customers it already has locked in. And Google will be a distant third, with impressive technology built on a fragile business foundation and a closed, Gemini-first strategy that runs counter to where enterprises are heading.

Lesson one: AWS proved scale isn’t a speed limit

The most remarkable part of the earnings season was that Amazon’s Amazon Web Services business has found a way to buck the law of large numbers. It’s growing 39.3% at its size. AWS revenue reached $46 billion, marking its sixth straight quarter of acceleration and its fastest growth in nearly five years, putting the business at an $184 billion annualized run rate, according to Amazon’s third-quarter release.

To put that in perspective, the last time AWS grew this fast was in the fourth quarter of 2021, when it posted 40% growth on $17.8 billion in quarterly revenue. AWS is now more than two-and-a-half times that size and is growing at essentially the same rate. The roughly $13 billion it added year over year is about three-quarters of what the entire business generated in a quarter back then.

The law of large numbers is supposed to catch every business. The largest cloud provider on the planet is accelerating instead.

Amazon Chief Executive Andy Jassy projected AWS growth of more than 40% in both 2027 and 2028. Operating income reached $17.8 billion at a 38.7% margin, and backlog climbed to $578 billion, up from $244 billion at the end of 2025. On the earnings call, Jassy indicated AWS could reach $1 trillion in as little as five years, giving Amazon two businesses that do north of a trillion dollars in revenue.

Bedrock: The broadest model shelf

Bedrock is the most underappreciated piece of Amazon’s strategy and, in the long term, the most important. Microsoft built its AI identity around OpenAI, and Google around Gemini. Both host third-party models, but each has a house favorite that its roadmap and pricing orbit around. AWS made model neutrality the product.

According to an AWS Builder Center guide, Bedrock offers more than 100 models from providers including Anthropic, OpenAI, Meta, Mistral, DeepSeek, Google, xAI and Nvidia. In the third quarter alone, Amazon added OpenAI’s GPT-6 family, Anthropic’s Claude Opus 5.5 and Sonnet 5.5, xAI’s Grok 4.7, and Moonshot’s Kimi K3, per Amazon’s quarterly release.

The open-weight wave makes this even more important. Bedrock already runs DeepSeek, Qwen, Mistral, Kimi, Gemma, Nemotron and OpenAI’s gpt-oss models as fully managed services, as AWS has detailed. As open models close the gap, enterprises will fine-tune smaller, cheaper models for specific tasks and reserve frontier models for the hardest problems. The platform that lets them route across all of them under a single security and governance model wins. A cloud built around a single-model company has a conflict of interest every time an open model gets good enough to replace it.

Customer spending is validating Bedrock’s value, as Fortune 500 spending on Bedrock rose nearly sevenfold year over year and token volume grew nearly 17-fold, Jassy said on the earnings call.

Silicon: The inference cost advantage

The second advantage is silicon, which is tightly linked to Bedrock. Most of Bedrock’s inference runs on Trainium, which lets AWS keep prices low while protecting margins, according to TIKR’s call coverage. Customers using Amazon’s custom silicon saved more than $2 billion in the third quarter alone, as noted in the call transcript.

The bigger news was Jassy’s commitment to sell Trainium outside AWS starting in late 2027. “We believe we’re going to have a very significant business selling chips externally,” he said, according to Yahoo Finance. Training is periodic, whereas inference is a perpetual cost that scales with every agent and transaction. Selling Trainium racks brings AWS’ price-performance to enterprise data centers, sovereign environments and neoclouds, and gives organizations that need on-premises inference a credible alternative to Nvidia.

AWS has the broadest infrastructure footprint, the broadest model selection, and silicon that lowers the cost per token. The downside is the bill: a 2026 capex of roughly $220 billion, per 24/7 Wall St., and negative $33.5 billion in free cash flow, according to TradingKey. That spending is backed by contracted demand. If the growth rate continues, investors should be happy to see Amazon spending ahead of demand, but all eyes will be on free cash flow.

Who builds on AWS: Builders, including digital natives, AI labs, independent software vendors and large enterprises with mature engineering teams seeking model choice, cost control and no dependence on a single model partner.

Lesson two: Microsoft’s moat is its customers

Of the three hyperscalers, Microsoft’s business is the most predictable because it’s very hard to disrupt and isn’t a disruptor. What it is now is likely what it will be in the future. Azure grew 48% to $33.2 billion, the first time Microsoft has disclosed Azure revenue in dollars, according to CNBC.

But the most notable data point concerned lock-in. Customers with more than 50,000 Copilot seats grew more than eightfold year over year, including Capgemini, Hitachi, Truist and the U.K.’s HMRC, according to Microsoft’s earnings call. That’s Microsoft’s true AI strategy: Upsell AI to the hundreds of millions of workers already living in Office, Teams and Windows. Switching costs are enormous, and Microsoft will capture a large share of enterprise AI spend regardless of how its models compare.

Google and others have tried to disrupt the Office franchise for decades but haven’t posed a serious threat. In the AI era, a company like Canva or Zoom could introduce an entirely new way to create documents and begin to erode that business, but until there is a material loss of share, that installed base seems safe.

Microsoft’s strategy has downsides. The company leans heavily on OpenAI, a partner that increasingly acts like a competitor. On the same earnings call, Microsoft Chief Financial Officer Amy Hood said some third-party capacity deliveries were delayed. Microsoft also didn’t update its total paid Copilot seat count, leaving open how deeply those seats are used.

Who builds on Microsoft: Large, Microsoft-standardized enterprises, especially in regulated industries and the public sector, prioritize knowledge-worker productivity within the tools they already use.

Lesson three: Google’s growth masks a fragile foundation

Google Cloud’s headline numbers were impressive. Revenue grew 103% to $30.7 billion, and the operating margin was 36.2%, according to highlights from Alphabet’s earnings call. But a strong quarter doesn’t necessarily signal durability, and the earnings exposed three structural problems.

First, look at what is driving growth. According to Alphabet’s investor relations, CFO Anat Ashkenazi said tensor processing unit system sales were a “meaningful contributor” to cloud revenue. Hardware sales are lumpy and concentrated among a few large buyers, which is a different business from recurring cloud consumption.

Second, the core business funding all of this is showing cracks. Search made up 49.7% of Alphabet’s revenue, below half for the first time in the company’s history. That trend is likely to continue given the diversity of AI platforms. A year earlier, it was 55.3%. Also, Search grew only 12%, down from 19% in Q1 and 17% in Q2. YouTube ad revenue grew 9.2% to $11.21 billion, short of analysts’ $11.43 billion expectation. That’s down 15% from a year earlier and 13% from last quarter, adding only $152 million over the last quarter.

Third, spending is outpacing cash. Capex reached $51.4 billion, while free cash flow was negative $9.1 billion. Alphabet expects capex to exceed $300 billion in 2027, according to CNBC. The core business’s decline is becoming significant as the capex arms race continues.

Google has world-class research, Gemini and TPUs. Against that, it has a Gemini-first strategy that sits awkwardly with the open-weight wave, an ad business under AI pressure, and a history of sunsetting products, which remains the biggest objection I hear from chief information officers. Also, Google has always struggled with its enterprise strategy.

In 2018, I wrote this post for CIO.com asking whether Google was ever going to get serious about enterprise, and nothing much has changed. It seems to roll out products that have promise but never makes them best-in-class. As an example, Google Meet is good but not great, and Docs, Gmail and Drive are the same. They’re low-cost, but there are always better alternatives. Why would we expect their AI strategy to be different?

Who builds on Google: AI-native startups, research-heavy organizations,and BigQuery shops willing to accept platform risk. It’s a harder sell to a regulated enterprise making a 10-year commitment.

What about IBM and Oracle?

Neither is likely to make a long-term dent in the hyperscaler market, for one simple reason: They lack the flywheel. Amazon has retail and advertising; Google has search; and Microsoft has Office and Windows, each massive cash-generating businesses that can fund a multiyear compute buildout. The AI infrastructure race is capital-intensive, and IBM and Oracle are running it without that engine.

Oracle’s numbers look impressive on the surface, with cloud infrastructure revenue up 121% to $7.4 billion in its September quarter and RPO at $664 billion. The problem is how it’s financing that growth. Capex hit $28 billion in a single quarter, free cash flow was negative $5.4 billion, and debt has reached $125 billion, according to CNBC. After a $20 billion equity sale, it’s now selling another $10 billion in bonds, as Yahoo Finance reported.

Its database business can’t fund this buildout, so Oracle is relying on borrowed money and customer prepayments. That’s a fragile way to compete with companies that can self-fund.

IBM has largely stayed out of the compute race. Revenue grew 5% to $17.1 billion, led by Red Hat, according to IBM’s earnings release. IBM’s presentation reports $7.8 billion in year-to-date free cash flow, roughly what Amazon or Microsoft spends on capex in two weeks. IBM will matter as a hybrid-cloud integration layer, not a primary AI platform.

What it means for IT leaders

• Separate productivity AI from application AI. For example, don’t let a Copilot licensing decision dictate where you build custom agents. That might be a fast path to agentic, but it could be the wrong decision long-term.
• Make cost per inference a board-level metric. As agentic workloads scale, custom silicon roadmaps matter more than model leaderboards.
• Insist on model choice. As open-weight models close the gap, build where you can swap models without rewriting applications. Open-weight models are nearly on feature parity with frontier models and are worth a look.
• Pressure-test vendor durability. A five-year AI platform decision requires a five-year partner. It’s often easy to choose a vendor you already have in-house, but the AI race is a marathon, and the wrong decision could set the company back if a change becomes necessary.

Final thoughts

The biggest takeaway from earnings season is that the hyperscaler AI race now has a predictable order. Amazon is growing nearly 40% at an $184 billion run rate by doing what it has always done: building an open platform from silicon to agents, offering the broadest shelf of frontier and open-weight models, and preparing to take its inference silicon beyond its own data centers. Microsoft is a strong second, monetizing its existing customer base. Google is a distant third, betting that superior technology can outlast disruption to its core business while asking customers to adopt a closed stack.

To be clear, when I say “distant third,” that’s a forward-looking statement. Currently, Google’s cloud business is only slightly smaller than Azure’s, so my basis for “distant third” is openness, customer trust, enterprise go-to-market and the ability to fund the buildout.

When AI becomes a line item in IT budgets, customers will gravitate toward the platform with the lowest inference costs, the broadest selection of models and the deepest operational track record. Openness wins platform wars.

Zeus Kerravala is a principal analyst at ZK Research, a division of Kerravala Consulting. He wrote this article for SiliconANGLE. 

Image: avnsuresh/Pixabay