Earnings Call Highlights: Amazon (AMZN) | Q3 FY2026 — AWS Up 39%, a $578B Backlog, Trainium Goes on Sale Outside AWS in Late 2027, and Management Calls for 40%+ Growth in Both 2027 and 2028

1. Executive Takeaway

Three signals mattered most for Amazon this quarter:

1. AWS growth is not a one-quarter fluke — it is a queue of signed orders. A $578B backlog, plus customers reserving hundreds of megawatts to a gigawatt of capacity well in advance, gives management what it calls good visibility into 2027, 2028 and part of 2029.

2. In-house silicon has gone from a cost tool to a second business. Most of Bedrock runs on Trainium, and Trainium will be sold outside AWS starting in late 2027. AWS's margin structure is starting to diverge from other clouds.

3. The investment is now big enough to require borrowing. Operating margin hit a record 13.9%, but capex is consuming all of the cash flow, and long-term debt rose roughly 2.6x in nine months.

AWS's bottleneck is no longer whether customers want capacity — it is how fast Amazon can get data centers, power and chips in place.

2. Revenue and Financials

Metric	Q3 FY2026	YoY	Notes
Revenue	$201.94B	+12.1%	+13% ex-FX; ~4 pts higher excluding the Prime Day shift into Q2
Gross margin	53.2%	+2.4 pp	Computed as (net sales − cost of sales) / net sales
Operating income	$28.08B	+61.2%	Operating margin 13.9%, a record
Net income	$20.69B	-2.4%	Year-ago quarter included a $9.5B Anthropic investment gain; non-operating was -$0.7B this quarter
Diluted EPS	$1.89	-3.1%	GAAP
AWS revenue	$45.98B	+39.3%	Operating margin 38.7% (39.4% in Q2)
North America revenue	$112.85B	+6.2%	Operating margin 7.7%
International revenue	$43.11B	+5.4%	+8% ex-FX, operating margin 3.7%
Advertising services	$19.64B	+10.9%	+12% ex-FX; ~850 bp hit from the Prime Day shift

▲ Amazon FY2026 Q3 key financials (revenue / gross margin / operating margin / net income / EPS / YoY / Q4 guidance / segment revenue) (Image source: Simple Tech Trend original; data from Amazon 8-K Exhibit 99.1 and the earnings call)

Read the numbers in two layers:

• Operations improved across the board. Operating income rose 61% and all three segments grew profit. North America jumped from $4.8B to $8.7B — partly a low base, since the year-ago quarter carried a $2.5B FTC settlement plus severance — but a 7.7% North America margin is in line with recent quarters on its own.

• The net income dip is accounting noise, not a weaker business. Last year's Q3 was lifted by a $9.5B Anthropic investment gain; this year non-operating items were a small loss. Reading the 3% EPS decline as weakness would be a mistake — operating income is the number to watch.

One AWS detail stands out: the sequential dollar increase shrank from $4.6B in Q2 to $3.8B in Q3. An analyst pressed on this in Q&A, and Jassy attributed it to uneven capacity additions from quarter to quarter rather than demand. It is the same picture we saw in the Microsoft Q1 FY2027 earnings call: a cloud provider's sequential growth now reflects how many racks were powered on that quarter.

3. Technology and Business Highlights

Trainium and Graviton: in-house chips become a second business

• The chips business grew triple digits, and revenue commitments for Graviton and Trainium each rose more than 20x YoY.

• Customers using Amazon's custom silicon saved more than $2B in Q3 alone.

• New generations are not cannibalizing old ones: after Graviton5 launched, backlog for the prior Graviton generation still more than doubled QoQ, and short-term capacity on prior-gen Trainium now sells at a premium to the first half.

• Amazon open-sourced TorchNeuron, so PyTorch code runs on Trainium without rewriting. That attacks the biggest barrier of the CUDA ecosystem: developer switching cost.

• Trainium will be sold outside AWS starting in late 2027.

Jassy laid out the trade-off of selling chips externally plainly: everyone builds advanced chips at the same foundry and gets a fixed annual allocation. Sell a chip and you recognize revenue immediately; keep it in your own cloud and it generates revenue and cash flow over a five-to-six-year life, a larger total. He said effectively only two companies in the world face this trade-off.

The other one is Google. We covered TPUs being sold as full systems in the Alphabet Q3 FY2026 earnings call; now Amazon has put its own timeline on the table. ASICs are no longer just an internal cost-down tool for hyperscalers — they are heading straight at GPUs for outside customers.

Bedrock: the fastest-growing multi-billion-dollar service in AWS history

• Fortune 500 spending on Bedrock rose nearly 7x YoY, and total token volume grew nearly 17x.

• Inference already accounts for "well over half" of AWS AI revenue and is still rising.

• Most Bedrock workloads run on Trainium. This is the key line of the quarter: Amazon is not renting GPU hours, it is wrapping its own silicon inside a service and selling tokens, keeping control of the margin.

Jassy expects six to eight "roughly as good" models over the next two to three years, with customers using several at once and switching often, which is why they need a multi-model platform. He even said Bedrock could grow as large as EC2.

Core cloud: AI pulls CPUs and storage along with it

• Core compute and storage revenue grew at its fastest rate in several years.

• The reason: post-training reinforcement learning and agent tool calls mostly run on CPUs, not AI accelerators. The more AI spreads, the more demand for Graviton and traditional cloud services.

• About 85% of global IT spend is still on premises, and "never moving to the cloud" workloads like mainframes and SANs are being migrated with AI-driven tools.

• AgentCore has tens of thousands of customers, and Q3 agent deployments were 2x the prior three quarters combined.

This is the same trend line as the Meta Q3 earnings call, where compute demand shifts from training to inference: inference and agent workloads are more heterogeneous than training, so CPUs, memory, storage and networking all have to grow together.

Retail: cost and speed

• Units grew nearly 15% YoY year to date, with paid units up 12% in Q3.

• Shipping costs rose 11%, below unit growth; excluding fuel and line-haul inflation they rose only 8%.

• The fulfillment network runs more than 1 million robots, and the robotic arm fleet doubled in 2026.

• Fresh groceries and everyday essentials grew about 3x faster than the rest of the store and are roughly half of all units.

4. Management Outlook

Q4 guidance: net sales of $240.0B–$247.0B, up 12%–16% YoY (including ~60 bp FX headwind); operating income of $28.5B–$33.5B versus $25.0B a year ago.

Longer term:

• AWS growth of 40%+ in each of 2027 and 2028, and a trillion-dollar revenue business within five to ten years.

• 2026 cash capex of about $220B, with 2027 "meaningfully more".

• Advertising growth returning to the "high teens" in Q4.

Reading the wording:

• "40%+ AWS growth in 2027 and 2028" is phrased as a belief, but the reason given is that large capacity reservations already extend into 2029. That is stronger than an ordinary expectation — the order book is backing the growth rate.

• On the "trillion-dollar business", Jassy admitted five to ten years cannot be predicted precisely and said it is "not a stretch to imagine". That is a vision, not guidance.

• On AWS margins, the CFO explicitly gave no guidance and named three headwinds: AI margins still below core, rising component costs such as memory, and depreciation.

One signal falls out of the arithmetic: net cash capex for the first three quarters of 2026 totals about $172.2B. If the full year really is "about $220B", Q4 is left with only about $48B, more than 30% below Q3's $75.9B. Management did not address this; it could be payment timing, or a conservative full-year figure. Next quarter will show where it actually lands.

5. Supply Chain and Customer Clues

Clue 1: the customer list is widening — and includes competitors. New AWS agreements include Meta, NVIDIA, OpenAI, Oracle, Salesforce and Qualcomm, among others. OpenAI models are on Bedrock, and Bedrock Managed Agents is powered by OpenAI. Cloud providers now compete and buy from each other at the same time.

Clue 2: owned or long-leased data centers are a margin moat. Jassy said neoclouds run about 3,500 bp worse operating margin than long-leased capacity, and most AWS capacity is owned or long-leased. Hyperscalers will keep building their own data centers rather than outsourcing heavily to neoclouds — sustained demand for construction, power and liquid-cooling equipment.

Clue 3: custom networking + custom chips = custom optical interconnect. The CFO credited margin gains to custom chips and custom networking. The bigger Trainium clusters get and the more Trainium is sold outside, the more scale-up and scale-out interconnect is needed — to specs AWS defines itself. In the Celestica Q3 FY2026 earnings call we saw 1.6T overtaking 800G next year and rack-scale CPO ramping in late 2027; rack and optical module demand from the ASIC camp is one of the drivers.

Clue 4: memory is a clear cost headwind. The forward-looking statements list memory chips as a resource and supply volatility risk, and both Jassy and the CFO cited rising component prices when discussing margins.

Clue 5: foundry allocation is the ceiling. The remark that everyone builds advanced chips at the same foundry on annual allocations means Trainium's upside depends not only on demand but on how much leading-edge process and packaging capacity Amazon can secure.

6. Conclusion

Three metrics to track next quarter:

1. Whether AWS growth reaches 40%: this is the baseline management set for 2027; continued acceleration in Q4 would mean capacity additions are keeping pace.

2. Actual Q4 capex: the full-year $220B implies only about $48B; a figure well above that would mean the 2027 ramp has started early.

3. AWS operating margin: 38.7% this quarter, down from 39.4% in Q2. The question is whether a rising AI mix, memory price increases and depreciation push it below 35%.

This quarter Amazon was not arguing how big AI demand is — it was arguing who can sell AI cheaper than everyone else and still earn more, using its own chips, its own data centers and its own network.

Trainium going external, Bedrock running on in-house silicon, and a 3,500 bp margin gap versus neoclouds all point to the same conclusion: cloud competition is shifting from "who can get GPUs" to "who has the better vertically integrated cost structure". For the supply chain, that means demand for ASICs, custom interconnect and self-built data centers will grow alongside GPUs rather than being displaced by them.

This article is for technology and industry trend analysis only and does not constitute investment advice.