How AWS uses custom chips to build a money-printing machine

• As AI workloads expand from pre-training to reinforcement learning and AI agents, the focus is shifting from peak compute performance to memory bandwidth and CPU capacity delivered per unit of total cost of ownership.
• AWS combines Trainium, Graviton, Nitro, and its networking infrastructure to reduce chip, power, cooling and data-transfer costs. Through Bedrock, it can also earn both infrastructure and token revenue by distributing models such as Claude.
• AWS revenue growth accelerated to 37% in 2Q26, while its operating margin approached 40%, suggesting that its core cloud profitability may already be the strongest among the Big Three. Trainium also extends the investment opportunity to HBM, TSMC’s advanced nodes, and CoWoS packaging.

Is the AI cloud just a money pit — endless capex with profits nowhere in sight? Quite the opposite!

AWS revenue has now accelerated for five consecutive quarters, growing 37% year-on-year in 2Q26 — the fastest pace in 18 quarters. More importantly, margins have kept improving even as revenue accelerates. How did this cloud giant, a latecomer to the generative-AI race, build such strong profitability? This article breaks down how AWS uses custom chips to build the cloud’s strongest money-printing machine.

Table 1: Amazon's custom-silicon family at a glance

Chip	Role	Latest generation	Process	Latest status
Trainium	AI training + inference accelerator	Trainium3	TSMC 3nm	Unveiled in late 2025; volume workhorse for 2026. Trainium4 already previewed
Graviton	General-purpose server CPU	Graviton5	TSMC 3nm	Generally available since June 2026; used by 98% of AWS's top 1,000 EC2 customers (company disclosure)
Nitro	Networking/virtualisation offload chip	Nitro v6	—	Standard across the entire EC2 fleet
Source: Amazon, iFAST Compilations

AWS Trainium3: behind Nvidia on specs — why is it still competitive?

Trainium3 is AWS's third-generation in-house AI accelerator, unveiled in late 2025. It is built on TSMC's 3nm process and carries 144GB of HBM3E high-bandwidth memory. Notably, the next generation, Trainium4, will support Nvidia's NVLink Fusion — a sign that even as AWS drives costs down with its own silicon, it intends to stay interoperable with Nvidia hardware and preserve deployment flexibility across its datacenters.

Related article: “NVIDIA | Beyond Chips: The Ecosystem Driving Its Next Growth Phase”

On absolute performance, Trainium3 clearly trails Nvidia's GB300: roughly half the compute and half the HBM capacity, and only about 60% of the memory bandwidth — on some specs it even trails Google's TPU. But as AI workloads extend from pre-training into reinforcement learning and AI agents, peak compute — while still important — is no longer the only yardstick of a chip's value.

Table 2: Trainium3 vs Nvidia GB300 vs Google TPUv7

Item	Trainium3	Nvidia GB300	Google TPUv7
Process	TSMC 3nm (N3P)	TSMC 4nm-class (4NP)	TSMC 3nm (N3E)
FP8 compute (dense)	2.52 PFLOPS	~5 PFLOPS	4.6 PFLOPS
HBM capacity	144GB HBM3E	288GB HBM3E	192GB HBM3E
Memory bandwidth	4.9 TB/s	8 TB/s	7.37 TB/s
Power	~1,000W	Up to ~1,400W	Undisclosed
Cooling & facility cost	Air-cooled (NL32x2) or liquid-cooled (NL72x2); high retrofit flexibility	Liquid-cooling only; heavier facility retrofits	Predominantly liquid-cooled (since TPUv3)
Procurement cost structure	In-house, near chip cost	Pays Nvidia's ~75% gross margin	In-house; co-designed with Broadcom
TCO per marketed FP8 PFLOPS	$0.31–0.38 /hr per PFLOPS (air-cooled SKUs $0.31–0.34; liquid-cooled NL72x2 $0.38)	~$0.55 /hr per PFLOPS	$0.35
TCO per memory bandwidth	$0.22 /hr per TB/s	$0.28–0.30 /hr per TB/s	~$0.22 /hr per TB/s
Source: Amazon, Nvidia, SemiAnalysis, iFAST compilations

AI training’s centre of gravity shifts to post-training: how RL reshapes HBM and CPU demand

Model developers used to improve their models mainly by scaling up pre-training and piling on compute. Increasingly, however, labs are finding that post-training still offers plenty of low-hanging fruit, so incremental compute demand is extending from pre-training into post-training — reinforcement learning (RL) in particular.

Zhipu’s GLM-5.3: GLM-5.3 keeps the same base model as GLM-5.2, yet delivers a big jump in logical-reasoning capability through better post-training — the reinforcement-learning (RL) mechanism in particular.

The compute-allocation trend: AI model companies are no longer relying solely on ever-larger pre-training runs — more compute is being redirected toward post-training.

Inside the RL loop: compute mechanics and hardware demands

Reinforcement learning requires the model to generate reasoning traces at scale, which the system then scores and verifies before updating the model. Generation continuously streams model weights and the KV cache through memory, raising the bar on HBM capacity and memory bandwidth; scoring and verification, meanwhile, lean heavily on CPUs.

Figure 1: GLM-5.2 and GLM-5.3 share the same base model, but better post-training — especially reinforcement learning — lifts model capability substantially

Source: Zhipu AI

Figure 2: The reinforcement-learning (RL) training loop — AI accelerator and CPU each do half the job

Source: iFAST Compilations

AWS Trainium3: the price-performance pick

Beyond post-training, labs are also increasingly focused on inference costs. With the rise of AI agents, models must hold long contexts and repeatedly call tools and query databases. Both trends make HBM capacity and bandwidth critically important.

When cloud providers evaluate AI chips, they do not simply compare single-chip performance. What they care about is the total cost of ownership (TCO) of the whole system — chips, power, cooling, CPUs, networking, and facility retrofits.

As post-training and AI agents take off, the yardstick is shifting accordingly: how much usable memory bandwidth each unit of TCO buys has become a key question for cloud providers.

The three core mechanisms by which AWS Trainium3 lowers TCO

• Lower chip cost: Trainium is designed in-house, and AWS buys HBM3E directly from memory makers — no Nvidia gross-margin premium to absorb.
• Lower power and cooling burden: Trainium3 draws roughly 1,000W, below the GB300's peak of ~1,400W, and supports air-cooled deployment, reducing spend on liquid-cooling gear and facility retrofits.
• Higher system efficiency: AWS co-designs the Trainium accelerator, Graviton CPU, Nitro, networking and the datacenter itself, reducing bottlenecks in CPU and data movement, lifting cluster utilisation and spreading system costs over more useful output.

Trainium's potential edge comes from the overall cost structure: as long as the savings on chips, power, cooling and systems outweigh the shortfall in effective bandwidth, each unit of TCO buys more effective bandwidth.

AWS Trainium3 vs Google TPUv7: deployment speed and ecosystem flexibility

On cost per TB/s of memory bandwidth, Google’s TPUv7 is broadly on par with Trainium3 — but the two differ markedly in deployment speed and business strategy. Trainium’s advantage lies in deployment speed and ecosystem lock-in. Its air-cooled SKUs can slot straight into existing datacenters, bringing capacity online faster; TPUs, by contrast, have been predominantly liquid-cooled since the third generation, and high-density deployments often require facility retrofits — while datacenters and power supply are precisely one of Google’s bottlenecks in scaling compute.

A broadening customer base: annualised chip revenue past US$25 billion

On the customer side, Anthropic has long used Trainium at scale for both training and inference; OpenAI has made a multi-year, multi-gigawatt compute commitment; and Uber, Pinterest and a growing list of AI start-ups are coming on board. Management has even disclosed that it is studying direct chip sales to external customers.

As the customer base broadens, Amazon's chip business has passed US$25 billion in annualised revenue, growing at a triple-digit percentage rate year on year.

Graviton: the underrated player

If Trainium tackles the cost of AI accelerators, Graviton addresses the other, often-overlooked half: CPU cost. As discussed, the shift toward post-training and reinforcement learning — plus the rise of AI agents — means code compilation, result verification, environment simulation, tool calls, and database queries all consume large amounts of CPU. Microsoft's Fairwater AI datacenter, for example, deploys hundreds of thousands of GPUs alongside tens of thousands of CPUs. In short, AI infrastructure is no longer just a GPU story; the CPU is becoming a key cost and supply variable.

Graviton’s technical edge: TSMC 3nm and co-design with the entire AI system

Amazon began designing CPUs for its own data centers well before the AI boom. It acquired Israeli chip designer Annapurna Labs in 2015 and launched the first-generation Graviton in 2018. That long accumulation of expertise lets AWS keep tuning its CPUs as AI workloads evolve.

Process and specs: The latest Graviton5 went into full production in mid-2026 on TSMC’s 3nm process, packing 192 cores with five times the memory capacity of its predecessor. Its advantage is not just the CPU itself but the fact that Graviton5 is co-designed with AWS’s entire AI system. Intel and AMD must remain compatible with a wide range of server configurations, forcing them to reserve extra interconnect lanes and general-purpose features.

Specialised integration: Graviton serves only AWS’s own datacenters, so unnecessary features can be stripped out, while dedicated Nitro hardware offloads networking, storage and virtualisation — handing more die area and power budget back to compute cores and memory.

This vertical integration matters most inside the AI rack. Trainium accelerators handle model computation while Graviton serves as the host CPU, managing task scheduling, data movement and result collection; reinforcement learning and AI agents then pile environment simulation, scoring, verification and tool calls onto the CPU as well. In other words, AWS designs not only its own AI accelerator but also the host CPU and the underlying system architecture — with no dependence on Nvidia, Intel or AMD — extending its cost advantage from the single chip to the entire AI rack.

By contrast, Microsoft's Cobalt and Google's Axion today mostly serve general-purpose cloud workloads such as websites and databases. Microsoft's AI racks still rely heavily on Intel or AMD CPUs, and Google's TPU clusters have not fully switched to Axion hosts. All three hyperscalers design their own CPUs, but at this stage only AWS has deeply integrated its CPU with its own AI accelerator — extending its cost advantage from Trainium to the host CPU and the full AI system.

Customer demand validates the strategy. Nearly 98% of AWS's top 1,000 EC2 customers already run Graviton; related customer commitments have tripled quarter on quarter; and Graviton5 is being adopted at roughly twice the pace of its predecessor at the same stage. Anthropic, OpenAI and Meta have also signed large CPU-capacity agreements with AWS.

Table 3: Custom CPUs at the big-three clouds

Item	AWS Graviton5	Microsoft Cobalt 200	Google Axion
Generation	5th (since 2018)	2nd (since 2024)	1st (since 2024)
Core architecture	Arm Neoverse V3	Arm Neoverse V3	Arm Neoverse V2 (C4A)
Process	TSMC 3nm	TSMC 3nm (N3P)	5nm-class (C4A)
Cores per chip	192	132	Up to 72 (C4A)
Commercial status	Generally available June 2026	Preview June 2026	C4A generally available
Host CPU of the AI rack	Yes — commands Trainium3	No — mostly x86	No — TPU clusters mostly x86
Primary role	General compute + AI system host	General cloud compute	General cloud compute
Source: Amazon, Microsoft, Google, SemiAnalysis, iFAST compilations

Networking and the datacenter: the part everyone overlooks

As noted above, the AI hardware race long ago expanded from single chips to entire data centers. Frontier models often need tens of thousands of chips running in concert, exchanging enormous volumes of data every second; once the network cannot keep up, expensive AI accelerators sit idly waiting for data, dragging down utilisation across the whole cluster. Engineers therefore cannot focus solely on building faster CPUs or accelerators — chips, networking, power and cooling have to be treated as one system. Only when every link works together does chip-level compute translate into real output. Trainium systems, for example, include backup NeuronLink network paths, so data can reroute even if individual links fail, and switches support live replacement without shutting down the node. Such choices do not make single-chip spec sheets shinier, but they cut downtime and lift cluster utilisation.

Trainium's cooling design follows the same logic. Most existing datacenters are air-cooled, and deploying liquid-cooled-only chips typically demands extra money and time for facility retrofits. With compute in short supply, operators want datacenters online as fast as possible. Trainium3 therefore ships in both air-cooled and liquid-cooled SKUs: the air-cooled version drops straight into existing facilities without waiting for liquid-cooling retrofits, lowering upfront investment and shortening time-to-revenue.

Deep in-house networking, split by transmission distance

The network layer is likewise deeply home-grown at AWS, with each tier handling a different transmission distance.

• NeuronLink and NeuronSwitch: high-speed interconnect between Trainium chips within a rack.
• Elastic Fabric Adapter (EFA): links racks together, combining multiple racks into large compute clusters.
• Nitro: offloads networking, storage and virtualisation so the host CPU is not consumed by infrastructure chores.

Through this chip-to-rack-to-cluster design, data bottlenecks between stages are reduced, letting Trainium spend more time computing rather than waiting for data — lifting utilisation across the entire cluster.

Bedrock × Anthropic: converting low-cost compute into token revenue

So far we have covered how AWS cuts costs through custom silicon and co-design. The next question is how AWS earns revenue.

The AI cloud has two business models:

• IaaS (Infrastructure as a Service): the cloud provider rents chips to model companies or large enterprises, charging by chips × hours. It is essentially selling raw compute for rent — and on Nvidia silicon, part of the profit has already been taken by the chip vendor.
• TaaS (Token as a Service): AWS serves end-enterprise customers directly through its Bedrock platform, with customers paying per input and output token. Beyond supplying the underlying compute, AWS controls the enterprise doorway, billing, compliance, and model distribution — so it participates in token revenue rather than just collecting hardware rent.

GPU clouds such as CoreWeave mainly run the IaaS model: they rent GPUs to model companies, which then use that compute to serve tokens to end customers. CoreWeave thus earns compute rent but struggles to share in Anthropic's downstream revenue growth.

AWS, by contrast, distributes Claude and other models directly through Bedrock. With its vast enterprise customer base, AWS handles model access, billing, compliance, and security in one place — letting it take a share of token revenue on top of the underlying compute income. Google and Microsoft are developing similar models, but AWS is currently the most aggressive and most mature player.

As model companies' revenue soars, AWS is a prime beneficiary. Anthropic's annualised revenue passed US$65 billion by the end of July 2026, and SemiAnalysis estimates that more than 40% of its 2Q26 revenue already came through cloud platforms. More than 100,000 customers use Claude through Bedrock. Given the depth of the partnership, Anthropic's growth is likely to translate into a meaningful token-revenue share for Bedrock.

On capital: Amazon has invested roughly US$13 billion in Anthropic to date and has committed up to US$20 billion more.

On compute: Anthropic has committed to purchasing over US$100 billion of AWS compute over the next ten years, at up to 5GW of capacity.

On silicon: Anthropic is deeply involved in Trainium’s design, giving AWS an anchor customer for the chip.

Anthropic is therefore not merely a customer but a deep partner: at the bottom of the stack it helps optimise Trainium and buys compute; at the top it feeds Bedrock high-value token revenue share.

Investors should note, however, that the partnership is not exclusive. Anthropic is also a large-scale Google TPU customer, and Claude is distributed through other cloud platforms as well.

Table 4: The Amazon–Anthropic deep binding

Layer	Nature	Key data
Compute	Preferred training partner	Project Rainier multi-gigawatt dedicated cluster; over 1 million Trainium2 chips in use; >US$100bn of compute purchases committed over 10 years
Silicon	Deep involvement in Trainium design	Trainium3's specs and design closely track Anthropic's needs
Revenue	Bedrock revenue-share partner	Over 100,000 customers run Claude via Bedrock
Capital	Three funding rounds; significant shareholder	~US$13bn invested to date; up to a further US$20bn committed
Source: Amazon, Anthropic, SemiAnalysis, iFAST compilations

Financial validation: the fastest revenue growth in 18 quarters — why AWS may be the most profitable of the big three

AWS revenue growth keeps accelerating — up 37% year on year in 2Q26, the fastest in nearly 18 quarters. Growth still trails Azure and Google Cloud, but AWS runs off a much larger base and started later in the generative-AI race: Azure was first to benefit from OpenAI and Copilot demand, and Google had its own TPUs to relieve GPU supply pressure. AWS's AI build-out came later, and its new AI racks only began converting into revenue through 2026. AWS is now in catch-up mode — its quarterly revenue increment already rivals that of Google Cloud, the fastest grower.

Figure 3: Big Three Cloud Providers: YoY Revenue Growth

Alongside the acceleration, AWS's profitability stands out. Its 2Q26 operating margin reached 39.4%, above Google Cloud's 35.6% and close to Microsoft Intelligent Cloud's 40.6%. Note, however, that Microsoft Intelligent Cloud includes not just Azure but high-margin Windows Server and SQL Server licensing and Enterprise Support, which flatter the segment margin — and as AI-infrastructure depreciation builds and the revenue mix shifts toward Azure, that margin may come down. Google Cloud's margin has surged from 17.8% in 1Q25 to 35.6%, but part of the group's shared AI R&D spend still sits at the group level, so the segment does not fully reflect model-training costs.

On disclosed figures alone, AWS's operating margin is still slightly below Microsoft Intelligent Cloud's. But allowing for Microsoft's high-margin software-licensing support and Google Cloud's incomplete absorption of shared AI R&D costs, AWS — holding a near-40% margin while revenue grows this fast — quite possibly already has the strongest pure-cloud profitability of the big three.

Figure 4: Operating margins across the big-three clouds

Investment implications: the core, plus the supply-chain extension

The AI race has entered a new phase. Beyond who owns the fastest accelerator, the market should watch who can supply memory bandwidth, CPU compute and datacenter capacity at lower cost — and monetise it through a better revenue model.

AWS has an edge on both sides: on cost, the co-design of Trainium, Graviton, Nitro, networking and the datacenter pushes down the total cost of ownership per unit of effective compute; on revenue, Bedrock controls the enterprise doorway and the deep Anthropic partnership gives AWS a share of token revenue.

Table 5: Amazon earnings forecasts and potential upside

Item	2025	2026E	2027E	2028E
EPS (USD)	6.99	10.29	11.65	12.00
EPS growth (YoY)	24.29%	47%	13%	3%
Potential upside (Based on a Fair P/E Multiple of 26x)				17%
Target price				312
Source: Bloomberg, iFAST Compilations. Data as of 19 August 2026

In-house design does not mean AWS does everything alone. AWS focuses on chip architecture and front-end design, while back-end physical design, wafer fabrication, advanced packaging, HBM and system interconnect are handled by specialist partners.

Table 6: Trainium's key supply-chain partners

Company	Role
Amazon	Architecture and front-end design; full-stack integration, monetised through Bedrock's TaaS model
Alchip (3661 TT)	Lead back-end physical design and packaging for Trainium3; leads both Trainium4 tracks
TSMC (TSM)	3nm wafer fabrication and CoWoS advanced packaging
SK Hynix / Micron	Main HBM3E suppliers for Trainium3 (succeeding Samsung, which supplied Trainium2)
Astera Labs (ALAB)	Trainium3 chip interconnect (scale-up): PCIe Gen6 switches and retimers inside the UltraServer
Credo (CRDO)	Rack-level high-speed copper (scale-out): AEC active copper cables linking servers and racks
Nvidia (NVDA)	Trainium4 plans to support NVLink Fusion for interoperability with Nvidia hardware
Source: Amazon, iFAST Compilations

Amazon has guided 2026 capital expenditure to roughly US$220 billion, most of it going to cloud and AI — and management has flagged rising memory prices as one reason capex is climbing. DRAM and HBM inflation feeds straight into cloud providers' infrastructure costs, while 3nm and CoWoS capacity remain expansion bottlenecks. AWS's long-standing partnership with TSMC helps coordinate capacity, but it cannot fully escape industry-wide supply-demand constraints.

Investors may consider the Global X Asia Semiconductor ETF (HKEX: 3119) as a one-ticket way to position for these opportunities.

Table 7: FactSet Asia Semiconductor Index earnings forecasts and potential upside

FactSet Asia Semiconductor Index	2025A	2026E	2027E	2028E
EPS (HKD)	19.3	51.7	70.6	79.7
EPS growth (YoY)		168.5%	36.4%	12.9%
Potential upside (Based on a Fair P/E Multiple of 20x)				90%
Target Price				1,594
Source: Bloomberg, iFAST compilations. Data as of 20 August 2026