Amazon's AI Empire: Inside the Anthropic Stake, AWS Surge, and Earnings Distortion

The evidence from April through 5 August 2026 presents artificial intelligence as the central force shaping Amazon’s strategy, valuation, earnings quality, and risk profile. Amazon is assembling an integrated AI ecosystem spanning AWS infrastructure, Trainium and Inferentia chips, Bedrock, enterprise inference, agentic applications, advertising, e-commerce, robotics, and frontier-model research ^{42,65,67,71,75,76}. Its investment in Anthropic connects three sources of value—AWS demand, chip utilization, and the mark-to-market value of an equity stake—while making reported earnings more volatile and less comparable across periods ^4,49,64,67.

The operating evidence is constructive. Amazon’s AI and chip activities have exceeded a $25 billion annualized revenue run rate, up from approximately $15 billion in the preceding quarter, while AWS remains the company’s principal profit engine and enterprise demand for AI appears strong ^{21,33,45,46,48,49,51,61,69}. Advertising provides a separate and more established monetization channel: second-quarter revenue reached $19.8 billion, an increase of 26%, supported by high-intent shopping behavior, first-party data, and the growing use of AI in campaign execution ^{61,62,66,68,81,82}.

Yet the headline second-quarter profit figure is a poor proxy for recurring operating performance. Amazon reported net income of $62.6 billion and earnings per share of $5.75, compared with expectations of $1.82, but approximately $53.4 billion of pre-tax, non-operating, unrealized income was primarily attributable to the Anthropic investment ^{33,45,46,49,63,68}. Adjusted analyses estimate operations-driven net income at roughly $20.8 billion after removing the Anthropic gain and other mark-to-market effects. A simpler calculation, subtracting the full $53.4 billion from reported net income, produces approximately $9.2 billion. The difference arises from the adjustment methodology and is itself a reason to reconcile reported results carefully ^67,68.

The Earnings Signal Must Be Normalized

Anthropic’s unrealized gain distorts Q2 comparability

Multiple sources corroborate that Amazon’s second-quarter net income and EPS were materially boosted by an unrealized mark-to-market gain on its Anthropic stake ^{34,45,46,63,64,67,68,71}. The most specific account identifies a $53.4 billion gain, classified as pre-tax, non-operating, and unrealized. It represented approximately 66% of second-quarter pre-tax income of $80.857 billion ^45,49,67,70. This was not cash, operating profit, or recurring earnings ^34,67. It therefore reduces the usefulness of headline EPS and complicates comparison with earlier periods ^34,49,63,67.

The more informative indicators are operating income, AWS growth, free cash flow, advertising, and adjusted earnings. Amazon generated approximately $27.5 billion of quarterly operating income—$27.46 billion in one reported formulation—which offers a more relevant measure of underlying progress than the $62.6 billion net-income headline ^49,68,95. Adjusted net income excluding the Anthropic gain and incorporating increased depreciation reportedly rose 48%, suggesting that the underlying business still delivered solid growth even though the reported result overstated recurring economics ⁹⁴.

This is not only a matter of historical presentation. A decline in Anthropic’s valuation could reverse a substantial portion of the gain and produce a large loss for Amazon ⁶⁷. The investment has consequently introduced earnings volatility beyond Amazon’s historical benchmarks ⁶⁷. The precise ownership percentage and financial scale of the stake are not consistently disclosed. One highly corroborated claim places Amazon’s ownership at 15%, while other reporting states that Amazon has not publicly disclosed the precise nature or size of the holding ^{13,33,64,71,98}. Amazon has reportedly invested $8 billion in Anthropic and may have a further $20 billion commitment contingent on commercial milestones, but the terms and accounting treatment remain incompletely documented ^14,38,68,78. These differences should remain unresolved disclosure issues rather than being reconciled through assumption.

Comparable accounting gains at other hyperscalers reinforce the sector-wide nature of the issue, but they should not be treated as direct evidence of Amazon’s operating economics. Claims that such gains represented two-thirds or nearly 80% of Alphabet’s earnings, together with reports that Microsoft recorded a $3.2 billion Anthropic-related gain, show that the Anthropic valuation cycle is affecting several hyperscaler income statements and may complicate cross-company comparisons ^17,18,26,37.

AI Demand and the Emerging AWS Model

A rapidly growing but company-defined revenue pool

The most consistently repeated operating datapoint is that Amazon’s AI and AI-chip businesses exceeded a $25 billion annualized revenue run rate in the second quarter of 2026 ^{33,45,46,51,61,67,69}. That followed a prior-quarter run rate above $15 billion ^{21,33,46,49,69}. The triple-digit growth rate and sequential increase are meaningful evidence of demand momentum, although Amazon does not disclose Anthropic’s precise contribution to AWS revenue or to individual revenue lines ^67,69.

The commercial model is broader than providing access to one foundation model. Amazon is positioning Bedrock as a governed enterprise distribution layer for models including Anthropic’s Claude, alongside Amazon Q, inference capacity, agentic applications, custom chips, and related infrastructure ^42,76. Claude Sonnet 5 is available through AWS GovCloud and distributed through Bedrock, while Claude Code can access Anthropic models through Bedrock rather than the direct Anthropic API ^85,89. Bedrock changes the procurement, auditability, governance, and data-residency calculus relative to direct API access; existing AWS spending commitments may also absorb Bedrock usage costs ⁸⁹. The relevant competitive choice is therefore not simply Amazon versus Anthropic, but direct Anthropic API access versus AWS-mediated consumption ⁸⁹.

This ecosystem can generate recurring cloud revenue and differentiation even if Amazon’s proprietary frontier models do not lead on every benchmark. The company continues to support Nova while reallocating resources toward a new frontier-model effort led by Pieter Abbeel, who joined through the Covariant acquisition ^75,90. The movement from broad experimentation toward more focused frontier-model competitiveness, integrated commercial products, and disciplined resource allocation is consequential ⁷⁶. Amazon’s AI capabilities also extend into e-commerce product-content management, shopping search, COSMO, Alexa for Shopping, Prime Video, recommendations, and advertising ^62,77,83,88.

Strategy rationalization is both discipline and evidence of uncertainty

Amazon created its AGI Lab in 2024 after hiring most of the team behind Adept ⁷³. It subsequently closed the lab, conducted layoffs, discontinued or reduced certain internal models, consolidated AI activity, and redirected engineering talent and scarce computing capacity toward higher-priority frontier objectives ^{73,74,75,76,90,91}. The company is reportedly winding down some costly proprietary-model efforts while preserving or expanding frontier-model research and the Nova platform ^43,90.

The appropriate interpretation is portfolio rationalization rather than retreat. The changes reflect industry-wide concern about the cost and uncertain effectiveness of proprietary models ⁴³, while preserving Amazon’s ability to monetize AI through infrastructure, Bedrock, inference, advertising, and enterprise applications. Management has articulated a pathway from present spending to future monetization and returns, and investors have responded more favorably when that pathway is explicit ⁸⁷. Management has even used the phrase “guaranteed returns” in discussing the infrastructure thesis, although this should be regarded as an assertion, not evidence that returns are contractually assured ⁵³.

The strategic tension is straightforward. Amazon must invest enough to capture demand without duplicating effort across internal models, third-party partnerships, custom silicon, and data-center capacity. Resource reallocation is therefore a potentially positive signal for capital discipline, but it also indicates that earlier AI initiatives did not all meet the company’s hurdle rates. The closure of the AGI Lab and reported reductions in flagship projects remain the principal counterweight to the bullish interpretation ^{73,74,75,76,90}.

Capital Intensity and the Test of Cash Returns

Demand is supporting the buildout, for now

Amazon is participating in a broad hyperscaler investment cycle alongside Alphabet, Meta, and Microsoft. Collectively, the group is committing hundreds of billions of dollars to data centers, servers, networking, memory, custom silicon, and related infrastructure ^{17,19,22,32,44,49,96}. Amazon reportedly spends tens of billions of dollars annually on data centers and Trainium and Inferentia chips. Its infrastructure program encompasses construction, servers, networking equipment, memory, and proprietary chips ^{33,47,49,65,69,87}. Amazon and Microsoft have also been described as collectively spending approximately $400 billion on AI, although this low-source-count estimate should not be treated as a precise company-level forecast ⁹⁷.

There is evidence that demand is beginning to validate the buildout. Amazon’s second-quarter results and management commentary were interpreted as showing that existing demand supports major hyperscaler infrastructure spending, and investors expressed confidence that AI investment is driving demand for Amazon’s services ^52,53. A reported $100 billion, 10-year Anthropic commitment to run models on AWS Trainium contributed to the sequential increase in AWS backlog, which reached approximately $496 billion in one account ^49,67. A $410 million compute agreement with Recursive Superintelligence, representing roughly 63% of that company’s disclosed funding, provides another example of demand for AWS capacity, although the arrangement reportedly contains no equity investment component ^20,28,72,80.

Free cash flow reveals the adjustment cost

The cash-flow burden is material. Amazon attributed a $25.8 billion year-over-year deterioration in free cash flow primarily to a $66.1 billion increase in property-and-equipment purchases related to AI infrastructure ⁶⁷. The company has issued bonds to help finance the buildout and has reportedly carried $119 billion of debt following an 81% increase ^17,44. Heavy investment binds cash in long-lived assets and creates execution, demand, and return-on-investment risks ^33,47,57,87. It also has an opportunity cost: capital directed toward AI cannot simultaneously be deployed to retail, logistics, or other AWS opportunities ⁹².

The same backlog and customer commitments used to support the bullish case may partly reflect Anthropic and other AI laboratories. One estimate attributes approximately 51% of relevant Amazon RPO exposure to OpenAI and Anthropic, while another alleges that a $500 billion AWS backlog may be concentrated primarily in those customers ^30,46,67,94. These isolated estimates should not be treated as verified disclosures. They nevertheless identify an important diligence question: whether AI demand reflects diversified enterprise adoption or a smaller number of heavily funded model developers. The source material suggests that Google’s backlog may be more concentrated in Anthropic than Amazon’s customer base, implying relatively better diversification for Amazon, but that comparison also rests on limited evidence ^17,96.

The Anthropic relationship may further blur the distinction between financing and operating demand. Reports that Amazon and Microsoft are investing in Anthropic while requiring it to purchase cloud services suggest that investment activity may be linked to reported cloud revenue ⁹⁶. This claim has low corroboration and should be investigated rather than accepted as proof of circular economics. Investors should examine customer funding sources, minimum commitments, capacity reservations, cancellation rights, and cash collection instead of relying on backlog headlines alone.

Anthropic: Strategic Asset and Concentrated Exposure

Amazon is both a major Anthropic investor and the company’s primary cloud infrastructure provider through AWS ^4,67. The relationship supports AWS demand, Trainium utilization, Bedrock distribution, and Amazon’s AI ecosystem narrative, while increasing the sensitivity of the Anthropic stake itself to valuation changes ⁶⁷. Amazon and Anthropic have reportedly entered into a $100 billion collaboration, alongside a previously announced agreement ^42,49. Anthropic has also agreed, according to one source, to run models on AWS Trainium for ten years ⁶⁷.

Anthropic’s reported valuation has risen to nearly $1 trillion, supported by multiple sources, and the company has reportedly raised $65 billion in funding ^{3,5,6,8,9,12,16,25,26,29,30,36}. Other estimates place annual revenue near $47 billion, with ARR potentially rising from $9 billion at the start of 2026 to approximately $74 billion by late July ^2,11,93. More speculative claims place a possible future valuation above $2 trillion or even $5 trillion, while one disputed estimate says Anthropic would require $600 billion of annual revenue to break even ⁹³. These figures describe the range of market expectations but should not be incorporated into an AMZN valuation without independent verification.

Anthropic’s own economics remain contested. Some claims suggest that it was profitable or approaching profitability in a recent quarter, with inference gross margin estimated at around 40% or overall gross margin around 70% ^{7,10,15,23,30}. Other commentary alleges that a 90% compute subsidy or onboarding discount from xAI, or token costs approximately five times public pricing, may have supported apparent profitability ³⁰. These claims are low-source-count and potentially contradictory. SpaceX has reportedly leased compute capacity to Anthropic, and Anthropic is exploring proprietary chips to diversify its infrastructure, developments that could reduce its dependence on AWS over time ^1,24,69.

Legal, regulatory, and reputational risks add another layer of exposure. Anthropic faced or reached a reported $1.5 billion settlement with authors concerning copyright and intellectual-property claims, and it challenged the Pentagon’s designation of the company as a supply-chain risk ^90,91. The vertical relationship between Amazon, an infrastructure provider, and a leading foundation-model developer may attract regulatory scrutiny ⁵⁵. Any deterioration in Anthropic’s valuation, commercial relationship, regulatory standing, or ability to finance infrastructure would therefore be a company-specific tail risk for Amazon ^68,71.

Advertising Offers the Clearest Independent Monetization Signal

Amazon’s advertising business generated $19.8 billion in second-quarter revenue, up 26%, making it a more tangible proof point than the Anthropic mark-to-market gain ^{61,62,68,81,82}. Advertising benefits from high-intent shopping behavior and first-party data, giving Amazon a differentiated set of signals for targeting and conversion ⁶⁶.

Amazon is applying AI across campaign setup, audience development, creative generation, bidding, product recommendations, keyword discovery, search-term analysis, budget allocation, and performance measurement ^66,68. It is also introducing agentic and conversational advertising through Alexa+ and Alexa for Shopping and expanding AI capabilities in digital advertising ^68,84. This creates a credible route from AI investment to monetization that does not depend exclusively on speculative model-company valuations.

The principal risks are advertiser adoption, execution, and the possibility that AI features fail to improve campaign performance sufficiently to justify their cost ⁸⁴. Even so, advertising growth, a reported 12% reduction in fulfillment costs, and broader e-commerce scale provide useful offsets to AI’s capital intensity ⁵⁸. Amazon’s online-store and third-party-seller businesses reportedly generate more than $403 billion in annual revenue, while advertising revenue reached $56.2 billion in 2024, underscoring the strategic value of applying AI to discovery, conversion, and monetization across the retail ecosystem ^58,66.

Valuation, Execution, and Governance Implications

Amazon’s share price responded favorably to the second-quarter report and the AI narrative, rising approximately 15.3% to $271.58 in one account. Other reports describe a strong rally, a gain of more than $400 billion in market value during the week, and a market capitalization above $3 trillion ^{40,41,50,52,59,60,69}. Shares were reported to be approximately 20% higher year to date and at record or all-time highs ^49,50,52,54. The reaction indicates that the market currently rewards evidence of AI demand and a credible long-term monetization framework, despite earlier technology-sector concerns that AI capital expenditure might prove excessive or uneconomic ^53,97.

AMZN is increasingly valued as a central participant in the AI investment cycle rather than solely as an e-commerce and cloud company ^52,87. A $10 trillion valuation scenario has been described, but its feasibility depends on Amazon’s ability to fund additional delivery investment and earn adequate returns on a very large capital base ⁶⁵. Amazon’s history of monetizing heavy reinvestment, including controversial investments in dominant franchises, supports management’s credibility but does not guarantee that the present AI cycle will produce the same result ³⁴.

The principal risk is not financial capacity. Amazon, Alphabet, and Meta are described as highly profitable and sufficiently cash-rich to finance aggressive AI investment, and Amazon reported $551.6 billion of stakeholder equity ^31,33. The risk is that returns arrive later, prove lower than expected, or remain concentrated in a few customers and partners. Amazon’s sensitivity to enterprise technology spending and energy prices increases the importance of demand durability and infrastructure operating costs ⁴².

The buildout also increases Amazon’s energy and environmental footprint because data centers and large-scale technology infrastructure consume substantial power ^39,61. Amazon reported 11% corporate growth alongside a 4% reduction in carbon intensity and has recorded six consecutive years of declining carbon intensity ⁸⁶. This is a favorable efficiency signal, but it does not remove the absolute energy, emissions, labor, supply-chain, and governance concerns associated with expanding data centers, semiconductors, robotics, satellites, logistics, and AI ⁵¹. Amazon’s 22.9% stake in X-Energy, valued at approximately $1.209 billion, may provide a strategic link to future energy supply but adds another investment exposure ³⁹.

Operational controls will become more important as deployment expands. One internal Amazon project using Anthropic’s Claude Sonnet incurred an additional $1.8 million of costs and exceeded its budget by 860%; other reported overruns totaled $541,000 and $134,000 ⁸⁸. These examples are isolated rather than representative of company-wide economics, but they demonstrate how inference and experimentation costs can escape control. Amazon has proposed consumption monitoring, automatic spending caps, usage alerts, and project-level cost attribution in response ⁸⁸.

The risk extends beyond direct overruns. Infrastructure may become obsolete if model architectures, chips, or workloads evolve more quickly than expected ⁸⁷. Large infrastructure commitments expose cash generation to execution, demand, and return-on-investment risk, while overinvestment could raise operating costs and weaken investor confidence if growth does not justify spending ^{27,33,34,42,47,49,87,88}. Amazon’s wider agenda—including robotics, satellites, grocery, pharmacy, rapid delivery, autonomous transport, and Prime Video—provides optionality but also creates potential capital-allocation dilution ^35,44,62,90. Expansion into adjacent markets is consistent with Amazon’s long-term orientation, but it increases the opportunity cost of directing scarce capital and engineering resources toward AI ^79,92. Other equity holdings, including Rivian, demonstrate that non-core investments can create material earnings volatility: Amazon recorded a $12.7 billion Rivian-related write-down in 2022 while continuing to hold an equity investment ⁵⁶.

Analytical Framework and Monitoring Priorities

Amazon should now be analyzed as an AI infrastructure platform with exposure across power, data centers, networking, chips, cloud services, foundation models, enterprise applications, advertising, and consumer interfaces. Its advantage lies less in owning the single best model than in combining infrastructure scale, AWS distribution, proprietary silicon, first-party commerce data, customer relationships, and partnerships with leading AI laboratories ^57,67,88.

The constructive case rests on operating leverage from AI demand. AWS should benefit as enterprise customers move from experimentation toward inference and agents; Trainium and Inferentia may improve economics and utilization; Bedrock can capture model consumption within AWS procurement and governance workflows; and advertising can monetize AI-enhanced discovery and campaign performance ^42,89. The reported $25 billion-plus AI run rate, strong enterprise demand, $100 billion Anthropic collaboration, and continued expansion of AI-enabled advertising are the principal supporting indicators ^45,61,67,68.

The countercase concerns return dispersion and earnings quality. AI infrastructure requires substantial upfront capital expenditure, depreciation, energy, and operating commitments before utilization and pricing are fully established. Changes in model architecture, customer funding, or cloud procurement could impair asset returns. Anthropic-related gains make current earnings appear stronger than recurring operations, while a valuation reversal could create a large reported loss. Customer concentration, potential circularity between investment and cloud commitments, and competition from Microsoft, Google, Oracle, Meta, OpenAI, Anthropic, and open-source models further complicate the outlook ^{35,46,71,94,96}.

The appropriate framework is therefore a sum-of-the-parts and cash-return analysis rather than a simple GAAP earnings multiple. AWS should be assessed on recurring cloud growth, backlog quality, inference utilization, margin, and capital intensity. Advertising should be assessed on revenue growth, monetization of first-party data, and AI-driven advertiser productivity. Retail and logistics should be evaluated separately, including the need for continued delivery investment. Anthropic should be treated as a volatile strategic investment and commercial dependency, not recurring operating income. Amazon’s third-quarter net-sales guidance of $197 billion to $202 billion provides a near-term operating anchor, but it does not resolve the longer-term return profile of AI capital expenditure ⁶⁸.

Four indicators deserve particular attention. First, AI revenue growth should be tested after adjusting for Anthropic-related commitments, with emphasis on whether it produces incremental AWS operating profit. Second, free cash flow should be monitored for stabilization as property-and-equipment spending moderates or utilization rises. Third, customer concentration in OpenAI, Anthropic, and other model developers should be tested against the isolated estimates. Fourth, the AGI restructuring should be evaluated by whether it produces higher returns per dollar of compute and engineering talent. Management has supplied a strategic rationale for the spending, and the market has rewarded that rationale; the next phase requires measurable cash returns rather than further narrative expansion ⁸⁷.

Several claims should remain explicitly discounted. Estimates of Anthropic’s profitability, inference margins, ARR, future valuation, compute subsidies, and break-even revenue are inconsistent and often supported by only one source ^30,93. Claims regarding 51% AI infrastructure exposure, a $500 billion AWS backlog concentrated in two customers, and the precise economics of cloud-linked investments are likewise insufficiently corroborated to serve as base-case assumptions ^30,94,96. The narrower and more durable conclusion is that Amazon has meaningful, growing AI exposure; Anthropic is strategically and financially material; second-quarter earnings were heavily distorted by an unrealized gain; and the investment thesis now depends on converting infrastructure scale into recurring, cash-generative returns.

Key Takeaways

• Normalize second-quarter earnings. The $53.4 billion Anthropic-related pre-tax gain drove most of the difference between Amazon’s $62.6 billion reported net income and underlying operating performance. Recurring economics should be assessed through operating income, AWS, advertising, and free cash flow ^67,68.

• Separate demand from returns. AI and chip activities have exceeded a $25 billion annualized run rate, but capital expenditure, depreciation, energy, customer concentration, and obsolescence risks remain substantial ^{33,45,46,47,51,61,67,69,87}.

• Treat Anthropic as both catalyst and tail risk. The relationship supports AWS, Bedrock, Trainium utilization, and Amazon’s AI positioning, while valuation declines, legal issues, partner dependence, or customer concentration could impair earnings and strategic momentum ^4,67,71,91.

• Give greatest weight to advertising as independent evidence. Revenue of $19.8 billion and growth of 26%, combined with high-intent commerce data and AI-enabled campaign tools, provide a more durable monetization signal than mark-to-market investment gains ^61,66,82.

Under current conditions, the evidence supports a constructive view of Amazon’s AI operating opportunity, but not an unqualified endorsement of its capital cycle. The short-run accounting gain is non-recurring; the long-run question is whether AWS, proprietary silicon, Bedrock, advertising, and related applications can earn normal—and eventually superior—returns on the capacity now being built.