Newsletter · · Ashutosh Agarwal

OpenAI's Compute Chief Says the Only Risk Is Building Too Slowly - The AI Capex Tracker - Week of July 17, 2026

The AI Capex Tracker for the week of July 17, 2026. OpenAI's own compute chief says demand far outstrips supply and the only real risk is building too slowly, while IBM's roughly 25% earnings-day drop hands the bears their first hard corporate sign of AI-hardware over-ordering.

The AI Capex Tracker

Week of July 17, 2026: OpenAI's Compute Chief Says the Only Risk Is Building Too Slowly

TL;DR

  • The best demand-side signal in weeks came straight from the buyer's chair. Sachin Katti, OpenAI's Head of Industrial Compute (and former Intel CTO), said flatly that "demand far outstrips compute supply today, so anything we can bring online, we consume immediately," and that his single biggest worry is not overbuilding, it's being physically unable to build fast enough. OpenAI's own tally: it will spend ~$50 billion on compute this year, inside an industry spending ~$700 billion. His proof the spend pays: "we tripled compute and we tripled revenue." (The MAD Podcast, Jul 16)

  • But the bears finally got a hard corporate data point instead of a podcast opinion: IBM. IBM pre-announced a revenue miss and fell ~25% in a day, and its own explanation is the tell, in the last weeks of June, customers "shift[ed] their quarterly CapEx spend towards servers, storage, and memory purchases to secure supply-constrained infrastructure ahead of expected price increases," and delayed software. Ben Emons of FedWatch Advisors called it a possible "canary in the coal mine" for double- and triple-ordering that ends in an inventory glut. (RiskReversal, Jul 15; DHUnplugged, Jul 15)

  • The whole fight has narrowed to one stock complex, memory, and it's now a valuation war, not a demand war. Peter Boockvar says "the bubble is clearly on the CapEx side," with Micron's 85% gross margin and ~10x sales screaming 1999-2000, and Chinese rivals (CXMT, Yangtze Memory) coming for the margin. The bulls counter that Micron's revenue is up ~350% year-on-year to a likely $50 billion-plus quarter, yet it trades at a forward P/E of 6-7, the market already pricing a ~70% profit collapse. (Wealthion, Jul 15; Buy Hold Rant, Jul 16)

What's new

A word on the window. This is a Friday recap sweeping roughly the last day, the July 16 podcasts, with a look back into the evening of July 15. Yesterday's issue leaned on the "spending forecasts keep going up" specialists (a16z, Carson, SaaStr). This cycle the conversation split cleanly in two: a rare on-the-record view from inside the largest buyer of compute on earth, and, for the first time in a while, a real company (IBM) putting a hard number on the "over-ordering" fear the bears keep circling. Ranked by where the dollars and the risk actually sit:

1. OpenAI's own compute boss says the only real risk is building too slowly. The MAD Podcast with Matt Turck, Jul 16. Sachin Katti, Head of Industrial Compute at OpenAI (previously CTO of Intel, professor at Stanford). This is an operator/insider, not a pundit, and that makes it the most important thing in the sweep.

Katti runs what he calls "one of the largest things humanity has ever built." His framing cuts directly against the bubble crowd: "Demand far outstrips compute supply today. So anything we can bring online, we consume immediately." Asked the obvious question, aren't you worried about overbuilding?, he flipped it: "Any time we have thought we have enough compute, [and decided] we can slow down, [it has] always negatively surprise[d] us like, oh, we should not have slowed down… our biggest worry is [that] the physical world does not move that fast." Factories, supply chains, and power can't add capacity fast enough, that, not a demand air-pocket, is what keeps him up at night. His durability argument is one line every PM should write down: "We tripled compute and we tripled revenue", spend and revenue have moved together, so far.

Three things a hedge-fund book should extract:

  • Power is the binding constraint, full stop. OpenAI now treats grid impact as "a hard commitment that we are not taking power away from the grid", funding new generation (gas, solar, hydro), transmission lines, transformers and substations wherever it builds. Where the grid is tapped out, "everyone's looking at behind the meter", meaning on-site power the data center owns rather than pulling from the utility, and today that means gas turbines, "the most dense, transportable form of energy… in the US." The catch: they're "bottlenecked" by turbine supply. On nuclear: "It can't come soon enough… the densest form of energy we can produce."
  • Cooling is now everywhere in the box. These chips "run really hot", liquid-cooled, because "you can't cool them with air", and it's not just the chips: "even the cables… the transformers that distribute the power become too hot," so they cool those too. Better cooling literally buys more compute: "the hotter the chip, the more memory bandwidth you get, the more flops you get." That is a straight tailwind for the thermal-management names.
  • OpenAI is now a chip company too (Jalapeno). Its custom silicon is optimized for one number, "tokens per watt", precisely "because the world is constrained by power today." And inference (running models for users) is now "perhaps even the majority" of compute, because "a lot of training is now inference" (synthetic data, post-training, test-time compute). Translation: the workload mix keeps tilting toward the thing that runs 24/7.

2. IBM is the bears' first hard print, the "canary in the coal mine." RiskReversal Pod, Jul 15. Ben Emons, FedWatch Advisors, with Guy Adami and Dan Nathan; corroborated by DHUnplugged, Jul 15.

IBM pre-announced (told the market its numbers early, which companies only do when something's off): preliminary revenue ~$17.2 billion, a ~$700 million miss, adjusted earnings ~$2.93 vs. ~$3.01 expected, a small miss that took the stock down ~25% and dragged ~400 points off the Dow. The reason is what matters. IBM's CEO said that in the last weeks of June, clients "shift[ed] their quarterly CapEx spend towards servers, storage, and memory purchases to secure supply-constrained infrastructure ahead of expected price increases", and delayed software. In plain terms: companies are so worried about not getting chips and memory that they're front-loading hardware buys and starving other budgets to do it. Emons, who owned IBM as a "defensive" way to play the memory boom, so this hurt, called it a possible "canary in the coal mine": "you're seeing double and triple ordering… at some point that moderates and… we might have just a lot of excess inventory of compute." He also sized the systemic exposure: SK Hynix and Samsung are ~40% of Korea's KOSPI index, and "semiconductor manufacturing is 60% of Taiwan's GDP", whole economies, not just stocks, are levered to this one trade. His verdict: "IBM today was a wake-up call."

"You're seeing double and triple ordering… at some point that moderates, and at some point we might have just a lot of excess inventory of compute." (Ben Emons, FedWatch Advisors, on RiskReversal, Jul 15)

3. The memory war is now bull-vs-bear on valuation, and China is the wildcard. Bear: Wealthion, Jul 15. Peter Boockvar, CIO of One Point BFG Wealth Partners. Bull: Buy Hold Rant, Jul 16. Retail investors, treat as opinion not operator color.

Boockvar's thesis is the cleanest bear framing this cycle: "The bubble is clearly on the CapEx side, not necessarily the technology", the spenders, not the users, are where the risk lives. His memory-specific warning: Micron is running an 85% gross margin "in what has historically been a highly cyclical space," and new supply doesn't hit until 2027-2028. His rule on cheap-looking cyclicals is worth memorizing: "never tell me what a PE ratio is on a cyclical stock and tell me it's cheap", because a low P/E on a cyclical usually means peak earnings, not a bargain. He'd rather price memory on sales, where Micron at its peak hit ~10x sales, the same 9-10x that DRAM makers touched in 1999-2000. And the part nobody's watching, per Boockvar: CXMT, the world's fourth-largest DRAM maker, was set to IPO in Hong Kong this week, with Yangtze Memory coming for the NAND (flash storage) side. China's playbook, in his Bezos paraphrase: "your profit margin is my opportunity", they chase share, not margin.

The bull rebuttal (retail, but the numbers are real): Micron's revenue has gone from ~$13B to ~$23B to ~$41B (two straight quarters up ~75%) and is expected above $50 billion this quarter, up ~350% year-on-year, yet the stock sits ~25% off its high at a forward P/E of 6-7. The bull's point: "Micron is getting discounted roughly 70% on its future profits because no one thinks it's sustainable." Same show on SK Hynix's new U.S. listing: it's an ADR (a U.S.-traded wrapper for a foreign stock, here at a 10-to-1 ratio) that adds currency and geopolitical layers "without a substantial discount," so they prefer Micron. So the two sides don't even disagree on the demand, they disagree on whether today's fat margins survive the next down-cycle.

4. The selloff was leverage, not fundamentals, and Burry's bear case still isn't proving out. Full Signal, Jul 16. Jack Kuznicki, macro strategist.

Kuznicki's read on the June/July chip drawdown: it was "leverage driven," not a demand crack. "Leverage drove it up and leverage is now driving it down, but the fundamentals are intact." He dates the top to mid-June, with the first crack on June 3, when Broadcom's guidance disappointed and knocked the interconnect/switch names (Broadcom, Marvell) before memory rolled over. He also flagged the crowding: a Bank of America survey found "83% of fund managers think it's the most crowded trade", near a record. On the most-cited bear thesis, Michael Burry's argument that GPUs wear out faster than the accounting assumes (chips depreciated over five years but really useful for ~three, which would understate costs and overstate profits), Kuznicki is blunt: "every single piece of data over the past seven months has just not been true" for that story; GPUs keep holding their value. His most provocative line is a warning for the ASIC trade: Broadcom and Marvell "design chips for other people… you're basically a consultant, and consultancies shouldn't trade at 60 times earnings." Where he's most bullish: Nvidia ("the one theme that has not worked… one of the stocks I'm the most bullish on") and the semiconductor-equipment makers (ASML, Lam Research), which have recurring service revenue that keeps paying even into a downturn.

5. The funding math is the real bear case, and it runs through the bond market. Something More with Chris Boyd, Jul 16. Chris Boyd and colleague (wealth advisors); corroborated by The Peter Schiff Show, Jul 15 (perma-bear generalist, weigh accordingly).

Boyd's number is the one to carry: hyperscaler capex ~$800 billion in 2026 (up from ~$400 billion in 2025) heading toward ~$1 trillion in 2027, and "up until recently, that's been funded from their own cash flow… now, recently, the numbers have gotten so big that they're borrowing." His striking stat: "about 40% of all net new [debt] issuance is going towards this." That collides with rates: with the government also borrowing heavily, quality AI-infrastructure debt may need to yield "6% to 7%," and lower-quality software/private credit already trades "through 12% yields… because they're not expected to be able to pay that paper in 2027, 2028." Higher yields to fund the build mean cheaper equities, the discount-rate math works against the very stocks doing the spending. Schiff sharpens the accounting version of the same worry: because capex isn't an operating expense, the hyperscalers "still have income," but if GPUs really last "two to three years" rather than the assumed "five or six," the spend is really a recurring cost of doing business, and "these big cash cows have turned into cash vacuums" now borrowing ">$100 billion" in the bond market. Both land on the same trigger: this breaks on funding stress, not on a demand print.

The debate

Steel-manning both sides of the $700 billion-plus 2026 hyperscaler capex thesis, this cycle the fight moved inside the companies (OpenAI's own compute chief vs. IBM's customers) and into the bond market.

Bull, the buyer says demand is insatiable, the spend is converting to revenue, and the selloff was mechanical. The single strongest bull fact this cycle is that it came from the demand side's own operator: OpenAI's Sachin Katti says demand "far outstrips" supply, everything built is "consumed immediately," and the company "tripled compute and tripled revenue", his only fear is building too slowly (The MAD Podcast, Jul 16). The chip-stock drawdown, per Jack Kuznicki, was "leverage driven, not fundamentals," and Michael Burry's GPU-depreciation bear case "has just not been true" for seven months, the chips keep holding value (Full Signal, Jul 16). And even the epicenter, memory, is cheap on the bull math: Micron up ~350% year-on-year but on a forward P/E of 6-7, meaning the market already assumes a profit collapse (Buy Hold Rant, Jul 16).

Bear, the spend is a cyclical top funded increasingly with borrowed money, and IBM just showed the over-ordering. The freshest bear evidence is a real corporate print, not an opinion: IBM missed and fell ~25% because customers hoarded servers and memory ahead of price hikes and delayed everything else, the "double and triple ordering" that Ben Emons warns ends in an inventory glut, a "canary in the coal mine" (RiskReversal, Jul 15). Peter Boockvar's memory-as-1999 framing (85% gross margins, ~10x sales, Chinese CXMT/Yangtze coming for the margin) says the fattest part of the trade is the most cyclical and the most exposed (Wealthion, Jul 15). And the funding stage flagged last week deepened: ~40% of all net new debt issuance now feeds this build, pushing required yields to 6-7% (12%+ for weaker credits) and compressing every equity valuation in the complex (Something More, Jul 16).

"The bubble is clearly on the CapEx side, not necessarily the technology." (Peter Boockvar, One Point BFG Wealth Partners, on Wealthion, Jul 15)

The synthesis. Hold both at once: the demand looks genuinely real (the largest buyer says it can't get enough, and revenue is tracking compute), while the financing looks genuinely fragile (more of the build now runs on debt at rising yields, and IBM shows customers straining their own budgets to secure supply). That's why the sell signals below are about funding, ordering behavior, and one specific corporate tell, not about token counts.

Sell signals to watch: a second name after IBM reporting that AI-hardware hoarding is crimping its other spend (the over-ordering spreads); memory pricing rolling over as CXMT/Yangtze supply lands (Boockvar's China tell); AI-infrastructure bond spreads widening or a failed/expensive raise (funding stress); any hyperscaler guiding flat or slowing capex on the late-July calls (still the master trigger); and Nvidia GPUs finally showing falling resale/rental value (the first real support for Burry's depreciation case).

Stocks in play

NVDA. Bull: the contrarian pick of the cycle, "the one theme that has not worked really is Nvidia," which is precisely why the macro strategist calls it "one of the stocks I'm the most bullish on," and OpenAI's insider view (demand insatiable, everything consumed immediately) underwrites the whole order book. Bear: its customers' own custom chips remain the structural threat, and it was a laggard even on the up days this week (it "watched from the sidelines" of the Jul 15 large-cap rally). Next: the read-through from TSMC's Jul 16 print, then Nvidia's own results in August. (Full Signal, Jul 16; Stock Market Today With IBD, Jul 15)

AVGO. Bull: still owns the lion's share of the custom-ASIC market that every hyperscaler needs to build its own Nvidia alternative. Bear: the freshest signal is negative, Broadcom's weak June 3 guidance is what started the whole interconnect/memory selloff, and Kuznicki's jab lands: a company that "designs chips for other people… is basically a consultant, and consultancies shouldn't trade at 60 times earnings." Next: custom-ASIC design-win cadence into the next print. (Full Signal, Jul 16)

AMD. Bull: the industry still wants a credible second source to Nvidia, and AMD's one-year run (~+274%) shows the market keeps paying for that optionality. Bear: mostly on the sidelines again, Kuznicki is a self-described "AMD skeptic" who simply "likes Nvidia so much more," saying the top chip-design talent sits at Nvidia and AMD but he'd rather own the leader. Next: quiet on the podcasts; watch for the next AMD AI event or an MI-series design win to reset the story. (Full Signal, Jul 16; Daily Stock Picks, Jul 15)

MSFT. Bull: one of the five anchoring the ~$800B build, and the hyperscaler pushing back hardest, it raised some consumer prices in direct response to memory-cost inflation rather than eating the whole hit. Bear: it's a clear laggard (down ~22% over the past year per the large-cap trackers), and it sits squarely in the funding squeeze if capex keeps climbing toward $1 trillion. Next: the FY26 Q4 capex commentary in late July, still the single guide that could move the whole complex if it comes in flat. (Full Signal, Jul 16; Daily Stock Picks, Jul 15)

GOOGL. Bull: the standout, the best-performing megacap of the past year (~+95%, roughly a double) with the only credible in-house alternative to Nvidia in its TPUs, and it just picked up a second win as Apple's chosen AI engine (Gemini on the iPhone), a low-cost way to extend its reach. Bear: it too is now funding growth with outside capital, and even Google is scrambling for power and sites. Next: late-July capex guide; any concrete move to sell TPUs externally. (Daily Stock Picks, Jul 15; Something More, Jul 16)

AMZN. Bull: a real second path to Nvidia via its own Trainium chips, and it was among the winners in this week's large-cap rally (it gapped above its 50-day average). Bear: same funding squeeze and same power gate as the group, and the $800B-in/$110B-out payback math applies squarely to its spend. Next: late-July earnings and the AWS capex guide. (Stock Market Today With IBD, Jul 15)

META. Bull: the retail bulls call it the most undervalued megacap after Nvidia, a forward P/E around 21 while growing faster than the group, and it's committed to keep spending and ship its own chip by September. Bear: it remains the ambiguous name, its plan to sell excess computing power is still read as a possible "over-ordered" tell, and its capex sits in the same debt-funded squeeze; note too that New York just became the first state to ban new data-center buildout, a reminder the siting backlash is real. Next: late-July earnings, watch the capex guide and whether "sell excess compute" is confirmed. (Buy Hold Rant, Jul 16)

Read-throughs

  • Power / thermal, the constraint, now confirmed from inside the biggest buyer. OpenAI's compute chief made "power, not chips" official: OpenAI is funding new generation and transmission wherever it builds, and where the grid is full it's going "behind the meter" with on-site gas turbines, but it's "bottlenecked" by turbine supply, and it wants nuclear that "can't come soon enough" (The MAD Podcast, Jul 16). The most actionable read for equipment: cooling is now pervasive, chips, cables, and even power transformers all need liquid cooling, and better cooling directly "buys" more compute (more memory bandwidth, more flops). That's a clean tailwind for the thermal-management complex (Vertiv, and the liquid-cooling supply chain), even though VRT and Eaton drew no name-specific commentary this cycle. Watch, too, the siting backlash, New York's new-data-center ban is the first state-level "no."

  • Memory, the epicenter, and the debate is valuation and China, not scarcity. Everyone agrees supply is tight into 2027-2028; the fight is whether the fat margins survive. Bear anchor: Micron's 85% gross margin and ~10x sales evoke 1999-2000, and Chinese entrants CXMT (fourth-largest DRAM maker, IPO'ing in Hong Kong) and Yangtze Memory "are coming for your gross margin" (Wealthion, Jul 15). Bull anchor: Micron's ~350% year-on-year revenue growth at a 6-7 forward P/E already prices a collapse (Buy Hold Rant, Jul 16). Two supporting prints: SK Hynix's $26.5B U.S. ADR listing (offered at $149, +13% day one, then "sunk hard" the next day, 7x oversubscribed) (DHUnplugged, Jul 15); and, a genuine risk-management datapoint, CoreWeave is "eyeing derivatives to hedge memory chip price risk," i.e. memory prices are now volatile enough to hedge like a commodity (Daily Stock Picks, Jul 15). And memory is leaking into the macro: Emons argues it's a live driver of "AI inflation" via the memory/information-processing lines of the CPI (RiskReversal, Jul 15). Most actionable: respect the tightness, but treat the fattest margins as the most cyclical, and put CXMT's IPO on your calendar.

  • Networking / optics, the relative-strength standouts. With interconnect names hit since Broadcom's June 3 guide, the two connectivity chip designers still holding up are Astera Labs ($62B market cap, holding its 50-day average while peers broke down, "outstanding fund ownership") and Credo, up ~140% over the past year at ~$240 and called "too cheap" (Stock Market Today With IBD, Jul 15; Daily Stock Picks, Jul 15). The contrarian pair to them is Marvell/Broadcom, where the "consultant multiple" skepticism now lives (Full Signal, Jul 16). Most actionable trade: the market is separating the connectivity leaders (ALAB, CRDO) from the merchant-ASIC designers (AVGO, MRVL), respect the divergence.

  • Semiconductor equipment, the "picks-and-shovels" hideout. Kuznicki's highest-conviction area is the equipment makers, because they keep earning through a downturn: Lam Research books ~30% recurring revenue servicing its machines, and ASML, which reported this week with volumes guided up ~30%, sells the only tools that make the smallest chips (Full Signal, Jul 16). Most actionable: if you want AI-capex exposure with a softer landing if the cycle turns, this is where the recurring revenue lives.

  • Utilities / nuclear, supportive but name-quiet. OpenAI's "nuclear can't come soon enough" is the clearest demand signal for the reactor theme, but no name-specific Vistra/Constellation/Talen commentary surfaced this cycle. Carry the positions; the demand narrative is intact, the fresh catalyst is not.

What changed vs last issue

Yesterday's issue (Jul 16, "AI capex now drives 45% of GDP growth") was about the forecasts going up even as the mood stayed sour, Carson's $470B to $740B revision, the token-index scare defused at the source, and a16z's "wall of money still to come." Today the conversation shifted from forecasts to evidence, and it cut both ways:

  • The demand case got a face and a title. Last cycle the "demand is fine" argument came from analysts (Silicon Data, a16z). Today it came from OpenAI's own Head of Industrial Compute, on the record: can't build fast enough, everything consumed immediately, compute and revenue tripling together. That's the strongest possible version of the bull case, from the buyer, not a bystander.

  • The bears finally got a hard corporate print: IBM. Every bearish item in prior issues was a podcast opinion or a market wobble. IBM is a real ~25% earnings-day drop, with management naming the mechanism, customers hoarding servers/memory ahead of price hikes and delaying software. This is the first concrete sign of the "double/triple-ordering to inventory glut" the bears have theorized. New, and material.

  • The selloff got re-diagnosed. Yesterday framed the June/July chip weakness as fear becoming price. Today Kuznicki reframes it as leverage unwinding (starting from Broadcom's June 3 guide), with fundamentals intact, and notes Burry's GPU-depreciation thesis still isn't showing up in the data after seven months. Conviction that memory is fundamentally cracking should tick down; conviction that it's positioning-driven should tick up.

  • The memory fight narrowed to valuation and China. New vs. yesterday: Boockvar's explicit 1999-2000 analog (10x sales) and the CXMT (Hong Kong IPO) / Yangtze Memory margin threat, a China angle the prior issue didn't carry. The bull counter (Micron ~350% YoY at a 6-7 P/E) is the mirror image.

  • The funding-stress theme hardened into a number. Last issue: "cash printers reaching for outside money" (Alphabet's ~$85B equity, Meta's share sale). Today: Boyd's "~40% of all net new debt issuance" feeds this build, with required yields at 6-7% (12%+ for weaker credits), a sharper, bond-market-specific version of the same worry, plus Schiff's depreciation framing (2-3 vs 5-6 year GPU life).

  • Deltas on the big numbers: 2026 hyperscaler capex framed at ~$800B (Boyd), consistent with Carson's ~$740B and SaaStr's $688B from yesterday; 2027 still ~$1T. New this cycle: OpenAI ~$50B of compute this year inside a ~$700B industry (Katti); IBM's ~$700M miss and ~25% drop; SK Hynix's $26.5B ADR (offered $149); Micron revenue $41B last quarter heading above $50B (+350% YoY); BofA survey, 83% of managers call AI the most crowded trade. TSMC reported Jul 16 (the "first hard number" we flagged), but no podcast in this window covered it, a coverage gap, not a signal; watch for reaction commentary in the next sweep. And still: no hyperscaler has guided capex flat, the master trigger remains unpulled into late-July earnings.

Next prints to watch: late-July hyperscaler earnings (MSFT, GOOGL, AMZN, META), where the one question that matters is whether any of them dares to guide capex flat, plus any second company after IBM reporting that AI-hardware hoarding is crimping its other spend, and CXMT's Hong Kong IPO as the first live read on Chinese memory supply.