Two popular versions of the bull case are already wrong. Amazon has not established frontier-model leadership over OpenAI, Anthropic or Google. And it is not enjoying an asset-light AI boom: net cash capital expenditure reached $128.3 billion in 2025 and approximately $96.3 billion in the first half of 2026. Management now expects roughly $220 billion for the full year.

One of Amazon’s most consequential AI investments predates the current boom by almost a decade.

In 2015 Amazon acquired Annapurna Labs. The AWS Nitro technical documentation attributes a family of AWS-designed components to the Annapurna team: Nitro hardware, Graviton CPUs, Trainium and Inferentia accelerators, Nitro SSD and the Aqua accelerator used in Redshift.

Nitro was the foundation. It moved virtualization, networking, storage and security functions into dedicated hardware, reducing the amount of work imposed on the host CPU and giving AWS tighter control over the server architecture. Graviton extended that control into general-purpose computing. Trainium targets machine-learning training; Inferentia targets inference.

Amazon does not need to beat Nvidia on every benchmark for this strategy to work. It needs to optimize the complete system better than it could by purchasing every component separately. Control over chip design, server configuration, memory, networking, compilers, scheduling and billing can lower the cost of producing a useful token and keep more of the economics inside AWS.

Amazon attaches large numbers to that opportunity. Its 2025 shareholder letter said Trainium2 offered roughly 30% better price-performance than comparable GPUs, Trainium3 improved on Trainium2 by 30–40%, and Trainium could eventually save tens of billions of dollars in annual capex while adding several hundred basis points to AWS operating margin. None of those figures yet represents independently measured, realized savings.

The silicon strategy is becoming commercial rather than experimental. Amazon reported in its Q2 2026 results that its chips business exceeded a $25 billion annualized revenue run rate. Yet that label combines Graviton, Trainium and Nitro. Amazon does not disclose revenue by chip family, utilization, gross margin or a standalone semiconductor profit-and-loss statement.

Software is the pressure point. AWS’s Neuron SDK supports PyTorch, JAX, Hugging Face and vLLM, and AWS has invested in OpenAI-compatible serving interfaces. At the same time, AWS acknowledges that specialized CUDA and Triton kernels may need to be rewritten for Trainium’s execution and memory model. A customer that has to rebuild its software stack may not experience the theoretical price advantage.

This creates a clean falsifier: if high-value customers continue to require Nvidia CUDA for most workloads, and Trainium adoption remains concentrated in Amazon, Anthropic and a handful of strategic partners, the custom-silicon moat will be narrower than the headlines suggest.

Training creates the spectacle. Inference creates the recurring bill.

A model may be trained or materially updated periodically. A production application invokes it repeatedly; when a customer searches, an employee asks a question, an agent retrieves a document or a recommendation appears. Each interaction can also consume databases, networking, security and monitoring.

But recurring usage does not guarantee recurring profit. Stanford’s 2025 AI Index found that the cost of querying a model at approximately GPT-3.5 performance fell from $20 to $0.07 per million tokens between November 2022 and October 2024—a decline of more than 280 times. Inference volume can soar while revenue per token collapses.

The bull case therefore requires workload growth to outrun price compression while custom chips protect AWS margins. It also depends on the surrounding system. An AI assistant that increases database, storage, security and application usage can enlarge the total AWS bill even as inference becomes cheaper.

Bedrock distributes that system. Amazon says six engineers built its Mantle inference engine in 76 days, Bedrock usage nearly doubled in March 2026, and first-quarter token volume exceeded all previous years combined. The missing numbers matter more: Amazon has not disclosed absolute token volume, revenue per token, model mix or contribution margin.

Amazon’s relationships with Anthropic and OpenAI provide demand visibility, but they also require analytical discipline.

Anthropic announced a new agreement in April 2026 committing more than $100 billion over ten years and up to five gigawatts of AWS capacity, spanning Trainium2 through Trainium4.

OpenAI and Amazon announced an expansion of their AWS agreement by $100 billion over eight years. Amazon’s 10-Q corroborates the commitments.

The headline values are enormous, but commitments are not recognized revenue or profit. The filings make clear that realization depends on customer usage and Amazon’s performance.

The relationships also complicate the capital-allocation story. Amazon’s Q2 2026 10-Q reports $28.7 billion of OpenAI preferred stock at 30 June and another $21.3 billion funded after quarter-end.

It also reports a $10 billion Q2 investment in Anthropic nonvoting preferred stock; its wider Anthropic holdings include preferred stock and convertible notes subject to an ownership cap.

The countercase is formidable. Microsoft has distribution through Microsoft 365, GitHub, Dynamics and security. Google has Gemini, Search, YouTube, TPUs and deep AI research. Oracle can win power-constrained, database-adjacent workloads.

Neoclouds such as CoreWeave can specialize in Nvidia capacity and compete on speed or price. The U.K. Competition and Markets Authority’s cloud investigation has already identified switching barriers and market power concerns around the largest cloud providers.

Amazon therefore does not need to be the only home of inference. It needs to capture a large enough share of production AI workloads, and enough adjacent consumption, to earn a return on its capacity.

Amazon’s most distinctive AI laboratory may not be a chatbot. It may be the warehouse.

Amazon said in June 2025 that it had deployed its one-millionth robot across more than 300 facilities. Its DeepFleet system is designed to coordinate the fleet, and Amazon says it can improve robot travel time by 10%.

The 10% figure comes from Amazon. The company has not published DeepFleet utilization, cost-per-package savings or a bridge from faster robot travel to fulfillment margin.

The system is broader than a single machine. Hercules moves up to 1,250 pounds of inventory. Proteus navigates around employees. Vulcan combines vision and touch. Amazon has been testing conversational instructions and expanding robotics in Europe.

This is physical AI connected to sensors, inventory, workers and delivery promises. It gives Amazon a live environment where software can be tested against real work every day.

If the system works, the benefit is not merely fewer labor hours. Better orchestration can place more selection closer to customers, reduce travel inside a fulfillment center, improve throughput and make same-day delivery viable in more regions.

Amazon said in Jassy’s 2025 shareholder letter that it had built more than 85 U.S. same-day delivery fulfillment centers carrying its top 90,000 SKUs and had delivered more than 500 million same-day units in 2026 thus far. Those figures are company-reported plans and operating metrics.

This is where robotics connects to customer obsession. Amazon already owns retail search, Prime, third-party sellers, fulfillment and advertising. Its FY2025 results said more than 300 million customers used Rufus to research, compare and buy products. Amazon separately says shoppers using the assistant are more than 60% more likely to purchase during that trip.

That is an association reported by Amazon, not proof that Rufus caused the purchase. The company does not disclose the resulting gross profit, advertising yield or inference cost.

The flywheel is plausible: AI improves product discovery; better discovery improves conversion; conversion increases the value of sponsored listings; denser demand improves the economics of regional fulfillment; better fulfillment strengthens Prime and seller economics.

But robotics also supplies a warning against promotional writing. In February 2026, Amazon updated its announcement to say it was no longer using Blue Jay in operations, although its technology would continue to support employees. Blue Jay’s withdrawal does not invalidate the robotics strategy. It shows that impressive demonstrations can be cancelled or repurposed before producing disclosed savings.

The labor question is unresolved too. Amazon reports a 34% five-year improvement in its Recordable Incident Rate. The Strategic Organizing Center’s OSHA-based analysis found Amazon’s 2024 warehouse injury rate was 66% higher than non-Amazon warehouses, with serious injuries nearly double. These are not like-for-like measures, and neither establishes that robotics caused the change.

The financial proof must come from cost per unit, units per labor hour, throughput, fulfillment margin, delivery density, inventory placement and depreciation. Robot counts are an input. They are not a return on capital.

“Customer obsession” can sound like corporate mythology. In the AI cycle, it has a concrete economic meaning: Amazon already owns many of the interfaces where AI can improve a transaction.

Rufus can help a customer decide. Alexa can become a shopping interface. A seller can use AI to write listings and forecast demand. An advertiser can generate creative and optimize a campaign. A warehouse manager can direct robots. A delivery planner can optimize routes. Each use case can improve a different part of the same ecosystem.

Amazon therefore does not need to discover one killer AI product. It can use AI to increase conversion, advertising yield, Prime frequency, seller retention, fulfillment density and employee productivity.

Amazon says its Q Developer Java migration saved 4,500 developer-years and was expected to reduce annual costs by $260 million. That company-reported case study is not a measured, company-wide productivity result. It does show how Amazon can capture an internal return without selling a new product.

The risk is that customer-facing AI reduces friction for consumers while weakening Amazon’s economics. A conversational assistant could make product comparison easier, including comparison with products sold elsewhere.

Generative search could reduce sponsored-listing visibility or shift discovery to a model provider. AI could also increase inference and customer-service costs faster than it increases conversion.

Adoption is the beginning of the measurement, not the end. The relevant outcome is incremental gross profit after inference costs and any effects on advertising, fulfillment and retention.

Amazon’s financial statements show a profitable engine financing an extraordinary buildout.

AWS revenue rose from $80.1 billion in 2022 to $128.7 billion in 2025. AWS operating income rose from $22.8 billion to $45.6 billion, while operating margin moved from 28.5% to 35.4%. In Q2 2026, AWS revenue reached $42.2 billion, up 37% year over year, and operating income reached $16.6 billion, a 39.4% margin. AWS is only about 18% of Amazon’s 2025 revenue, but it generated roughly 57% of consolidated operating income.

Advertising adds a second growth and monetization engine, although Amazon does not disclose its standalone operating margin. Advertising revenue reached $68.6 billion in 2025, up from $37.7 billion in 2022, and reached $19.8 billion in Q2 2026, up 26% year over year. That revenue growth can help cushion the consolidated investment cycle while AWS expands.

Now the paradox: Amazon’s trailing-twelve-month operating cash flow reached $161.4 billion through June 2026, yet Amazon-defined trailing free cash flow was negative $7.6 billion after $169.0 billion of net property-and-equipment purchases. AWS net property and equipment rose from $190.1 billion at the end of 2025 to $263.8 billion in just six months.

That free-cash-flow definition does not include equity investments. The same 10-Q reports $39.8 billion of acquisition and other investment cash payments in the first half of 2026, including OpenAI and Anthropic funding, plus the further $21.3 billion OpenAI investment after 30 June. The capex-to-cash-flow bridge is necessary, but it is not Amazon’s whole capital-allocation story.

The backlog is large. Long-term performance obligations rose to approximately $496 billion at June 30, 2026, from $244 billion at the end of 2025, with a 6.4-year weighted-average remaining life. But those commitments are primarily AWS, usage-dependent and not a guarantee of margin.

The next risk is depreciation. Amazon reports that server and networking equipment generally has a useful life of five to six years, while data-center buildings can last more than 30 years. Much of the 2026 server and network spend will begin depreciating as it is placed into service. A cash-flow recovery can therefore coexist with an income-statement margin squeeze if utilization lags or equipment becomes obsolete quickly.

Net income is also a poor shortcut. Q2 2026 net income included $53.4 billion of non-operating income, primarily from the revaluation of Amazon’s Anthropic investment. That is not operating proof. The operating evidence is AWS growth, AWS margin, customer usage, backlog conversion, capex productivity and free cash flow.

Amazon closed at approximately $258.63 on 21 August 2026, implying an equity value of roughly $2.8 trillion. That valuation is the final filter. Amazon can execute well and still produce a mediocre shareholder return if the market has already priced in the cash-flow recovery.

Yahoo Finance placed Amazon at roughly 22.6 times forward earnings as of 19 August 2026, but that apparently simple multiple deserves caution. Forecast definitions differ across providers, capital intensity is rising, and trailing net income is distorted by the Anthropic revaluation. For Amazon, the more revealing debate is what normalized 2028–2029 operating cash flow and maintenance capex will look like after the present buildout.

Public investors and analysts are already underwriting that recovery. Pershing Square’s 2026 investor presentation argued that inference demand would rapidly absorb Amazon’s planned data-center expansion through 2027. In August, Morgan Stanley maintained an Overweight rating and raised its price target to $335. These are judgments, not operating evidence. The shares outperform over the next two to three years only if AWS utilization, retail leverage and free-cash-flow conversion exceed the expectations already embedded in the price.

The strongest bear case has five parts.

First, Amazon may be overbuilding. Customer commitments create visibility, but they may be delayed, under-used or margin-dilutive. If AWS growth slows while capex remains near $200 billion, the return on capital deteriorates.

Second, model economics may be deflationary. Open-weight models, distillation, caching and quantization can make inference much cheaper. Total tokens may rise while provider revenue per token falls faster.

Third, custom silicon may be operationally impressive but commercially narrow. If Trainium is economical only when Amazon and Anthropic fund the porting work, third-party demand will not provide the scale needed to create a true external moat.

Fourth, rivals have assets Amazon cannot copy easily. Microsoft owns the productivity workflow. Google owns first-party models and research. Nvidia owns the default software ecosystem. Neoclouds can offer focused capacity. The best AI company may not be the best AI infrastructure allocator.

Fifth, physical AI could become a costly automation program without a disclosed margin bridge. Robotics faces safety, labor, maintenance, inventory and regional-density constraints. Amazon’s operational complexity is itself a risk.

The public investor debate reflects the same uncertainty. A Levin Capital memo presents a strongly bullish case in which negative 2026 free cash flow is a timing issue and AWS growth accelerates.

The Henry Fund report is useful because it shows how backlog, advertising and robotics can support the cycle while acknowledging slower AI adoption as a central uncertainty. These are investor models, not consensus or company guidance.

These scenarios are from our investigation’s financial model. They are not guidance and should not be read as valuation estimates. The base case is not “AI goes perfectly.” It assumes Amazon remains heavily capital intensive in 2027 and only later converts the investment into cash.

Model note: These are illustrative operating cases only. Free cash flow is modeled as consolidated revenue multiplied by an assumed operating-cash-flow margin, less cash capex; displayed values are rounded. No tax, interest, debt funding, dilution, terminal value, discount rate, share count, net debt or equity multiple is modeled. They are not an intrinsic valuation, price target or recommendation.

Over the next two to three years, five bridges matter more than the slogans:

The decisive evidence must appear in three places at once: the income statement, the cash-flow statement and Amazon’s operating metrics. Backlog can grow without cash. A robot fleet can grow without margin. A chip run rate can grow without standalone profit. The thesis needs all three forms of proof.

Amazon does not need to build the world’s best model. It needs the world to use more AI.

Every successful model can create demand for compute, storage, databases, networking, security and monitoring. Amazon can supply those services through AWS, lower their cost through custom silicon, distribute them through Bedrock and use the same technology inside retail, advertising and logistics. Few companies can attack the opportunity from so many directions.

That breadth is Amazon’s advantage. The price of that advantage is an enormous capital commitment.

The investment case now depends on two conversions. The first is financial: capex must become revenue, margin and free cash flow. The second is operational: customer-facing AI and robotics must become measurable retail and fulfillment profit.

If both conversions appear, Amazon will not need to win the model race. It will have built the infrastructure and operating system around many of its winners. If they do not, the same breadth that makes the thesis attractive will have produced one of the most expensive capacity bets in corporate history.

The author owns shares in Amazon. This is not investment advice or a recommendation to buy or sell securities. Historical figures are drawn from Amazon filings and earnings releases. Forward scenarios are illustrative assumptions prepared for this article, not consensus estimates or company guidance. Amazon management claims and counterparty commitments are identified as such; they have not been independently audited unless explicitly described as filed financial data. Research cut-off: 23 August 2026.