Amazon’s $496 billion backlog, up $132 billion in one quarter, is proof spending is already spoken for
Amazon Web Services has already sold more cloud capacity than it can build. That is the central fact inside Thursday’s Q2 2026 earnings report — and it reframes every other number in the release, including the record $220 billion capital expenditure commitment Amazon announced alongside it.
AWS posted $42.2 billion in revenue for the quarter ended June 30, 2026, a 37% year-over-year increase and the division’s fastest growth rate in 18 quarters, as Amazon’s Q2 2026 earnings release confirmed. The result surpassed Wall Street’s consensus estimate of roughly 31% growth by a margin most analysts called decisive. But the more consequential disclosure came on the earnings call, when CEO Andy Jassy announced that contracted cloud commitments — revenue from signed contracts not yet recognized — had jumped from $364 billion to $496 billion in a single quarter. That $132 billion single-quarter increase substantially exceeded the comparable backlog growth Google Cloud recorded over the same period, according to available analysis.
The size of that backlog is not merely a financial metric. It is evidence that enterprise demand for AWS infrastructure has reached a point where signed contracts already exceed Amazon’s near-term capacity to deliver. Jassy said as much, directly: “We will still not have enough capacity to meet all the demand we have in 2026. And I believe this dynamic will also be true in 2027, too. In fact, the demand we already have for 2028 is striking.”
Amazon shares climbed more than 10% in after-hours trading on Thursday.
Memory Costs Push the Spending Target from $200B to $220B
Amazon had entered 2026 with capital expenditure guidance of approximately $200 billion — itself already the largest infrastructure commitment in the company’s history. On Thursday’s call, Jassy raised that figure to $220 billion.
The specific driver Jassy named was memory costs. High-bandwidth memory — the specialized chip architecture that sits atop every AI accelerator GPU — has seen manufacturing capacity diverted away from conventional DRAM and toward AI data center use, because memory makers earn an estimated three to five times more revenue per wafer producing it. That supply shift has pushed memory contract prices up sharply in 2026, adding roughly $20 billion to Amazon’s projected infrastructure bill.
Jassy said Amazon remains on pace to double its power capacity by the end of 2027 compared to 2025. Most of the cloud computing capacity planned for 2027 has already been reserved by customers, he added, and a significant portion of 2028 capacity had also received commitments. Asked by a Wells Fargo analyst how the $496 billion RPO figure changes Amazon’s 2028 planning, Jassy confirmed the company expects to keep adding significant data center capacity for the next two to three years.
That scale of forward commitment matters for enterprise technology buyers. An organization that has not yet signed a multi-year AWS capacity agreement now faces a market where the lion’s share of near-term availability has been allocated in advance by customers who moved earlier.
The Number Behind the Number: What $496B in Signed Contracts Actually Means
Amazon’s remaining performance obligations — the formal accounting term for contracts signed but not yet converted into recognized revenue — reached $496 billion at the end of Q2, growing at triple-digit rates year over year, as Amazon’s Q2 2026 earnings release confirmed.
To put the scale in context: AWS now runs at a $169 billion annualized revenue rate, which Jassy noted would place it 24th on the Fortune 500 list if it were a standalone company. The $496 billion backlog is nearly three times that annualized revenue rate — which means Amazon has contracted work equivalent to almost three years of current AWS revenue that it has not yet billed or delivered.
Alphabet’s Google Cloud carried a comparable $467.6 billion in remaining performance obligations as of March 31, 2026, per its SEC filing. Amazon’s $132 billion single-quarter jump in backlog substantially exceeded the growth in Alphabet’s comparable figure over the same period, according to available analysis. The distinction matters for enterprises weighing cloud platform commitments: both major hyperscalers are running backlogs that signal sustained multi-year demand, but AWS’s rate of backlog accumulation in the most recent quarter was sharply higher in absolute dollar terms.
AWS Margin Expansion: The Trainium Thesis Showing Results
An easily missed detail in the Q2 release is the movement in AWS operating margins. AWS generated $16.6 billion in operating income in Q2 2026, a 64% increase from $10.2 billion a year earlier, as the Q2 2026 earnings release shows. The operating margin reached 39.4% — up from 32.9% in Q2 2025, a 650-basis-point year-over-year expansion.
That margin expansion, occurring simultaneously with a major CapEx increase, is the clearest available evidence that Amazon’s Trainium custom chip strategy is generating real financial results at scale. Rather than buying Nvidia GPU capacity at market prices that reflect the same memory shortage inflating Microsoft’s and Meta’s capex costs, Amazon built Trainium — a custom AI accelerator chip architecture optimized for the matrix-multiplication operations that dominate transformer training and inference, as covered in prior analysis of Amazon’s silicon strategy.
The third generation, Trainium3, is manufactured on TSMC’s 3-nanometer process. A full 144-chip UltraServer delivers 362 MXFP8 petaflops of aggregate compute — performance that matches Nvidia’s Blackwell NVL72 at rack scale, at roughly 50% lower total cost of ownership according to third-party analysis. The chip connects through a NeuronLink-v4 interconnect using a NeuronSwitch-v1 all-to-all switched fabric that offloads inter-chip communication overhead from compute cores.
The tradeoff is real: Trainium only delivers its performance and cost advantages through Amazon’s proprietary Neuron SDK. Enterprises whose AI workloads are built for Nvidia’s CUDA ecosystem face a porting cost to migrate — and that friction is why Nvidia’s hold on AI compute spending has not collapsed despite Amazon’s cost-advantage claims. But for customers already running on Bedrock or investing in new AI workflows, the $220 billion CapEx commitment largely funds infrastructure where Amazon controls the cost at the silicon level.
The two businesses that most directly reflect the Trainium and Bedrock investment — AWS’s AI services arm and its chips business — each crossed $25 billion in annualized revenue run rates in Q2, both growing at triple-digit year-over-year rates, according to the Q2 2026 earnings release. Anthropic and OpenAI, which Jassy described as the two leading AI labs in the world, have each made multi-year, multi-gigawatt capacity commitments on Trainium.
Graviton5 and Bedrock: The Infrastructure Story Beyond AI Chips
AWS’s server CPU business reinforced the custom-silicon theme. Graviton5, released to general availability in Q2, delivers up to 30 to 40% better price-performance than comparable instances and up to 25% better compute performance than Graviton4, as detailed in the Q2 2026 earnings release. Graviton is now used by 98% of the top 1,000 EC2 customers, and Graviton5 is growing roughly twice as fast as Graviton4 did at the same point in its lifecycle.
On Amazon Bedrock — the managed foundation model service that allows enterprise customers to deploy AI models from Amazon, Anthropic, Google DeepMind, Meta, and others through a single platform — hundreds of thousands of customers now use the service, and Amazon added more Bedrock customers in the last six months than in the first two years after launch, with customers spending more in Q2 than in all prior quarters combined, as the Q2 2026 earnings release notes. The Q2 release noted new model additions including OpenAI’s GPT-5.6, Anthropic’s Claude Opus 5, Google DeepMind’s Gemma 4, and SpaceXAI’s Grok 4.3.
The Company-Wide Picture: First $200B Quarter, With Context Required
Amazon’s total net sales reached $200.6 billion in Q2 2026, a 20% year-over-year increase and the first time the company crossed $200 billion in revenue in a single quarter, according to the Q2 2026 earnings release. Operating income increased 43% to $27.5 billion. Advertising services grew 26% to $19.8 billion.
The headline net income figure — $62.6 billion, or $5.75 per diluted share, more than triple the year-ago period — requires important context. Amazon’s own Q2 2026 earnings release confirms that $53.4 billion of that net income came from a non-operating, non-cash gain, primarily from marking the company’s investment in Anthropic to a higher paper value. The underlying operating story — $27.5 billion in operating income, up 43% year over year — is nevertheless robust on its own terms.
Free cash flow on a trailing 12-month basis swung to an outflow of $7.6 billion, from an inflow of $18.2 billion in the comparable prior-year period, as the Q2 2026 earnings release discloses. Long-term debt on the balance sheet reached $128.9 billion as of June 30, 2026, up from $65.6 billion at year-end 2025, reflecting the bond market activity Amazon has used to fund its infrastructure commitment.
The Infrastructure Race AWS Is Running Against
Amazon’s CapEx raise to $220 billion does not exist in isolation. The four major hyperscalers — Amazon, Microsoft, Google, and Meta — are collectively projected to spend approximately $725 billion on AI infrastructure in 2026, up 77% from roughly $410 billion in 2025, almost entirely to build out AI training and inference capacity. Google Cloud grew 82% year-over-year in its most recent quarter; Microsoft Azure grew 43% in its fiscal fourth quarter.
AWS’s 37% growth came on a substantially larger revenue base than either competitor, and it was paired with what appears to be the largest absolute increase in contracted backlog among the three major cloud platforms, according to available analysis.
For Q3 2026, Amazon guided for revenue between $197 billion and $202 billion, as the Q2 2026 earnings release states. The company noted that comparisons are complicated by the timing of Prime Day, which was shifted to June this year from its traditional July slot; excluding that timing impact, third-quarter year-over-year growth would be nearly 400 basis points higher.
The harder question — whether 2027’s infrastructure demands will require yet another upward revision to the $220 billion guidance — is one Jassy declined to answer definitively. What he was direct about: the backlog growing at triple-digit rates and the assertion that Amazon expects to add significant data center capacity for the next two to three years.
Frequently Asked Questions
What is AWS’s revenue backlog, and why does it matter?
AWS’s remaining performance obligations — contracts customers have signed but Amazon has not yet billed or delivered — reached $496 billion at the end of Q2 2026, up from $364 billion at the end of Q1, a $132 billion single-quarter increase, as CNBC’s earnings coverage confirms. This figure matters because it represents pre-sold capacity: cloud computing commitments that exist on paper but have not yet been converted into infrastructure Amazon can deliver. A backlog growing at triple-digit rates while the company simultaneously warns it cannot meet current demand signals that signed customer agreements are outrunning Amazon’s ability to build the data centers and chips needed to fulfill them. For enterprise buyers, it is an early warning that the window to secure competitive multi-year pricing and capacity agreements may be narrowing.
Will Amazon have enough cloud capacity to meet demand in 2027?
Based on Jassy’s own statement on the Q2 2026 earnings call, the direct answer is: probably not, as CNBC reported. Jassy said he expects the supply-demand imbalance to persist through 2027 and noted that demand for 2028 is already “striking.” Amazon is on pace to double its power capacity by the end of 2027 compared to 2025, and the company said most of the capacity planned for 2027 has already been reserved. Organizations that have not yet committed to multi-year AWS contracts may find the available uncommitted pool has already been allocated by earlier-moving enterprise customers.
Why did Amazon raise its 2026 capital spending from $200B to $220B?
Jassy named one specific cause: higher memory costs, as CNBC reported from the earnings call. High-bandwidth memory — the specialized chip architecture embedded in every AI accelerator — has seen manufacturing capacity diverted from consumer DRAM toward AI data center production, because memory makers earn significantly more revenue per wafer on the AI-grade product. That supply shift has pushed memory contract prices sharply higher in 2026, adding roughly $20 billion to Amazon’s projected infrastructure bill. Amazon’s Trainium custom chip strategy is designed to partially mitigate this exposure, since Trainium’s architecture can route around some of the premium memory dependency that inflates Nvidia GPU costs — but the wider market for server memory components is still subject to the same supply constraint.
What is Amazon Trainium, and does it actually save money at scale?
Trainium is Amazon’s proprietary AI accelerator chip, built to run AI model training and inference workloads as an alternative to Nvidia GPUs, as prior TechTimes coverage of AWS’s silicon strategy documents. The current generation, Trainium3, is manufactured on TSMC’s 3-nanometer process and delivers comparable compute performance to Nvidia’s best rack-scale systems at roughly 50% lower total cost of ownership according to third-party analysis. The strongest indicator that the cost savings are real rather than claimed: AWS operating margins expanded to 39.4% in Q2 2026, up from 32.9% a year earlier — a 650-basis-point increase — even as the company raised its overall capital expenditure. AWS is gaining margin share while spending more, which is the financial signature of a chip cost advantage flowing through the income statement. The tradeoff is a software porting requirement: Trainium only delivers its advantages through Amazon’s proprietary Neuron SDK, not through Nvidia’s CUDA ecosystem, which means enterprises with existing CUDA-based AI workflows face a real migration cost before they can access the savings.
https://www.techtimes.com/articles/322572/20260731/aws-backlog-hits-496b-amazon-raises-ai-spend-220b-capacity-runs-short.htm
