Newsletter · 2026-07-27 · Ashutosh Agarwal

AI Capex Reckoning Arrives as Wall Street Demands Hyperscalers Prove the Returns - The AI Capex Tracker - Week of July 27, 2026

The weekend reaction to Google's capex-driven negative cash flow, Nvidia's supply-constrained bull case, and New York's new data-center moratorium, for the week of July 27, 2026.

The AI Capex Tracker

Week of July 27, 2026: AI Capex Reckoning Arrives as Wall Street Demands Hyperscalers Prove the Returns

TL;DR

Over the weekend, the market's new rule hardened into a slogan: prove it. After Google spent so much last quarter that its cash flow went negative for the first time since 2004, the burden of proof has flipped from the skeptics to the spenders. NewEdge Wealth's Cameron Dawson framed the whole race as a "prisoner's dilemma" (everyone has to keep spending because the first one to cut admits it overspent) and warned that Wall Street already expects Google's free cash flow to swing from roughly plus $13 billion this year to minus $13 billion next year. (RiskReversal Pod, Jul 24)

Nvidia's Jensen Huang, speaking from Seoul, handed the bulls their single best counter-argument: there is not enough of anything, and there won't be for years. He said Nvidia is "constrained in just about every part of the supply chain" (memory, land, power, even construction workers) and that the industry can only "double each year" and no faster "for a decade." He backed it with a $500 billion-plus deal with Korea's SK Group and said Nvidia has to buy "a lot of system memories" to build "$1 trillion worth" of its next-generation systems. (Bloomberg Tech, Jul 25)
The bottleneck now has a political face, and the week ahead is the real test. New York's governor ordered a one-year freeze on big new data centers while the state studies whether the grid can handle them, and the demand-side wall is spreading beyond Texas. And Microsoft, Amazon, Meta (plus Apple) all report next week, into a Federal Reserve meeting where the market suddenly sees a real chance of a rate hike. (Power & Politics, Jul 24; The Exchange, Jul 24)

What's New

A note on the window: this is a Monday recap sweeping Friday evening plus the weekend. There were no blockbuster new earnings, this was the market digesting Google's Wednesday-night print and getting itself set for what several podcasts called a "make-or-break week." The debate got sharper, the bears got louder and more specific, and a genuinely new bottleneck appeared: politics. Ranked by where the real money sits:

The weekend verdict on Google: the burden of proof has flipped, and the "prisoner's dilemma" is the frame to internalize. (RiskReversal Pod, Jul 24, Cameron Dawson, Chief Investment Officer at NewEdge Wealth.)

Dawson gave the clearest anatomy of why a great quarter got sold. On the surface Google looks cheap at about 15 times earnings, "but you have to appreciate that 80% of its earnings in the second quarter came from the gain on Anthropic and SpaceX," a one-time $99 billion mark-up that won't repeat. Strip it out and, she argued, "there's a very easy scenario where Google has negative earnings growth next year." The funding math is the real worry: Wall Street's consensus has Google's full-year free cash flow at positive $13 billion in 2026 flipping to negative $13 billion in 2027, and "you have to ask the question of who's going to pay for it?" She flagged that a recent Amazon bond offering "got much less demand than what was expected," and cited an estimate of $1.65 trillion in debt tied to the data-center build, much of it in off-balance-sheet vehicles that, she noted dryly, we're supposed to have "learned our lesson" about since Enron.

Her central metaphor is worth taping to the monitor:

"We've been thinking about hyperscaler CapEx as a rather unfortunate prisoner's dilemma… everybody colluding… saying we all have to keep spending because we've spent so much. And if one of us stops spending, then it's going to suggest that maybe we spent too much… All you have to do is see one person defect… and see their stock pop on a CapEx cut. And then that could call into more questions about this whole AI infrastructure trade."

Cameron Dawson, NewEdge Wealth, on RiskReversal Pod, Jul 24

Why it moves the thesis: for the whole cycle, the bull trigger was "the number keeps going up." Dawson's framing says the next big move could come from the opposite, the first hyperscaler brave enough to cut capex, whose stock pops, forcing everyone to re-examine what's already been spent. She also punctured the favorite bull comfort blanket (that AI is "so early" at 2% heavy adoption) with market history: railroads, telephone, radio and the internet all kept adopting while the stocks that built them fell, because "you are early in the adoption curve" does not mean "you are early in the investment curve."

The bull's best rebuttal came, fittingly, from the man selling the shovels: Jensen Huang, live from Seoul. (Bloomberg Tech, Jul 25; Bloomberg Intelligence, Jul 25, Jensen Huang, CEO of Nvidia, interviewed by Bloomberg's Ed Ludlow.)

Huang's answer to the "over-building" fear is that there is nowhere near enough of anything:

"We don't have enough bits. We're constrained in HBM memories, LPDDR memories. We're constrained in just about every part of the supply chain. We're even constrained now with land and power and construction workers to set up the data centers. I think this is one of the areas that is going to make sure that we continue to build out in a throttled way… for a decade."

Jensen Huang, Nvidia, on Bloomberg Tech, Jul 25

His key structural claim: "I think we have the ability as an industry to double each year. But we're going to have a hard time growing much faster than that," because, unlike phones and PCs, "it's really, really hard to scale up land, power, and shell." He put commercial weight behind it: a partnership with Korea's SK Group worth "over $500 billion" in memory purchases and supercomputer sales as SK builds out 2 gigawatts of AI factories; large purchase agreements with SK Hynix because "in order to build $1 trillion worth of [Nvidia's Rubin-generation] systems, you're going to have to buy a lot of system memories"; and a $1 billion investment in Korean AI cloud Naver. His decade-long framing: the semiconductor industry "is probably going to have to be 10 times larger than it is today," because soon computers won't just serve a billion people, they'll serve "100 billion agents and billions of robots." (He also made his first-ever post on X to share an open-models letter co-signed with Microsoft's Satya Nadella, a reminder these giants still cooperate even as they compete.) Why it matters: HBM (high-bandwidth memory, the fast memory stacked next to an AI chip) being sold out "for a decade" is the cleanest, most literal version of the bull case: if the constraint is physics, not demand, then today's capex isn't over-ordering, it's rationing.

The bear chorus got louder, and, crucially, more specific about the exit. (Planet MicroCap, Jul 25, Dan Rasmussen, founder of Verdad Advisers; Wealthion, Jul 24, Kevin Muir ("The MacroTourist"); The Peter Schiff Show, Jul 26, Peter Schiff.)

Rasmussen offered two "eternal truths" that cut straight at the trade. First: "CAPEX spending always disappoints… because the expectations were too rosy and because competitors invested on the same thesis," so "it's almost certitude that this stuff is going to crash at some point." Second, the tell for timing: "the next stage in a bubble is equity issuance. And we're about to see equity issuance at a scale we have never seen before… pretty close to the top." He's blunt that semis aren't a new permanent infrastructure layer: "semiconductor market is cyclical… oil's really valuable too, but oil's still cyclical because it's capital intensive. You just can't escape those laws of nature."

Muir made the sharpest positioning point: the people who know these businesses best "are actually selling, not buying," Google raised capex by an amount that "was just shocking… they didn't need to do that." His advice for actual AI bulls is counter-intuitive and worth chewing on: don't buy the builders, buy the users. "If you really truly are an AI bull… you should be buying the companies that are going to benefit from all that AI, as opposed to the companies who are going to build it," because massive overcapacity means everyone else gets to rent AI for far less than it cost to build. His cautionary tale is the dot-com fiber build: the NASDAQ fell ~75% and took "something like 14 years to break even," and the big fiber companies ("like 45% of the Canadian index") simply "do not exist now." He, too, flagged the hidden leverage: "there's all sorts of SPVs, special purpose vehicles, that are hiding a lot of the debt."

Schiff went furthest, calling the roughly $500 billion spent last year and ~$750 billion projected this year classic malinvestment that won't earn its return as Chinese competition commoditizes AI, and pointed to Oracle, down about 41% year-to-date, as the market already punishing a heavy spender with nothing yet to show for it.

A genuinely new bottleneck appeared, and it's political, not physical. (Power & Politics, Jul 24, Kathy Hochul, Governor of New York.)

Hochul announced an executive order placing a one-year moratorium on new large data centers (50 megawatts or more, "as much power as 50,000 homes") while the state studies whether the grid can absorb them. The scale problem in her own words: New York has 30 pending applications, and in one county alone (St. Lawrence) five applications "would consume as much power as one nuclear reactor," the same nuclear power she's trying to build to serve Micron's $50,000-job fab in Syracuse. She's not offering incentives, either: data-center developers, she said, "basically admitted we don't need the tax breaks… we're making enough money." She frames it as a straight trade-off: "if I have a choice of powering 50,000 jobs… or powering an enormous facility that creates a couple of jobs or a dozen jobs, I think I'm going to go with the job creators." Why it matters: Texas already showed the physical limit; New York now shows the political one. When developers "don't need tax breaks," communities and governors start asking why they should host the load at all. This is a rising, under-priced risk to the pace of the buildout, and, by extension, to the demand read for everything upstream.

The bull rebuttal, in the operators' and owners' own words. (All-In, Jul 24, Chamath Palihapitiya and David Friedberg.)

The All-In hosts staked out the clean bull case on Google. Chamath's anchor stat: Google's return on invested capital has averaged "32%" over its two decades public, "when you are a machine… that compounds money at 32% over a 20-year average, you give these guys the benefit of the doubt." Friedberg's conclusion was even blunter: "if you want to bet on AI, I think the best public market stock to own is Google," because in the "worst, worst, worst case scenario… they have the lowest cost infrastructure in the world to run other people's models." They also noted Tesla's capex surged 142% to a planned ~$25 billion, and, like Google, went free-cash-flow negative, a reminder the spend is broadening beyond the classic hyperscalers. Why it matters: the bull case no longer rests on "revenue will show up soon." It rests on decades of capital discipline and the argument that owning the lowest-cost compute is a moat in itself, regardless of which model wins.

The Debate

Steel-manning both sides of the (now roughly $1.2-trillion-in-2027, $1.4-trillion-in-2028) hyperscaler capex thesis. Google gave us the first hard data point last week; the weekend was spent arguing over what it meant.

Bull: the constraint is physics, the discipline is real, and the returns compound. Huang's "constrained in just about every part of the supply chain… for a decade," backed by a $500 billion SK Group deal and $1 trillion of systems to build, is the strongest evidence that the spend is rationing scarce capacity rather than over-ordering (Bloomberg Tech, Jul 25). Google's 32% two-decade return on capital argues management has earned the benefit of the doubt, and owning the lowest-cost compute is a moat even in the bear case (All-In, Jul 24). Even a bear concedes the counter-move: Dawson notes Meta once got its multiple "chopped to bits," found cost discipline, and became "a great buying opportunity," the same script could rescue any hyperscaler that shows restraint (RiskReversal Pod, Jul 24).

Bear: the cash flow is going negative, the returns are unproven, and history rhymes. The hardest new facts are financial: Street consensus sees Google's free cash flow flipping to negative $13 billion in 2027, an Amazon bond deal already drew weak demand, and roughly $1.65 trillion of data-center debt is piling up in hard-to-see vehicles (RiskReversal Pod, Jul 24). Rasmussen's "capex always disappoints" and "semis are cyclical, you can't escape those laws of nature," plus his warning that record equity issuance signals the top, is the cleanest bear logic yet on the market's most-crowded trade (Planet MicroCap, Jul 25). Muir's dot-com fiber analog (NASDAQ down ~75%, "14 years to break even," the builders wiped out while the users thrived) argues you can be a raging AI technology bull and still short the builders (Wealthion, Jul 24). And Petrides warns the discipline may be forced, not chosen: if investors stay frustrated "they'll speak with their fingers and hit the sell button" until management is pushed "into submission to either put up or shut up," exactly as happened to Meta in 2022 (The Rundown, Jul 26).

The synthesis. Both sides now agree on the mechanism and disagree only on the timing. Huang says the constraint is real for a decade; the bears say the equity-and-debt issuance to fund it is peaking right now. The swing factor is next week: three of the four biggest spenders report into a market that has openly changed its scoring, and into a Fed meeting where a rate hike is suddenly on the table.

Sell signals to watch:

A "defection." Per Dawson's prisoner's dilemma, watch for the first hyperscaler to cut or soften capex and see its stock pop, that would flip sentiment fastest, because it implies the others over-spent.
Another negative free-cash-flow line next week. Google went first at negative territory; if Microsoft, Amazon or Meta join it, "negative FCF" stops being a Google story and becomes an industry story.
A failed or weak bond deal. Amazon's soft demand was the first crack; a hyperscaler struggling to place debt to fund the build is the funding-market tell (Dawson).
Record equity issuance. Rasmussen's "next stage in a bubble is equity issuance… pretty close to the top."
Regulatory pushback spreading. New York just froze 50MW+ projects; more moratoriums or permit denials would throttle the demand read for chips and power gear alike (Hochul).

Stocks in Play

NVDA. Bull: Huang's "constrained in just about every part of the supply chain… for a decade," plus a $500 billion SK Group deal and $1 trillion of Rubin-generation systems to build, says demand is rationed, not saturated (Bloomberg Tech, Jul 25). Bear: it's the crowded global proxy for a trade the market just started punishing, and Rasmussen calls semis inescapably cyclical with a crash "almost certitude" once over-investment shows (Planet MicroCap, Jul 25). Next: next week's Microsoft/Amazon/Meta capex guides (the demand read for GPUs), then Nvidia's own August quarter.

AVGO. Bull: the "own-the-silicon" theme keeps getting validated, Google's in-house chips are central to its lowest-cost-infrastructure bull case, which underlines that hyperscalers treat custom accelerators as core (All-In, Jul 24). Bear: no fresh operator color on Broadcom itself this cycle, so if the group re-rates on the "returns unproven" fear, the custom-silicon names ride the same beta down. Next: the hyperscaler capex guides next week that fund the custom-chip ramp.

AMD. Bull: quiet on the podcasts this cycle (no fresh MI-series datapoint) but any credible second source to Nvidia benefits as the overall compute pie keeps expanding and supply stays "constrained… for a decade" (Bloomberg Tech, Jul 25). Bear: with no company-specific catalyst, it trades on group sentiment, and the group just entered a "prove it" regime. Next: the next MI-series design-win confirmation; read-through from next week's hyperscaler guides.

MSFT. Bull: the "guide higher" camp expects it to reaffirm and extend spending when it reports next week (the outcome that would floor the whole complex) and one podcast noted Microsoft has been more measured than peers, "not doing this" all-in stare-down, which could read as discipline (Big Technology, Jul 24). Bear: it sits squarely in the "does it match Google's raise, and does it go free-cash-flow negative too?" spotlight (The Rundown, Jul 26). Next: earnings next week, the capex guide and the free-cash-flow line are the numbers that matter.

GOOGL. Bull: a 32% two-decade return on capital, an 82% cloud quarter, and (in the worst case) "the lowest cost infrastructure in the world to run other people's models"; Friedberg calls it "the best public market stock to own" for an AI bet (All-In, Jul 24). Bear: 80% of Q2 earnings came from a one-time $99 billion gain, capex is up at $195-205 billion, free cash flow already went negative for the first time since 2004, and consensus sees it hitting negative $13 billion in 2027 with a live question about whether Gemini has slipped to "copilot status" (RiskReversal Pod, Jul 24; Big Technology, Jul 24). Next: the Microsoft/Amazon/Meta prints that either validate or isolate its capex raise.

AMZN. Bull: AWS is a primary compute home for Anthropic and remains a core beneficiary of a supply chain that's sold out for years (Bloomberg Tech, Jul 25). Bear: a recent Amazon bond offering "got much less demand than what was expected," the first crack in the funding market that pays for all this (RiskReversal Pod, Jul 24). Next: earnings next week, watch the AWS capex guide and whether the free-cash-flow line follows Google negative.

META. Bull: a top-tier spender riding the same "constrained for a decade" supply backdrop, with a history (2022-23) of finding cost discipline and re-rating hard when it chooses to (RiskReversal Pod, Jul 24). Bear: it remains the poster child for the debt-funded, priced-for-perfection critique, and reports into a market that will punish another negative-free-cash-flow raise (The Rundown, Jul 26). Next: earnings next week, the capex guide and free-cash-flow line.

Read-Throughs

Power / grid: the physical bottleneck is now the industry's own headline worry, and "announced ≠ built." Adam Selipsky, the former AWS CEO now running infrastructure firm Helix, gave the sharpest operator warning of the sweep: data centers have gone from "20, 100, maybe 150 megawatts" three or four years ago to "routinely… a gigawatt or two," and the certainty of getting power is "one of the most critical and probably highest risk elements today that I personally believe is being mispriced… by the market." His prediction: "you'll see a lot more data centers that have been announced so prominently in the public eye actually not getting completed." The bridge fixes cost a fortune (on-site fuel cells or reciprocating engines run "six, seven, eight [times]" the cost of an efficient large-scale gas plant), which is exactly why he partnered with Vistra for gas generation and EPC know-how (Tech Disruptors, Jul 24). Corroborating from the field: Iron Mountain's Chris Pennington said on-site natural gas plants of "300, 400, 500 megawatts" are now common, but come with "10, 15, maybe 20 year" contracts that lock in gas costs, "a short-term solution that has potential long-term impacts" (Inside Data Centre, Jul 24). Most actionable trade: stay long the power/grid/cooling supply chain (VRT, ETN, gas-turbine and behind-the-meter names), the constraint is worsening, "bring your own power" keeps winning, and the gap between announced and buildable capacity is a filter, not a fear.
Regulation: the new, actionable read. New York's one-year moratorium on 50MW+ data centers is the first hard political brake, and the governor's tell (developers "don't need the tax breaks… we're making enough money") signals communities will increasingly ask why they should host the load (Power & Politics, Jul 24). Watch for copycat moratoriums; each one is a small negative to the demand read for chips and power gear, and a positive for whoever can offer self-contained, behind-the-meter, community-friendly power.
Memory / networking: Huang just reframed it as the tightest link in the chain. Being "constrained in HBM memories, LPDDR memories" and signing "large purchase agreements" with SK Hynix to feed "$1 trillion worth" of systems is the cleanest read that memory (Micron, SK Hynix, Samsung) is the most rationed slice of the build (Bloomberg Tech, Jul 25). The caution: the money just rotated out of memory. Muir described the whip-saw (money rushed into "chips and memory stocks," then "within 2 weeks them all having corrected 30% plus" as it rotated back toward the Mag 7), a reminder these are momentum trades layered on a real fundamental (Wealthion, Jul 24).
Optics / networking: still a coverage gap. For another cycle, no podcast in this window carried substantive color on the optical names (Coherent, Lumentum, Fabrinet) or interconnect silicon (Marvell, Astera Labs, Credo). A calendar gap to close on a future sweep, not evidence the pressure has eased.
The rotation read. The through-line is unchanged and intensifying: money is leaving the crowded, priced-to-perfection compute names and hunting for either the physical bottleneck (power, grid gear) or, per the bears, the users of cheap future AI capacity (biotech, financials) rather than its builders. Whether that's healthy broadening or the first crack depends entirely on next week's prints.

What Changed vs Last Issue

Last issue (Jul 24, "Google raised capex to $205B. Shares fell anyway.") reported the print itself. This issue is the weekend reaction, and the reaction hardened the new regime and, importantly, added two things last issue didn't have: a coherent bear consensus and a political bottleneck.

The "prove it" rule got a mechanism. Last issue's headline was that raising capex is no longer rewarded. This issue names the specific move that could break the trade: Cameron Dawson's "prisoner's dilemma," where the first hyperscaler to cut capex sees its stock pop and calls everyone else's spending into question. The bull trigger and the bear trigger have swapped places.
The cash-flow hole got a 2027 number. Last issue's fresh bear fact was Google's first negative FCF quarter (-$5.9 billion). This issue extends it to a full-year forecast: Street consensus sees Google flipping from roughly +$13 billion (2026) to -$13 billion (2027), with a weak Amazon bond deal and ~$1.65 trillion of data-center debt as the funding-market evidence.
The bulls found their anchor: supply, not demand. Last issue the bull case leaned on Kurian's "rent-now-own-later." This issue it's Jensen Huang saying the whole industry is "constrained… for a decade" and can only "double each year," backed by a $500 billion SK Group deal and $1 trillion of systems to build. If true, capex isn't over-ordering, it's rationing.
A brand-new bottleneck: politics. Last issue's constraint was physical (the Texas grid). This issue adds a political one (New York's one-year moratorium on 50MW+ data centers) and a striking admission that developers "don't need the tax breaks." Regulatory pushback is now a live, spreading risk to the pace of the build.
The bear bench got deeper and more specific. Last issue leaned on Noble and Irons. This issue adds Rasmussen ("capex always disappoints," semis cyclical, equity issuance = top), Muir ("buy the users, not the builders," SPV debt, 14-years-to-break-even fiber analog), Schiff (Oracle -41% YTD as the cautionary tale), and Petrides ("put up or shut up," Meta-2022 redux).
Macro re-entered the frame. Rates and oil, ignored last issue, are back: the 10-year is above 4.6%, oil is up ~$20/barrel off the July 3 low, and the market now prices a ~38% chance of a Fed hike next Wednesday and effectively 100% by September (Dawson, Petrides). A hike into a debt-funded capex boom is not a friendly backdrop.
Deltas on the big numbers: aggregate hyperscaler capex framed at $1.2 trillion (2027) → $1.4 trillion (2028) (Dawson); Google FY2026 capex $195–205B, FCF consensus +$13B (2026) → -$13B (2027); ~$1.65 trillion data-center debt; Nvidia–SK Group deal >$500B, $1 trillion of Rubin-generation systems to build; industry can only ~2x per year for a decade (Huang); semiconductor industry to be ~10x larger in a decade (Huang); Tesla capex +142% to ~$25B, also FCF-negative; NY moratorium on 50MW+ projects, 30 applications statewide, industry planning ~$6 trillion of data centers over 5–10 years (Hochul); Oracle ~-41% YTD (Schiff); dot-com NASDAQ -75%, 14 years to break even (Muir).
Still open: whether Microsoft, Amazon and Meta match Google's raise and whether any also flips free-cash-flow negative; whether a hyperscaler "defects" with a capex cut; whether the Amazon bond wobble becomes a broader funding-market problem; the optics/networking coverage gap (now several cycles); and how far the regulatory brake spreads.

Next prints to watch: Microsoft, Amazon, Meta and Apple next week (do they match the raise, and does anyone else go FCF-negative?), the Fed meeting Wednesday (a live hike risk into rising oil and yields), and then Nvidia's own August quarter as the demand read for everything upstream.