AWS Accelerates to 37% and the Backlog Hits $496B; Strip the One-Timers and North America Retail Margin Went Backwards

Results vs. Consensus

Amazon beat on every operating line and beat its own guidance by a wide margin, but the headline EPS number is not an operating result. Of the $80.9B of pre-tax income, $53.4B came from non-operating other income, primarily marks on the Anthropic investment. No consensus compiler models unrealized investment gains, so the reported +216% EPS surprise compares two different things. The operating beat underneath is real and is concentrated almost entirely in AWS.

Year-over-Year Comparison

Sequential Comparison

Quality of Beat

• Revenue: Clean and organic, with one large timing caveat. Growth was 20% both as-reported and excluding foreign exchange, so this is not a currency artifact. But Prime Day moved into Q2 this year for most large countries including the U.S., having sat entirely in Q3 in 2025. Management quantified the effect only on the Q3 side (about 400 bps). Applying a symmetric adjustment to Q2, underlying growth was closer to 15.6% than 19.6%, which is still an acceleration off Q1 but a far less dramatic one. Paid units grew 17% and third-party seller mix held at 61%, so the unit engine is genuinely healthy underneath the timing noise.
• Margins: The consolidated story is entirely AWS. AWS operating income of $16,621M was 61% of consolidated operating income on 21% of revenue, and AWS margin of 39.4% cleared the top of the published estimate range (30.9%–38.2%) even before adjusting for the derivative gain. Excluding that gain, AWS margin was roughly 38.0%, which management framed as 520 bps of year-over-year expansion. North America is the offset: excluding the tariff refund, segment operating income was about $8,523M on $116,177M of revenue, a 7.3% margin against 7.5% in Q2 2025, and 13% growth rather than the reported 21%. International margin was flat year over year at 4.1%.
• EPS: Not comparable as reported. Removing the $53,415M of other income and taxing the remainder at a normalized 20% produces about $2.01 of underlying EPS. Applying the identical method to Q2 2025 ($20,857M pre-tax less $1,117M of other income, taxed at 20%, on 10,806M shares) gives $1.46, so underlying EPS grew roughly 38%. That is a strong result and a genuine beat against the $1.82 consensus, but it is one-sixth of the headline. Note the effective tax rate itself is distorted: the reported 22.5% rate sits on top of $17,691M of deferred taxes, while cash taxes paid fell 44% year over year to $2,655M.
• Free cash flow: The weakest line in the release and the one that deteriorated most. Trailing-twelve-month free cash flow of ($7,604M) compares to +$18,184M a year ago, a swing driven by trailing capex of $169,007M (+64%) outrunning operating cash flow of $161,403M (+33%). Operating cash flow is not the problem; capital intensity is.

Segment Performance

AWS

This is the quarter's entire story and it is a very good one. Revenue of $42,232M grew 36.8% as reported and 37% excluding foreign exchange, the fifth consecutive quarter of acceleration and the fastest rate since 2021. The sequential dollar add of $4,645M is the datapoint that matters more than the percentage: it is roughly 80% larger than any prior sequential increase in AWS history, which means the absolute capacity being monetized is stepping up, not just the growth optics against an easy comparison.

"Revenue growth of 36.7% year-over-year, accelerating for the fifth straight quarter, our fastest growth in 18 quarters back when AWS was less than half its current revenue size. We added over $4.6 billion in revenue quarter-over-quarter, about 80% more than our largest increase ever. Our backlog stands at $496 billion, growing triple digits year-over-year." — Andy Jassy, CEO

Margin is the second surprise. At 39.4% reported and roughly 38.0% excluding the derivative gain, AWS printed above the top of the entire published range of sell-side estimates. The mechanism management described is unglamorous and therefore credible: capacity optimization, software and process efficiency, custom silicon displacing merchant parts, and fixed-cost discipline. The bear framing entering this print was that AI workloads carry structurally lower margins than core cloud, and this quarter is direct evidence against it, at least at current utilization.

Assessment: AWS is now a business growing 37% at a 38% clean operating margin on a $169B run rate. That combination does not exist elsewhere in the megacap complex. The mix shift alone is a multi-year consolidated margin tailwind: every point of revenue mix moving from retail to AWS is worth roughly 30 points of incremental segment margin. The one thing this quarter does not resolve is what happens to that margin when depreciation from the 2026 capex vintage lands in 2027.

North America Retail

Reported results look strong: revenue up 16%, operating income up 21% to $9,123M, margin of 7.9%. The footnote changes the read. The roughly $600M of tariff-related refunds sits entirely in this segment, and it is explicitly non-recurring; management characterized it as "the significant majority of refunds we expect to receive." Excluding it, segment operating income was about $8,523M, growth was 13% rather than 21%, and margin was roughly 7.3% against 7.5% in Q2 2025. On a clean basis, the North America margin went backwards year over year.

Management's own cost commentary is consistent with that reading. The segment absorbed higher fuel costs tied to Middle East conflict and higher line-haul rates from driver capacity constraints, partially offset by an FBA fuel and logistics surcharge implemented in April. The productivity story is intact but it is running against cost inflation rather than compounding on top of it.

"We continue to lower our overall cost to serve, even as we face heightened transportation costs driven by fuel inflation from the conflict in the Middle East and higher line haul rates from driver capacity limitations. Excluding the impact of higher fuel and line haul rates, shipping costs grew more slowly than worldwide unit growth at a pace that is relatively consistent with last quarter." — Brian Olsavsky, CFO

Assessment: The underlying fulfillment algorithm still works: shipping cost growth below unit growth, on management's own ex-fuel framing. But the reported margin is being carried by a one-time refund, and the segment is now the drag on the consolidated story rather than the second engine the Q1 print suggested it was becoming. This is the pillar to watch next quarter, and one more clean-basis backslide would change the rating.

International Retail

Revenue of $42,197M grew 15% excluding foreign exchange with operating income of $1,717M, up 15%. Margin held at 4.1% against 4.1% a year ago. Growth accelerated from 11% ex-FX in Q1 to 15% here, which is the more interesting number, and it came without margin give-back. This segment has stopped being a credibility problem and has become a small, steadily improving contributor. It is 21% of revenue and 6% of operating income, so it does not move the thesis in either direction.

Assessment: Fine, and quietly better than it looks. The acceleration to 15% ex-FX with flat margin suggests the selling-fee and selection investments in Europe and Brazil are being absorbed rather than funded out of the P&L. No thesis implication yet; a second quarter of accelerating growth with stable margin would start to matter.

Advertising

Advertising is inside the North America and International segments rather than broken out, but at $19,809M it grew 26% excluding foreign exchange after four consecutive quarters pinned at 22%. That is the first acceleration in the ad line in a year, and management tied it directly to agentic shopping surfaces rather than to sponsored-products volume alone. On a trailing basis the ad business is now running above $75B.

"Increasingly more shoppers are discovering products in our agentic and conversational experiences, including in Alexa+ and Alexa for Shopping. Shoppers who click a sponsored prompt convert to a sale 48% more often and spend 21% more on average than those who don't." — Andy Jassy, CEO

Assessment: The most underpriced line in the company, and the acceleration is the strongest evidence yet that agentic shopping is additive to ad monetization rather than a threat to it. The conversion and basket-size statistics on sponsored prompts imply the new surfaces monetize better than the old ones, which is the opposite of what the search-disruption bear case predicted. No margin disclosure, so this remains an inference rather than a measurement.

Key KPIs

Key Topics & Management Commentary

Overall Management Tone: The most forward-leaning call of the past five quarters, and unusually specific where prior calls were framework-level. Management pre-empted the return-on-capital question with an unprompted, quantified explanation of the capex cycle before a single analyst asked, and volunteered both one-time benefits rather than letting the Street find them. Where the tone thinned noticeably was on funding: the direct question about sources of capital for the buildout drew a deflection, and neither the doubling of long-term debt nor the $24.4B of cash that went into non-marketable investments during the quarter was addressed anywhere on the call.

1. AWS: The Fifth Acceleration, and What Is Actually Driving It

The natural skeptical read on a fifth straight acceleration is that it simply reflects capacity coming online, which would make it a supply story with a ceiling rather than a demand story with runway. Management pushed back on that framing with a specific mechanism: AI workloads are pulling core consumption with them, because post-training, reinforcement learning, and agentic tool use run predominantly on CPUs rather than accelerators.

"Growth in AI drives core because post-training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators. This is an advantage for AWS, as our Graviton chip is the strongest CPU chip, offering up to 30%-40% better price performance than other options." — Andy Jassy, CEO

The supporting evidence is in the mix. AI revenue run rate crossed $25B, up from over $15B one quarter earlier, while total AWS revenue added $4.6B sequentially. If AI alone accounts for roughly $2.5B of that sequential add, core is still contributing over $2B a quarter, which is a faster core-only growth rate than AWS reported in aggregate two years ago. The CFO made the same point in plainer language: "As customers invest in AI, we see a corresponding increase in core consumption."

Assessment: The flywheel argument is now supported by disclosed numbers rather than asserted. This matters because the durability of the AWS re-rating depends on whether the acceleration is a capacity artifact that normalizes or a demand shift that compounds. The evidence this quarter points to the latter, and it is the strongest single support for maintaining the rating.

2. The $496B Backlog and the 2028 Demand Signal

Backlog went from $364B excluding the Anthropic deal in Q1 to $496B, growing triple digits year over year. Against trailing-twelve-month AWS revenue of $148.4B, that is a 3.3x forward book. The more striking disclosure was on timing: capacity for 2027 is largely reserved, and management described 2028 demand as already visible.

"Apart from what we've talked about in 2026, the lion's share of capacity in 2027, we're adding a lot of capacity, as I mentioned just a few minutes ago, is largely reserved, and we have quite a bit of capacity that's already been reserved for 2028. ... In fact, the demand we already have for 2028 is striking." — Andy Jassy, CEO

Management also framed the long-run ambition more aggressively than at any prior point, moving from "a few hundred billion dollars" to a trillion-dollar aspiration. That is a statement about direction rather than a forecast, and should be read as such, but the backlog is not: it is contracted revenue, and it is the single hardest number underwriting the capex program.

Assessment: The backlog is what separates this capex cycle from a speculative one. A company spending $220B against a $496B contracted book with capacity sold out through 2027 is making a working-capital decision, not a bet. The residual risk is concentration: management has not disclosed how much of the $496B sits with the two AI labs versus the enterprise base, and the Q1 "reasonable breadth" language was not repeated this quarter.

3. Custom Silicon: Chips and AI Both Cross $25B

The chips business now runs above $25B annualized, growing triple digits, with Graviton deployed at 98% of the top 1,000 EC2 customers and Graviton revenue commitments up nearly three times sequentially. Trainium carries multi-year, multi-gigawatt commitments from both Anthropic and OpenAI, plus a lengthening list of smaller adopters.

"We are unusually well-positioned for this AI inflection, given our leading price-performance chips in both AI with Trainium and CPU with Graviton. ... Graviton is used by 98% of our top 1,000 EC2 customers. The revenue commitments have increased nearly three times quarter-over-quarter, and Graviton5 is growing nearly 2x faster as Graviton4 did." — Andy Jassy, CEO

One framing change deserves attention. Last quarter management quantified the chips business two ways: an internal run rate and a $50B standalone equivalent "if our chips business was a standalone business and sold chips produced this year to AWS and other third parties as other leading chip companies do." This quarter only the direct run rate was given, with no standalone-equivalent figure. The number moved from over $20B to over $25B on the comparable basis, so the growth is real, but the more flattering framing was retired without comment.

Management also went out of its way to reaffirm the Nvidia relationship, noting customers "will run on Nvidia for as long as we can foresee." That is the correct posture commercially and a useful corrective to the idea that Trainium displaces merchant silicon rather than supplementing it.

Assessment: The silicon pillar is confirmed and strengthened. The margin evidence is now visible in the AWS line rather than resting on management assertion: a 38% clean segment margin while growing 37% is what a cost-advantaged compute stack looks like from the outside. The retired $50B framing is a minor negative signal on disclosure discipline, not on the business.

4. Capex Goes to $220B, and Memory Is the Reason

The single largest guidance change on the call was the capital plan moving from approximately $200B to approximately $220B, attributed entirely to component cost inflation rather than to incremental capacity.

"We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from our prior estimate of about $200 billion. Even at that amount, 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." — Andy Jassy, CEO

This is a $20B cost increase for the same physical footprint, which is a straightforward margin and return-on-capital headwind unless it is recovered in pricing. Pressed on exactly that, management confirmed that existing contracts are locked and that only new agreements reprice: "The deals that you sign, those will be the prices and those will be the agreements that we have over the duration of that contract." With capacity through 2027 largely reserved under contracts signed before the memory spike, the recovery window on this cost increase is slow.

Assessment: The most concrete negative in the print, and it is being under-weighted because it arrived alongside a 37% AWS number. A $20B cost increase absorbed largely inside a pre-contracted book compresses returns on the 2026 vintage specifically. Amazon's answer is that it is still capacity-constrained, so the units are worth building at the higher price. That is probably right, and it is still worse than the plan management laid out ninety days ago.

5. The ROIC Framework: What Management Is Asking the Street to Underwrite

Management opened with an unprompted, unusually quantified defense of the capital cycle, separating the spend into two components with different economics.

"For servers and networking equipment, on average, it takes a little less than three years to break even on that investment. The servers currently have a useful life of at least five to six years, and most of our AI capacity these days is being contracted for at least five-year terms. ... For our data centers, which have 30-plus-year useful lives, we should get at least five to six generations of server economics." — Andy Jassy, CEO

The claim is falsifiable in a way the Q1 version was not. Last quarter the framing was a six-to-24-month monetization lag; this quarter it is a specific sub-three-year break-even against five-to-six-year asset lives and five-year customer contracts. Management also conceded the near-term consequence directly, saying the company will "encounter free cash flow headwinds until these data centers come online."

Assessment: Getting the specific break-even period on the record is the most useful disclosure of the call, because it converts an article of faith into a testable claim. If servers break even inside three years against five-year contracted terms, the unit economics work and the current free-cash-flow trough is a timing phenomenon. The one thing the framework does not address is the reported P&L: break-even on a cash basis says nothing about what accelerating depreciation does to operating margin in 2027.

6. The $53.4B Anthropic Mark, and the Larger One Hiding in Equity

Net income of $62,647M includes $53,415M of pre-tax non-operating other income, attributed primarily to the Anthropic investment. That alone would be the story in most quarters. The larger number is in the equity section: other comprehensive income included $41,988M of net unrealized gains on available-for-sale debt securities, net of $13,695M of tax, implying roughly $55.7B pre-tax. Accumulated other comprehensive income went from $28,230M at December 31 to $66,287M, and "Other assets" grew from $122,607M to $284,132M over the same six months.

Neither Anthropic mark was mentioned on the call. Not in the prepared remarks, not in the CFO's segment walk, and not in any of the six analyst exchanges. A gain larger than the company's entire trailing-twelve-month operating income passed through the financial statements without a single sentence of commentary.

Assessment: The accounting is defensible and the economics are probably real, but the silence is a governance signal worth logging. Marks work both ways: the same methodology that produced roughly $109B of combined pre-tax gains this quarter will produce losses of comparable scale if private AI valuations compress, and the Street has no framework for that because management has never provided one. This is the second consecutive quarter the mark has gone unexplained.

7. Free Cash Flow Turns Negative, and the Cash Number Is the Flattering One

Trailing-twelve-month free cash flow of ($7,604M) against +$18,184M a year ago is the headline. The supplemental disclosure is more informative: the increase in property and equipment acquired but not yet paid was $10,700M in the quarter and $29,267M over the trailing twelve months, against a $1,600M decrease in the year-ago quarter. Adding that back to trailing net capex of $169,007M puts accrual-basis capital spending near $198.3B, which against $161.4B of operating cash flow implies free cash flow closer to ($36.9B).

Assessment: Free cash flow is worse than the Q1 model contemplated and worse than the headline suggests. Our Q1 estimate put FY26 free cash flow in a ($5B) to +$5B band; the trailing figure is already below that at the halfway mark and the second half carries a heavier capex load. This does not break the thesis, because the backlog underwrites the spend and operating cash flow is compounding at 33%, but it removes any remaining argument that the capex is self-funding on a current-period basis.

8. Retail Margin: The Pillar That Cracked

The Q1 thesis held that retail margins were inflecting as fulfillment automation scaled, and the Q1 print supported it with North America operating income of $8.3B against $5.8B a year earlier. This quarter breaks the pattern once the tariff refund is removed. Clean North America margin of roughly 7.3% against 7.5% in Q2 2025 is the first year-over-year decline in this coverage.

The productivity investment continues: robotic arm fleet expected to more than double in 2026, facilities being retrofitted with current-generation technology, inventory placement and consolidation-rate gains. But management's own hedge was notable, and more cautious than anything said in Q1.

"While operating margin may fluctuate and progress may not always be linear, we take a deliberate approach to achieving sustained long-term improvement in our cost to serve." — Brian Olsavsky, CFO

Assessment: One quarter of clean-basis margin decline against a genuine external cost shock (fuel, line-haul rates) is not a broken thesis, and the surcharge response was appropriate. But this pillar moves from confirmed to at risk. The reason it matters disproportionately is that it was the pillar that made AMZN more than an AWS proxy. Without it, the entire investment case rests on cloud.

9. Everyday Essentials: The Same-Day Perishables Unlock

The most substantive retail disclosure was operational rather than financial. Amazon appears to have solved the perishables problem by routing fresh grocery through existing same-day facilities rather than building dedicated formats, and the engagement data is striking.

"We have finally found something that is a real needle mover for us in offering perishables in our same-day facilities that we're now able to offer same-day perishables in 2,300 cities around the U.S. If you look in those cities, nine of the top 10 best sellers in those geographies are perishables. The number of monthly active perishable customers has increased 50% since the start of this year." — Andy Jassy, CEO

Same-day orders containing perishables average over three times more units per order, which is the number that converts this from a category expansion into a basket-economics story. Amazon is already the second-largest U.S. grocer at over $150B of gross merchandise sales, so the incremental frequency is landing on a large base.

Assessment: The most credible new source of retail operating leverage in the print, and the strongest argument that the margin pillar is at risk rather than broken. Higher units per order against a fixed same-day delivery cost is exactly the density mechanism that made the fulfillment network profitable in the first place. It is not visible in this quarter's segment margin, and it should be within two to three quarters if the mechanism is what management describes.

10. Alexa, Agents, and the Application Layer

Amazon is pushing up the stack on both sides of the house. On the consumer side, Alexa for Shopping reached over 350 million users in the last twelve months with active users nearly doubling and interactions up over 5x year over year. On the enterprise side, Amazon Q, Kiro, Connect, Transform, and the newly launched Continuum represent a deliberate move from selling infrastructure to selling applications, with a named customer list including 3M, Allianz, AstraZeneca, BMW, Exxon, and the NFL.

Management was explicit that this is customer-driven rather than strategy-driven, describing Q as an internal tool that external enterprises asked for after Amazon built it for itself. The same is true of Kiro, which tripled in usage sequentially and is claimed to be up to 50% more cost-effective than alternatives.

Assessment: Strategically sound and financially invisible. None of these carry disclosed revenue, so they are optionality rather than model inputs. The relevant near-term read is that the application layer raises AWS switching costs at exactly the moment when inference is becoming commoditized, which supports the durability of the core growth rate more than it supports any new revenue line.

11. The Balance Sheet: Debt Doubles, and $24.4B Leaves Quietly

Total assets crossed $1 trillion for the first time, reaching $1,095,689M against $818,042M at December 31. Long-term debt roughly doubled from $65,648M to $128,894M, with $66,998M of proceeds raised in the first half, and quarterly interest expense rose 155% to $1,314M. Separately, the investing section shows $24,359M in the quarter and $39,767M in the first half under "acquisitions, net of cash acquired, non-marketable investments, and other," against $1,700M in the year-ago quarter.

That last line is a real cash outflow of the same order as a quarter of AWS revenue, and it sits outside capex entirely. It was not mentioned in the prepared remarks or in any answer.

Assessment: Amazon is now funding the buildout with debt and deploying substantial additional cash into non-marketable investments, and neither was volunteered. The leverage is easily serviceable against $161B of operating cash flow, so this is a disclosure concern rather than a solvency one. But an investor modeling cash needs from capex alone is missing roughly $40B of first-half outflow.

Guidance & Outlook

Amazon cleared its own Q2 revenue guide by $1.6B above the high end and its operating income guide by $3.5B above the high end, a 14.4% overshoot on the profit line. That is the widest beat against guidance in the coverage period and it is the context in which the Q3 range should be read.

The Q3 revenue guide looks alarming at first glance: 9% to 12% growth against the 20% just delivered. Management identified both drivers explicitly. Prime Day shifted into Q2 this year having sat entirely in Q3 in 2025, worth nearly 400 bps to the Q3 comparison, and foreign exchange is a further 80 bps headwind.

"Excluding the impact of Prime Day in both 2025 and 2026, third quarter 2026 year-over-year growth would have been nearly 400 basis points higher. ... Second, the Q3 guidance anticipates an unfavorable impact of approximately 80 basis points from the year-over-year changes in foreign exchange rates based on current rates." — Brian Olsavsky, CFO

Implied quarter-over-quarter trend: Management quantified only the Q3 side of the Prime Day shift. Applying a symmetric adjustment to Q2, underlying growth was roughly 15.6% in Q2 against roughly 14.7% at the Q3 guide midpoint and 16.1% at the high end. The underlying trend is therefore approximately flat in the mid-teens, not the nine-point cliff the headline comparison implies. That symmetric adjustment is our estimate, not a company figure.

Street at: Consensus entering the print sat at roughly $196.9B of revenue and $1.82 of EPS. The Q3 revenue guide midpoint of $199.5B lands close to where the Street already was for Q2, which means the sequential setup is effectively flat on the top line. The operating income guide is the more interesting number: a $24.5B midpoint is 41% above the year-ago quarter and implies a 12.3% margin against 9.7%, so the profit trajectory is guided to keep expanding even through the Prime Day give-back.

Guidance style: Amazon guides operating income in a deliberately wide band, $4B here, that brackets a conservative midpoint. The company has now exceeded the high end of its own operating income guide, not merely the midpoint, and did so by 14%. On that pattern the realistic Q3 outcome sits at or above the $26.5B high end rather than at the $24.5B midpoint, which would be roughly $2B ahead of where a midpoint-anchored model lands.

Analyst Q&A Highlights

AWS Margin at 39% and Whether It Holds

The opening question went straight to the quarter's most surprising number, and framed it against the prevailing assumption that AI workloads dilute cloud margins near term. Management's answer separated the result into its components and, notably, volunteered the ex-item figure rather than defending the headline. The hedge was consistent with prior quarters: margins fluctuate with investment levels and product mix.

Q: "Brian, many have assumed that AI workloads would be lower margin, at least near term. Can you just talk about the drivers of the 39% AWS operating margin in 2Q and just how we should think about sustainability?"
— Doug Anmuth, J.P. Morgan

A: "Yes, we're pleased with the growth in both revenue and also margin expansion that we had in AWS in Q2, especially given the size of our business. You're seeing, despite the large investments, AWS margins have continued to remain strong, and we're up 650 basis points year-over-year. 520 basis points if you exclude the derivative accounting gain that I mentioned. We've said before, these margins will fluctuate. They're based on a number of factors, including our investment levels Mix of products, mix of AI versus non-AI. I would say that the profitability you're seeing from AWS isn't random. It's a result of disciplined efficiency gains, capacity optimization, which we benefited quite a bit from in Q2, and always closely managing our fixed costs."
— Brian Olsavsky, CFO

Assessment: Management answered honestly and did not claim sustainability, which is the correct posture. The unprompted disclosure of the 520 bps ex-item figure is a credibility gain. What went unanswered is the 2027 depreciation question sitting behind it, and no one on the call asked.

Whether Amazon Needs Its Own Frontier Model

A structural question about the AI stack: if the model layer is where value accrues, is a company that resells other people's frontier models structurally disadvantaged? The answer was a two-part position that has shifted since prior quarters. Management argued the business does not require a frontier model, then disclosed that Amazon is building one anyway, with cost control, prioritization, and speed as the stated reasons rather than competitive necessity.

Q: "Andy, strong Amazon Bedrock traction with customers spending more in the quarter than in all the prior quarters combined. When you think about the full stack offering, does Amazon need its own leading model toward the frontier?"
— Doug Anmuth, J.P. Morgan

A: "My view of it is that AWS and Amazon can have a wildly successful business without its own frontier model. A lot of that is because there is not going to be one model to rule the world. ... All that said, we are pursuing our own frontier model, and we're doing it for a few reasons. First of which is it just gives us additional control over cost. ... My view of it is that within the next few years, you're going to have at least a half dozen models that are comparably good to each other. They'll all be in Bedrock, one of them will be ours."
— Andy Jassy, CEO

Assessment: The commoditization thesis is self-serving but probably correct, and it is the strategically comfortable position for whoever owns the distribution layer. The concrete new information is the frontier-model commitment, which carries training costs that are not separately disclosed and are presumably inside the technology and infrastructure line that grew 22% this quarter.

Whether the Acceleration Is Demand or Just Capacity Landing

The most important skeptical question of the call, and the one that determines whether the 37% growth rate is extrapolable. If the acceleration simply reflects capacity coming online, growth is supply-gated and will normalize; if demand is the constraint, it compounds. Management answered by listing demand-side drivers first and treating capacity as one contributor among several, then reaffirmed the power-capacity commitment made two quarters ago.

Q: "Just thinking about the AWS acceleration, was that really driven by a lot of capacity coming online in the quarter? You guys have been more open than your peers on gigawatts you're adding. Any help you can give us on how much you might be adding in the second half versus the first half, also how you're thinking about 2027?"
— Justin Post, Bank of America

A: "Yes, we're adding a lot of capacity, there are a lot of other reasons why it's growing. We're on pace with the capacity build that we talked about a few quarters ago, where we said we expect to have double the power capacity by the end of 2027 that we had in 2025, and we continue to be on that track."
— Andy Jassy, CEO

Assessment: A partial answer. Management reaffirmed the aggregate power commitment but declined the specific request for first-half versus second-half capacity additions, which is the number that would let the Street separate supply from demand. The refusal is understandable competitively and unsatisfying analytically. The demand-side argument rests on the backlog, which is disclosed, so the question is answerable indirectly.

Whether the Long-Lived Data-Center Spend Can Slow in 2027

The question every capex-cycle investor needs answered: is there a visible off-ramp, or does the spending compound indefinitely? Management did not offer one. The answer instead reframed the demand curve as barbelled, with AI labs and a handful of runaway applications at one end, enterprise cost-avoidance use cases at the other, and the large middle of existing production workloads still largely unconverted.

Q: "As you look into 2027, you look at the demand that's coming, et cetera, are you at a point where you're going to be able to start to slow that long-lived data center spend at all in 2027, or is that just too soon where you're still going to have to be opening up new data centers for the next two, three, four years as you look into 2027?"
— Brian Nowak, Morgan Stanley

A: "We have so much demand right now. Apart from what we've talked about in 2026, the lion's share of capacity in 2027, we're adding a lot of capacity, as I mentioned just a few minutes ago, is largely reserved, and we have quite a bit of capacity that's already been reserved for 2028. ... I don't know if the trajectory of that middle part of the barbell will be the same wildly steep trajectory that we've seen with the current barbell AI labs piece. We have a lot of demand in front of us, and we're going to invest in this business to continue to be the significant market segment leader that we are today."
— Andy Jassy, CEO

Assessment: The answer is no, and it was given plainly. Investors should model elevated capital intensity through at least 2028. The barbell framing is the most intellectually honest thing said on the call, because it concedes uncertainty about the largest segment of future demand rather than extrapolating the AI-lab growth rate across the enterprise base.

Selling Trainium Outside AWS

The largest piece of unpriced optionality in the story, and the language has firmed since last quarter without becoming a commitment. Prior guidance described a good chance of selling racks within a couple of years; this quarter management described active conversations with customers who want the chips independent of the cloud.

Q: "In the past 90 days or so, the company talked publicly about selling Trainium at some point to third-party data centers. How do you think about when you could do that and just sort of the ROIC on that versus core AWS loads?"
— Brian Nowak, Morgan Stanley

A: "We do have an increasing number of customers who are interested in us providing the Trainium chips to them, separate from our cloud, and we're actively having those conversations and exploring, and I expect there's a real chance we'll do that in the future."
— Andy Jassy, CEO

Assessment: The demand-side evidence firmed (customers are asking) while the timing language loosened (from a couple of years to the future). The return-on-capital half of the question, which is the harder one given that internal Trainium deployment captures the full AWS margin stack, went unaddressed. Treat this as optionality with no date attached rather than a modelable revenue line.

Funding the Buildout

With capex guided to $220B against negative trailing free cash flow, the funding question is now first-order, and it drew the thinnest answer of the call. Long-term debt had already roughly doubled since December and interest expense had risen 155% year over year, neither of which was raised by management on the call.

Q: "Brian, just given the demand signals you guys are both talking about and plans for additional capacity, what are your current thoughts on sources of capital for the build-out over the next couple of years?"
— Colin Sebastian, Baird

A: "You've seen us issue debt this year. We have a lot of options available to us as we continue to fund this growth that we're seeing in AWS. We'll continue to look at all the options and make the appropriate decision at the right time, but nothing to share today."
— Andy Jassy, CEO

Assessment: A deflection, and the clearest one on the call. The question was directed to the CFO and answered by the CEO in two sentences with no quantification of debt capacity, target leverage, or whether equity is contemplated. Given that the balance sheet moved $63B of incremental long-term debt in six months, this deserved a real answer and did not get one.

Backlog Durability and Pricing Against Cost Inflation

The sharpest analytical question of the call: if remaining performance obligations have grown to roughly two and a half times the level at which the current capacity plan was set, does the plan need to move again, and can pricing recover the memory-driven cost increase? The answer confirmed the uncomfortable half, which is that signed contracts are fixed and only new agreements reprice.

Q: "Your RPO reported as two and a half times that of the third quarter of 2025 when you gave us the doubling of capacity comments for year-end 2027. How does that RPO number and the massive expansion there impact your outlook for future capacity? ... Could you talk about how your pricing strategy at AWS incorporates future cost inflation? Do your longer-term contracts allow for stable return profiles despite cost inflation?"
— Ken Gawrelski, Wells Fargo

A: "The deals that you sign, those will be the prices and those will be the agreements that we have over the duration of that contract. New agreements that you sign, you always take into account what your costs are and how you ultimately build a price that you agree to with your customers. I think it's no secret right now to any company in the world that there are inflated prices right now on some of the components like memory and hard drives and SSDs."
— Andy Jassy, CEO

Assessment: The exchange establishes that the $20B memory-driven capex increase lands against a book of contracts signed at pre-inflation prices, with recovery available only on new business. That is a quantifiable drag on the return profile of the 2026 and 2027 capacity vintages specifically. It is the most important thing learned in Q&A and it received no follow-up.

What They're NOT Saying

1. 2027 depreciation: Two consecutive calls have now passed without management framing what depreciation from the 2026 capex vintage does to reported operating margin in 2027. Approximately $220B of capital entering service against a five-to-six-year server life implies a step-change in the depreciation line, and the cash-basis break-even framework management offered does not speak to the P&L at all. This is the single largest gap between the bull framing and the reported financials.
2. Backlog concentration: The $496B backlog was disclosed with a growth rate but no breadth commentary. Last quarter management characterized the book as having "reasonable breadth" beyond a small handful of customers; that language was not repeated. With Anthropic and OpenAI both carrying multi-gigawatt Trainium commitments, the share of backlog attributable to two counterparties is a material unknown.
3. The Anthropic marks: Roughly $53.4B through the income statement and roughly $55.7B pre-tax through other comprehensive income, and not one sentence about either on the call. No methodology, no sensitivity, no indication of how future periods will be communicated. The asymmetry matters: the same accounting produces losses of similar magnitude if private AI valuations compress.
4. The $24.4B of non-marketable investment outflow: The cash flow statement shows $24,359M in the quarter and $39,767M in the first half under acquisitions and non-marketable investments, against $1,700M in the year-ago quarter. This is a cash commitment roughly half the size of quarterly capex, entirely outside the capex guidance, and it was never mentioned.
5. Amazon Leo economics: Last quarter's guidance carried an explicit $1B incremental cost drag for Leo in Q2 and commercial service on track for Q3. This quarter the update was that the constellation is close to 400 satellites and service begins "this year," with no cost quantification and a softer timing commitment. The financial disclosure went backwards.
6. The retired $50B chips framing: Last quarter management sized the chips business two ways, including a $50B standalone-equivalent figure. This quarter only the direct run rate was given. The comparable number grew, so the change is presentational rather than substantive, but dropping the more favorable framing without comment is unusual.
7. Free cash flow guidance: Still none, now two quarters into a negative trailing figure. Management addressed the framework and conceded free cash flow headwinds directionally but has never put a magnitude or a trough date on it.
8. Headcount and the AI-productivity story: Employees rose 20,000 sequentially to 1,595,000, up 3% year over year, which is the fastest growth in several quarters and sits awkwardly against last quarter's prominent internal-AI-productivity narrative. Neither the reversal nor its relationship to that narrative was addressed.

Market Reaction

• Pre-print setup: AMZN closed at $235.50 on July 30, up 2.0% year to date and 2.3% over the trailing twelve months, and down 1.2% over the trailing 30 days. The S&P 500 was up 8.7% year to date over the same window, so the stock entered the print having underperformed the index by roughly seven points for the year while AWS growth roughly doubled. The 52-week closing range was $198.79 to $274.99.
• After-hours move: Shares advanced 9.1% to $257.06 in extended trading following the release and the call.
• Reaction session (July 31): The stock opened at $264.99, a 12.5% gap, traded a $262.06 to $273.22 range, and closed at $271.58, up 15.3% or $36.08. That is the largest single-session advance since 2012. Volume of 125.9M shares was 2.4 times the 30-day average. The S&P 500 rose 0.7% on the session.

The move built rather than faded, which is the notable structural feature. After-hours priced a 9% reaction; the regular session added six more points and closed near the day's high. That pattern is the opposite of the Q1 print, where an initial after-hours advance decayed into a 0.77% next-day close, and it indicates the buy side worked through the release overnight and concluded the AWS acceleration was the operative fact rather than the capex number.

The attribution across the wire coverage was consistent and narrow: this was an AWS re-rating. Not one major outlet led with the $5.75 EPS print, which is the correct reading given that the figure is an unrealized investment mark. What repriced was the return-on-capital question. A market that had spent three quarters asking whether $200B of annual capex would ever produce revenue got a 37% growth rate, a $496B backlog, and a 38% clean segment margin in the same release, and paid for it.

Context tempers the enthusiasm slightly. Despite the largest one-day move in fourteen years, the stock closed at $271.58 against a pre-print 52-week closing high of $274.99, so it did not make a new closing high. The 15.3% recovers part of a year-long de-rating rather than extending a run, which is why the setup after the move is less stretched than a headline percentage of that size would normally imply.

Street Perspective

Debate: Is 37% AWS Growth Extrapolable or a Capacity Artifact?

Bull view: Five consecutive quarters of acceleration on a base that has roughly doubled since the last comparable growth rate cannot be a comp artifact. The $496B backlog is 3.3 times trailing AWS revenue and grew triple digits, capacity is largely reserved through 2027, and management describes visible 2028 demand. The constraint is supply, and supply is being added on a disclosed schedule.

Bear view: Growth that is supply-gated has a ceiling set by the capex schedule, not by demand, which means the deceleration point is knowable and arrives when the power-capacity doubling completes at the end of 2027. Management declined to give first-half versus second-half capacity additions, which is the disclosure that would settle the question.

Our take: The bull has the better of it, but on the backlog rather than the growth rate. Contracted revenue of $496B against $148B of trailing revenue is the fact that makes the trajectory durable regardless of which side of the supply-demand question is binding. The bear point about the 2027 capacity cliff is real and under-discussed, but it is a 2028 question and the market is not paying for 2028 today.

Debate: Does Negative Free Cash Flow Matter, or Is It Purely Timing?

Bull view: Operating cash flow compounded 33% to $161.4B. The entire free-cash-flow deficit is capital entering service against a contracted book, with servers breaking even in under three years against five-year customer terms and thirty-year data-center lives. This is the 2014-2017 AWS build in compressed form, and that cycle ended with multi-year free-cash-flow expansion.

Bear view: The reported ($7.6B) understates the burn. Accrued-but-unpaid equipment grew $29.3B over the trailing twelve months, putting accrual-basis free cash flow near ($37B), and the company is now funding the gap with debt that doubled in six months. Add $39.8B of first-half non-marketable investment outflow and the cash commitment is far larger than the capex guidance describes.

Our take: The bear has the better arithmetic and the bull has the better economics. The accrual adjustment is correct and materially changes the burn, and it is not in most models. But the break-even framework management put on the record this quarter is specific enough to be falsifiable, and the backlog underwrites it. The honest position is that free cash flow gets worse before it gets better and the trough is probably 2027 rather than 2026. That is a duration problem, not a solvency or economics problem.

Debate: Is the Retail Margin Story Broken or Interrupted?

Bull view: The clean-basis margin decline is entirely explained by identifiable external cost shocks, fuel inflation and line-haul rate increases, against which management has already implemented a surcharge. The underlying algorithm still works on management's own framing: excluding fuel and line-haul, shipping costs grew more slowly than unit growth. Robotics deployment doubles this year and the same-day perishables mechanism is a genuine new density lever.

Bear view: The 7.9% reported margin required a one-time tariff refund to reach, and without it the segment went backwards year over year. Shipping costs grew 19% against 17% unit growth on a reported basis. Headcount rose 20,000 sequentially. The automation narrative has been running for three years and the clean margin is now lower than it was a year ago.

Our take: Interrupted, with a shorter leash than the bull case deserves. The external cost attribution is credible and the perishables density mechanism is real and new. But this pillar was the reason to own AMZN as something other than an AWS proxy, and it has now produced one quarter of clean-basis decline. Two would change the rating. The perishables basket economics should be visible in segment margin within two to three quarters if the mechanism is what management describes.

Debate: How Should the Anthropic Mark Be Treated in Valuation?

Bull view: The mark reflects a real economic gain on a genuinely valuable asset. Anthropic's enterprise value has risen with its capacity commitments, Amazon's stake is large, and the combined pre-tax gains flowing through income and other comprehensive income this quarter exceed the company's entire trailing operating income. It is an asset the market is not valuing separately.

Bear view: It is an unrealized mark on a private company, taxed on paper via $17.7B of deferred taxes while cash taxes fell 44%, with no disclosed methodology and no sensitivity analysis. The same accounting produces symmetric losses if private AI valuations compress, and management has provided no framework for that scenario across two consecutive quarters of silence.

Our take: Exclude it from earnings entirely and treat the stake as a balance-sheet item, which is what we have done. Building it into an earnings multiple imports the volatility of private AI valuations into a cloud-and-retail valuation, and the fact that the Street's reaction ignored the EPS line entirely suggests the market agrees. The asset is real; it is simply not earnings.

Model Update Needed

Valuation impact: We are changing the earnings basis. GAAP EPS is now dominated by unrealized investment marks and is not usable as a multiple input, so we move to an ex-mark basis. On FY27 revenue of roughly $950B, a consolidated operating margin of 13.5% to 14.0%, and net interest and tax at current trajectories, we estimate FY27 ex-mark EPS of approximately $9.20 on roughly 11.0B diluted shares. At 32 to 34 times, that supports a fair-value range of $295 to $315 against the $271.58 close, or 9% to 16% upside. The bull case, in which Trainium is sold outside AWS on a disclosed timeline, supports $350 and above. The bear case, in which 2027 depreciation compresses consolidated margin toward 12% and takes ex-mark EPS to roughly $8.00, supports $240 at 30 times, or 12% downside. The asymmetry favors ownership but by less than it did at initiation, because the market has now paid for a meaningful part of the AWS re-rating in a single session.

Thesis Scorecard Post-Earnings

Overall: Thesis strengthened where it matters most and narrowed at the same time. The two AWS-linked bull pillars are materially stronger than at initiation, and two of the four original bear points are now actively challenged. But the retail-margin pillar moved to at risk on a clean-basis decline, the free-cash-flow bear point moved from emerging to materializing, and a new bear point on 2027 depreciation has been added. Three months ago this was a three-engine story trading at a capex discount. Today it is an AWS story with two supporting engines, one of which stalled, and the capex discount has been partially paid.

Action: Maintaining Outperform, with a shorter leash. The $496B backlog, the sub-three-year server break-even now on the record, and a 38% clean AWS margin at 37% growth are enough to underwrite the spending and the rating at $271.58 against a $295 to $315 base-case range. Two things would move us to Hold: a second consecutive quarter of clean-basis North America margin decline, or 2027 depreciation guidance that compresses consolidated margin faster than AWS mix can offset. Both are answerable within two quarters, and management has now declined to address the second one twice.