In This Article
Key Takeaways
- Amazon reported second-quarter 2026 results on July 30, 2026: net sales of $200.6 billion, up 20 percent, and operating income of $27.5 billion.
- On the call, CEO Andy Jassy said the company now expects roughly $220 billion in cash capital expenditure for 2026, up from a prior estimate of about $200 billion. Second-quarter cash capex was $53.1 billion.
- The stated reason for the increase was the higher cost of memory. Not more buildings, not more accelerators. The same plan at a higher price.
- TrendForce projected in July that server DRAM contract prices would rise 13–18 percent quarter over quarter in 3Q26, with long-term agreements capping increases for some large cloud buyers and shifting the pressure onto everyone else.
- Jassy said most 2027 capacity is already reserved, with some 2028 capacity spoken for, and that Amazon is on pace to double its power capacity by the end of 2027.
Capital expenditure guidance is usually the least interesting number in a cloud earnings report. It moves when a company decides to build more. On July 30, 2026, Amazon moved its 2026 number up by about $20 billion and said the build plan had not changed. The bill had.
What Amazon reported
For the quarter ended June 30, 2026, Amazon posted net sales of $200.6 billion, a 20 percent increase from $167.7 billion a year earlier, and operating income of $27.5 billion against $19.2 billion in the same quarter of 2025. AWS segment net sales were $42.2 billion, up 37 percent, which the company put at a $169 billion annualized run rate.
Andy Jassy, president and CEO, said in the release: “AWS is booming, growing 36.7% year-over-year in Q2—our fastest growth in 18 quarters—and our AI and Chips businesses each eclipsed run rates of more than $25 billion.”
One number in the release deserves to be set aside before it confuses anything. Net income came in at $62.6 billion, or $5.75 per diluted share, and the bulk of that was a $53.4 billion non-operating pre-tax gain tied to the carrying value of Amazon's stake in Anthropic. That is a revaluation of an investment, not cash the business earned. Read the $27.5 billion operating line instead.
The extra $20 billion
Cash capital expenditure was $53.1 billion in the second quarter, which CFO Brian Olsavsky tied primarily to AWS and generative AI. Then came the guidance change. Jassy told analysts the company now believes it will spend approximately $220 billion in cash capex in 2026, and Olsavsky attributed the move from the prior estimate of about $200 billion to the higher cost of memory.
Sit with the shape of that. A $20 billion swing in a single line, announced mid-year, caused by component pricing rather than by a decision to build anything new. Amazon has held the $200 billion figure since February. Nothing about the number of buildings changed between February and July. What changed is what it costs to fill them.
Memory is a bigger share of an AI server than most people who have not priced one would guess. A general-purpose cloud instance runs on a host with a normal complement of DRAM. An accelerated training node carries a much larger pool of host memory plus stacks of high-bandwidth memory sitting on each accelerator package. Two kinds of memory, both scarce, both bid on by every hyperscaler at once.
Why memory got expensive
The supply side has been visible all year. In a July 9, 2026 note, TrendForce projected server DRAM contract prices would rise 13 to 18 percent quarter over quarter in the third quarter of 2026, and described a market that has split in two. Several U.S. cloud service providers signed multi-year long-term agreements that restrict suppliers from raising prices to them. From the third quarter on, TrendForce said, the increases land on customers without those agreements, and on any incremental supply bought outside contract by customers who have them.
The demand side showed up in supplier results the same week Amazon reported. Samsung Electronics announced second-quarter figures on July 30, 2026: consolidated revenue of KRW 171.5 trillion, an all-time quarterly high and a 28 percent sequential increase, with operating profit of KRW 89.5 trillion. Its Device Solutions division, which contains the memory business, accounted for KRW 127.5 trillion of revenue and KRW 89.2 trillion of operating profit, and Samsung said the memory business set records for both quarterly revenue and operating profit.
Put the two together and Amazon's capex revision stops looking like a surprise. It is the buyer's side of a supplier's record quarter.
Who absorbs the 3Q26 server DRAM increase
| Buyer | Position | Exposure per TrendForce |
|---|---|---|
| Large CSP with a multi-year LTA | Contract restricts supplier price increases | Capped on contracted volume |
| Same CSP buying above contract | Incremental supply sold outside the LTA | Exposed at market |
| Everyone without an LTA | Spot and standard contract pricing | Primary source of 3Q26 increases |
Capacity is reserved years out
The other thing Jassy said on the call matters as much as the capex number. Apart from 2026, he said, the lion's share of 2027 capacity is largely reserved, and a meaningful amount of 2028 capacity has already been reserved as well. He also said Amazon remains on pace with a plan discussed several quarters ago to double its power capacity by the end of 2027.
Reserved is doing real work in that sentence. It means the capacity most buyers think of as a utility, available on request, is increasingly a forward market. Large customers are committing to future quarters the way an airline commits to jet fuel. What remains on demand is the residual after those commitments are met.
Amazon's own silicon
The Jassy quote pairs an “AI” business with a “Chips” business and puts each above a $25 billion run rate. Amazon did not break either out as a reportable segment, and the release does not define the boundaries, so treat those as company-supplied framings rather than audited lines. Jassy also said on the call that demand for Trainium, Amazon's training accelerator, is very high.
A second credible source of accelerators changes the negotiating table for everyone, which is the interesting part. It does not, however, route around the problem in this story. Custom silicon needs high-bandwidth memory too, from the same three suppliers, in the same tight market.
Why it matters
The following is our analysis, not reported fact.
Your hardware quotes are going to look wrong. If a refresh, a lab build, or an on-premises inference cluster was budgeted on 2025 memory pricing, the quote you get this quarter will not match it, and the gap will be concentrated in RAM and storage rather than in the accelerator. Re-price before committing to a schedule, and separate the accelerator line from the memory line so the variance is visible to whoever approves it.
The LTA split is the practical lesson. Volume buyers under contract are insulated; the increase lands on everyone else. Most teams reading this are on the exposed side of that line, including through their vendors. Ask a hardware supplier whether a quote is held or floating, and for how long. A 30-day quote in this market is a different product from a 180-day quote.
Reserved capacity beats optimistic capacity planning. When a provider says most of 2027 is already spoken for, a 2027 project that assumes on-demand availability of a specific accelerator in a specific region is carrying a risk nobody has written down. If a workload genuinely needs a named chip in a named region, treat that as a procurement item with a lead time, not a configuration flag.
Memory pressure rewards efficiency work that was always worth doing. Quantization, smaller task-specific models, batching, and caching all reduce memory footprint, and their value rises with the price of what they save. Our guides on choosing a model tier and context economics cover the mechanics. The economic case for both got stronger this quarter without anyone changing a benchmark.
Do not read capex as a demand signal in isolation. This is the specific trap in the headline. A $20 billion increase driven by input prices tells you about the memory market, not about how much more AI is being served. Reading it as demand growth overstates the underlying trend, which is exactly the mistake that makes the next comparison quarter look worse than it is.
What the release does not tell us
Amazon did not publish a memory-cost breakdown. The attribution is a statement by executives on a call, not a disclosed line item, and there is no way from public filings to check how much of the $20 billion is memory price versus mix, timing, or anything else. Guidance is also guidance. The $200 billion figure held for roughly five months before it moved, and the $220 billion figure carries the same status.
Nor does one buyer's experience generalize cleanly. Amazon sits on the protected side of the contract line TrendForce described. Firms without that buying power face a steeper curve than the one implied by a hyperscaler raising a budget it can afford to raise.
On the currency conversions we skipped
Samsung reports in Korean won. Dollar equivalents circulating in coverage depend on the rate and date chosen, so we have quoted the won figures from Samsung's own release and left the conversion out.
Where the compute money is actually going
A plain-language walk through the AI buildout: what the capex numbers buy, who supplies it, and which parts are constrained.
Read the compute boom explainerSources: Amazon.com Announces Second Quarter Results (July 30, 2026); Amazon Q2 2026 earnings release, SEC EDGAR Exhibit 99.1; Amazon Q2 2026 earnings call transcript; CNBC on the capex revision; TrendForce, server DRAM contract prices in 3Q26 (July 9, 2026); Samsung Electronics second-quarter 2026 results (July 30, 2026). Analysis and framing by Precision AI Academy.
Common questions
Is $220 billion a final number? No. It is management's current expectation for cash capital expenditure in 2026, given on the July 30 call. The prior expectation was about $200 billion and it moved once already this year.
Does this mean cloud prices go up? Amazon did not say that, and published list prices did not change with this report. What the report establishes is that the input cost of building capacity rose. Where that lands is a separate decision.
Which memory is short, DRAM or HBM? Both markets are tight. The specific projection cited here is for server DRAM contract prices, which TrendForce put at 13 to 18 percent quarter-over-quarter growth in the third quarter of 2026.
Why was Amazon's net income so much larger than its operating income? A $53.4 billion non-operating pre-tax gain on its Anthropic investment. It is a change in carrying value, not operating performance, and it should not be read as either.