Andy Jassy stood in front of investors and told them the bill was going up. Capex for 2026 would hit $220 billion, not the $200 billion he'd guided in February and held steady in April. Then he told them the money still wouldn't be enough.
"Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026," Jassy said. "And I believe this dynamic will also be true in 2027 too. In fact, the demand we already have for 2028 is striking."
Then he named the reason the number moved: memory prices.
The stock reportedly rose more than 10% in extended trading, per CNBC.
Read that sequence again. A company raised its spending forecast by $20 billion, admitted the increase was driven by input costs it doesn't control, warned it would stay capacity-constrained for at least two more years — and the market cheered.
Now here's the figure that didn't make the after-hours celebration. Amazon's trailing-twelve-month free cash flow swung to an outflow of $7.6 billion. A year ago that same measure was a $18.2 billion inflow. That's a roughly $26 billion reversal, and it is a trailing-twelve-month figure that has now turned negative.
Last week in this space we flagged the open question of whether Amazon and Meta would push free cash flow negative under AI capex. Meta gave a light forecast and its stock fell 9%. Amazon crossed the same line — and got a reportedly 10% pop instead, though other sources put the settled reaction closer to 7-8%.
That number was AWS. Cloud revenue grew 37% to $42.23 billion, beating the $40.54 billion consensus and marking the fastest growth in 18 quarters — reportedly its fastest pace since 2021. The growth rate accelerated from 28% in Q1 and blew past the 31% analysts modeled.
And it's profitable growth, not a land grab. AWS delivered $16.62 billion in operating income against a $13.62 billion consensus, at a reportedly 36.8% margin. That single segment now accounts for nearly 61% of Amazon's entire operating profit.
The AI and chips businesses each cleared a $25 billion annualized run rate, more than doubling year over year. AWS is now a $169 billion run-rate business. On its own it would rank 24th on the Fortune 500. The engine works. That was never really the question.
The question is how you pay for the engine while it's still spooling up. Capex in the June quarter alone hit $54.21 billion — up 68% from $32.1 billion a year earlier, and above the $49.35 billion Street estimate. That's the mechanical driver of the cash-flow flip. Revenue is rising fast; spending is rising faster.
Jassy's defense of that spending is the strongest part of his case, and it deserves to be taken seriously rather than waved away.
He walked investors through the timing. Amazon starts spending on a data center roughly two years before it opens, which creates a window where cash flows out but revenue hasn't arrived yet. Once operational, he said, those facilities generate revenue for 30 years.
The servers inside them, he added, typically recoup their cost in under three years and then keep generating profit for another two or three. It's the same accounting logic Mark Zuckerberg offered a day earlier: the outflow is front-loaded, the return is long-tailed.
If you believe that framing, the negative free cash flow isn't a wound — it's the visible cost of laying track ahead of the train. The $496 billion AWS backlog of contracted-but-not-yet-delivered work is the collateral for that argument. There is demand; it's signed; it just hasn't come online.
The negative free cash flow isn't a wound — it's the visible cost of laying track ahead of the train. The $496 billion AWS backlog is the collateral.
But look closely at what's forcing the number higher, because it complicates the tidy story. The $200 billion-to-$220 billion jump wasn't Amazon choosing to build more. Jassy said memory prices were the "prime factor." DRAM prices have reportedly run up sharply industry-wide on hyperscaler demand.
That matters because it decouples the capex line from Amazon's own decision-making. In the clean version of Jassy's defense, you spend more because you've decided to build more capacity for demand you can see. In this version, part of the $20 billion increase buys you the same capacity at a higher price.
Rising input costs don't extend the 30-year data-center life or shorten the three-year server payback. They just make the front-loaded outflow bigger without any promise the back-loaded return scales to match. That's the crack in the argument.
The whole structure rests on one assumption: demand holds through 2028. Jassy leaned on exactly that — the 2028 demand is "striking," he said. The 30-year life and sub-three-year payback only pencil out if the workloads keep arriving to fill the capacity Amazon is prepaying for.
So the real test isn't whether AWS is booming. It plainly is. Rivals confirm the tide — Google Cloud grew 82%, Azure 43% — and Synovus's Dan Morgan summed up the relieved consensus: "There were concerns about market share losses on AWS, but that's been put to bed now."
The test is whether the cash gap widens faster than the backlog converts. Every quarter that memory prices push capex up while revenue recognizes on its own two-year lag, the outflow deepens. The $7.6 billion burn isn't the bottom if the input-cost pressure Jassy described persists into 2027, as he said it would.
Amazon can carry a negative free cash flow number for a long time. It is a $200-billion-a-quarter revenue company with a 61%-of-profit cloud engine and reportedly $53.4 billion in pretax income primarily from its Anthropic investment padding this quarter's $62.6 billion net income. Liquidity is not the issue.
The issue is what investors chose to price on Thursday. They priced the 37% and the reportedly $16.62 billion. They did not price the $26 billion swing in trailing cash flow, or the fact that the reason for the extra spend is a cost Amazon can't engineer away.
That's the deferred reckoning. Meta got punished for the same cash dynamics because its revenue guide gave the market nothing to look at instead. Amazon handed investors a revenue beat and a booming backlog, and they took the trade — celebrating the top line while quietly financing the slide beneath it.
The bet embedded in a 10% pop is that the 2028 demand Jassy reportedly called "striking" shows up on schedule, and that memory prices stop climbing before the outflow does. Both have to hold. Neither is fully in Amazon's hands.
For now the numbers say the engine is the best it's been in four years. They also say, in a line nobody put on the chyron, that it's running on cash it isn't yet generating.