COMPANY PROFILE / AI & MACHINE LEARNING
Amazon Web Services (AWS)
Using its own chips and cloud to secure the compute of Anthropic and OpenAI through long-term contracts and investment
In 2026, Amazon’s filings recorded expansions of roughly US$100 billion each in compute contracts with Anthropic and with OpenAI. The contract backlog rose from about US$244 billion at the end of December 2025 to about US$496 billion at the end of June 2026. This profile looks not at all of Amazon but at AWS: Trainium and Inferentia, Amazon Bedrock, and the relationships with Anthropic and OpenAI, and sets out from primary sources what it is investing in, who it works with and what it is aiming for. Contracts, investments and datacenter plans reflect what was published as of 15 October 2026 and are treated as not yet confirmed.
- AWS in 30 seconds, and where it stands in AI and machine learning
- The logo and how the name is written
- Where the company sits in AI and machine learning
- Scale (limited to AWS)
- Resources committed to AI and machine learning
- What it has done so far
- What comes next
- What the company is aiming for in AI and machine learning
- Map of AI and machine learning partnerships
- Technologies, products and services
- Risks this company carries
- Glossary, references and claim-to-source audit
1. Amazon Web Services (AWS) in 30 seconds, and where it stands in AI and machine learning
NASDAQ: AMZNFiscal year ends on 31 December
Chips & memory (secondary)Designs its own chips and offers them on its own cloud
Amazon BedrockAI chips and a model-serving platform
AWS is not a maker of AI models but a company that rents out the compute for AI (cloud and its own chips). In 2026, it signed investments and large compute contracts with two leading model developers, Anthropic and OpenAI, at around the same time Sourced. Amazon’s filings describe the expansion with Anthropic as “more than US$100.0 billion over 10.0 years” and the expansion with OpenAI as “US$100.0 billion over 8.0 years”, and say that both include obligations related to the performance of AWS’s own chips.
However, AWS does not disclose AI-only revenue or profit Sourced. We can read AWS’s overall growth and its contract backlog, but this profile does not estimate how much of it is for AI.
2. The logo and how the name is written
A note on names: in Amazon’s 10-K the business is reported as a segment called AWS. The legal entity is “Amazon Web Services, Inc.”, treated as a wholly owned subsidiary of the listed parent Amazon.com, Inc., but its financial figures are disclosed in Amazon.com, Inc.’s filings. The chips AWS designs itself, Trainium (training and inference) and Inferentia (inference), and the model platform Amazon Bedrock are separate products Sourced.
Official website: https://aws.amazon.com/3. Where the company sits in AI and machine learning
4. Scale (limited to AWS)
Amazon also runs retail and advertising, but this profile covers only the AWS segment. Figures come from the Form 10-K for the year to December 2025 and the Form 10-Q for the period to June 2026 Sourced.
| Measure (US$ million) | FY2023 (year to Dec 2023) | FY2024 (year to Dec 2024) | FY2025 (year to Dec 2025) |
|---|---|---|---|
| Amazon total net sales | 574,785 | 637,959 | 716,924 |
| AWS segment net sales | 90,757 | 107,556 | 128,725 |
| AWS segment operating income | 24,631 | 39,834 | 45,606 |
| AWS operating margin (operating income over sales; our calculation) | 27.1% | 37.0% | 35.4% |
| AWS net additions to property and equipment | 24,843 | 53,267 | 96,496 |
| AWS depreciation and amortization | 12,531 | 13,320 | 21,450 |
| 2026 (Form 10-Q) | Second quarter (April to June) | First half (January to June) |
|---|---|---|
| AWS net sales (US$ million) | 42,232 (up 37% year on year) | 79,819 (up 33%) |
| AWS operating income (US$ million) | 16,621 | 30,782 |
| AWS net additions to property and equipment (US$ million) | 48,604 | 90,120 |
| Amazon-wide cash capital expenditures (US$ billion) | 53.1 | 96.3 |
| Other basics | Details |
|---|---|
| Employees (31 December 2025) | About 1,576,000 across Amazon (full-time and part-time). AWS headcount is not given in the 10-K |
| AWS property and equipment, net | US$190.1 billion at the end of December 2025 and US$263.8 billion at the end of June 2026 (US$110.7 billion at the end of December 2024) |
| Definition of the AWS segment | 10-K: “global sales of compute, storage, database, and other services for start-ups, enterprises, government agencies, and academic institutions” |
The figures above are for AWS as a whole and do not show the size of AI alone. AWS also sells compute, storage and database services that have nothing to do with AI. Net additions to property and equipment also include equipment not used for AI and non-cash acquisitions (such as equipment not yet paid for). The 10-K says spending on technology and infrastructure is expected to increase over time to support AI and machine learning initiatives, but it does not give an AI-only amount Sourced.
5. Resources committed to AI and machine learning
AWS does not disclose headcount, R&D spending or capital expenditure for AI alone. The table shows what could be confirmed from filings and official announcements.
| Resource | What could be confirmed | Type |
|---|---|---|
| People | About 1,576,000 across Amazon. Headcount for AWS, chip development and AI staff is not disclosed | Sourced |
| Money (capital expenditure) | AWS net additions to property and equipment: US$24.8 billion in 2023, US$53.3 billion in 2024, US$96.5 billion in 2025 and US$90.1 billion in the first half of 2026 (10-K, 10-Q). Amazon-wide cash capital expenditures were US$96.3 billion in the first half of 2026; the 10-Q says “the majority” supports AWS growth and expects them to increase in 2026. The AI share is not disclosed | Sourced |
| Outside investments | Anthropic: US$8.0 billion in convertible notes from the third quarter of 2023 to the fourth quarter of 2025; US$10.0 billion in Series G and Series H nonvoting preferred stock in the second quarter of 2026; and a facility of up to US$20.0 billion (the ceiling fell to US$15.0 billion after the Series H purchase, and availability grows as compute-delivery milestones are met). OpenAI: US$15.0 billion in Series C preferred stock in the first quarter of 2026 and a commitment to buy US$35.0 billion more, of which US$13.7 billion was invested in the second quarter and the remaining US$21.3 billion after 30 June (10-Q) | Sourced |
| Universities, research institutions and public funding | Dartmouth: in December 2025, a partnership with AWS and Anthropic to bring AI to the whole campus (AWS press release; an adoption collaboration, not research funding). Public funding such as NAIRR is not in the 10-K or the sources used here | Not found |
| Datacenter power (MW or GW) | AWS’s own owned or under-construction MW is not in the 10-K or 10-Q. Anthropic says it secured up to 5 GW of capacity on AWS, and OpenAI says it will use about 2 GW of Trainium capacity; these are customer-side contracts and announcements, and none says whether the figure is IT load or utility capacity | Not yet confirmed |
Contractual commitments as disclosed (30 June 2026, Amazon-wide, using the 10-Q’s own definitions)
| Item (US$ million) | Total |
|---|---|
| Long-term debt, principal and interest | 220,309 |
| Operating lease liabilities | 116,350 |
| Finance lease liabilities, including interest | 16,660 |
| Leases not yet commenced | 137,214 |
| Unconditional purchase obligations (energy, equipment, content, software and others) | 130,065 |
| Total (also includes financing obligations and other commitments not listed above) | 650,034 |
This is part of the 10-Q’s table of principal contractual commitments (Amazon-wide); no AWS-only breakdown is given. Free cash flow (operating cash flow less purchases of property and equipment, net of proceeds and incentives) was negative US$7.6 billion for the twelve months to June 2026 (positive US$18.2 billion a year earlier), and operating cash flow over the same period was US$161.4 billion. In the first half of 2026 financing inflows from debt were US$82.4 billion, and the 10-Q says Amazon expects to undertake additional financing activities in 2026 Sourced. These are obligations as disclosed, not an assessment of Amazon’s finances.
Amazon’s 10-K and 10-Q disclose “performance obligations, primarily related to AWS, associated with commitments in customer contracts for future services that we expect to fulfill but have not yet been recognized” (contracts with an original term of more than one year). The company does not call this “RPO”; this is its own wording. It was about US$244 billion at 31 December 2025 (weighted-average remaining life 4.1 years) and about US$496 billion at 30 June 2026 (6.4 years) Sourced. The company says timing and amount of recognition depend on customer usage and AWS’s performance, and can extend beyond the original contract term.
Three cautions. (1) It is not an AI-only figure. (2) In the same section the 10-Q mentions the expansions with OpenAI (US$100.0 billion over 8.0 years) and Anthropic (more than US$100.0 billion over 10.0 years), but does not say how much of each sits in the backlog. (3) With a longer remaining life, a large amount can still be recognized as revenue well into the future. This profile does not treat the figure as a measure of AI demand Not yet confirmed.
6. What it has done so far
| When | What happened | Why it matters |
|---|---|---|
| 2023 | Began working with Anthropic and started investing through convertible notes (10-Q; Anthropic’s announcement) | The starting point of pairing investment in a model maker with compute supply |
| 29 October 2025 | Announced that Project Rainier is running: a cluster of nearly half a million Trainium2 chips, in use by Anthropic. It said Claude was expected to run on more than one million Trainium2 chips by the end of the year. Sites include Indiana | A very large cluster built on its own chips |
| 2 December 2025 | Announced general availability of Trainium3 UltraServers (AWS’s first 3nm AI chip). It named Anthropic among customers and said Amazon Bedrock was already serving production workloads on Trainium3 | A new generation of its own chip put into service |
| 3 December 2025 | Partnered with Dartmouth and Anthropic to bring AI to the whole campus | An adoption case in higher education |
| 27 February 2026 | Strategic partnership with OpenAI. Amazon will invest US$50 billion (US$15 billion first, and US$35 billion more when conditions are met). OpenAI and AWS will expand their existing US$38 billion agreement by US$100 billion over 8 years, and OpenAI committed to consume about 2 gigawatts of Trainium capacity. AWS becomes the exclusive third-party cloud distribution provider for OpenAI Frontier | Investment and a compute contract with a second leading model maker |
| 20 April 2026 | Expanded contract with Anthropic. Anthropic committed to spend more than US$100 billion on AWS technologies over ten years, securing up to 5 GW. Amazon invested US$5 billion and may invest up to US$20 billion more (Anthropic’s announcement) | Deepened ties with its largest customer and investee in AI |
| 31 July 2026 | Filed the Form 10-Q for the period to June 2026: AWS sales up 37% in the second quarter, contract backlog of about US$496 billion, and the remaining OpenAI investment made after 30 June | The contracts with model makers appear in the filings |
7. What comes next
All of the above are plans, contracts or targets announced by the company or its counterparty, and none is guaranteed Not yet confirmed. Amazon’s 10-Q says cash capital expenditures are expected to increase in 2026. Anthropic says “Claude Platform on AWS” is coming soon, and the OpenAI announcement said a jointly built “Stateful Runtime Environment” was expected to launch “in the next few months”, but we could not confirm its status since then Not yet confirmed.
None of the following appears in the sources used here Not yet confirmed.
AWS’s own datacenter power (MW) / the number of Trainium chips shipped / the company that manufactures Trainium / AI-only revenue and profit / the share of the backlog held by Anthropic and OpenAI / the launch date of the Stateful Runtime Environment. Figures from press reports or earnings-call remarks are not used here.
8. What the company is aiming for in AI and machine learning
This part includes points the company has not stated directly. They are inferences from primary sources, set out with their basis.
| What can be read | Primary-source basis | Type |
|---|---|---|
| Anchoring on its own chips, it secures the long-term compute demand of the big model makers | The 10-Q says the expansions with both Anthropic and OpenAI include “contractual obligations related to the performance of AWS chips”. OpenAI will use about 2 GW of Trainium, and Amazon’s CEO is quoted saying Anthropic’s commitment to run on Trainium “for the next decade” reflects progress on custom silicon (Anthropic’s announcement) | Inference |
| It pairs investment with compute contracts to keep customers close | For both OpenAI and Anthropic, the investment announcements and the contract expansions came within a few months of each other in 2026. The 8-K describes the equity letter agreement and the AWS commercial and joint-collaboration agreements as separate arrangements | Inference |
| It aims to be a platform for several models, not tied to one | Bedrock is said to be used by more than 100,000 organizations and serves Anthropic’s Claude. The February 2026 agreement also covers making services based on OpenAI models available on AWS | Inference |
| It is investing in AI capacity ahead of demand | AWS net additions to property and equipment in 2025 were about 1.8 times the prior year (our calculation), and the 10-Q expects cash capital expenditures to increase in 2026 | Sourced |
| It also puts weight on the cost of inference, not only training | AWS describes Trainium3 as built for “token economics” in agentic, reasoning and video generation applications, and Inferentia as inference-only | Inference |
The February 2026 OpenAI announcement (Exhibit 99.1) puts the investment, the compute contract, joint model development and the use of Trainium into a single announcement. The 8-K, however, discloses the letter agreement under which Amazon is obliged to buy US$35.0 billion of shares as an arrangement separate from the AWS commercial and joint-collaboration agreements Sourced.
From this structure we read the investment as being pursued together with securing the customer Not yet confirmed. But the terms of each contract (conditions for investing, termination rights and so on) are not set out in full in the official documents, and we cannot confirm whether the investment and the contracts are conditional on one another.
9. Map of AI and machine learning partnerships
10. Technologies, products and services
| Product or facility | Scale and specifications (developer’s stated values) | Stage | Date and source |
|---|---|---|---|
| Trainium2 / Project Rainier | A cluster of nearly half a million Trainium2 chips. An UltraServer combines four servers of 16 chips each, 64 chips in all, linked by NeuronLink. Anthropic says it uses over one million Trainium2 chips (Anthropic’s announcement) | In operation | AWS article (29 October 2025); Anthropic announcement (20 April 2026) |
| Trainium3 | AWS’s first 3nm chip. Per chip: 2.52 petaFLOPS at FP8, 144 GB of HBM3e, 4.9 TB/s of memory bandwidth. An UltraServer holds up to 144 chips: 362 petaFLOPS at FP8, 20.7 TB of HBM3e, 706 TB/s aggregate bandwidth. AWS states up to 4.4 times the compute, about 4 times the memory bandwidth and 4 times the energy efficiency of Trainium2 UltraServers | Available (generally available 2 December 2025) | AWS press release, AWS What’s New, AWS product page |
| Trainium4 | Said to bring much higher FP4 performance and more memory bandwidth and HBM capacity. No numeric specifications are given | Delivery expected from 2027 (not yet confirmed) | Official announcement of 27 February 2026 (Exhibit 99.1) |
| Inferentia / Inferentia2 | Inference chips. Inferentia2 is stated to deliver up to 4 times the throughput and up to 10 times lower latency than the first generation. Inf2 instances have up to 12 chips | Available (no newer generation appears in the sources used here) | AWS product page (checked 15 October 2026) |
| Amazon Bedrock | A platform for building generative AI applications and agents. The company says more than 100,000 organizations use it. It serves Anthropic’s Claude, and Anthropic says “Claude Platform on AWS” is coming soon | Available | AWS product page; Anthropic announcement |
| Amazon Nova | Amazon’s own family of models. It is mentioned in the OpenAI announcement only as available to Amazon developers, and no specifications or usage figures appear in the sources used here | Available (details not confirmed) | Official announcement of 27 February 2026 (Exhibit 99.1) |
In AWS’s own description, the Trainium3 design focuses on how to secure bandwidth for compute Sourced. Each chip carries 144 GB of HBM3e (4.9 TB/s), with a two-level on-chip SRAM between the HBM and the compute engines that AWS says keeps data movement to a minimum. Besides FP8 it supports the low-precision MXFP8 and MXFP4 data types and hardware-accelerated 4-bit-weight quantization. Because reading model weights often limits inference, this reads as an effort to cut cost through memory structure and numeric precision.
On cooling, AWS’s article says one of the Project Rainier sites (Indiana) will make maximum use of outside air and use no water for cooling from October to March. For new datacenter components that combine advances in power, cooling and hardware, it projects up to 46% less mechanical energy use and 35% less embodied carbon in concrete (company projections; no third-party verification appears in the sources used here). See also our explainers on HBM, AI accelerators, and power and cooling for AI datacenters.
11. Risks this company carries
| Risk | Details | Type |
|---|---|---|
| Dependence on specific model makers | The contract expansions are each on the order of US$100 billion for Anthropic and for OpenAI. Both are AWS customers and Amazon investees (the 10-Q says these are variable interests in entities Amazon does not consolidate, with maximum exposure generally limited to the carrying value and future funding commitments) | Sourced |
| Swings in the value of the investments | Private holdings are carried at Level 3 fair value, which moves with the investee’s funding rounds. Other income (expense), net was US$53.4 billion in the second quarter of 2026, of which about US$50.5 billion was an upward adjustment to Anthropic preferred stock. It does not enter AWS operating income | Sourced |
| Capital expenditure and cash flow | Free cash flow for the twelve months to June 2026 was negative US$7.6 billion. The 10-Q expects cash capital expenditures to increase in 2026 | Sourced |
| Size of future obligations | Principal contractual commitments total US$650 billion (Amazon-wide), including US$137.2 billion of leases not yet commenced and US$130.1 billion of unconditional purchase obligations. These are disclosed obligations, not an assessment of Amazon’s financial health | Sourced |
| Supply constraints | The 10-K says it may rely on limited or single sources of supply, and the forward-looking statements in the February 2026 announcement cite “resource and supply volatility, including for memory chips” | Sourced |
| Timing of backlog recognition | Amounts depend on customer usage and AWS’s performance, and the remaining life is 6.4 years. A large contract can still be recognized as revenue well into the future | Sourced |
| AI regulation and intellectual property | The 10-K says use of AI may lead to more claims of infringement, and that regulations governing AI may apply | Sourced |
| Delivery of plans | Trainium4 delivery in 2027, up to 5 GW of capacity and use of the facility depend on delivery milestones and conditions, and none is guaranteed | Inference |
12. Glossary
- Trainium
- AWS’s own AI chip, used for training and inference. The current generation is Trainium3 (3nm).
- Inferentia
- AWS’s own inference chip.
- UltraServer
- AWS’s configuration that links several servers at high speed into one large machine.
- Amazon Bedrock
- A service for using generative AI models and an agent platform on AWS.
- Project Rainier
- A very large compute cluster of Trainium2 chips that AWS built with Anthropic.
- Contract backlog (10-K wording)
- Unrecognized performance obligations in customer contracts for future services, primarily AWS, with an original term over one year.
- Nonvoting preferred stock
- Shares without voting rights that rank ahead in dividends and liquidation. Amazon’s Anthropic investment is mainly in this form.
- Level 3 fair value
- A fair value that cannot be read from market prices and is set with the company’s own valuation model.
13. References
- Amazon.com, Inc. Form 10-K for the fiscal year ended December 31, 2025 (filed with the U.S. SEC on 6 February 2026). AWS segment sales, operating income and capital additions, performance-obligation backlog, Anthropic disclosures, contractual commitments, risk factors and headcount. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm
- Amazon.com, Inc. Form 10-Q for the quarterly period ended June 30, 2026 (filed with the U.S. SEC on 31 July 2026). AWS results for the first half of 2026, the US$496 billion backlog, investments in and contract expansions with Anthropic and OpenAI, contractual commitments and free cash flow. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm
- Amazon.com, Inc. Form 8-K (filed 27 February 2026; Items 1.01, 7.01 and 8.01). The US$35.0 billion equity commitment to OpenAI and the related AWS commercial arrangement. https://www.sec.gov/Archives/edgar/data/1018724/000110465926021050/tm267374d1_8k.htm
- Amazon.com, Inc. Exhibit 99.1 to that Form 8-K, the press release “OpenAI and Amazon Announce Strategic Partnership” (27 February 2026). The US$50 billion investment, about 2 gigawatts of Trainium, and the Trainium4 delivery timing. https://www.sec.gov/Archives/edgar/data/1018724/000110465926021050/tm267374d1_ex99-1.htm
- Anthropic Official announcement “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute” (20 April 2026). The contract as described by Anthropic. https://www.anthropic.com/news/anthropic-amazon-compute
- Amazon Web Services Official press release “Trainium3 UltraServers Now Available” (2 December 2025). Availability of Trainium3 UltraServers and named customers. https://press.aboutamazon.com/2025/12/trainium3-ultraservers-now-available-enabling-customers-to-train-and-deploy-ai-models-faster-at-lower-cost
- Amazon Web Services AWS What’s New, “Announcing Amazon EC2 Trn3 UltraServers” (2 December 2025). Trainium3 specifications (developer's stated values). https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/
- Amazon Web Services Official article “AWS activates Project Rainier” (29 October 2025). A cluster using nearly half a million Trainium2 chips. https://www.aboutamazon.com/news/aws/aws-project-rainier-ai-trainium-chips-compute-cluster
- Amazon Web Services AWS Trainium product page (checked 15 October 2026). Description of the Trainium3 architecture (developer's stated values). https://aws.amazon.com/ai/machine-learning/trainium/
- Amazon Web Services AWS Inferentia product page (checked 15 October 2026). Positioning of Inferentia and Inferentia2 (developer's stated values). https://aws.amazon.com/ai/machine-learning/inferentia/
- Amazon Web Services Amazon Bedrock product page (checked 15 October 2026). The company's statement on the number of organizations using it. https://aws.amazon.com/bedrock/
- Amazon Web Services Official press release “Dartmouth becomes first Ivy League institution to partner with AWS and Anthropic” (3 December 2025). https://press.aboutamazon.com/aws/2025/12/dartmouth-becomes-first-ivy-league-institution-to-partner-with-aws-and-anthropic-to-launch-ai-at-institutional-scale
- Amazon Web Services Official website (to confirm company details; the logo was taken from AWS’s official static-content site, a0.awsstatic.com). https://aws.amazon.com/
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| AWS segment sales, operating income, capital additions and depreciation (FY2023 to FY2025), Amazon total sales, headcount, and the US$244 billion backlog | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm |
| AWS results for the second quarter and first half of 2026, the US$496 billion backlog, cash capital expenditures | Sourced | Reference 2 https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm |
| Investments in Anthropic (US$8.0 billion of notes, US$10.0 billion of Series G and H, the facility, carrying values) and in OpenAI (US$15.0 billion, US$13.7 billion, the remaining US$21.3 billion) | Sourced | Reference 2 https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm |
| The US$35.0 billion equity commitment to OpenAI and the AWS commercial and joint-collaboration agreements | Sourced | Reference 3 https://www.sec.gov/Archives/edgar/data/1018724/000110465926021050/tm267374d1_8k.htm |
| Partnership with OpenAI: US$50 billion investment, US$100 billion expansion of the US$38 billion agreement, about 2 GW of Trainium, Trainium4 delivery from 2027 | Sourced | Reference 4 https://www.sec.gov/Archives/edgar/data/1018724/000110465926021050/tm267374d1_ex99-1.htm |
| Expanded contract with Anthropic: over US$100 billion over ten years, up to 5 GW, Amazon investing US$5 billion and up to US$20 billion more | Sourced | Reference 5 https://www.anthropic.com/news/anthropic-amazon-compute |
| Availability of Trainium3 and named customers (2 December 2025) | Sourced | Reference 6 https://press.aboutamazon.com/2025/12/trainium3-ultraservers-now-available-enabling-customers-to-train-and-deploy-ai-models-faster-at-lower-cost |
| Trainium3 specifications and comparisons (developer’s stated values) | Sourced | Reference 7, 9 https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/ https://aws.amazon.com/ai/machine-learning/trainium/ |
| Project Rainier (nearly half a million Trainium2 chips, Indiana, cooling and efficiency projections) | Sourced | Reference 8 https://www.aboutamazon.com/news/aws/aws-project-rainier-ai-trainium-chips-compute-cluster |
| Positioning of Inferentia and Inferentia2 | Sourced | Reference 10 https://aws.amazon.com/ai/machine-learning/inferentia/ |
| Number of organizations using Bedrock | Sourced | Reference 11 https://aws.amazon.com/bedrock/ |
| Partnership with Dartmouth | Sourced | Reference 12 https://press.aboutamazon.com/aws/2025/12/dartmouth-becomes-first-ivy-league-institution-to-partner-with-aws-and-anthropic-to-launch-ai-at-institutional-scale |
| Contractual commitments (10-Q) and free cash flow | Sourced | Reference 2 https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm |
| AWS operating margin (operating income over sales) and net additions to property and equipment at about 1.8 times the prior year | Our calculation | Our calculation from the figures in Reference 1 |
| Classification used in the type chart (cloud & infrastructure primary, chips & memory secondary) | Our calculation | Our own classification, based on the descriptions in References 1, 6 and 7 |
| AI-dedicated headcount, AI capital expenditure and AI-only revenue | Not yet confirmed | None disclosed; not estimated here (not in References 1 or 2) https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm |
| AWS’s own datacenter power (MW), Trainium shipment counts, the chip manufacturer | Not yet confirmed | Not in References 1, 2, 6, 7 or 8. Press figures are not used https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm |
| The share of the backlog held by Anthropic and OpenAI | Not yet confirmed | Not in Reference 2. Not estimated here https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm |
| Trainium4 delivery in 2027, up to 5 GW, about 2 GW of Trainium, use of the facility, launch of the Stateful Runtime Environment and Claude Platform | Not yet confirmed | Reference 2, 4, 5 https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm https://www.sec.gov/Archives/edgar/data/1018724/000110465926021050/tm267374d1_ex99-1.htm https://www.anthropic.com/news/anthropic-amazon-compute |
| Reading: anchoring on its own chips to secure long-term demand | Inference | Our own reading of the 10-Q wording on chip performance obligations and the CEO’s quoted remark |
| Reading: pairing investment with compute contracts | Inference | Our own reading of the timing of the announcements in February to April 2026 |
| Reading: aiming to be a platform for several models | Inference | Our own reading of the Bedrock description and the OpenAI collaboration agreement |
| Reading: weight on the cost of inference | Inference | Our own reading of the Trainium3 and Inferentia descriptions |
| Risk that delivery of plans depends on milestones and conditions | Inference | Our own reading of the descriptions of delivery milestones and conditions |
Last updated: 15 October 2026 (date checked) / Troy Technical
All figures in this article are based on documents Amazon.com, Inc. filed with the U.S. Securities and Exchange Commission and on official announcements from AWS, Amazon and Anthropic. Market-research estimates and press figures are not used. Chip and service performance figures are the developer’s stated values, not independent measurements. Passages marked “Inference” are our reading of primary sources and are not statements by the company. Contracts, investments and datacenter plans reflect what had been published as of 15 October 2026.