Company Deep Dive · AWS · October 10, 2026
AWS built the market it now has to defend. This deep dive examines whether cloud scale, custom silicon, and enterprise distribution are enough to keep AWS at the center of the AI infrastructure stack.
AWS did not need AI to be the largest cloud provider in the world. It built that position over nearly two decades on compute, storage, and enterprise trust. The AI era is now testing whether that position is an asset or a liability — whether AWS's scale lets it absorb the AI buildout on its own terms, or whether AI is rewriting the rules of cloud competition faster than AWS's lead can compound.
This is not a standard company overview. It is an examination of whether AWS can convert cloud scale, custom silicon, data-center capacity, networking, software, and enterprise distribution into one of the primary control points of the AI infrastructure stack — or whether that stack ends up controlled elsewhere.
Core Question
Can AWS maintain its cloud leadership while becoming a central infrastructure platform for the AI economy?
1. AWS Market Position
AWS remains the largest cloud infrastructure provider by revenue, holding roughly 28% of global cloud infrastructure spending in the second quarter of 2026, ahead of Microsoft Azure at about 21% and Google Cloud at about 15%, according to Synergy Research Group's tracking. That share has drifted down from around 30% a year earlier — not because AWS is shrinking, but because Azure and especially Google Cloud are growing faster off smaller bases.
The more telling number is absolute growth. AWS posted $42.2 billion in quarterly revenue in Q2 2026, up 36.7% year-over-year — its fastest growth rate in 18 quarters — putting the division on a $169 billion annualized run rate. Operating income reached $16.6 billion at a 39.4% margin, up sharply from the prior year. For a company of AWS's size, re-accelerating growth while expanding margin is a signal that AI demand is adding to the business rather than merely replacing lower-margin workloads.
Cloud's strategic importance has not diminished as AI workloads grow — it has deepened. Training and running large models requires enormous, reliable compute, storage, and networking capacity that most enterprises have no interest in building themselves. AWS's existing footprint of data centers, availability zones, and managed services gives it a running start that a greenfield AI infrastructure provider does not have.
The CODEW Lens: A declining share number paired with accelerating absolute growth is not evidence of a weakening business. It is evidence of a bigger market growing faster than even its largest incumbent can keep pace with on a percentage basis.
2. AI Infrastructure Strategy
AWS's AI infrastructure strategy rests on four pillars: raw data-center scale, Nvidia GPU capacity, custom silicon, and the networking and storage layer that ties it together.
| Capital spending | Amazon raised its full-year 2026 cash capex guidance to roughly $220 billion, up from $200 billion, with the bulk directed at AWS data centers and AI capacity. |
| GPU infrastructure | AWS continues to deploy Nvidia's latest GPU generations (including B200/B300-class instances) alongside its own silicon, keeping Nvidia-based capacity as the default option most customers reach for first. |
| Custom silicon | Trainium (training) and Inferentia (inference) chips, run through AWS's Neuron software stack, are being deployed at a scale AWS describes as heading toward more than a million accelerators. |
| Networking & storage | High-bandwidth interconnects, purpose-built AI networking, and S3 as a unified data layer underpin both training clusters and inference serving. |
On the Q2 2026 earnings call, CEO Andy Jassy said AWS would not have enough capacity to meet all of 2026's demand even at $220 billion in spending, that most 2027 capacity is already reserved, and that booked demand for 2028 is already notable. That is a company describing itself as supply-constrained in a buildout most competitors are also racing to complete.
The CODEW Lens: AWS is not choosing between Nvidia and custom silicon. It is running both at scale simultaneously, using Trainium to pull down its own cost structure while keeping Nvidia capacity available for customers and workloads that need it.
3. The Custom Silicon Advantage
Hyperscalers build their own AI chips for the same reason they built Graviton: owning silicon lets them capture margin that would otherwise go to a chip vendor, and it gives them a second supply source when demand for a single vendor's hardware outstrips what that vendor can produce.
AWS's Trainium3, unveiled at re: Invent in December 2025 and reaching general availability in early 2026, is built on a 3nm process and is positioned as AWS's most direct answer yet to Nvidia's training hardware. AWS's own benchmarks claim a meaningful generational jump over Trainium2 and a 30–40% cost advantage over comparable GPU-based instances for workloads natively ported to the Neuron software stack — figures that are AWS-reported rather than independently verified. By Jassy's account on the Q1 2026 earnings call, AWS's custom-silicon business had already passed a $20 billion annualized revenue run rate, with more than $225 billion in Trainium-related revenue commitments on the books.
The clearest proof point is Anthropic's Project Rainier, a large-scale Trainium2 training cluster built in partnership with AWS. A frontier AI lab choosing to train on custom silicon rather than exclusively on Nvidia GPUs is a stronger signal of Trainium's real-world viability than any AWS-published benchmark.
The trade-off is portability. Trainium runs on AWS's Neuron SDK, not Nvidia's CUDA, which means workloads optimized for Trainium are effectively locked into AWS. That is a feature for AWS's retention strategy and a cost for customers who want to preserve the option of moving workloads elsewhere. AWS has also signaled openness to selling Trainium chips to third-party data centers outside its own cloud — a move that, if pursued, would turn custom silicon from an internal cost advantage into a standalone hardware business competing more directly with Nvidia.
The CODEW Lens: Custom silicon does not need to beat Nvidia on raw performance to matter. It only needs to be good enough for a meaningful share of workloads at a lower cost, since every workload AWS shifts onto Trainium improves its own margin and reduces how much leverage any single GPU vendor holds over its cost structure.
4. Bedrock and the AI Software Layer
If custom silicon is AWS's hardware bet, Amazon Bedrock is its software bet. Bedrock gives enterprises managed access to foundation models from multiple providers — including Anthropic, Meta, Mistral, and, as of 2026, OpenAI's GPT-5.5 through a new Bedrock Managed Agents arrangement — through a single API, without customers needing to manage underlying infrastructure. AWS reports more than 225,000 active Bedrock customers, including over 80% of the Fortune 100.
The more consequential 2026 development is Bedrock AgentCore, a framework-agnostic runtime for building, deploying, and operating AI agents in production. Unlike the earlier Bedrock Agents offering, AgentCore supports third-party frameworks — LangChain, LangGraph, LlamaIndex, CrewAI — rather than locking developers into AWS-native tooling, while still routing the underlying compute, memory, guardrails, and observability through AWS infrastructure.
This is the clearest expression of AWS's attempt to control the full stack from infrastructure to application: AWS does not need to win the foundation-model race to capture AI value, as long as training happens on AWS-adjacent silicon, inference runs on AWS compute, and the agents built on top of those models are deployed, governed, and monitored through AWS's own runtime. Hosting rival labs' models on Bedrock — rather than competing only with an in-house model — is a deliberate bet that infrastructure neutrality wins more enterprise workloads than model exclusivity would.
The CODEW Lens: By positioning Bedrock as the neutral marketplace for other companies' models rather than a showcase for its own, AWS is betting that owning the infrastructure and distribution layer is more durable than owning any single model.
5. Enterprise Distribution
AWS's deepest AI advantage may not be technical at all — it is distribution. Industry surveys consistently show AWS with the broadest base of enterprises running significant production workloads, ahead of Azure and well ahead of Google Cloud, reflecting nearly two decades of enterprise relationships, procurement agreements, and compliance certifications that are expensive and slow for a competitor to replicate.
That distribution matters disproportionately in AI because of data gravity — the tendency for compute to migrate toward where data already lives, rather than the other way around. Enterprises with years of operational data already sitting in AWS tend to default to running inference and agent workloads in AWS as well, since moving large datasets across clouds is costly, slow, and operationally risky. AWS executives have pointed to this directly, arguing that because more enterprise data resides in AWS than anywhere else, AI inference naturally follows it there.
This also shows up in how AI is lifting AWS's broader, non-AI business. Jassy has described a feedback loop in which AI spending drives demand for conventional compute as well — reinforcement learning post-training and agentic tool use run substantially on CPU infrastructure, not just AI accelerators — meaning AI workloads are pulling incremental spend into AWS's core cloud business, not just its AI-specific product lines.
The CODEW Lens: Existing enterprise relationships are a harder asset to displace than a faster chip or a cheaper API. A competitor can match AWS's hardware roadmap long before it can match AWS's existing footprint inside enterprise IT organizations.
6. Competitive Landscape
AWS is not fighting on a single front. Each competitor is pressuring a different part of its position.
| Microsoft Azure | Holds roughly a 21% share, growing faster than AWS on the back of OpenAI's models and deep enterprise software bundling through Microsoft 365 and Copilot. |
| Google Cloud | The fastest-growing of the three, up roughly 82% year-over-year in Q2 2026 to a record ~15% share, driven by TPU-based infrastructure, Gemini, and enterprises rebuilding data platforms from scratch around AI. |
| Nvidia | Simultaneously AWS's largest AI hardware supplier and, through custom silicon, the dependency AWS is actively trying to reduce. |
| Oracle Cloud | A distant share player overall, but has carved out a meaningful position in large AI training capacity deals with major model developers. |
| Alternative infrastructure | GPU-cloud specialists (such as CoreWeave) and chip challengers (Cerebras, Groq, SambaNova, AMD's Instinct line) chip away at specific workloads rather than competing for the full stack. |
AWS's relative strength is breadth and maturity: the widest service catalog, the deepest enterprise compliance footprint, and the largest base of production workloads. Its relative weakness is that Microsoft Azure and Google Cloud are both growing meaningfully faster off smaller bases, and both have tighter vertical integration with a single leading model family (OpenAI for Azure, Gemini for Google) that AWS's multi-model Bedrock strategy deliberately avoids replicating.
The CODEW Lens: AWS's percentage growth rate will likely keep lagging Google Cloud's for some time simply because of base-size math. The more useful comparison is whether AWS's absolute dollar growth and margin trajectory hold up — and so far, in 2026, they have.
7. AI Economics
The economics of AI infrastructure differ from traditional cloud in ways that matter for AWS's margin profile. Training is capital-intensive and bursty — large clusters run flat-out for weeks or months, then sit idle or get repurposed. Inference is the opposite: steadier, higher-volume, and far more sensitive to cost-per-token, since it runs continuously against live production traffic rather than in discrete training runs.
This is exactly where custom silicon is meant to pay off. Trainium and Inferentia are narrower-purpose than general GPUs, which limits their flexibility but lets AWS tune cost and power efficiency more aggressively for its own workload mix. AWS's own figures put Trainium's cost advantage at 30–40% over comparable GPU instances for well-optimized, Neuron-native workloads — though independent, apples-to-apples benchmarking against Nvidia's latest generations remains limited as of mid-2026, so these figures should be read as AWS's own claims rather than settled fact.
On the headline numbers, AI is not pressuring AWS's profitability — it is expanding it. AWS's operating margin reached 39.4% in Q2 2026, up roughly 650 basis points year-over-year, even as the company absorbs the depreciation of a rapidly growing capital base. Both AWS's standalone AI business and its chips business individually crossed $25 billion in annualized revenue run rate in the same quarter. Jassy has defended the scale of capital spending by pointing to long asset lives — data centers that monetize for 30-plus years and servers that can break even in under three years against multi-year customer contracts.
The counterweight is rising input costs. Amazon specifically cited higher memory, hard-drive, and SSD prices as a driver of its capex increase to roughly $220 billion for 2026, and free cash flow turned negative in the quarter as spending outran operating cash generation. Margin expansion and negative free cash flow are both true at the same time — a reminder that reported profitability and cash economics can diverge sharply during a capital buildout of this size.
The CODEW Lens: The real test of AI economics at AWS is not this year's margin print — it is utilization three to five years out, once today's capacity is built and today's growth rate inevitably normalizes.
8. Strategic Risks
AWS's path to becoming a central AI infrastructure platform is not guaranteed. Several risks sit directly in its way.
| 01 | Nvidia dependency — most AI compute on AWS still runs on Nvidia GPUs, leaving AWS exposed to Nvidia's pricing, allocation decisions, and roadmap even as it builds Trainium capacity. |
| 02 | Faster-growing rivals — Google Cloud's TPU-and-Gemini integration and Azure's OpenAI relationship are both capturing a disproportionate share of new AI workloads, even if neither yet rivals AWS's absolute scale. |
| 03 | Infrastructure spending requirements — sustaining ~$220 billion-plus in annual capex requires demand to keep materializing roughly on schedule; a demand slowdown would land on a cost base that is much harder to shrink than to build. |
| 04 | Custom-chip execution risk — Trainium's economics depend on continued generational improvement and real-world adoption beyond AWS-internal and closely partnered workloads like Anthropic's; chip programs can stumble on yield, performance, or software maturity. |
| 05 | Open-source models — as strong open-weight models proliferate, the software differentiation Bedrock offers narrows, pushing competition back toward infrastructure price and performance, where rivals are investing just as aggressively. |
| 06 | Customers building alternatives — the largest AI labs and the most sophisticated enterprises have both the incentive and, increasingly, the capital to build or co-develop their own infrastructure rather than depend entirely on any single hyperscaler. |
The CODEW Lens: None of these risks is disqualifying on its own. Together, they describe a company that has to keep executing well on multiple fronts simultaneously — hardware, software, pricing, and capacity — with very little room for a prolonged misstep in any one of them.
The CODEW Verdict
AWS is not trying to win the AI race by building the best model. It is trying to win it by making itself the default place AI runs — and on the 2026 numbers, that bet is working.
Cloud leadership gave AWS the capital, the customer base, and the data-center footprint to compete in AI from a position of strength rather than catch-up. Trainium and Inferentia are reducing AWS's reliance on any single GPU vendor without abandoning Nvidia capacity customers still want. Bedrock and AgentCore are positioning AWS as the neutral infrastructure layer underneath other companies' models, rather than betting the business on a single in-house model competing for frontier status. And enterprise distribution — the hardest asset of all to replicate — keeps data, and increasingly inference, gravitating back toward AWS by default.
None of that guarantees AWS controls the AI infrastructure stack indefinitely. Google Cloud and Azure are both growing faster off smaller bases, Nvidia dependency has not disappeared, and the capital intensity of this buildout leaves little margin for a sustained demand shortfall. But as of late 2026, AWS looks less like an incumbent playing defense and more like the infrastructure layer the rest of the AI economy is building on top of — which is precisely the position this research question set out to test.
The CODEW Stat
$169B run rate · 28% share · $220B 2026 capex AWS generated $42.2 billion in Q2 2026 revenue, up 36.7% year-over-year — its fastest growth in 18 quarters — at a 39.4% operating margin, putting the division on a $169 billion annualized run rate even as its global cloud infrastructure share eased from roughly 30% to 28%. AWS's AI and custom-chip businesses each individually cleared $25 billion in annualized revenue, while Amazon raised full-year 2026 capex guidance to approximately $220 billion to keep pace with AI demand that, by its own account, continues to outstrip supply through at least 2028.
Editorial Note
AWS: The Cloud Giant at the Center of the AI Infrastructure Race examines whether AWS's cloud scale, custom silicon, enterprise distribution, and software layer position it as a central control point of the AI infrastructure stack. It is the first entry in The CODEW's AI Infrastructure research flywheel, feeding into deeper coverage of custom silicon, inference economics, and competitive positioning across Amazon, Nvidia, Microsoft, and Google.
Educational content only. Not investment advice. Analysis is based on public company disclosures, earnings calls, industry research (including Synergy Research Group data as reported by third parties), and original editorial judgment. Figures reflect company and analyst reporting as of the article's last-updated date and are subject to revision. This is not an exhaustive account of AWS's business.
Reviewed by Erwin Castro
on
Saturday, October 10, 2026
Rating:
No comments: