Amazon's AWS Carries the Company. Its Biggest Bet Hasn't Proven Out Yet.

Here is the number that should stop you before anything else does: AWS is 18% of Amazon's revenue and 57% of its operating income.

Retail, the part of Amazon most people actually think about when they hear the name, is 82% of revenue and only 43% of profit. Blended retail margin is about 6%. AWS runs at 35%.

Amazon is not, in profit terms, a retailer. It is a cloud company that also happens to run the world's largest store.

That single fact is the spine of everything below: the capex bet, the valuation, the competitive gap, the regulatory overhang. If you remember nothing else, remember that AWS carries this company, and AWS is now in the middle of the biggest, least-provable bet of its history.

The number that got worse while the number that matters got better

Free cash flow fell 70.7% year over year in fiscal 2025, from $38.2 billion to $11.2 billion. Capital expenditure nearly doubled, from $83.0 billion to $131.8 billion, funding AI data centers and Amazon's own AI chips. For 2026, management has guided to roughly $200 billion in capex, a figure now large enough that trailing four-quarter capital spending exceeds trailing four-quarter operating cash flow for the first time since 2022.

Operating cash flow itself is fine, up 6.5% to $148.5 billion. This isn't a profitability problem. It's a "we are spending almost everything we make, plus some debt, on a bet that has not paid off yet" problem.

Whether this capex is value-creating or value-destroying is not resolvable from the data available today. It has to be tracked, not assumed.

That is the honest state of the evidence, and it is worth sitting with rather than rushing past. AI accelerators are widely estimated to have a real economic life of two to three years. Amazon depreciates them over five to six.

If that mismatch is real, the strong margins AWS is reporting right now are partly an accounting illusion, one that a wave of catch-up depreciation will eventually correct. Alphabet's own CFO has already flagged depreciation "meaningfully increasing" in 2026 industry-wide. Amazon faces the identical mechanical lag on its own buildout.

The counter-evidence: AWS just had its best quarter in nearly four years

Here's what complicates the bear case, and it complicates it a lot. In Q1 2026, AWS revenue grew 28% year over year, the fastest pace in 15 quarters. Segment operating margin expanded to 37.7%, actually above the FY2025 level of 35.4%, not below it, the opposite of what a margin-compression story would predict.

Total company operating margin hit a record 13.1%. AWS's annualized run rate (revenue for the quarter multiplied by four, a standard way to read one quarter's pace as a yearly number) is now $150 billion, and its AI-specific revenue alone is running above $15 billion a year.

If the AI capex were quietly destroying AWS's margin structure, this is not what that would look like. This is what it looks like when new capacity comes online and immediately finds paying demand.

The honest caveat: one strong quarter doesn't retire the depreciation-lag argument, it just delays the moment you can trust the answer. The 2026 capex wave has barely started depreciating. Q1's margin strength happened before most of the $200 billion in new spending is even generating a depreciation charge. The next four to six quarters, not this one, are the real test.

AWS is winning on size and losing on speed, and the gap is what matters

Absolute numbers still favor Amazon. AWS remains the largest cloud infrastructure provider by revenue, with roughly 28 to 30% global share, ahead of Microsoft Azure (21 to 25%) and Google Cloud (13 to 14%), depending on which research firm's estimate you use.

Growth rate tells a different story:

Provider	Q1 2026 revenue growth	Global share
AWS	28% (versus 19% trailing-year)	28-30%
Microsoft Azure	40%	21-25%
Google Cloud	63%	13-14%

Azure is compounding at roughly double AWS's rate, on a smaller base, aided by the Microsoft 365 and OpenAI bundling advantage that a pure infrastructure vendor like AWS structurally cannot replicate. Google Cloud, still the smallest of the three, is growing faster than either rival, powered by its own TPU silicon and captive Gemini model family.

Absolute share leadership sustained by a shrinking growth-rate gap is a very different situation from absolute share leadership sustained by a widening one. This is the single competitive metric worth checking every quarter, more than any headline share number.

Amazon's answer to Nvidia: build your own chips

If there is one place the bull case looks concrete rather than aspirational, it's custom silicon. Amazon's Trainium, Inferentia, and Graviton chip lines have crossed an estimated $20 billion annualized revenue run rate and are growing at triple-digit rates.

Graviton, the general-purpose chip, is used by 98% of the top 1,000 EC2 customers by spend. That's not a marketing chip. That's genuine, broad adoption. Trainium2, the AI training chip, reportedly delivers about 30% better price-performance than comparable GPU instances and sold out. Trainium3, shipping since early 2026, improves on that by another 30 to 40% and is described as nearly fully subscribed.

Why this matters: Nvidia is estimated to capture 55 to 60% of all hyperscaler AI capex dollars, the single most profitable link in the entire AI infrastructure chain. Every dollar Amazon can route through its own silicon instead of Nvidia's is a dollar of margin Amazon keeps rather than hands upstream. Custom silicon is Amazon's only real structural lever against the capex supercycle eroding its own margins from the inside.

The Anthropic relationship: real upside, real circularity

AWS's backlog (contracted future revenue that hasn't yet been recognized as sales) reached $364 billion in Q1 2026. Of that, $225 billion is Trainium commitments and more than $100 billion is tied to Anthropic, the AI lab Amazon has invested up to $33 billion in total ($8 billion already, up to $25 billion more committed).

Anthropic, in turn, has committed to spend over $100 billion on AWS over ten years.

Read that again slowly. Amazon's equity investment helps fund Anthropic's ability to buy AWS compute. That AWS compute purchase inflates AWS's backlog figure.

And Amazon's equity stake in Anthropic gets marked up as Anthropic's business grows, partly on the strength of the very AWS relationship Amazon is subsidizing. It's a loop, not a straight line, structurally similar (if smaller) to the circular financing arrangements market observers have already flagged between Microsoft and OpenAI, and Nvidia and OpenAI.

None of this makes the relationship fake. Anthropic is a real, fast-growing company and the commercial logic (lock in a frontier-model tenant, use its committed spend to de-risk your own capex) is coherent.

But a meaningful share of AWS's forward growth narrative now depends on the health of one customer that Amazon itself is capitalizing. If Anthropic's business falters, both the equity mark and a chunk of the AWS backlog take the hit simultaneously. This is a real, currently unquantified risk, not a resolved one.

Valuation: priced for the good outcome, not cheap enough to survive the bad one

Amazon trades at 34.4x core, trailing earnings (using FY2025's organic EPS of $7.17, stripped of a one-time $16.8 billion Anthropic mark-to-market gain that inflates the unadjusted GAAP trailing multiple to a misleadingly cheap-looking 29.5x). Price to book is 6.5x.

Company	Trailing P/E	Segment context
Amazon (core)	34.4x	AWS reaccelerating, capex supercycle unresolved
Microsoft	22.8x	Azure growing 40%
Alphabet	26.9x	Google Cloud growing 63%
Walmart / Costco / MercadoLibre	39-48x	Retail and e-commerce comps

Amazon trades at a premium to its two closest cloud peers and a discount to retail peers, which is roughly what you'd expect from a business that's a blend of both. It doesn't scream cheap or expensive against this set.

What matters more is a three-year outlook run through three scenarios, modeled out with exact figures rather than rough guesses: bear case, capex fails to earn its keep and the market re-rates Amazon toward a retail-like multiple, implies -26.7% (to $180.60). Base case, AWS growth and margin roughly hold, implies +43.3% (to $353.40). Bull case, AWS reaccelerates toward its 2024 margin peak, implies +98.8% (to $490.10).

At today's price, the market is already paying close to what the base case would justify. That leaves meaningful upside if the bull case shows up, and real, uncushioned downside if the bear case does instead. This is not a stock priced for disappointment.

Three fronts of risk that don't touch the core thesis yet, but could

Regulatory. The FTC's federal antitrust case, alleging Amazon used anti-discounting and marketplace practices to maintain monopoly power, goes to a bench trial on February 9, 2027. No resolution is likely before late 2027 at the earliest, given appeals; this is a multi-year overhang, not a near-term catalyst.

More immediate: in June 2026 the European Commission issued a preliminary designation that would classify AWS itself, not just Amazon's marketplace, as a gatekeeper under the EU's Digital Markets Act, meaning AWS would face specific, enforceable rules on interoperability and data access, with fines of up to 10% of global revenue for non-compliance. This is the first time cloud infrastructure specifically, not just a marketplace, has faced this kind of obligation, and a final decision is expected around October 2026. Amazon has operated under similar marketplace-gatekeeper rules in the EU for two years without material earnings impact, which is some evidence the company can absorb compliance-style regulation better than structural remedies, but cloud gatekeeper status has no real precedent to calibrate against.

Labor. In April 2026 the NLRB ordered Amazon to bargain with the Amazon Labor Union-Teamsters at its Staten Island JFK8 facility, the first legally enforceable bargaining order in the company's history. In dollar terms, this affects roughly 5,500 workers today. The exposure that matters is precedent: Amazon employs more than one million US warehouse workers, and a first successful order tends to invite more of them.

Management, outside the circle of competence. Andy Jassy is a genuinely strong operator inside AWS and logistics, the business he built. Outside it, the record is weaker: One Medical has been a persistent drag since Amazon acquired it (Amazon has not disclosed its segment results since, itself a transparency gap), and Project Kuiper, Amazon's satellite broadband answer to Starlink, had 394 satellites in orbit against the 578 needed for its first phase, with an FCC deadline of July 30, 2026 that appears to have been missed or requiring an extension.

Jassy has been willing to restructure or shrink ventures outside AWS rather than defend them indefinitely, which reads as discipline rather than empire-building, but it's still a mixed track record worth pricing in. One more small but telling data point: the $10 billion buyback authorization from 2022 has about $6.1 billion left unused and has barely moved in years, while the share count keeps rising from stock-based compensation, meaning the buyback is offsetting dilution rather than actually returning capital to shareholders.

The one piece of good news that has nothing to do with AI

Amazon also picked up a genuine tailwind that has nothing to do with capex or cloud: the closure of the de minimis import loophole. The US eliminated duty-free treatment for low-value shipments in 2025, and the EU follows with a per-parcel levy from July 1, 2026, raising landed costs for Temu, Shein, and TikTok Shop by an estimated 20 to 40%, platforms whose entire price edge depended on that exemption. Amazon's own low-price answer, Amazon Haul, runs on domestic, already-tariffed fulfillment and is largely unaffected. For once, a regulatory shift cuts in Amazon's favor rather than against it.

Where this leaves an investor

Amazon is a strong business at a fair price, not a strong business at a bargain price. That distinction matters more here than in most stocks you'll read about this year, because the price already assumes the AI capex bet works out roughly as planned.

The evidence right now is genuinely split. AWS margin and growth both just hit multi-year highs. Custom silicon is real and scaling. The de minimis reversal disarms Amazon's cheapest competitors. Against that: free cash flow has been crushed, the depreciation wave from $200 billion in 2026 spending hasn't fully landed, AWS's growth rate still trails both rivals, and a meaningful share of AWS's forward story runs through one circularly-financed customer.

This is not a moment for high conviction in either direction. If you already own it, hold and watch two numbers: AWS operating margin needs to hold above 36 to 37% for two consecutive quarters, evidence the AI capex is actually being absorbed profitably rather than merely deferred onto a future income statement. And watch for the AWS backlog-to-revenue conversion, whether those bookings actually show up as recognized revenue, to become independently verifiable rather than resting on Amazon's own disclosure of a number partly inflated by its own Anthropic investment.

If you're putting new capital in, size it for the -26.7% bear case, not the +43.3% base case. The market has already paid for the good outcome. It has not yet had to price the bad one.

Sources

This article is LLM-assisted, disclosed per site policy, and does not constitute financial advice or a recommendation to buy or sell any security.

Key facts