Google Cloud vs AWS vs Azure: Revenue, Run Rate, TPUs, and the Race to No. 1

AWS is still the largest cloud. Google Cloud is the fastest-growing one, and AI chips are the reason. In Q3 2026, AWS made $46.0 billion, Azure $33.2 billion, and Google Cloud $30.7 billion. Google Cloud grew 103% year over year, more than twice the rate of Azure (48%) or AWS (39%).

This post explains how a cloud "run rate" is calculated and why it can mislead. It shows what is inside each company's number, how much of Google's growth may be TPU hardware, and how Google's TPU compares to Nvidia. It ends with a 10-year model of when Google Cloud can pass Azure and AWS, and what that does to Alphabet as a business.

Note: this is analysis, not investment advice. The reported numbers come from each company's Q3 2026 filings and earnings calls. The 10-year numbers are a scenario model, and we label them as such.

The Key Numbers (Q3 2026)

All three companies reported on October 28 and 29, 2026. The table uses calendar Q3 2026 (July to September). For Microsoft, that is fiscal Q1 2027.

Sources: Amazon Q3 2026 release, Microsoft FY27 Q1 release, Alphabet Q3 2026 release, and the three earnings calls linked below.

Q3 2026 is the first quarter in which Microsoft reported Azure revenue in dollars. Before September 2026, Microsoft gave only a growth rate. That makes this the first clean three-way comparison in years.

How a Cloud Revenue Run Rate Is Calculated

A revenue run rate shows the yearly pace of a business at its latest quarter. The formula is:

Annualized run rate = latest quarterly revenue x 4

For Google Cloud: $30.709 billion x 4 = $122.8 billion.

A run rate is not a forecast. It is also not actual revenue for a year, and it is not contracted annual recurring revenue (ARR). For a business that grows fast, the run rate is much larger than what it earned in the last 12 months.

Three numbers answer three different questions:

1. Calendar 2025 revenue answers "how big was it last year?" Google Cloud: $58.7 billion.
2. Trailing 12 months (TTM) answers "how much did it earn in the last year?" Add the last four quarters: $17.66B + $20.03B + $24.77B + $30.71B = $93.2 billion.
3. Run rate answers "how fast is it selling now?" $122.8 billion.

The measure you pick changes the story. On 2025 revenue, Google Cloud was 46% of the size of AWS. On the Q3 2026 run rate, it is 67% of AWS. Both are correct. The second one shows how fast the gap is closing.

Analysts also use "net new annualized revenue." This is the change in run rate from one quarter to the next: (Q3 revenue - Q2 revenue) x 4. For Google Cloud, ($30.709B - $24.768B) x 4 = $23.8 billion. For AWS it was $15.0 billion, and for Azure $15.0 billion. This metric shows momentum. It swings a lot, and one large hardware delivery can move it.

Revenue and Growth, Quarter by Quarter

Two years ago, AWS was 2.4 times the size of Google Cloud. Now it is 1.5 times the size. Azure was 1.4 times Google Cloud. Now it is only 1.08 times.

The growth chart shows why. AWS growth doubled in a year, from 20% to 39%. Azure went from 40% to 48%. Google Cloud tripled its growth rate, from 34% to 103%.

All three are accelerating at a size where most businesses slow down. The cause is the same for each: AI demand for compute is larger than the supply of chips, power, and data centers. Each company says it is capacity-constrained. Microsoft CFO Amy Hood said "customer demand continues to exceed available supply" on the FY27 Q1 call.

Among the three, AWS had 50% of combined revenue in Q3 2024. In Q3 2026 it had 42%. Google Cloud went from 21% to 28%. Azure stayed near 30%.

What Is Inside Each Number

The three numbers do not measure the same business. Read them with these differences in mind.

Microsoft changed its definition in September 2026. Under the new structure, Azure "becomes more purely our consumption-based platform and infrastructure business," according to its segment recast filing. GitHub cloud and Security Copilot moved out of Azure. That makes Azure the cleanest pure-infrastructure number of the three.

Google Cloud is the broadest. It includes Workspace (Gmail, Docs, Meet for business), which is subscription software. Since 2026 it also includes TPU systems sold as hardware. Alphabet's Q3 2026 10-Q says these product sales "generally consist of the sale of TPU systems comprising hardware, software" and services. It does not give the dollar amount.

How Much of Google Cloud's Growth Is TPU Hardware?

Alphabet does not split the number. On the Q3 2026 call, management said GCP was the largest contributor to growth and that TPU systems sales were a meaningful contributor. It also said growth without TPU systems sales was very strong. Google Cloud chief Thomas Kurian named Anthropic, Apple, Oracle, and Crux AI as TPU systems customers. Nearly 100 customers use them.

There are other clues. The 10-Q lists inventory "primarily of hardware related to TPU systems for sale to enterprise customers." Free cash flow was -$9.1 billion in Q3, partly because of that inventory build. The backlog includes $16.8 billion from a multi-year agreement with Crux AI, a related-party joint venture.

We can test the effect with a sensitivity. Assume TPU systems sales were zero a year ago. This is the strictest case, because it puts all hardware revenue into this year's growth.

1. Take Q3 2026 Google Cloud revenue: $30.71 billion.
2. Remove a TPU systems amount, for example $5 billion: $25.71 billion is left.
3. Compare to Q3 2025 revenue of $15.16 billion: $25.71 / $15.16 = 1.70.
4. Growth without hardware is 70%.

Hardware would need to be about a quarter of Google Cloud revenue for the rest of the business to grow slower than Azure. Management said GCP, not TPU hardware, was the largest contributor. So the core cloud business is most likely growing faster than both Azure and AWS.

Hardware still matters for durability. A cloud contract earns revenue every month. A TPU system sale earns revenue once, when Google delivers it. To grow next year, Google must sell more systems than this year. Alphabet expects "a meaningful revenue ramp in 2027" for TPU systems, so this effect will get larger before it gets smaller.

Profit and Backlog

Growth is not useful if it loses money. Here Google improved the most. Google Cloud's operating margin rose from 23.7% to 36.2% in one year. Operating income tripled, from $3.6 billion to $11.1 billion. AWS still makes more profit: $17.8 billion, at a 38.7% margin.

Both margins will face pressure. Alphabet said it started to rent third-party compute at the end of Q3 and expects "margin headwinds in the near term." All three companies will add large depreciation costs as new data centers come online.

Backlog is revenue that customers have contracted for but Google, Amazon, or Microsoft has not yet delivered. Accountants call it remaining performance obligations (RPO). It is the best signal of future revenue.

Google Cloud's backlog was $560.8 billion at September 30, 2026, up from $513.9 billion at June 30. Alphabet expects to recognize "just over 55%" of it in the next 24 months. That is about $310 billion over two years, or about $155 billion a year. That is more than the current run rate of $122.8 billion, before any new contracts.

AWS reported a $578 billion backlog on its Q3 2026 call, up from $496 billion in Q2. Microsoft's $677 billion is for all commercial contracts, including Microsoft 365, so it is not comparable to the other two.

TPU vs Nvidia GPU: Why Google's Chip Matters

A TPU (Tensor Processing Unit) is Google's own AI chip. It is an ASIC (application-specific integrated circuit): a chip designed for one type of work, here the matrix math in neural networks. An Nvidia GPU is a programmable processor. It can run many types of parallel work and has the CUDA software ecosystem.

The current TPU is Ironwood (TPU v7). Google's Ironwood announcement lists these specifications:

SemiAnalysis compared Ironwood to Nvidia in its report TPUv7: Google Takes a Swing at the King. Its main findings:

• Specs are close to Nvidia. Peak FLOPs and memory bandwidth are about 10% below the Nvidia GB200.
• Cost is much lower. For Google, the all-in total cost of ownership (TCO) per Ironwood chip is about 44% lower than a GB200 server. With Google's margin added, external prices are up to about 30% lower than GB200 and 41% lower than GB300.
• Pods are very large. One Ironwood pod connects up to 9,216 chips. The network uses optical circuit switches between racks of 64 chips.
• Google now sells TPUs to other companies. The Anthropic agreement covers 1 million TPUs. About 400,000 are sold as finished racks through Broadcom. About 600,000 are rented through Google Cloud, which SemiAnalysis estimates at $42 billion of backlog.
• Software is the gap. Google is building a native PyTorch backend and vLLM support. The XLA compiler and multi-pod runtime are still closed source.

This is the reason Google Cloud can grow faster than its rivals. Google designs the chip, builds the data center, trains Gemini on it, and sells it as cloud capacity or as hardware. Lower cost per chip means Google can price lower or earn a higher margin than a cloud that buys Nvidia at Nvidia's margin.

AWS and Microsoft also build chips. AWS said revenue commitments for its Trainium and Graviton chips grew more than 20 times in a year, and it will sell Trainium outside AWS from late 2027. Microsoft CEO Satya Nadella said Maia 200 delivers "up to 1.7x more tokens per watt" in Teams and Copilot. Google's TPU is the only one with ten years of production use and external sales at gigawatt scale today.

Vertical Integration: Who Owns Which Layer

Vertical integration means one company owns more layers of the product, from chip to application. Each layer it owns is a margin it does not pay to a supplier.

Google is the only one of the three that owns a frontier model and a mature AI chip and trains the model on that chip. Amazon owns the deepest set of custom infrastructure chips, but its leading model partner is Anthropic. Microsoft owns the largest enterprise software distribution, but much of its AI demand comes from OpenAI.

The 5 to 10 Year Outlook

Nobody can forecast cloud revenue for 10 years. A model can still show what each outcome requires. Our model starts from the Q3 2026 run rate and applies a growth rate for each year. Growth slows as each business gets larger.

AWS uses Amazon's own outlook of more than 40% growth in 2027 and 2028. Azure starts near Microsoft's Q4 guidance of 49-50%. The Google base case cuts its growth from 103% to 60% in 2027 to allow for hardware sales that do not repeat.

Three results from the model:

1. Google Cloud passes Azure first. In the base case, it passes Azure in 2027. In the bull case, it passes AWS in 2030.
2. AWS reaches $1 trillion. At its guided growth, AWS passes a $1 trillion run rate around 2035. Amazon CEO Andy Jassy said AWS could become a trillion-dollar business within 5 to 10 years.
3. The bear case is not small. Even if Google Cloud growth falls to 40% next year, it reaches a $571 billion run rate by 2036.

A Formula for When Google Cloud Passes AWS

The model hides one simple rule. If one business is smaller but grows faster, the time to catch up is:

Years to catch up = ln(size ratio) / ln((1 + faster growth) / (1 + slower growth))

"ln" is the natural logarithm. AWS is 1.497 times the size of Google Cloud. Azure is 1.080 times.

The rule of thumb: a 10-point growth lead closes the AWS gap in about 5 to 6 years. A 5-point lead takes about 10 years. The Azure gap is so small that almost any lead closes it within two years.

Google Cloud Can Pass Azure in Q4 2026

Microsoft guided Azure growth of 49-50% in constant currency for Q4 2026. Azure made $24.13 billion in Q4 2025. So Azure revenue should be about $36.0 to $36.2 billion in Q4 2026.

To pass that, Google Cloud must grow 17.5% from Q3 to Q4. It grew 23.7% in Q2 and 24.0% in Q3, quarter over quarter. Alphabet said TPU systems sales will "continue to scale in Q4." On reported revenue, Google Cloud is on track to become the No. 2 cloud this quarter or next.

Remember the definitions. Google Cloud includes Workspace and TPU hardware, and Azure does not. On pure cloud infrastructure, Azure will stay ahead for longer.

What This Means for Google as a Business

Google Cloud is changing what Alphabet is. Alphabet has been an advertising company for 25 years. Cloud is now the fastest-growing part, and it will soon be the largest part.

Segment operating income here is Google Services plus Google Cloud, before Alphabet-level costs. Source: Alphabet Q3 2026 release.

If Search grows 11% a year and Google Cloud follows our base case, Cloud passes Search in revenue around 2029.

Search: More Queries, Less Revenue per Query

Google Search has more use than before, but it earns less on each query. On the Q3 2026 call, Sundar Pichai said query growth "is the highest it's been in the last 5 years." Search revenue grew 12%, slower than in Q2. If queries grow faster and revenue grows slower, revenue per query is falling.

Chief business officer Philipp Schindler gave four causes for the slowdown:

AI Mode is Google's chat-style search. Its queries are about 3 times longer than classic searches. It shows fewer ads, so each query earns less. Google is now expanding AI Mode ads to international markets and adding ads inside multi-turn conversations. Schindler said click-through rates "are already similar" for sponsored and organic product listings in AI Mode. Google did not say how large the revenue gap per query is.

Q4 will also be harder. CFO Anat Ashkenazi said Q3 started to lap the Search acceleration of late 2025. The rest of Google Services did not make up the difference: YouTube ads grew 9% and Network ads grew 1%.

This is why Cloud matters more each quarter. If Google closes the AI Mode ad gap, the query growth turns into revenue growth. If it does not, Search grows slowly while Cloud carries Alphabet's growth.

Cash: The Spending Comes Before the Return

Operating income grew 30%, but the stock fell about 10% after hours. Earnings per share missed estimates for a reason outside the operating business. Other income and expense was -$8.2 billion, "due primarily to unrealized loss from our investment in SpaceX." The bigger concern is cash:

Q3 free cash flow also paid for the TPU inventory build and a $5.2 billion payment on a 2018 European Commission fine.

Depreciation is the next risk. About 65% of Q3 infrastructure spend went to servers and network equipment, which Alphabet depreciates over 6 years. Google also says its next TPUs, TPU 8t and 8i, offer "up to 2.7x better training performance" than the previous generation. Chips that lose that much value each generation may not stay useful for 6 years. Shorter lives would mean higher depreciation and lower reported profit.

Alphabet says it expects a return on invested capital "meaningfully above our cost of capital." It did not give the number.

TPU Systems: A New Hardware Business

TPU systems are full racks of TPUs with networking, storage, and software. An analyst asked if a gigawatt of TPU systems costs about $35 billion, and how much of that becomes revenue. Ashkenazi said "we don't provide specific financials at the product level."

The cash cycle is clear:

1. A customer signs a contract. It goes into backlog.
2. Google builds inventory for that contract only.
3. Google delivers the systems and records the revenue.
4. Cash comes in on delivery.

Google sells these systems directly and through partners. Kurian said Apple is "scaling pretraining of their foundation models on the latest generation of TPUs." Google also launched Crux AI, a joint venture with Blackstone that offers TPU systems to customers, and a partnership with Oracle.

The demand signals are strong. Alphabet's own models process about 32 billion tokens per minute, up 40% in one quarter and more than 4 times in a year. Customers who used Gemini Enterprise in Q3 2025 "more than quadrupled their spend" a year later. Gemini 4 Argon usage ramped 3.5 times faster than Gemini 3 in its first week. The Gemini app has 1 billion users.

The Anthropic Stake as a Source of Cash

Google is also one of Anthropic's largest investors. The Anthropic S-1 says Google will hold more than 5% of Anthropic in non-voting Class N shares. 2025 court filings put Google's stake at about 14%, with a cap of 15%.

At the reported IPO valuation range of $1.5 to $2 trillion, a 14% to 15% stake is worth about $210 billion to $300 billion. That is about one full year of Alphabet's 2026 capital expenditure.

One option for Alphabet is to sell some or all of the stake after the IPO. A sale could pay for 12 to 18 months of capex and fund share buybacks with the rest. There are costs:

1. Taxes. The gain over Google's cost is taxable, so the cash is less than the market value.
2. Timing. An IPO lock-up usually blocks insider sales for about 180 days.
3. The customer link. Anthropic is one of the largest TPU customers. The S-1 lists about $114.5 billion of firm Google Cloud commitments. A full exit could weaken that relationship.
4. Upside. If Anthropic keeps growing at its current rate, the stake could be worth much more later.

A partial sale keeps the customer link and still covers a large part of the capex. Alphabet has not said it plans to sell.

Four Changes to Alphabet

1. Revenue becomes less dependent on ads. Ad revenue rises and falls with the economy. Contracted cloud revenue, backed by a $561 billion backlog, is more stable.
2. The business needs much more capital. Search needs few data centers per dollar of revenue. Cloud needs many.
3. Cash flow gets more uneven. TPU systems sales need inventory before delivery, so cash leaves before revenue comes in.
4. AI spending serves both sides. The same TPUs serve Gemini in Search and AI Mode, and sell to cloud customers. Google pays for that compute at its own cost, not Nvidia's price.

The risk is concentration. A small number of AI labs, such as Anthropic, buy a large share of AI compute. If one large customer slows its spending, Google Cloud growth slows with it. Our analysis of the Anthropic S-1 shows how large those compute commitments are, and that Google is one of the main suppliers.

Five Numbers to Watch Each Quarter

Check these five numbers in each earnings report to see which path Google Cloud is on:

1. Google Cloud growth without TPU systems sales. If Alphabet ever reports it, compare it with Azure. Above 48% supports the bull case.
2. Backlog growth. Google Cloud added $47 billion of backlog in Q3. If quarterly additions stay above $40 billion, revenue growth has fuel.
3. Operating margin. Watch whether Google Cloud stays above 30% while it rents third-party compute.
4. AWS growth versus Amazon's guidance. If AWS stays above 40%, the time for Google to pass AWS goes up.
5. Azure versus Google Cloud revenue in Q4 2026. The first crossover is the clearest proof of the trend.

Choosing a Cloud for Your AI Work

The cloud race matters to buyers, not only to investors. Price per token and access to capacity now differ by provider and by chip, and they change each quarter.

OpenNash puts forward-deployed engineers on your team. We work with Fortune 2000 companies to put AI into production at the lowest cost per task, on the cloud you already use or a better one. A typical engagement:

1. Get your data ready: clean, permissioned, and connected to the cloud where it lives.
2. Measure cost per task for each workload on each cloud and chip, including TPU and Trainium options.
3. Build evals first, so you can change models or providers without a drop in quality.
4. Ship to production with cost controls and usage reporting per team.
5. Train your team to run and change the system without us.

At the end, you own the code, the evals, and the cost reports, with no lock-in to one provider.

Book a call with an OpenNash forward-deployed engineer or email us at hello@opennash.com. In the first session, we find the AI workload with the fastest return and estimate its cost on each of the three clouds.