How Microsoft, Google and Amazon Are Competing for AI Infrastructure

The three biggest cloud providers are fighting the same war on several fronts at once: who has the most capacity, the best chips, the largest customers, and the power to keep it all running. The contest isn’t only about winning today’s cloud market. It’s about who will own the foundation that AI is built on. For more coverage of this race, see our AI infrastructure and data center news.

The scoreboard

Start with where the three stand. According to Synergy Research Group’s Q2 2026 data, as summarized by Articsledge, Amazon leads cloud infrastructure services with a 28 percent share, Microsoft has 20 percent, and Google has 15 percent, in a market worth $143.4 billion for the quarter. A year earlier, AWS held about 30 percent.

Share is only half the story. The market itself grew about 43 percent year over year, the fastest pace in eight years, so a provider can grow strongly and still lose share. Shattered.io’s roundup of Q2 earnings puts AWS growth at around 37 percent, Azure at 43 percent, and Google Cloud at 82 percent. Google’s rise is the headline: TechnologyChecker reports Google Cloud revenue of $24.8 billion in the quarter and a committed-but-unbilled backlog of $514 billion, up from $106 billion a year earlier.

A caution on these numbers: market-share trackers differ, and some sites republish figures that don’t match the original reports. Check the Synergy release and each company’s earnings before quoting a specific share.

The spending race

Behind those results is an enormous investment race. One 2026 tracker lists Amazon’s capex guidance at about $220 billion, Alphabet’s at $195 to $205 billion, and Microsoft’s at about $175 billion, with Microsoft’s figure reflecting a lease-accounting change rather than a cut. FactSet notes that this spending is outrunning cash flow, pushing hyperscalers toward debt and other external financing.

Why spend so much? We covered the motives in why Big Tech companies are building their own data centers: control over design, cost at scale, speed to capacity, and independence. In a market where customers have committed to more capacity than exists, the provider that can bring new capacity online first wins the contract.

Front 1: Custom chips

Each company is now competing partly through its own silicon, and each has a different approach.

Google has the longest track record. Its TPUs have been in development for a decade, and the eighth generation is split into separate training and inference chips, as we described in why Google designs its own AI hardware.

Amazon began with Annapurna Labs and its Graviton CPUs before moving to AI accelerators. Our piece on how Amazon is building its own AI silicon covers Trainium and Project Rainier, which uses nearly 500,000 Trainium2 chips for Anthropic according to Data Centre Magazine.

Microsoft is the latecomer on custom AI silicon but moving fast. Its Maia 200 is in production in Iowa and Arizona, according to Data Center Dynamics, and we examined the strategy in why Microsoft is building its own AI chips.

All three still buy large numbers of Nvidia GPUs, so the chip contest is about leverage and cost, not replacing Nvidia. As we noted in why Nvidia is expanding beyond GPUs, Nvidia is responding by selling whole systems.

Front 2: Anchor customers

The biggest AI labs are the prizes, because a single customer can fill gigawatts of capacity. AWS and Anthropic announced a commitment of more than $100 billion over ten years covering several generations of Trainium. Google, meanwhile, counts Anthropic as the anchor external TPU customer, committed to up to one million chips according to one 2026 guide. And Microsoft’s Maia 200 is reportedly of interest to Anthropic as well.

That overlap shows how the market is evolving. Frontier labs increasingly spread workloads across providers and chip types, so the clouds compete on price, performance per watt, and availability rather than relying on lock-in alone. Clarigital also reports that OpenAI agreed to use about 2 gigawatts of Trainium capacity in February 2026, which would make Amazon’s chips a platform for more than one lab.

Front 3: Power and sites

The quiet battleground is electricity. As we explained in why AI data centers need much more power, grid connections can take five years or more, and the companies that secure power first can build first. Hyperscalers are signing nuclear agreements, funding on-site generation, and choosing sites by grid capacity rather than only by fiber access. Enverus tracks more than 10 GW of nuclear power purchase agreements signed across the largest tech companies, a pool that includes Amazon, Google, and Microsoft.

Front 4: Software and ecosystems

Hardware wins deals, but software keeps customers. Microsoft has the enterprise relationships and the Copilot products, Google has Gemini and its AI research, and Amazon has the broadest cloud customer base. Each is trying to make its chips easy to adopt, whether through a Maia SDK with PyTorch and Triton support, Google’s TPU tooling, or AWS’s Neuron software. The company that makes switching painless gains more business.

The risks

The race carries serious risk for all three. Spending is front-loaded, returns come later, and investors have reacted nervously when guidance rises: CNBC reported that Amazon, Meta, and Microsoft shares fell after Alphabet raised its 2026 capex forecast. If AI demand slows, all three could end up with expensive, underused capacity. Power constraints, equipment shortages, and local opposition to new data centers could also slow the build-out.

The bigger picture

Microsoft, Google, and Amazon are competing for AI infrastructure on every layer at once: capacity, chips, customers, power, and software. No single player is likely to dominate all of them. Amazon still leads in scale, Google has the fastest momentum, and Microsoft has the deepest enterprise reach and a close relationship with the labs it supplies. For readers following the race, the metrics to watch are backlog growth, capacity coming online, and how much of each company’s AI workload runs on its own silicon. Follow our AI hardware coverage as the competition develops.