Newsletter · · Ashutosh Agarwal

AWS Sells Out Through 2028 as Hyperscaler Capex Heads Toward 1.5 Trillion - The AI Capex Tracker - Week of August 4, 2026

A synthesis of what investor podcasts said about AI capital spending for the week of August 4, 2026, including AWS CEO Matt Garman confirming Amazon's data-center capacity is pre-sold through 2028, a re-drawn capex scoreboard heading toward 1.5 trillion dollars in 2027, and the bear case migrating from stock valuation to a credit-cycle argument.

The AI Capex Tracker

Week of August 4, 2026: AWS Sells Out Through 2028 as Hyperscaler Capex Heads Toward 1.5 Trillion

TL;DR

  • The head of Amazon's cloud business went on the record and gave the bears their clearest answer yet. AWS CEO Matt Garman said Amazon's 2026 capital spending (capex, the money spent building data centers and buying chips) is now $220 billion, up $20 billion, will be "bigger next year," and that much of the capacity is "already spoken [for] through the end of 27 and even through much of 28" on five-year customer commitments. In plain terms: they are pre-selling the buildings before they finish them. "Today demand still significantly outstrips supply." (Bloomberg Talks, Aug 3)
  • With every hyperscaler now reported, the running total got re-drawn, and it went up. The big spenders (Amazon, Microsoft, Google, Meta, plus Oracle and Tesla) have spent ~$580 billion over the trailing twelve months, are on track for ~$900 billion in 2026, and look headed toward ~$1.5 trillion in 2027, a number that is starting to bend free cash flow negative at more than one of them. (Chip Stock Investor, Aug 3)
  • The stocks snapped back hard, and the bears quietly moved the argument to credit. Amazon touched a $3 trillion market cap for the first time and the chip complex bounced, as the July crash is increasingly read as one over-levered fund being liquidated rather than the thesis breaking (Market Mondays, Aug 4). Meanwhile a viral essay, argued on two separate podcasts this cycle, reframes the whole boom as a debt-fueled real-estate cycle, not a stock-market one, with Nvidia's reported $250 billion financing backstop for an OpenAI data center as exhibit A. (What Bitcoin Did, Aug 3; Market Maker, Aug 3)

What's new

A note on the window: this is a Tuesday issue, so it sweeps roughly the last 24 hours, podcasts published August 3 and 4. The prior issue (Monday, Aug 3) explained why the tape whipsawed, the forced liquidation of a ~$20 billion, ~4x-levered fund, and flagged that the real stress had moved into the credit market. Today's podcasts do two things: they put an actual operator (the AWS CEO) in front of a microphone to answer the demand question directly, and they re-anchor the capex numbers now that the last hyperscaler has reported. Ranked by where the dollars of profit-and-loss impact are biggest.

1. The AWS CEO answered the "who's actually buying this?" question in his own words. Bloomberg Talks, Aug 3, Matt Garman, CEO of Amazon Web Services. This is a genuine operator, not a pundit, the person who actually signs the contracts.

Garman confirmed Amazon's 2026 capex is $220 billion, up $20 billion, "mostly AI" but also inflated by "higher memory pricing." He expects to "keep investing in CapEx next year as well." The line the bulls will quote for weeks:

"As Andy mentioned, much of our capacity is already spoken [for] through the end of 27 and even through much of 28. And so as we're investing, we're getting five-year commitments from customers… today demand still significantly outstrips supply." Matt Garman, AWS

Other useful color: AWS's AI business is at a $25 billion annual revenue run-rate, and the mix "keeps shifting more and more towards inference," meaning away from one-time model training and toward the day-to-day running of AI (inference), which is the recurring, sticky revenue. Amazon's own Trainium chips are "largely sold out through the end of next year," and can cut a customer's inference bill "20%, 30%" versus renting Nvidia GPUs. He even left the door open to selling Amazon's chips outright someday ("something that we would definitely consider in the future"). Why it moves numbers: the single loudest bear question is "where is the demand to justify $900 billion of spend?" The person running the biggest cloud just said the capacity is pre-sold two-plus years out on multi-year contracts. That is about as direct a rebuttal as the operator side can give.

2. The capex scoreboard got re-drawn now that everyone has reported, and it stepped up. Chip Stock Investor, Aug 3, Nicholas and Kasey Rossolillo.

With all the hyperscaler reports in hand, they tallied ~$580 billion of capex over the trailing twelve months across Amazon, Microsoft, Google, Meta, Oracle and Tesla, rising to "let's just call it 900 billion, perhaps a little bit more" for full-year 2026, and, on current trajectory, "possibly headed towards 1.5 trillion in 2027." Set that against the revenue it is supposed to feed: trailing cloud revenue across the majors plus Oracle is "nearly 400 billion," with AWS at ~$148 billion (approaching a $200 billion annual pace), Microsoft's cloud at ~$138 billion, Google Cloud at ~$78 billion (still growing 80%-plus year over year), and Oracle at $34 billion. Their honest bottom line: this is "not a sustainable rate," it is already pushing "a couple of these companies… into quarterly negative free cash flow" (Amazon has now joined Oracle), and nobody yet knows whether revenue "organically catches up to the infrastructure spending." Why it matters: last issue framed spend as "$800 billion over the next 12 months trending toward a trillion." The frame this week is firmer and higher, ~$900 billion this year, ~$1.5 trillion next.

3. Amazon's rebound to $3 trillion made the "chips are back" case. Market Mondays, Aug 4, a retail-investing show hosted by equity-research voices (pundit commentary, not operators).

Their read of the earnings run: capex "was still at the forefront and it hadn't slowed down," and, the key shift, Microsoft "for the first time all year" showed "revenue come back from this CapEx spend," with Azure cloud growth accelerating. Then Amazon closed it out: ~$200 billion of revenue in the quarter, AWS growing to ~$42 billion (versus a ~$40 billion estimate), and the stock hitting a $3 trillion market cap for the first time. Their framing of the July crash, the episode is literally titled "a $35 billion lesson: no one is bigger than the market," is that it was a leverage story, not a demand story, and the setup into the back half is "all signs point to yes, this story is going to continue." Treat the specific price targets they throw around as enthusiasm, not analysis, but the through-line (capex up, cloud revenue now visibly following it, chips bouncing) is corroborated across the operator podcasts above.

4. The bear case consolidated around one viral essay: this is a credit cycle, not a stock cycle. What Bitcoin Did, Aug 3, Luke Gromen, macro strategist; Market Maker, Aug 3, UK markets show.

Two independent podcasts this cycle leaned on the same anonymous Substack essay ("Groundbreaker," titled "The Second Derivative: Why No One Understands the AI Boom"). Its argument, in plain English: stop analyzing AI like the dot-com stock bubble and start analyzing it like the 2007-08 mortgage-credit bubble. Data centers are being built with mountains of debt, and, like housing, the trouble starts not when demand falls but when its rate of growth merely slows, because that is when the borrowing can no longer be refinanced. Gromen's escalation of the stakes: "80%, 90%… of GDP growth in this country over the last 12 to 18, 24 months has been driven by AI-related spend," so if it unwinds "now you're in a recession," the federal deficit could balloon by "$1.8 to $3 trillion" on top of the existing ~$2 trillion, and, he argues, that is precisely why "the Fed can't let the AI bubble burst." He also flagged that Google "went cash flow negative, first time in its history." The Market Maker hosts put the mechanism bluntly:

"This is not a Ponzi scheme. But… you continually need more money coming in in order to keep the wheels of this circular flow of financing going. And if the money stops going in, then the whole thing falls apart."

Their exhibit A for that circularity: the reported news that "Nvidia is in talks with OpenAI to guarantee $250 billion in financing for a data center in Ohio," Nvidia backstopping its own customer's purchases with its own credit rating, on which "NVIDIA share price went down 4.5%." Why it matters: the bear thesis has fully migrated from "the stocks are too expensive" to "the financing is circular and the credit market is where it breaks." That is a more serious, and more falsifiable, claim.

5. AMD wrote another compute check, and reports this week. Last Week in AI, Aug 3.

AMD "has committed up to $5 billion to Anthropic," with Anthropic set to "deploy up to two gigawatts of AMD's Instinct MI450 AI GPUs and their new Helios rack-scale system… in 2027." Anthropic now has compute deals with "SpaceX AI… Google… Amazon, and now… AMD," a customer so hungry for chips it is partnering with everyone at once. The hosts' framing of why this matters for the whole complex: AMD is one of the only players that "potentially [has] the ability to compete with Nvidia, which no one else really does" outside Google's TPUs. Why it moves numbers: it is a second concrete anchor tenant behind AMD's data-center roadmap, and it lands right as AMD reports earnings this week.

The debate

Steel-manning both sides of the ~$900-billion-in-2026 (toward ~$1.5 trillion in 2027) hyperscaler capex thesis. This cycle the two sides stopped arguing about the same thing: the bulls are now arguing about demand and contracts, and the bears are arguing about debt and financing. They are almost talking past each other, which is itself the tell.

Bull, the demand is contracted, not speculative, and the leaders are converting it. The strongest bull point came from an operator, not a pundit: AWS is pre-sold "through the end of 27 and even through much of 28" on "five-year commitments," and "demand still significantly outstrips supply" (Bloomberg Talks, Aug 3). On The Six Five, Aug 3, Daniel Newman put it as "the market is not funding CapEx, it's funding contracted demand," citing Microsoft's $678 billion backlog (booked future revenue), Amazon's $496 billion backlog, 82% Google Cloud growth, and Andy Jassy's line that "once revenue growth outpaces incremental CapEx growth, the returns become compelling," with AWS growing 37% against a 10% capex raise and Azure growing 43% at ~$41 billion a quarter. On The Circuit, Aug 3, two semi analysts called Amazon "the most rational" spender and noted Jassy confirmed the payback on a data center is now "under three years." And crucially, the bulls argue the July crash was mechanical: Google and Amazon "did the exact same things on opposite ends of that event and got completely different results," which only makes sense if a forced fund liquidation, not fundamentals, drove the tape (The Six Five, Aug 3).

Bear, the returns aren't visible yet, the financing is circular, and this is a credit cycle. Patrick Moorhead's counter on The Six Five was precision itself: the debate isn't revenue, it's conversion, "Meta accelerated revenue 28% and got shot anyway by raising its CapEx," and "even the credit market is repricing the trade in real time," pointing to BlackRock's Meta-backed data-center bond priced at "more than $12 billion at 7.5%, 275 basis points over treasury" (The Six Five, Aug 3). Chip Stock Investor, bulls at heart, conceded the capex-to-revenue ratio is "not sustainable" and that free cash flow has turned negative at some names (Chip Stock Investor, Aug 3). And the "Second Derivative" credit framing, cited independently by Luke Gromen and the Market Maker hosts, reframes the whole thing as a debt-fueled real-estate cycle vulnerable to a mere slowdown in growth, with Nvidia's reported $250 billion OpenAI financing backstop (NVDA -4.5% on the news) as the circular-financing smoking gun (What Bitcoin Did, Aug 3; Market Maker, Aug 3). Hanging over both: cheaper Chinese open-weight models (Kimi K3, competitive at a fraction of the cost) could shrink the amount of infrastructure the world actually needs, leaving "a bunch of white elephants" (Market Maker, Aug 3).

The synthesis. The bull got the better of it this cycle, because the bull produced an operator (Garman) with contracts and the bear produced an essay and a widening credit spread. But note what the bear case has become: it no longer needs demand to collapse. It only needs the rate of growth to slow enough that the next big data-center bond can't be refinanced cheaply. That is a lower bar, and it is why the funding market, not the earnings, is now the thing to watch.

Sell signals to watch:

  • An ugly or failed bond deal. The BlackRock/Meta data-center bond at 7.5% (275bps over Treasuries) is the early tremor. The first hyperscaler or neocloud that can't place paper at a sane price is the crack.
  • Widening credit-default-swap spreads on the spenders. Gromen flagged Oracle's and Microsoft's insurance-against-default costs "starting to rise." Watch whether that continues.
  • The second derivative of capex growth slowing. Per the "Second Derivative" thesis, the danger point is when spending growth merely decelerates, not when it falls.
  • A monetization no-show. Meta raised capex and gave no clear plan to earn it back, and got punished. The next name that spends without a revenue story is vulnerable.
  • A fresh leg of forced selling. The "$35 billion lesson" fund unwind may not be fully finished.

Stocks in play

NVDA. Bull: still the compute bottleneck, AWS calls itself "one of NVIDIA's absolute largest customers in the world," and Nvidia has effectively "won" AI training "for the foreseeable future" (The Circuit, Aug 3). Bear: now the face of the circular-financing worry, its reported $250 billion backstop for an OpenAI data center knocked the stock down 4.5%, and as the market shifts from training to inference, its near-100% share "has to go down," the only question is how far (Market Maker, Aug 3; The Circuit, Aug 3). Next: its late-August quarter, the demand read for the entire complex.

AVGO. Bull: the "own-the-silicon" custom-chip theme keeps riding hyperscaler in-housing; named in passing as a partner in the ASIC race. Bear: no standalone Broadcom signal this cycle, so it trades on the complex's beta right after a violent forced-selling episode. Next: read-through from Amazon's and Google's custom-chip ramps. Otherwise quiet in the podcasts this cycle.

AMD. Bull: wrote another anchor deal, up to $5 billion into Anthropic, which will deploy up to 2 gigawatts of MI450 GPUs and the new Helios rack-scale system in 2027; described as the one credible non-Nvidia, non-Google challenger (Last Week in AI, Aug 3). Bear: GPU market share is still small and contested (roughly 10-14%), and shipment/ramp timing on MI450 and Helios remains the show-me item (Market Mondays, Aug 4). Next: its Q2 earnings this week, watch data-center GPU revenue and any new hyperscaler commitment.

MSFT. Bull: the quarter turned the story, CFO Amy Hood pledged Microsoft would be "free cash flow positive in 2027," Azure growth accelerated to ~43% at ~$41 billion a quarter, and Copilot finally showed traction (30 million paid seats, net seats doubling quarter over quarter, weekly engagement "on par with Outlook and Teams") (The Six Five, Aug 3). Bear: it went into the print down ~30% on the year, and the credit market flagged Microsoft's default-insurance cost "starting to rise" (The Six Five, Aug 3; What Bitcoin Did, Aug 3). Next: whether the "capex discipline plus accelerating cloud" formula holds, and the promised 2027 free-cash-flow inflection.

GOOGL. Bull: the fastest-growing big cloud, up ~82% year over year, and a multi-stream business that can fund the build internally (The Six Five, Aug 3). Bear: it posted its first-ever negative free cash flow, raised its capex outlook to ~$195-205 billion, and the stock fell despite a revenue beat, the market punishing spend it can't yet see a return on (Market Maker, Aug 3; The Six Five, Aug 3). Next: the 2027 capex number, and whether core profit can grow into it.

AMZN. Bull: the standout of the cycle, AWS re-accelerated to its biggest growth in 18 months (~$42 billion in the quarter), grew margins despite higher memory costs (helped by its own Trainium and Graviton chips), and management says AWS alone "can double from here" toward $1 trillion in annualized revenue; the stock touched a $3 trillion market cap (Market Mondays, Aug 4; Chip Stock Investor, Aug 3). Bear: it is the single biggest spender at $220 billion of capex (up $20 billion), it dipped into negative quarterly free cash flow, and a chunk of the raise is just higher memory prices, not more compute (Bloomberg Talks, Aug 3; Chip Stock Investor, Aug 3). Next: whether AWS growth keeps outpacing the capex raise as the 2027 spend climbs.

META. Bull: still a ~3-billion-user advertising machine with the balance sheet to keep spending. Bear: the flunk of the cycle, described as "kind of a train wreck," with a "horrible" guide, questions about whether it overbuilt capacity (and is now trying to lease the excess out), no clear plan to monetize the compute, and a first-party-silicon effort that has "lost their path," implying higher cost per token than rivals (The Six Five, Aug 3; The Circuit, Aug 3). Next: an actual monetization plan for the compute it is building, the market wants specifics.

Read-throughs

  • Custom silicon vs. Nvidia, the inference split is the trade. The clearest structural point this cycle: training is Nvidia's ("won… for the foreseeable future," near-100% share), but inference, the day-to-day running of models, is "much more heterogeneous," and there the hyperscaler, not the customer, decides which chip runs each workload "for the best margins" (The Circuit, Aug 3). Amazon says Trainium can save "20%, 30%" on inference and is "sold out through the end of next year" (Bloomberg Talks, Aug 3). Most actionable trade: as the workload mix tips to inference, custom-ASIC names (AVGO, plus the in-house Trainium/TPU stories) gain share at the margin, the question for NVDA is not if its inference share falls but how far (70% is one outcome, 40% is a very different one).
  • Memory, still the tightest link, but light on fresh operator color this cycle. Amazon explicitly blamed part of its capex raise on "higher memory pricing" (Bloomberg Talks, Aug 3), and one futurist-investor flagged the AI memory shortage running "through end of 2026" plus a reported ~$250 billion, decade-long manufacturing investment (Strategy Sunday, Aug 3). The overhang the bears keep raising: the Chinese memory maker CXMT's blockbuster IPO signals more supply coming, which would eventually pressure prices. Most actionable trade: memory (Micron, SK Hynix, SanDisk, Western Digital) remains the highest-beta way to play the whole trade in either direction, and several of those names report this week, the next real print on pricing.
  • Power / grid, genuinely quiet in the podcasts this cycle. No new data-center-power, ERCOT/Texas, gas-turbine or nuclear-PPA commentary surfaced in this narrow two-day window (the one energy podcast that came up was about grid batteries, not AI load). That is a coverage gap, not a signal that the constraint eased, the supply-chain analysts still describe capacity as simply "not there." VRT, ETN, Vistra, Constellation and Talen were quiet in the podcasts; the demand backdrop they sell into is unchanged.

What changed vs last issue

Last issue (Monday, Aug 3, "A $20B fund blew up. Chips cracked. Microsoft soared.") swept the weekend and explained the July whipsaw as a forced fund liquidation, with the real stress migrating into credit (Snider) and Chanos's incremental-return collapse (100% to 25%) as the marquee bear number. Today's ~24-hour sweep adds the operator rebuttal and re-anchors the numbers.

  • An operator finally answered the demand question directly. Last issue's bull case leaned on secondhand reads of the earnings calls. This issue has the AWS CEO on the record: capacity pre-sold "through the end of 27 and even through much of 28" on "five-year commitments," and "demand still significantly outstrips supply." That is the cleanest operator-side counter to the credit bears we've had.
  • The capex frame stepped up. Last issue: "~$800 billion over the next 12 months, trending toward a trillion." This issue, with everyone reported: ~$580 billion trailing twelve months, ~$900 billion for 2026, and ~$1.5 trillion penciled for 2027.
  • The rebound got confirmed and named. Last issue reported the snap-back and Microsoft's record day. This issue: Amazon touched a $3 trillion market cap for the first time, the chip complex bounced, and the July drop is now widely read as one ~$35 billion levered-fund unwind rather than the thesis breaking.
  • The bear case fully migrated from valuation to credit, and found a shared text. Last issue had Snider ("all about credit") and blowing-out CDS spreads. This issue: two independent podcasts built their bear case on the same "Second Derivative" essay (AI as a 2008-style debt/real-estate cycle), and Gromen escalated the macro stakes ("80-90% of GDP growth… driven by AI-related spend," a potential $1.8-3 trillion deficit hit if it unwinds, "the Fed can't let the AI bubble burst").
  • Fresh, concrete data points: AWS AI run-rate $25 billion (shifting to inference); Amazon capex $220 billion (+$20B); Microsoft backlog $678 billion, Amazon backlog $496 billion; AWS +37% vs a 10% capex raise, Azure +43% at ~$41B/quarter; Amy Hood's "free cash flow positive in 2027" pledge; Copilot 30 million paid seats; Google's first-ever negative free cash flow and ~$195-205B capex outlook; Nvidia's reported $250 billion OpenAI Ohio financing backstop (NVDA -4.5%); AMD's up-to-$5 billion Anthropic deal (2 GW of MI450 + Helios in 2027); Anthropic run-rate cited at ~$74 billion, OpenAI $30 billion-plus; BlackRock/Meta data-center bond at 7.5% (275bps over Treasuries).
  • Still open: whether Nvidia's late-August quarter confirms demand for the whole complex; whether the next big AI bond clears cleanly or widens; whether Meta ever produces a monetization plan; whether cheaper Chinese open-weight models (Kimi K3) start shrinking the required build; and whether the forced-selling episode is truly finished.

Next catalysts to watch: AMD's Q2 earnings this week (data-center GPU revenue, new anchor customers); the memory prints this week (SanDisk, Western Digital, the next read on pricing); Nvidia's late-August quarter (the demand read for everything upstream); and the next hyperscaler or neocloud debt raise, the funding market is now the swing factor.