Amazon Q3 2026 Capex, AWS Growth, and 2027 Outlook

Amazon’s Q3 2026 print is a capacity story financed with debt: AWS grew 39% to a $184 billion run rate with a $578 billion backlog, while cash capex hit $75.9 billion in the quarter and the company still expects about $220 billion for 2026—and “meaningfully more” in 2027. The release attributes the free-cash-flow swing to AI infrastructure, but management did not break spend into Trainium, Nvidia GPUs, and power.

Quarterly cash capex, and what is not in that number

Amazon’s cash capex is purchases of property and equipment net of proceeds from sales and incentives—not gross cash outlays, and not finance-lease additions. On that definition, CFO Brian Olsavsky said Q3 cash capex was $75.9 billion and “primarily relates to AWS and AI.” The cash-flow statement shows the gross figure underneath that: purchases of property and equipment of $77.259 billion, against $1.381 billion of proceeds from sales and incentives, which nets to about $75.9 billion.

Finance leases are disclosed separately and are not in cash capex. Property and equipment acquired under finance leases, net of remeasurements and modifications, was $1.472 billion in the quarter (versus $977 million a year earlier). Principal repayments of finance leases were $433 million and sit in financing cash flow, not investing. Operating-lease assets acquired were $7.407 billion. Adding finance-lease commencements to cash capex would overstate cash spent; the economic commitment is real, but the cash leaves later as lease principal.

• Gross purchases of PPE: $77.259 billion in Q3 2026 versus $35.095 billion in Q3 2025.
• Nine-month gross purchases: $175.670 billion; proceeds from sales and incentives $3.482 billion, so net cash capex year-to-date is about $172.2 billion.
• Trailing-twelve-month net purchases of PPE: $210.7 billion, up $94.8 billion year over year. The release says that increase “primarily reflects investments in artificial intelligence.”
• Trailing-twelve-month free cash flow swung to an outflow of $42.4 billion from an inflow of $14.8 billion a year earlier. Operating cash flow still rose 29% to $168.3 billion. The gap was funded in part by $48.102 billion of long-term debt proceeds in the quarter. Long-term debt was $173.1 billion at September 30, 2026, versus $65.6 billion at December 31, 2025.

What this means for anyone modeling Amazon: use the $75.9 billion cash-capex figure when comparing to the $220 billion full-year guide. The $77.3 billion gross number that circulated in headlines is before incentives. Finance leases add roughly $1.5 billion of non-cash asset additions this quarter, not another $1.5 billion of cash capex.

Full-year 2026 guide and the 2027 direction

The company reiterated, in both the release and on the call, that it continues to expect 2026 cash capital expenditures of approximately $220 billion, “the majority for AWS.” That figure was raised earlier in the year from about $200 billion, with management previously tying the increase to higher memory costs. It was not raised again on October 29.

The arithmetic of the guide is the tension. About $172 billion of net cash capex is already in the first nine months, so holding $220 billion implies roughly $48 billion in Q4—a sharp slowdown from $75.9 billion in Q3. Management did not walk through that deceleration. Outside commentary (XTB) flagged the guide as conservative relative to the run rate and noted the company gave no indication of when the cash-burn trend reverses. Amazon did not raise the 2026 number on this call, so the official figure remains approximately $220 billion.

For 2027 the language is directional only. Jassy: “We still expect 2026 cash CapEx of approximately $220 billion, the majority for AWS, and expect to spend meaningfully more in 2027 with a clear line of sight to attractive financial returns, as I laid out on our last call.” No dollar range, no growth rate, and no split of incremental spend between chips, power, and buildings. Fourth-quarter operating guidance is $28.5–$33.5 billion of operating income on $240–$247 billion of sales; that does not constrain 2027 capex.

What this means for competing hyperscalers: Amazon is the only one of the three large clouds, on this week’s prints, that both held a very high 2026 number and explicitly pointed 2027 higher without a ceiling. Microsoft trimmed its fiscal-2026 capex view to about $170 billion; Alphabet held $195–$205 billion and also said 2027 would be significantly higher. Amazon’s incremental message is that the binding constraint is still supply, not a desire to protect free cash flow.

AWS growth, backlog, and where AI shows up in the P&L

AWS net sales rose 39% (39.3% in the release) to $46.0 billion, the fastest growth in nearly five years, to a $184 billion annualized run rate. Operating income was $17.8 billion, a 38.7% margin, versus $11.4 billion a year earlier. That is about 63% of company operating income of $28.1 billion. Core compute and storage grew at their fastest rates in several years, so the acceleration is not only an AI overlay on a flat base.

Backlog is the forward indicator management wants investors to use. Jassy said AWS backlog is $578 billion, growing triple digits year over year and “well diversified”—more than double a year earlier, per contemporaneous reporting of the call. He tied the multi-year growth claim to that book: “we believe that AWS can grow north of 40% in each of 2027 and in 2028, and be a trillion-dollar revenue business in the next 5 to 10 years.” The earnings release used “over 40%” for both years and the $184 billion run rate.

On the mix of AI revenue, the only quantitative split disclosed is inference versus training, not chip vendor versus custom silicon. Jassy: enterprises are “still very early in using inference at scale, and yet inference already accounts for well over half of AWS’s AI revenue and growing,” with “a lot of training revenue as well.” Bedrock token volume grew nearly 17 times year over year; Fortune 500 spend on Bedrock was nearly 7 times higher. Chips revenue is growing triple digits, and revenue commitments for Graviton and Trainium each increased more than 20 times year over year. Customers using custom silicon saved more than $2 billion in Q3 alone. Trainium will be offered outside AWS late in 2027.

What this means for the growth math: 40%+ growth on a $184 billion run rate for two years is a capacity-addition statement as much as a demand statement. The $578 billion backlog is contracted future revenue, not recognized revenue, and management said the largest blocks—hundreds of megawatts, sometimes up to a gigawatt—are reserved well in advance, which is why they claim visibility into 2027, 2028, and some of 2029.

AI share of spend: Trainium, Nvidia, power

Amazon did not disclose a percentage of capex going to Trainium, Nvidia GPUs, or power. The closest official characterizations are qualitative and should not be turned into a mix model.

• The year-over-year increase in purchases of property and equipment “primarily reflects investments in artificial intelligence” (earnings release).
• Q3 cash capex of $75.9 billion “primarily relates to AWS and AI” (Olsavsky).
• The majority of the ~$220 billion 2026 cash capex is for AWS (Jassy).
• The company is “installing an AI accelerator about every 14 seconds,” and capacity is being absorbed about as fast as it is installed (Jassy). That phrase covers accelerators generally; it is not a Trainium-versus-Nvidia count.
• AWS expanded its Nvidia GPU portfolio in the quarter (CNBC), and new agreements cited in the release include Meta, NVIDIA, and OpenAI, among others. No dollar allocation was given.
• Power is treated as a constraint and a cost, not a disclosed capex line. Guidance assumptions explicitly exclude energy-derivative remeasurements. Component-cost pressure called out on the call was memory, not a power-price figure.

The mechanism management is selling is vertical integration, not a GPU rental book. Jassy contrasted AWS with “primarily hosting GPUs or renting short-term data center capacity,” and said most capacity is owned or long-leased, which he argued carries better operating margins than colocation-heavy models. Trainium and Graviton are the cost lever inside that owned fleet: price-performance of 30–40% better on Graviton versus x86 alternatives, TorchNeuron so existing PyTorch runs on Trainium without a rewrite, and external Trainium sales only in late 2027—so near-term chip spend still mostly serves AWS’s own capacity, not a merchant chip business.

What this means for anyone trying to size the AI bill: “primarily AI” and “majority AWS” are not the same as “90% GPUs.” Fulfillment, retail logistics, and non-AI AWS still sit inside the $220 billion. Without a segment capex table, the honest range is: most of the dollars are AWS infrastructure, the incremental dollars versus last year are mostly AI, and the Trainium/Nvidia/power split was not disclosed on October 29.

Capacity, power, and returns: what Jassy and Olsavsky actually said

On the October 29 call, capacity comments were about demand absorption and reservation lead times, not a restatement of the “sold out” line. Jassy said overall demand for AWS capacity “remains exceptionally strong,” that an accelerator is going in about every 14 seconds, and that customers are taking capacity about as fast as it is installed. In Q&A he said the big blocks—“hundreds of megawatts or sometimes up to a gigawatt”—are reserved well in advance because “you can’t leave a gigawatt of capacity hanging around,” which is the basis for “very good visibility into ’27 and ’28 and some of ’29.” He also said Amazon keeps some capacity for on-demand EC2 and Bedrock rather than selling every megawatt forward; that reservation policy was stated more fully six days earlier, not as a new numeric disclosure on the earnings call.

The sharper power and sold-out language is from October 23, on the a16z podcast, reported by Reuters—not from the earnings release. Jassy said 2026 capacity is sold out, 2027 is “largely sold out,” and “a good chunk of 2028 is sold out,” while AWS still holds some capacity for on-demand, Bedrock, and EC2. On bottlenecks: “There’s only so much data center and power today. There’s only so much memory, and I don’t think that shortage ends in 2027.” He also said Amazon “could invest a lot more and it wouldn’t be a bad use of capital.” Treat that as management’s pre-print view of the constraint set (space, power, memory), confirmed in direction by the call’s “meaningfully more in 2027” and the gigawatt-reservation comments, but not as a figure audited in the 8-K.

On returns, both executives argued payback is shortening even as depreciation and memory costs rise. Jassy: operating margins on AI are “rapidly improving… well ahead of where they were on core at a similar time of evolution, even with heightened component costs like memory and significant depreciation from the CapEx we’re investing. We continue to accelerate our CapEx payback” and have “strong return on invested capital.” He called AI “a once-in-a-lifetime opportunity” and said the company will invest to remain the leader rather than ration a 40% growth path. Olsavsky: the investments, “especially in AI,” are meant to drive “long-term revenue and free cash flow, as well as return on invested capital.” He also said the 38.7% AWS margin is the product of efficiency, capacity optimization, fixed-cost control, and a rising share of custom chips—and that margins will still fluctuate with AI mix and component costs. Neither gave a payback period, an IRR, or an ROIC percentage on this call. (On the prior quarter’s call, Jassy had said servers and networking break even in a little less than three years on average; that specific timing was not repeated as a new Q3 metric.)

What this means for capital allocation: the returns argument is “payback is moving left, backlog is contracted, custom silicon improves the cost stack,” not a published ROIC bridge. The cash-flow statement is the counterweight—TTM free cash flow of negative $42.4 billion and $48 billion of new long-term debt in one quarter—so the return case only works if the $578 billion backlog converts and the 40% growth years actually show up in revenue before depreciation and interest dominate.

How to use the figures

Item	Figure	Source treatment
Q3 cash capex (net of incentives)	$75.9 billion	Olsavsky on the call; equals gross PPE purchases minus proceeds
Q3 gross purchases of PPE	$77.259 billion	Cash-flow statement in the Oct. 29 release
Q3 finance-lease asset additions	$1.472 billion	Supplemental cash-flow disclosure; not cash capex
2026 cash capex	Approximately $220 billion, majority AWS	Release and Jassy; unchanged
2027 capex	“Meaningfully more” than 2026; no number	Jassy
AWS revenue / growth	$46.0 billion, +39.3%	Release
AWS backlog	$578 billion, triple-digit growth	Jassy on the call
AWS 2027–28 growth	North of / over 40% each year	Jassy; also in the release
AI share of capex	Not quantified by Trainium / Nvidia / power	“Primarily” AI for the YoY PPE increase; majority of 2026 cash capex is AWS

Confidence is high on the cash-flow and guidance figures because they are in the October 29 release and cash-flow statement. Confidence is high on the call quotations from multiple transcripts (Investing.com, AlphaStreet, Benzinga), which agree on the $75.9 billion, $578 billion, 40%+, and “meaningfully more” lines. Confidence is lower on any implied Q4 capex of ~$48 billion—that is arithmetic from the guide minus year-to-date net spend, not a company forecast. There is no sourced split of spend across Trainium, Nvidia GPUs, and power.

---

Recent Findings Supplement (October 2026)

Amazon reported Q3 2026 results on October 29, 2026, highlighting accelerated AWS growth amid heavy AI-driven capex that turned trailing twelve-month (TTM) free cash flow deeply negative.[1][2]

Quarterly cash capex (purchases of property and equipment): $77.3 billion in Q3 (vs. $35.1 billion YoY and well above analyst estimates around $57–58 billion). TTM net purchases of PP&E reached approximately $210.7 billion (up 82% YoY), driving TTM free cash flow to an outflow of $42.4 billion (vs. +$14.8 billion prior year). The surge is attributed primarily to AI infrastructure investments.[3][4]

Full-year 2026 expectations: Reaffirmed at approximately $220 billion in cash capex (majority directed to AWS and AI). Q4 operating income guided at $28.5–33.5 billion.[1][2]

2027 directional commentary: Management expects to spend “meaningfully more” than $220 billion on capex, with continued acceleration of payback periods and a “clear line of sight to attractive financial returns” and strong ROIC on the investments.[5][6]

AWS revenue growth and backlog: AWS net sales reached $46.0 billion (+39.3% YoY, fastest pace in nearly five years), achieving a $184 billion annualized run rate. Contracted backlog stood at $578 billion (more than double YoY, triple-digit growth, well diversified).[1][7]

AI-related spend and commentary: Capex is heavily weighted toward AI infrastructure (including Trainium custom chips, Nvidia GPUs, power, data centers, and memory). Specific percentage breakdowns of spend allocation (e.g., Trainium vs. Nvidia vs. power) were not disclosed. Jassy noted installing an AI accelerator roughly every 14 seconds, with capacity snapped up immediately by customers. Inference already accounts for well over half of AWS AI revenue (and growing), with substantial training revenue as well. Custom chips (Graviton and Trainium) delivered >$2 billion in customer savings in Q3; Graviton offers 30–40% better price-performance. Strong Trainium demand prompted plans for external sales late in 2027.[8][9]

Capacity constraints, power availability, and ROIC comments (Jassy and Olsavsky): Jassy stated that 2026 capacity is sold out, 2027 is largely sold out (reserving some for on-demand/EC2/Bedrock), and a good chunk of 2028 is committed. Key bottlenecks include data center space, power availability, and memory (HBM for AI vs. DRAM tradeoffs); he does not expect the shortage to end in 2027. On the earnings call, Jassy described AI as a “once-in-a-lifetime opportunity” that will transform Amazon’s scale, revenue, and free cash flow, with strong conviction on returns despite near-term cash burn. Olsavsky highlighted efficiency gains, capacity optimization, custom chips improving cost structure/margins, and accelerating breakeven/payback periods, while noting AI mix exerts some margin pressure. Both emphasized disciplined investment with visibility into attractive long-term ROIC.[10][11][6]

These details come exclusively from the October 29, 2026 earnings release, 8-K, and call commentary (plus contemporaneous reporting). No other significant post-April 30, 2026 updates on these specific metrics were identified.