Amazon is buying time, power, and chips before it buys revenue

TL;DR

• Amazon reported Q2 2026 net sales of $200.6 billion — its first $200 billion quarter — and raised full-year capex guidance from $200 billion to $220 billion.
• AWS revenue grew 37% year-on-year to $42.2 billion, the fastest pace in 18 quarters. AWS operating income surged 64% to $16.6 billion, with margins expanding to 39.4%.
• Trailing-12-month free cash flow swung to a $7.6 billion outflow as property and equipment purchases rose $66.1 billion. Long-term debt reached $128.9 billion.
• CEO Andy Jassy warned that capacity will still fall short of AI demand through 2027, with "striking" demand already visible for 2028.
• The four largest hyperscalers — Amazon, Microsoft, Google, and Meta — collectively plan to spend roughly $725 billion on capex in 2026, up 77% from $410 billion in 2025.

The quarter, in numbers that matter

Amazon's Q2 2026 earnings, reported 31 July, delivered a set of figures that are individually impressive and collectively signal an industrial transformation. The headline numbers:

That net income figure — $62.6 billion, or $5.75 per share — needs an asterisk. Roughly $53.4 billion of pre-tax non-operating income was primarily linked to Amazon's investment in Anthropic. Strip that out, and the cleaner operating signal is AWS: $42.2 billion in sales growing at 37%, with margins expanding even as the company pours capital into infrastructure.

Sources: Amazon Q2 2026 earnings release; Wall Street Journal; Reuters; CNBC; App Economy Insights.

What Amazon actually bought with the extra $20 billion

The capex increase is not simply "more GPUs." CEO Andy Jassy attributed the $20 billion revision primarily to higher memory chip costs — a detail that corrects the familiar "Big Tech buys chips" story. The constraint has become a system: accelerators, high-bandwidth memory (HBM), networking fabric, electricity, land, cooling, construction labour, and the specialised staff who can commission a hyperscale data centre.

Amazon's spending spans AI infrastructure, robotics, semiconductors, and satellites (Project Kuiper). The company's custom chip business — Graviton, Trainium, and Nitro — passed a $20 billion annual revenue run rate last quarter. Jeff Bezos described it this week as becoming a fourth pillar of the company, alongside Marketplace, Prime, and AWS.

Jassy was explicit on the earnings call: even at $220 billion, Amazon will not have enough capacity to meet AI demand in 2026 or 2027. Demand for 2028 is, in his word, "striking." He added that AWS could "very possibly be a trillion-dollar annual revenue business for us in time, with very appealing accompanying free cash flow and return on invested capital."

Those are company claims. But AWS's 37% growth — and a contracted backlog that reached $496 billion — offers more evidence of paid demand than a model demo or a press release ever could.

The hyperscaler context: $725 billion and rising

Amazon is not spending alone. Across the industry, the four largest hyperscalers collectively plan to spend roughly $725 billion on capital expenditures in 2026, up 77% from $410 billion in 2025, according to Tom's Hardware and earnings reports:

• Amazon: $220 billion (raised from $200B)
• Microsoft: Azure grew 43% in its most recent quarter; capex is tracking above $80 billion
• Google: Google Cloud delivered 82% growth in Q2 2026; capex is tracking above $90 billion
• Meta: Significant AI infrastructure spend, though with a different revenue model (advertising, not cloud)

This is not a normal investment cycle. It is a land grab for compute capacity in a market where demand is outstripping supply and the lead time for a new data centre — from site selection through power agreements, construction, and commissioning — can run three to five years.

The financial trade-off: cash flow vs. capacity

The defining finance trend of the AI cycle is visible in Amazon's numbers: profitable companies are converting operating cash flow into long-lived infrastructure before revenue arrives.

Amazon's operating cash flow rose 33% to $161.4 billion over the trailing 12 months. Free cash flow nonetheless moved from an $18.2 billion inflow a year earlier to a $7.6 billion outflow. The driver: property and equipment purchases rose $66.1 billion year-over-year.

Long-term debt on the balance sheet reached $128.9 billion as of 30 June 2026, up from $65.6 billion at year-end 2025 — reflecting the bond market activity Amazon has used to fund its infrastructure commitment.

This is not distress. It is deliberate capital allocation. Amazon is borrowing at investment-grade rates to build assets it expects to generate returns over decades. The question is whether those returns materialise at the scale and pace the spending implies.

The AWS margin story: Trainium is showing results

An easily missed detail in the Q2 release is the movement in AWS operating margins. AWS generated $16.6 billion in operating income, a 64% increase from $10.2 billion a year earlier. The operating margin reached 39.4% — up from 32.9% in Q2 2025, a 650-basis-point expansion.

Part of that expansion reflects the growing contribution of Amazon's custom silicon. Trainium chips, designed for AI training workloads, reduce Amazon's dependence on third-party GPU suppliers and improve the unit economics of AI infrastructure. As Trainium deployment scales, AWS can offer competitive AI compute pricing while maintaining or expanding margins — a structural advantage that pure GPU resellers cannot replicate.

The demand signal: who is committing

Amazon is making these investments based on contracted demand from some of the largest AI companies. OpenAI and Anthropic have signed commitments to AWS worth $138 billion and more than $100 billion, respectively, for the coming years, according to GeekWire.

"We're not investing approximately $200 billion in capex in 2026 on a hunch," Jassy wrote in his April shareholder letter.

The $496 billion AWS backlog — representing contracted future revenue not yet recognised — is the strongest available evidence that the demand is real, not speculative. But backlog is not revenue. It is a commitment to spend, and commitments can be renegotiated if the AI market shifts.

Hype check

Three claims should be resisted:

1. "The capex increase proves AI is a guaranteed winner." It proves Amazon sees enough contracted and prospective demand to prefer the risk of overbuilding to the risk of refusing customers. Those are not the same claim.
2. "Free cash flow will recover quickly." It may, but the timing depends on when the infrastructure currently being built begins generating revenue — and on whether AI workload demand continues to grow at current rates.
3. "Amazon's AI business is already profitable at scale." Amazon does not disclose AI-specific revenue, costs, or margins. The AWS margin expansion is encouraging but does not isolate AI infrastructure returns.

Stakeholder landscape

Cloud customers may benefit from more capacity and competitive pricing in the medium term but will remain exposed to scarcity, power constraints, and changing price-performance dynamics in the near term. If you are negotiating an AWS commitment, do it now — capacity scarcity gives the seller leverage.

Memory, networking, power-generation, and data-centre suppliers gain strategic importance. The HBM market, dominated by SK Hynix and Samsung, is a particular bottleneck. Any disruption to memory supply chains will flow directly into hyperscaler capex and delivery timelines.

Investors need to separate operating earnings from investment gains (the Anthropic contribution to net income is non-recurring) and judge returns on deployed infrastructure over years, not quarters. The metrics that matter: AWS revenue growth, capex-to-revenue ratio, free-cash-flow conversion, backlog quality, and the estimated useful life of AI equipment.

Competitors — particularly Microsoft and Google — are making the same bet with different balance sheets. The hyperscaler race is now a contest of who can deploy capital most efficiently, not who has the best model.

Employees should note that Amazon cut AGI jobs even as it raised infrastructure spending to $220 billion. The AI investment is in silicon and concrete, not headcount.

Recommendations

For technology leaders negotiating cloud contracts: Treat capacity commitments, data residency, inference unit economics, and exit options as board-level commercial terms. The scarcity is real — lock in capacity now if you need it in 2027.

For investors: Track AWS growth rate, capex-to-revenue ratio, free-cash-flow conversion, backlog quality, and the useful life of AI equipment. Do not rely on headline net income, which includes non-recurring investment gains.

For AI companies building on cloud infrastructure: Diversify across providers if possible. A single-provider commitment at current scarcity levels creates concentration risk.

For policymakers: The concentration of AI infrastructure in three US companies — Amazon, Microsoft, and Google — raises questions about supply-chain resilience, competition, and national infrastructure dependency that go beyond antitrust.

Uncertainty ledger

• Amazon does not disclose AI-specific capex, capacity utilisation, or return by product line. The AI infrastructure story is inferred from aggregate numbers and management commentary.
• The demand outlook is management guidance, not an independent forecast. Jassy's "striking" 2028 demand is a claim, not a verified pipeline.
• The useful life of AI infrastructure is uncertain. If next-generation chips make current infrastructure obsolete faster than depreciation schedules assume, returns will be lower than projected.
• The Anthropic investment gain that inflated Q2 net income is non-recurring. Future quarters will not benefit from the same accounting treatment.
• The memory chip price increase that drove part of the capex revision could reverse, but the capacity expansion it funds is permanent.

Bottom Line

Amazon's $220 billion plan says the AI market has moved from experimentation to industrial provisioning. The winners will not be defined only by the best models. They will be defined by who can deliver reliable compute, power, and network capacity at a price customers can keep paying — and who can do it while competitors are still waiting for their data centres to come online.

Sources

• Tier 1: Amazon Q2 2026 Earnings Release — primary filing
• Tier 1: Wall Street Journal — "Amazon Shares Jump as Cloud Sales — and Spending — Accelerate" (30 July 2026)
• Tier 1: Reuters — "Amazon crushes Q2 earnings with $200.6B revenue, AWS growth accelerates" (31 July 2026)
• Tier 1: CNBC — "Amazon beats Q2 earnings on 37% AWS growth, raises capex to $220B" (31 July 2026)
• Tier 1: Investopedia — "Amazon Plans to Spend $20 Billion More on AI. Wall Street Loves It" (30 July 2026)
• Tier 2: App Economy Insights — "Amazon: The CapEx Equation" (31 July 2026)
• Tier 2: GeekWire — "Amazon earnings preview: Wall Street looks for more cloud growth as AI spending hits a record" (29 July 2026)
• Tier 2: Tom's Hardware — Big Tech capex spending for 2026 across hyperscalers