Newsletter · · Ashutosh Agarwal

Eisman Says Buy the Suppliers Not the Giants as the Economy Leans on AI Capex - The AI Capex Tracker - Week of July 14, 2026

The AI Capex Tracker for the week of July 14, 2026. Apollo's Torsten Slok told Steve Eisman that AI capex now accounts for about half of US growth even as hyperscaler free cash flow drains toward zero, sharpening the shovels-over-miners case, while a factor investor, a bond manager and a veteran strategist made the debate genuinely two-sided again.

The AI Capex Tracker

Week of July 14, 2026: Eisman Says Buy the Suppliers Not the Giants as the Economy Leans on AI Capex

TL;DR

  • The scariest bear case this cycle is no longer "the stocks are pricey," it is "the whole US economy now leans on this one trade, and the companies doing the spending may have no lasting edge." Apollo's chief economist Torsten Slok told Steve Eisman that the AI build-out alone is adding about 1 percentage point to US growth, roughly half of all the growth the economy is producing this year, while the big cloud companies' spare cash is "literally dropping down toward zero" and could go negative. Eisman's takeaway: if these are giant, capital-hungry businesses with no moat (nothing stopping customers from flipping between ChatGPT, Gemini and Claude), "I'd rather buy Cisco that's going to supply you." (The Real Eisman Playbook, Jul 13)

  • The single cleanest bubble number got sharper: about $755 billion of AI spending going in this year against under $50 billion of AI revenue coming out. One host called that gap "the single clearest, most measurable warning sign of the whole bubble case," and said the number to watch is not the stock price but whether that revenue gap starts to close in the coming earnings reports. (Providence Financial Retirement Show, Jul 13)

  • But the pushback got real, too, and the debate is genuinely two-sided again. A factor investor argued the spending "as a percentage of their overall company is actually not that much" and does not hit the danger threshold; a bond investor said he is happy to lend to these "incredibly high quality" balance sheets at 6–7% because "what are the odds Microsoft is not going to pay its debt? Probably not that high"; and a veteran strategist noted earnings are genuinely exploding, up about 20–22% this season. (Two Quants and a Financial Planner, Jul 13; Bloomberg Surveillance, Jul 13)

What's New

A note on the window first. This is a Tuesday recap, so it sweeps roughly the last 24 hours (Monday, July 13). It was a thin, macro-heavy day: the specialist chip shows (Nvidia's Rubin, AMD's Helios roadmap, Broadcom's custom chips) and the named memory and optics analysts were all quiet, and no fresh single-name semiconductor color cleared the bar. What did clear the bar was the best macro-strategist episode we have had in a while, and, unusually, a strong power-infrastructure operator. Ranked by where the dollars and the risk actually sit:

1. Apollo's Torsten Slok reframes the whole trade for Steve Eisman: the economy now rides on AI capex, and the spenders may have no moat. The Real Eisman Playbook, Jul 13 pairs host Steve Eisman (of The Big Short) with Torsten Slok, chief economist of Apollo. Both are named, serious voices, not anonymous pundits. This is the episode to sit with.

Slok's core point is that AI has become the load-bearing wall of the US economy. Of the roughly 2% growth the economy is producing this year, he attributes about 1 point to the AI spending boom (the data centers plus the energy to run them), 0.9 to last year's tax bill, and 0.3 to reshoring factories. In plain terms: strip out AI and you strip out half the country's growth. He adds that this same build-out is now adding about 0.3 points to inflation, because the chips, the construction labor and the energy are all getting more expensive, one reason he thinks the Federal Reserve has "zero" chance of cutting rates this year and the market is now pricing hikes in September and December.

Then the two vectors that matter for the trade. First, the money. "Capex" is simply the cash a company spends building things, here, data centers. Slok notes free cash flow (the cash left over after a company runs its business and pays for that building) at the big cloud players "has absolutely gone from being very, very high and literally is dropping down over the next six, 12 months towards zero," and "might even begin to go negative." Eisman fills in the scale with Google: it spent about $80 billion on AI last year and funded it from its own cash flow; this year it is spending about $190 billion and just raised $85 billion in equity. The build is still funded, but no longer out of pocket.

Second, and to Eisman the more important point: moats. A moat is a durable edge that keeps competitors and customers from walking away. Eisman's worry is that the hyperscalers may not have one, "people flip from Gemini to Claude to ChatGPT," so you have "massive companies spending trillions of dollars for something that may have no moats. And that's not a recipe for longevity." His analogy is the one to internalize:

"It's kind of like comparing airlines to TransDigm. Airlines is a terrible business because it's very capital intensive and you have no pricing power. And TransDigm, which supplies parts to airlines, is a great business... I'd rather buy Cisco that's going to supply you." (Steve Eisman)

Slok's refinement: some hyperscalers will end up with moats and some will not, because compute demand is "almost unlimited" but the price of that compute, the price of tokens, keeps falling toward zero, with Chinese open-source models already at maybe "1% of what is the price of a token from a US model." He also drops the scariest single-company line of the day: Oracle has a $600 billion backlog, but half of that backlog is OpenAI. Why it matters: this is the shovels-vs-miners thesis, restated by a legendary short-seller and a top macro economist on the same show. If the moat is genuinely thin, the trade is to own the suppliers (chips, power, networking) rather than the giant, cash-burning platforms, and to watch cash flow, not the headline profit.

2. The capex-to-revenue gap got a crisp figure, and a clear "number to watch." Providence Financial Retirement Show, Jul 13, hosted by Anthony Saccaro, is a financial-planning show; treat it as informed commentary, not operator insight.

Saccaro laid out the bubble math as cleanly as anyone has: hyperscalers are on track to spend about $755 billion on AI this year, with forecasts that have already climbed "from $650 billion to $725 billion, with some projecting $785 billion," all "before the revenue shows up." Against that, he pegs total AI industry revenue at under $50 billion. His framing:

"That's a massive gap between what's being poured in and what's coming back out... $755 billion going in, and less than $50 billion coming out. Every other argument, valuations, concentration, debt, is really just a symptom of this core problem. Money is being spent on a promise, not a proven return."

His actionable tell is the useful part: do not watch the stock price, watch "whether that revenue gap actually starts closing in coming earning reports." If it does not, "this is the exact setup that precedes the dot-com collapse, when infrastructure got built years ahead of the demand that was supposed to justify it." Why it matters: it gives you one specific thing to monitor as late-July hyperscaler earnings land, is revenue finally catching up to the spend, or not.

3. The bull side found its footing, three different lenses that all say "calm down." Two Quants and a Financial Planner, Jul 13 is a financial-planning show that ran three substantive guest clips.

This is the counterweight to the bears, and it is worth taking seriously. Matt Zenz, a factor investor (someone who studies which measurable traits historically predict returns), pushed back on the panic directly. History says companies that massively over-invest, growing their asset base 70–100%, essentially doubling in size, tend to deliver lousy future returns. But the hyperscalers are not close: "A lot of these AI companies, Google, Microsoft... as a percentage of their overall company, it's actually not that much... These companies that are doing these data-center type things actually aren't hitting that threshold." His conclusion is refreshingly honest: "there's not enough data to say one way or another." As one host put it, "this is the first way I've seen this framed as it's not the biggest thing in the world."

A bond investor on the same show flipped the debt worry on its head. Yes, Microsoft, Meta, Amazon and SpaceX are all tapping the bond market to fund the build, but from a lender's seat that is fine: "Almost all of these major companies are incredibly high quality tech companies with very, very strong balance sheets... you get a 6% to 7% type yield... What are the odds that this is going to get so bad that Microsoft is not going to pay its debt? Probably not that high." His memorable framing: this "might be the same as '99" for equity investors, but "different for bond investors," in the fiber-optic boom, plenty of borrowers defaulted; this time the borrowers are the strongest balance sheets on earth.

And veteran strategist Jim Paulsen supplied the raw bull fact: forward earnings estimates are at a record high, up about 4% just in the last three months when the norm is for estimates to be cut into a season. His caveat is the important nuance, though: "earnings momentum is great, until it isn't," and markets historically roll over from peak earnings momentum, with momentum stocks surging right at the end (as they did in 2000, and as they are doing again now). Why it matters: after weeks of one-way bearishness on the podcasts, this is genuine, well-argued pushback, the spending may simply not be as reckless as the fear implies.

4. A rare, high-quality power-operator read: "you've got to bring your own power," and gas turbines are winning. Redefining Energy, Jul 13 features David McCullough, founder of infrastructure investor Quinbrook. This is a genuine operator: his data-center developer, Rowan, has gone from zero in 2021 to the "number three data center developer in the United States," and recently announced a deal with Blackstone.

McCullough's blunt summary of the industry: "The only thing that gets you a lease with any of the compute firms is power and timing of power." And the biggest shift underway is that the grid cannot deliver fast enough, so developers now have to supply their own electricity on site, "behind the meter," meaning power generated at the facility that bypasses the public grid. Because AI cannot wait years for a grid connection, that on-site power is increasingly truck-mounted gas turbines and backup generators, not slower-to-mature long-duration batteries: "The urgency of the power needs... means you don't have time for a lot of that evolution... That's why you're seeing this move to short-term [gas turbines]." He notes the real innovation now is the software and control systems that stitch multiple on-site power sources together, and that some hyperscalers are building their own internal power teams. Why it matters: this is the cleanest operator confirmation yet that speed-to-power, not the grid, is the binding constraint, and it points straight at the equipment that sits "between the meter and the chip" (power electronics, cooling) and at anyone who can deliver dispatchable on-site generation fast.

5. The first real siting backlash lands on a governor's desk: New York's data-center moratorium. Closed! NYC's Real Estate Podcast, Jul 13 features State Senator Kristen Gonzalez, the prime sponsor of New York's data-center bill.

Gonzalez described a bill, passed by both chambers and now awaiting Governor Hochul's signature, that puts a one-year "yellow light" pause (her words) on new hyperscale data centers of 20 megawatts and above. The numbers explain the alarm: there are more than 28 hyperscaler proposals in New York that would add roughly 9,000 megawatts to the grid, about a third of the entire ~31,000 MW the state uses today, and the state grid operator says that even if nothing new is added, there is already a demand-supply gap coming by 2030. The bill would also create a separate electricity and water rate class so those costs do not land on ordinary consumers' bills, and require an environmental review of water use. She cited Virginia, the data-center capital, where, she said, roughly $53 million was invested per long-term job created, versus a ~$332,000 average. Why it matters: community and political opposition to siting is no longer hypothetical; it is real legislation that slows the build in populous states, which, as McCullough argues, pushes even more of the build toward behind-the-meter, self-supplied power in friendlier jurisdictions.

The Debate

Steel-manning both sides of the $700B-plus 2026 hyperscaler capex thesis, the trade that Torsten Slok says now accounts for roughly half of all US economic growth.

Bull, the spending is rational, funded, and not actually that extreme. The factor lens says it plainly: as a share of these enormous companies, the capex "is actually not that much" and does not reach the level where over-investment historically wrecks returns; there is "not enough data" to call it a disaster (Two Quants, Jul 13). The credit lens agrees from a different angle, the borrowers are the highest-quality balance sheets in the market, paying 6–7% because supply is heavy, and the odds Microsoft cannot pay are low; this is "different for bond investors" than 1999 (Two Quants, Jul 13). Earnings are genuinely exploding, up about 20–22% this season, driven largely by margins, with estimates rising into the prints rather than being cut (Bloomberg Surveillance, Jul 13). And even Slok concedes the demand is real: compute demand is "almost unlimited." The build is funded, expensively, via debt and equity issuance, but funded.

Bear, the earnings lean on one trade, the cash is draining, and the spenders may lack a moat. Slok's macro warning is the sharpest: half the economy's growth now depends on this capex, hyperscaler free cash flow is heading toward zero or negative, and if the businesses being funded have no moat (customers flip between models freely, token prices race to zero, Chinese models undercut at about 1% of the price), then "that's not a recipe for longevity," buy the suppliers, not the platforms (The Real Eisman Playbook, Jul 13). The spending-vs-revenue gap is stark, about $755B in, under $50B out, and "money is being spent on a promise, not a proven return" (Providence Financial, Jul 13). Even the bull data carries a warning: earnings momentum peaks before markets roll over, and momentum stocks are surging exactly as they did at the 2000 top (Two Quants, Jul 13). And the concentration is a fragility in itself: BNY Mellon's Geoffrey Yu estimates roughly 80% of US growth is now coming from the AI capex trade, which is what separates the US from the rest of the world; if that demand cools without the consumer picking up the slack, "that's where we get a few more wobbles" (Bloomberg Surveillance, Jul 13).

"Free cash flow for the hyperscalers has... literally [dropped] down over the next six, 12 months towards zero. And it might even begin to go negative." (Torsten Slok, Apollo, on The Real Eisman Playbook, Jul 13)

A newer, slow-burn worry, the software credit wall. Slok flagged software as "the number one sector that's vulnerable" in credit: high debt, thin ability to service it, and a maturity wall (a cluster of debt coming due at once) in 2028–29 for loans written in 2021–22, just as AI puts the long-term value of many software businesses in doubt. He notes about $500 billion of the ~$2 trillion private-credit market is software, and public names like Salesforce and ServiceNow are already down 50–60% from their peaks. It is not systemic, the loans sit on lightly levered lenders, not GFC-style 40-to-1 banks, but it is a real read-through for anyone long levered software into 2028 (The Real Eisman Playbook, Jul 13).

Sell signals to watch: the capex-to-revenue gap not closing in the late-July hyperscaler prints (Providence's "number to watch"); any hyperscaler actually cutting forward capex (none has); hyperscaler free cash flow confirmed turning negative; token pricing cracking or Chinese models visibly taking share; power deals that "quietly assume power they have not actually locked down" starting to slip on delivery dates; and early signs of stress at the 2028–29 software debt wall.

Stocks in Play

NVDA. Bull: still the default AI chip, and Slok concedes compute demand is "almost unlimited." Bear: Eisman's no-moat framing, "I'd rather buy Cisco that's going to supply you," plus token prices racing toward zero and Chinese models at about 1% of the price, both squarely aimed at Nvidia's end-demand. Next: Q2 earnings in August; no fresh Rubin roadmap color this cycle. (The Real Eisman Playbook, Jul 13)

AVGO. Quiet on the podcasts this cycle, no fresh operator color on custom silicon after yesterday's Apple-deal and downgrade chatter. Next: custom-ASIC design wins; watch for the next capex-guide read-through.

AMD. Quiet on the podcasts this cycle, no fresh MI450X/Helios roadmap color. The data-center share-shift story (passing Intel) is intact from prior issues but had no new evidence today. Next: AMD Advancing AI Day, July 2026.

MSFT. Bull: the credit market's favorite, "what are the odds Microsoft is not going to pay its debt? Probably not that high," at a 6–7% yield. Bear: it sits inside the software complex Slok calls the most vulnerable corner of credit, and it is now funding the build with debt rather than pure cash flow. Next: FY26 Q4 capex commentary, late July. (Two Quants, Jul 13; The Real Eisman Playbook, Jul 13)

GOOGL. Bull: vertically integrated with its own TPU silicon, and can clearly fund the build, it raised $85B in equity to do it. Bear: that is the point, capex leapt from about $80B (self-funded) to about $190B (now needs outside capital), and Slok will not guarantee every hyperscaler ends up with a moat. Next: July capex guide. (The Real Eisman Playbook, Jul 13)

AMZN. Bull: the strongest balance sheet in the group and a credit the bond market is happy to own. Bear: the clearest case of capex outrunning internal cash, it is in the issuance wave Slok and Eisman describe, and "there's no rainbow in year four" on maintenance spending still hangs over it from prior issues. Next: late-July earnings. (The Real Eisman Playbook, Jul 13)

META. Quiet on the podcasts this cycle, no fresh color after yesterday's $2.3B depreciation change and the compute-rental debate. Next: late-July earnings; watch the depreciation footnote for any further useful-life stretch.

Read-throughs

  • Power and thermal (VRT, ETN), the clearest actionable read of the day. Quinbrook's McCullough says the binding constraint is speed-to-power, and with the grid too slow, developers must "bring your own power," increasingly truck-mounted gas turbines and backup generators, stitched together by control software. That points straight at the power-delivery and cooling gear that sits "between the meter and the chip" (Vertiv, Eaton) and at fast, dispatchable on-site generation. The New York moratorium reinforces it: siting friction in populous states pushes more of the build toward self-supplied, behind-the-meter power. Most actionable: stay long the power-equipment and behind-the-meter enablers; treat grid-dependent, permitting-heavy projects with more caution. (Redefining Energy, Jul 13; Closed! NYC, Jul 13)

  • Own the shovels, not the miners. Both Eisman ("buy Cisco that's going to supply you") and a private-equity host laying out the "AI capital stack" land on the same conclusion: the durable money is in the infrastructure underneath, senior secured infrastructure debt, private credit, and the scarce physical inputs (land, power generation, grid interconnection rights), while the bubble lives in the application- and model-equity layer at the bottom. His line: "Nobody can print a permitted gigawatt of power... deliverable in the next 18 months." Treat the PE-host framing as informed opinion, not operator data, but it rhymes exactly with the operator (Quinbrook) and the short-seller (Eisman). (The Real Eisman Playbook, Jul 13; Making Billions, Jul 13)

  • Memory, networking and optics (Micron, SK Hynix, Marvell, Astera Labs, Credo, Coherent, Lumentum). Quiet on the podcasts this cycle, after yesterday's Micron cash-flow and SK Hynix listing news, no fresh operator or specialist color cleared the bar today. Carry the positions.

  • Software, a new caution flag. Slok singles out software as the most vulnerable credit sector, with a 2028–29 maturity wall meeting AI's threat to these businesses' long-term value; about $500B of private credit is software, and Salesforce and ServiceNow are already down 50–60% from peaks. Read-through: be wary of levered software borrowers into 2028, distinct from the mega-cap platforms. (The Real Eisman Playbook, Jul 13)

  • Rates backdrop. Worth holding in mind: Slok says the Fed has "zero" chance of cutting this year and the market is pricing hikes in September and December, with inflation around 3.5%, a headwind for the long-duration growth names leading this trade. (The Real Eisman Playbook, Jul 13)

What Changed vs Last Issue

Yesterday's issue (Jul 13, "Strip the accounting and the AI boom shrinks") was about the bears finally doing the arithmetic, Boockvar and Gave showing that about 28% headline profit growth becomes about 17% ex-"other income" and mid-single digits ex-semis, plus the Meta depreciation trick, Micron's cash machine, the SK Hynix listing, and AMD passing Intel. Today, three things moved:

  • The bear case widened from "the earnings are flattered" to "the whole economy is exposed." Yesterday's critique was an accounting one. Today Apollo's Torsten Slok makes it a macro one: AI capex is about half of US GDP growth (BNY's Geoffrey Yu independently pegs it at about 80% of the growth impulse), hyperscaler free cash flow is heading toward zero or negative, and the spenders may lack a moat. That is a bigger, structurally scarier claim than a depreciation footnote.

  • A brand-new, slow-fuse risk appeared: the 2028–29 software credit wall. Not previously on our radar, about $500B of private credit in software, maturing into 2028–29 just as AI questions those businesses' terminal value. One to file for later, not this quarter.

  • The debate turned genuinely two-sided. Yesterday leaned almost entirely bearish. Today brought three credible bull and neutral lenses, the factor view (capex is not extreme relative to firm size), the credit view (happy to lend to these balance sheets at 6–7%), and Federated's Stephen Auth on about 20–22% earnings growth. This is healthier than a one-way bearish chorus and lowers our conviction that a break is imminent.

  • Power flipped from "quiet on operators" to a real signal. Yesterday the power beat had no operator color. Today Quinbrook, a top-three US data-center developer, gave a clean, actionable read (bring-your-own-power, gas turbines winning on speed), and New York's data-center moratorium landed on the governor's desk as the first hard siting backlash.

  • Deltas on the big numbers: 2026 hyperscaler capex now cited at about $755B (Providence), consistent with the $700B-plus anchor and edging toward $785B in some forecasts; the AI capex-to-revenue gap is quantified as about $755B in vs under $50B out. Google's capex figures got concrete, about $80B last year to about $190B this year plus an $85B equity raise. And the rates regime shifted the wrong way for growth: from expecting cuts to the market now pricing two hikes. No hyperscaler has cut capex; the memory and single-name chip counters are unchanged (quiet today).

Next prints to watch: the SB6 large-load queue deadline in Texas falls tomorrow (Jul 15), a real-world test, flagged in prior issues, of how much of that pipeline is actually buildable; then late-July hyperscaler earnings (MSFT, AMZN, META, GOOGL) that will show whether the about $755B of spending is finally buying visible revenue, Providence's "number to watch."