The August 2026 expansion of the AWS–Nvidia partnership makes explicit a structural feature of Amazon’s artificial-intelligence strategy: AWS’s near-term capacity programme is closely coupled to Nvidia’s technology roadmap 36. AWS has agreed to add 2 million Nvidia GPUs to its data centres 21,23,30,39, supplementing an earlier commitment exceeding 1 million chips 21. Deliveries and deployment are planned across AWS facilities during 2027 and 2028 21,23.
This is more than an unusually large equipment purchase. In the short run, it converts a portion of uncertain AI-infrastructure demand into a pre-committed supply relationship. Nvidia receives order-book visibility extending into 2027–2028 23, while AWS secures access to the hardware that advanced-model customers commonly request as their default platform 21. The interesting question is not whether this commitment is large, but whether its persistence reflects a temporary shortage of substitute capacity or a more durable dependence within the cloud-production system.
Scale, scope, and the limits of what is disclosed
The announced order spans Nvidia’s Blackwell Ultra, Rubin, and Rubin Ultra generations 23. Its economic scale is plausibly measured in tens of billions of dollars when networking, storage, and cooling are included 21. Analyst extrapolations place the GPU component at roughly $70 billion of upfront Nvidia revenue, or approximately $35,000 per GPU and two years of capital expenditure 23; a separate probability-weighted estimate reaches the same $70 billion figure 23.
These estimates should not be mistaken for disclosed contractual terms. Accounts of the arrangement state that no financial commitment, contract terms, or delivery timetable has been specified 31,39. The 2-million-GPU headline is therefore the principal hard number available 31. What is firmer is the direction of demand: Nvidia said that demand exceeded expectations in the five months after Amazon’s earlier commitment 21. The announcement itself came through an Nvidia newsroom release 39, and Nvidia’s shares reacted positively after hours 39.
The partnership’s scope also extends beyond accelerators. It includes networking hardware, open models, CPUs, data-processing software, and robotics platforms 21,23. AWS’s relevant stack incorporates GB200 and GB300 NVL72 systems, the Grace Blackwell Superchip, NVLink-C2C, fifth-generation NVLink, the Transformer Engine, and DGX Cloud 35. P6 and P6e systems provide up to 1.8 TB/s of NVLink bandwidth per GPU 35.
This broader integration changes the character of the relationship. Nvidia is not merely supplying GPUs; it is becoming a strategic platform partner 21, with additional points of influence over customer choices 21 and wider control of the software environment 21. AWS gains a broader systems catalogue, a means to meet customer demand, and an avenue for physical-AI initiatives 21, as well as a packaged environment for advanced workloads 21. Nvidia, for its part, gains a distribution channel reaching beyond conventional data-centre training and inference 21.
Why Nvidia Retains Near-Term Pricing Power
The immediate economics follow from the limited elasticity of substitution for frontier workloads. Customers building advanced models continue to favour Nvidia because their software ecosystems are already optimized for its architecture 21. During critical product launches, they have also been willing to pay premium prices rather than incur delay 21. This preserves Nvidia’s pricing power 21, illustrated by a 15% Rubin price increase 17.
The current phase of AI development consequently favours Nvidia’s integrated platform 21. Its position in frontier training appears more defensible than in inference 28. This distinction matters: an accelerator may be technically substitutable in principle, yet commercially difficult to substitute when software compatibility, networking, operational tooling, and launch timing are considered together.
The demand evidence is consistent with this reading. AWS revenue reached $42.2 billion in the second quarter of 2026 34, growing 37% 11,12,13,34. Amazon’s North American segment generated $29.6 billion of operating profit in 2025, at an approximately 6.9% margin 32. Elsewhere in the cloud system, Google Cloud recorded 11 consecutive quarters of expanding profit margins 3,7,8,9,22. At the application layer, Zhipu’s API gross margin moved from negative 0.4% to positive 24.6%, although this remains materially below mature-software gross margins of 70–80% 28.
These observations do not prove that every GPU deployment will earn an adequate return. They do, however, indicate that consumption-based growth at Amazon, Microsoft, and Alphabet is being treated as validation of Nvidia’s end-customer demand and margin headroom 15. AI demand is explicitly identified as the driver of large data-centre expenditures on Nvidia hardware 38.
Amazon’s Hedge: Capacity Today, Independence Over Time
Amazon’s strategy is coherent precisely because it contains a tension. AWS is expanding its Nvidia exposure while continuing to invest in custom Trainium chips 21 and in tools intended eventually to reduce that dependence 21. Cloud providers seek proprietary silicon because Nvidia captures a significant portion of the value created by AI infrastructure 21. Nvidia’s largest customers remain determined to develop alternatives 21, and Nvidia’s share of cloud infrastructure could decline even amid overall market growth if custom silicon attains sufficient scale and software compatibility 21.
There is evidence that this internal hedge can improve Amazon’s economics where it is adopted. AWS’s custom-silicon approach, including Graviton5, is described as margin-accretive relative to x86 29. Yet the short-run and long-run positions must be distinguished. In the short run, AWS requires merchant silicon at scale to serve workloads whose software and operational requirements remain centred on Nvidia. In the longer run, Trainium, Graviton, and associated software may enlarge Amazon’s bargaining power, preserve more infrastructure value within AWS, and reduce single-supplier exposure.
The adjustment path is not frictionless. Each incremental Nvidia order strengthens the supplier Amazon ultimately seeks to constrain 21. Moreover, some market participants view neoclouds and Nvidia reference designs as superior to AWS’s own infrastructure 33. Thus, the Nvidia hardware that attracts customers into AWS may simultaneously make competing environments more legible and attractive. The partnership therefore improves AWS’s current service capacity while potentially raising the cost of achieving strategic independence later.
Cost Inflation and the Economics of Capacity
The infrastructure organism is also constrained by its component base. SK Hynix, dominant in HBM3 and a key Nvidia supplier 16, is itself heavily dependent on Nvidia as its primary customer 16. It is projected to move from a 7.73 trillion won operating loss in 2023 2,16 to 47.2 trillion won of operating profit in 2025 16. SLC NAND contract prices are forecast to rise a further 120–170% in the second half of 2026 10,34, driven by AI edge computing, networking, and automotive demand 34. Samsung reported record semiconductor-unit sales 34, even as component costs reduced margins in its smartphone and television divisions 34.
These figures identify memory and manufacturing costs as a near-term pressure point for ecosystem margins. Rising input costs threaten the roughly 75% gross margin that anchors the economics 1,15,24,26. Price increases and a richer product mix may partly offset that pressure 15. Nvidia’s 15% Rubin price increase was itself partly a response to HBM4-driven margin compression 17. Whether these mechanisms preserve the approximately 75% gross-margin line 15 is therefore the point at which Amazon’s capacity commitments and Nvidia’s supplier economics most directly meet.
Mutual Dependence and the Financing Structure
Dependence runs in both directions. Three direct customers account for 21%, 17%, and 16% of Nvidia’s quarterly revenue, exceeding half in aggregate 15. Demand has not yet broadened beyond a small group of hyperscalers 15, although direct-customer concentration may overstate concentration among ultimate buyers 15, and roughly half of Nvidia’s revenue now comes from non-hyperscalers 24.
Nvidia also carries $279 billion in obligations, more than half associated with HBM4 memory 17, and raised $25 billion through a high-grade bond sale in June 2026, its first such issuance since 2021 4,5,14. Around this core lies a more elaborate financing structure. Nvidia effectively lends to companies unable to qualify for traditional financing 15; compute is being securitized with pension funds and insurers holding the liabilities 37; Nvidia participates in a $500 billion consortium with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR 23; and large customers’ profits fund order backstops 24.
This structure may extend the capacity build-out, but it also binds participants more closely to a common demand cycle. For Amazon, the stated tail risk is that an AI-cycle downturn would produce a highly correlated drawdown for Nvidia and Amazon 23. Nvidia’s supplier concentration is separately identified as a systemic dependency risk 20,31. The utilisation arithmetic deserves similar caution: operating these GPUs at 70% utilisation for five years at 2026 rates is considered unrealistic 23. The relevant uncertainty is therefore not merely whether demand is high today, but whether utilisation, pricing, and financing can remain mutually consistent across the full deployment horizon.
Nvidia’s Results as an Ecosystem Signal
Nvidia’s latest results function as a market test of the broader AI-infrastructure thesis. Second-quarter revenue was $96.2 billion, up 106% year over year 18,24,27, while data-centre revenue reached $89.0 billion, up 117% 19,21,25. Revenue exceeded the roughly $92.17 billion LSEG consensus 24, and earnings per share of $2.22 surpassed the $2.10 expectation 24. Management guided to approximately 70% growth for fiscal 2028, describing the outlook as supply-constrained 24, compared with Street expectations of roughly 44% 24.
The countervailing evidence is that growth is decelerating from 106% 24, and the size of Nvidia’s guidance beats has diminished even as its 13-quarter streak continues 24. The market nonetheless responded with a near-9% share-price gain 24 and an approximately $441.5 billion one-day increase in market capitalisation, the second-largest single-day dollar gain in U.S. market history 24.
Valuation remains a less settled matter. The shares trade near $220 24, below their $236.54 52-week high 6,24, and the forward multiple is below the 35–40 times recorded at the same point in 2024 and 2025 24. A forward earnings multiple of 18.2 times appears inexpensive relative to the IT-sector median of 22.4 times only if forward EPS of $11.73, nearly twice trailing EPS of $6.25, is realised 15. At the same time, the 21.5-times-sales multiple stands at the 89th percentile of peers 15 and embeds substantial profitability assumptions 15 that presume near-perfect execution 24.
Market structure moderates some of the apparent volatility. Nvidia’s heavy index weighting structurally limits upward volatility 15. Market makers harvest theta in a top-five options name with an expected move of about 6% 15, while a 10% move in Nvidia’s share price is equivalent to roughly AMD’s entire market capitalisation 15. These mechanics do not settle the underlying economic question, but they help explain why the market’s response should not be read as a simple measure of industrial certainty.
Implications for AWS
Under current conditions, the AWS–Nvidia commitment provides AWS with assured access through 2027–2028 to the hardware most readily demanded for advanced AI workloads 21,23, while giving Nvidia multi-year order visibility 23. This is a rational short-run allocation where substitute platforms remain constrained by software, systems integration, and customer time horizons.
The longer-run equilibrium is less determined. Amazon’s custom-silicon investments offer a route to better margins and greater independence 21,29, but the scale of its Nvidia purchases also reinforces Nvidia’s technological and commercial position 21. Meanwhile, costly memory inputs 10,15,34, Nvidia’s large obligations and financing arrangements 17,23,24,37, and the possibility of correlated weakness in an AI downturn 23 create structural vulnerabilities worth monitoring.
The central issue, then, is not whether Amazon’s Nvidia relationship is beneficial in the aggregate. It plainly expands AWS’s present capacity and product range. The more exact question is whether AWS can use that capacity to cultivate sufficient custom-silicon scale, software compatibility, and customer acceptance before the dependence inherent in the current arrangement becomes more durable than the strategic hedge designed to offset it.