Newsletter · · Ashutosh Agarwal

Memory Selloff Tests the AI Trade as Hyperscalers Turn to Bond Markets - The AI Capex Tracker - Week of July 15, 2026

AI capex and semiconductors newsletter for the week of July 15, 2026. A record one-day drop in SK Hynix led a memory-driven selloff across AI chips, as bulls led by Gavin Baker argued there are no dark GPUs while bears zeroed in on doubling depreciation and a fresh wave of AI bond issuance from Nvidia, Amazon and others.

The AI Capex Tracker

Week of July 15, 2026: Memory Selloff Tests the AI Trade as Hyperscalers Turn to Bond Markets

TL;DR

  • The fear finally turned into a real selloff, and it started in memory. South Korea's market tumbled Monday, SK Hynix had its single worst day on record, and the damage spilled into US-listed chip names, all just the second trading day for SK Hynix's brand-new US shares, which had set a first-time listing record on Friday. The three big memory-and-foundry names (TSMC, Samsung, SK Hynix) are now worth a combined ~$4.4 trillion and carry roughly the same weight in emerging-market indexes that the Magnificent Seven carry in the S&P 500, so when they wobble, everyone feels it. (Bloomberg Tech, Jul 14; Squawk on the Street, Jul 13)

  • The bulls came out swinging with the best rebuttal we've heard this cycle. Gavin Baker, who lived through the 2000 crash, put it simply: that bust was defined by "dark fiber", cable laid in the ground and never switched on, 97% of it at the peak, whereas today "there are no dark GPUs." Every chip that gets built is running flat out. He also notes Nvidia trades near 40x earnings versus Cisco's 150-180x in 2000, and the biggest AI spenders have actually raised their returns on capital since the spending began. (The a16z Show, Jul 14)

  • But the bear chorus got louder and more specific, and it has moved to the balance sheet. Veteran tech bear Fred Hickey laid out an "A.I.-mageddon" case built on depreciation and collapsing prices, and multiple shows zeroed in on the same new tell: the biggest, richest companies on earth are now borrowing to fund the build. Nvidia, SpaceX and Amazon have issued roughly $75 billion of bonds in just the past few weeks, and one estimate floating around is up to $10 trillion of AI-related debt over the next several years. When the cash-printers start passing the hat, that's a stage change. (Thoughtful Money, Jul 14; Rebel Capitalist News, Jul 14)

What's New

A note on the window first. This is a Wednesday recap, so it sweeps roughly the last day (Tuesday, July 14), plus the Monday selloff that dominated it. And unlike yesterday's thin, macro-heavy day, this one had a clear headline: the market finally sold the AI trade, memory led it down, and the bulls and bears both showed up with their sharpest arguments in weeks. Ranked by where the dollars and the risk actually sit:

1. The selloff arrived, and it started where the bulls always said the crack would show, memory. Bloomberg Tech, Jul 14, host Ed Ludlow with Bloomberg infrastructure reporter Dina Bass and Swissquote senior market analyst Ipek Ozkardeskaya. Squawk on the Street, Jul 13 for the origin.

Here's what happened. On Monday, South Korea's market fell hard, SK Hynix, the company that pioneered the special "high-bandwidth memory" (HBM) chips that sit next to Nvidia's GPUs, dropped about 15% in Seoul, its biggest one-day fall ever, with Samsung sliding too. That weakness carried into the US session, hitting SK Hynix's newly listed American shares (down about 6%) and dragging the whole AI-chip group with it. The timing is what makes it sting: SK Hynix had only just set a record for a first-time US share sale on Friday, pricing at $149; by its second day of US trading, investors were already heading for the exits. Swissquote's Ipek called it plainly, a classic "buy the rumor, sell the fact," with "a correction underway" and downside risks that "prevail as long as the volatility remains this high."

Why this is the story and not just a bad day: the reason memory ran so hard is that it looked like it had stopped being a boring, up-and-down commodity business. As Bass explained, customers who "need them desperately" have been signing multi-year supply deals, "a new thing in that business", and SK Hynix's own chairman told Bloomberg, "I have some confidence that the demand will grow. Our supply capacity never going to catch up." That's the bull case in one sentence. Monday was the market asking whether it believes it. And there's a real read-through for everyone: these three chip names (TSMC, Samsung, SK Hynix), now worth ~$4.4 trillion together, "have the equivalent weighting in emerging markets indices as the Mag 7 does in the S&P 500." Big funds looking to "rotate out" of them is exactly how a memory wobble becomes a market-wide one.

"Investors again questioning whether the rally has simply run too far, too fast." (Ed Ludlow, Bloomberg Tech, Jul 14)

One more genuinely useful data point buried in the same show: TSMC's June sales rose 68% year-on-year, and the Street's math implies the June quarter was up about 36%. TSMC reports Thursday (Jul 16), the next real check on whether the demand is still there or the doubters are early.

2. The best bull rebuttal of the cycle: "There are no dark GPUs." The a16z Show, Jul 14, Gavin Baker, managing partner and chief investment officer of Atreides Management, interviewed by a16z's David George.

If you only listen to one thing this week to steel-man the long side, it's this. Baker invested through the 2000 crash and makes the cleanest historical contrast yet. That bust, he argues, was really a telecom bust defined by "dark fiber", fiber-optic cable laid in the ground and never lit up because it was useless without the gear on either end. "At the peak of the bubble, 97% of the fiber that had been laid in America was dark." Today, he says, "there are no dark GPUs", read any technical paper and the problem is GPUs "melting" from being run so hard. In plain terms: in 2000 the industry built capacity nobody used; today the capacity is maxed out the moment it's plugged in.

His other points are worth writing down. On valuation: Cisco peaked at 150-180x its trailing earnings in 2000; Nvidia is "more like 40 times." On whether the spending is paying off: "the return on invested capital of the biggest spenders on GPUs, who are all public," has risen "call it a 10-point increase" since they ramped up, "there's no debate that thus far the ROI on AI has been really positive." He sizes the buffer, too: the hyperscalers together throw off roughly $300 billion of free cash a year and sit on ~$500 billion of cash, versus roughly $40-50 billion to light up one gigawatt of Nvidia-based data center. And he flags where he thinks the real casualties will be, not Nvidia, but the custom in-house chips: "In the next 3 years, I think you'll see a bunch of high-profile ASIC programs canceled, especially if Google starts selling TPUs externally." Why it matters: this is a named, credible investor directly rebutting the exact bubble the bears are calling, on the same day the bears were loudest, and his ASIC warning is a specific, bearish read-through for the custom-silicon trade (Broadcom, and the hyperscalers' own chips).

3. Fred Hickey's "A.I.-mageddon": the profits are an illusion that depreciation is about to expose. Thoughtful Money, Jul 14, Fred Hickey, longtime tech skeptic and author of the High-Tech Strategist newsletter, with host Adam Taggart.

Hickey has watched tech cycles for "nearly 5 decades," and his case is the most detailed bear argument of the sweep. The core idea is an accounting mismatch. Suppliers like Micron book the revenue now, Micron went from negative gross margins in 2020 and 2023 to 85% today, and SanDisk is up 700% this year, while the buyers (the hyperscalers) spread the cost of those chips over five to six years as depreciation. So supplier earnings look spectacular and hyperscaler costs look small, "and that has created this gigantic gap." The bill is now starting to arrive: Hickey says Google's depreciation expense is set to roughly double, from about 17% to 35% of revenue by 2028.

He piles on evidence that the demand side is already cracking on price. He cites Chamath Palihapitiya asking his own CTO on a podcast what the return on token spending was, answer: "5% max," while "costs are doubling every 45 days." He points to Microsoft pulling most of its Claude Code licenses, Tesla putting $200 weekly limits on employee AI use, and Coinbase halving its AI budget, all shifting to cheaper models. The scariest number for American model-makers: Chinese open-source models have jumped to about 46% of usage share (per OpenRouter), up from an 11% average over the prior year, with DeepSeek delivering "90% of your tasks at 1.5% of the cost." He notes Meta is cutting prices 75%, and that only Google and Microsoft are currently free-cash-flow positive among the big spenders. Why it matters: this is the bear case with receipts, it names the companies cutting spend, quantifies the Chinese price threat, and gives a concrete date (the 2028 depreciation ramp) for when the reported profits stop flattering the picture.

4. The stage change everyone noticed: the cash-printers are now borrowing. Rebel Capitalist News, Jul 14, host George Gammon (informed pundit, not an operator). The Peter Schiff Show, Jul 15, Peter Schiff (pundit).

Two skeptic-side shows landed on the same, genuinely new signal, and it's worth separating from the general doom. For years the AI build was funded out of the hyperscalers' own enormous profits. That era is ending: free cash flow across the group has turned, in Gammon's words, "steeply negative," so the companies are now tapping the bond market. He flags a Wall Street Journal figure, a "quarter-trillion-dollar onslaught of AI bonds", and notes Nvidia, SpaceX and Amazon issued roughly $75 billion of bonds in just the last few weeks, with high-end estimates of up to $10 trillion of AI-related debt to come. His blunt framing: demand isn't the question, the economic model is, "if you have a business where you're selling $20 bills for $10, you're going to have a lot of demand."

Schiff sharpens the accounting angle the bulls dislike most: because AI spending is booked as capital expenditure, it doesn't dent the earnings number Wall Street watches, but GPUs "become obsolete very quickly," with a real useful life "more like two to three years" versus the five-to-six-year depreciation schedules companies use. If that's right, the spending is less like building a railroad and more like an ongoing operating cost that never stops. His warning is that these firms went from being lenders (Meta and Google parking cash in Treasuries) to massive borrowers, which "is also what's pushing up interest rates, crowding out the market," until "at some point, the bond market is going to break." Why it matters: treat the pundit theatrics with a grain of salt, but the underlying fact, the strongest balance sheets on earth are now issuing debt to keep the build going, is real, corroborated across shows, and is the cleanest new sell-signal to monitor.

5. Where's the return? A sober, data-heavy look at AI's ROI problem. AI to ROI, Jul 14, Ray Rike (Benchmarket) and Peter Buchanan (The New Plan). Tom Bilyeu's Impact Theory, Jul 14, reacting to bear analyst Ed Zitron.

Underneath the market drama, the more important long-run question is whether customers are actually getting their money's worth, and the data is not encouraging yet. Rike and Buchanan cite a Ramp report showing average monthly AI token spend across ~50,000 companies has risen 13x since January 2025, while only 27% of enterprise executives say AI has met their return expectations. Their case study is the tell: Uber rolled Claude Code out to ~5,000 engineers in December, adoption jumped from 32% in February to 84% in March, and the company "consumed their entire annual budget by April," with the COO reporting no measurable customer benefit or faster product releases. They add that only 18% of coding token spend "translates into shipped products that reach actual users," and that one company reportedly spent $500 million in a single month on tokens from one vendor.

On the more provocative end, Bilyeu reacted to Ed Zitron's claim that AI is "a $10 billion industry pretending to be a trillion dollar one." Zitron's specifics are the useful part: OpenAI "burned $20.9 billion in 2025" (per audited financials the FT reported), its costs rise in line with revenue with no proven path to better margins, and Oracle is "building 7.1 gigawatts of capacity just for one customer" (OpenAI) while flagging in its own annual report the risk it "might not get paid." Notably, Bilyeu himself disagreed with Zitron's doom, he thinks the technology is genuinely transformative, especially in medicine and coding, which makes the segment a fair fight rather than a one-sided rant. Why it matters: this is the demand-side version of the bear case. The spending is real; the measurable payback, so far, mostly isn't, and that's the gap late-July earnings need to start closing.

The Debate

Steel-manning both sides of the $700B-plus 2026 hyperscaler capex thesis, on the day the market actually took a swing at it.

Bull, this is nothing like 2000, the chips are all in use, and the returns are real. Gavin Baker's frame is the anchor: no dark GPUs (versus 97% dark fiber at the 2000 peak), Nvidia at ~40x versus Cisco's 150-180x, and a ~10-point rise in the big spenders' returns on capital since the build began, "thus far the ROI on AI has been really positive" (The a16z Show, Jul 14). The buffer is enormous, ~$300 billion of annual free cash flow and ~$500 billion of cash across the group. Semiconductor specialist Ben Pouladian adds that this has been an earnings-driven move, not a valuation bubble: Nvidia's revenue went from $27 billion in 2023 to ~$250 billion over the last twelve months, yet its forward multiple actually trades "at a discount to its past 5 or 10 years," and "intelligence" isn't a commodity the way telecom bandwidth was, so the Cisco comparison misses (Monetary Matters, Jul 14). And the memory bulls have a genuine structural argument: multi-year supply contracts are new to a business that used to be purely cyclical, and supply "never going to catch up" with demand (Bloomberg Tech, Jul 14).

Bear, the profits are flattered, the demand is subsidized, and the cash is running out. Fred Hickey's accounting case is the sharpest: supplier earnings are inflated because buyers depreciate the chips over 5-6 years, and that hidden cost is about to surface (Google's depreciation roughly doubling to ~35% of revenue by 2028), while demand is propped up by prices being cut toward zero and Chinese models taking ~46% usage share at 1.5% of the cost (Thoughtful Money, Jul 14). The funding case is new and corroborated: free cash flow has gone "steeply negative" and the group is now issuing tens of billions in bonds ($75B in recent weeks from Nvidia, SpaceX and Amazon), which pundits warn will keep pushing rates up until "the bond market is going to break" (Rebel Capitalist News, Jul 14; The Peter Schiff Show, Jul 15). The ROI case rounds it out: token spend up 13x since early 2025 but only 27% of executives seeing the returns they expected, and Uber burning a whole year's AI budget in four months with nothing measurable to show (AI to ROI, Jul 14). Macro strategist Marco Papic adds the "second-derivative" point: even if capex keeps rising, its growth rate is slowing, from roughly $200B to $400B to $800B to $900B, and physical limits (copper, steel, electricity, labor) mean "you cannot actually build $3 trillion worth of data centers" (Stansberry Investor Hour, Jul 14).

"There are no dark GPUs." (Gavin Baker, Atreides Management, on The a16z Show, Jul 14)

The single best synthesis came from Jason Ware, the chief investment officer at Albion, who is neither a permabull nor a permabear. On memory, he warned: it's arguably still a commodity business, and "you want to buy commodity type stocks when the PEs are high and sell them when the PEs are low", Micron trading at ~9-10x forward earnings is the market telling you it doesn't fully believe these are durable, moaty businesses. His trigger to watch is the one that matters most: "If Microsoft gets on their next earnings call, Satya Nadella tells everybody, hey, our CapEx is flat for the next quarter… that's going to get really interesting." Given the huge run these stocks have had, "it doesn't even have to be a pivot. It just has to be a slowdown" for them to get hit hard (The Rundown, Jul 13).

Sell signals to watch: any hyperscaler signaling flat or slowing forward capex on the late-July calls (Ware's trigger); the capex-to-revenue gap failing to close as those results land; AI-bond spreads widening as issuance climbs; memory contract prices rolling over (Pouladian pencils in a "freak out" around mid-2027 as new supply lands); Chinese model share continuing to climb; and Google's depreciation ramp starting to bite the reported numbers.

Stocks in Play

NVDA. Bull: the cleanest anti-bubble case of the cycle, "no dark GPUs," ~40x earnings versus Cisco's 150-180x, and rising returns on capital at its customers; revenue $27B (2023) → ~$250B (last twelve months) with the multiple actually below its own history. Bear: it got caught in Monday's chip rout, its costs are said to rise in line with revenue, and, new this cycle, even Nvidia is now issuing bonds (reportedly ~$20B, for liquidity) as part of the broader debt wave. Next: Q2 earnings in August; TSMC's print Thursday is the near-term read-through. (The a16z Show, Jul 14; Monetary Matters, Jul 14)

AVGO. Bull: Broadcom is the enabler letting Meta and others build Nvidia-alternative systems, "going to market together" with AMD on open-standard Ethernet fabrics. Bear: Baker's warning lands squarely here, "a bunch of high-profile ASIC programs" could be canceled in the next 3 years, "especially if Google starts selling TPUs externally," which would shrink the custom-silicon opportunity Broadcom is priced for. Next: custom-ASIC design-win cadence; watch for any hyperscaler cooling on its in-house chip plans. (The a16z Show, Jul 14)

AMD. Bull: "AMD will always be kind of the second source and you need a second source", a structural role that survives even if custom chips disappoint. Bear: same ASIC-skepticism cloud as Broadcom, since AMD is also the plug-in behind Broadcom's fabric pitch. Next: AMD's AI event in San Francisco "in a few weeks" (Advancing AI Day) is the catalyst to reset expectations. (The a16z Show, Jul 14; Monetary Matters, Jul 14)

MSFT. Bull: Jason Ware thinks the fear driving it down, that "AI is going to eat software", is "absurd on its face" given how entrenched and moaty its enterprise business is. Bear: it's being lumped into the software de-rating, its once-prized OpenAI tie has "become a liability," it's pulled most of its Claude Code licenses, and Hickey pegs it down ~20% on the year, one of the worst-hit of the Mag 7. Next: FY26 Q4 capex commentary, late July, and it's the name Ware says would rock the whole market if it guides capex flat. (The Rundown, Jul 13; Thoughtful Money, Jul 14)

GOOGL. Bull: Baker's favorite structural position, "arguably the leading AI company today," with Gemini taking "15 or 20 points of traffic share in the last 2 or 3 months," and the TPU is "by far… the only alternative to Nvidia for training." If Google starts selling TPUs externally (Anthropic is rumored to want "tens of billions"), it reshapes the chip market. Bear: Hickey's depreciation math bites hardest here, expense rising from ~17% to ~35% of revenue by 2028, and its capex is increasingly funded with outside capital. Next: July capex guide. (The a16z Show, Jul 14; Thoughtful Money, Jul 14)

AMZN. Bull: Baker calls its Annapurna group "arguably the most talented silicon team at any hyperscaler," and expects Trainium 3 to be a big step up, a real second path to Nvidia. Bear: it's squarely in the new bond-issuance wave (part of the ~$75B in recent weeks), and Hickey says Amazon is spending close to 100% of its cash flow on data centers, with a reported $500M one-month token bill as a cost-discipline red flag. Next: late-July earnings. (The a16z Show, Jul 14; Rebel Capitalist News, Jul 14)

META. Bull: its stock soared last week as the market set up for a bigger capex budget; its ~5-gigawatt, ~$50B data-center shell (partnered with Entergy) shows it can secure the power, and it's leasing its own compute rather than trying to become a cloud provider. Bear: Hickey flags Meta cutting AI prices 75% and, like Amazon, spending near 100% of cash flow on the build; it will likely need to tap the equity market to keep funding it. Next: late-July earnings; watch the capex guide and the depreciation footnote. (Squawk on the Street, Jul 13; Thoughtful Money, Jul 14)

Read-throughs

  • Memory (Micron, SK Hynix), the epicenter this cycle, and now cuts both ways. The bull structural case is real (multi-year contracts, supply that "never going to catch up"), but Monday showed how violently these trade, SK Hynix down ~15% in a day, its new US shares selling off in their second session. Jason Ware's discipline is the one to hold: it may still be a commodity, so a high multiple is a warning, not comfort, and Micron at ~9-10x forward earnings says the market doubts the durability. Pouladian pencils in a memory-price "freak out" around mid-2027 as new supply lands. Most actionable: respect the structural demand but size for the volatility, and treat any hint of a capex slowdown as the trigger. (Bloomberg Tech, Jul 14; The Rundown, Jul 13; Monetary Matters, Jul 14)

  • Power / thermal (VRT, ETN) and on-site generation, still the cleanest physical read. Energy Innovation's Robbie Orvis put hard numbers on the bottleneck: order a new grid-scale gas turbine from GE and "they'll say, OK, see you in seven, eight years." So data centers are grabbing "old jet engines, basically anything they can find", he lives in Northern Virginia and describes neighbors "breathing the diesel generator fumes" because developers are "searching for any power they can find" and "the cost is not an issue." His firm models wholesale power rates rising ~74% by 2035, with PJM interconnection queues taking ~7 years. That points straight at fast, on-site power (turbines, gensets, and fuel-cell challengers, Pouladian flagged Bloom Energy for ~80-day deployments) plus the power-delivery and cooling gear between the meter and the chip. Most actionable: stay long the speed-to-power enablers; treat grid-dependent, permitting-heavy projects with caution. (Renewable Rides, Jul 14; Monetary Matters, Jul 14)

  • Networking / optics (Marvell, Astera Labs, Credo, Coherent, Lumentum). One useful thread from Pouladian: as memory prices bite, "memory pooling" (a technology called CXL) is getting real attention, Astera Labs "was left for dead and then shot up 400%," with Credo and Marvell in the same space. Optics remains "interesting" for moving data faster between GPUs, but Coherent and Lumentum specifically were quiet this cycle. Carry the positions. (Monetary Matters, Jul 14)

  • Custom silicon / ASICs, a fresh caution flag. Baker's call that "a bunch of high-profile ASIC programs" get canceled in the next three years, especially if Google sells TPUs externally, is a direct challenge to the custom-chip trade (Broadcom and the hyperscalers' in-house efforts). Read-through: the safest chip exposure may still be the merchant leader (Nvidia) plus the genuine second source (AMD), with the bespoke-ASIC names carrying more roadmap risk than the market assumes. (The a16z Show, Jul 14)

  • The bond market, the new thing to watch. The clearest new risk isn't a stock, it's a spread. With hyperscaler free cash flow turning negative and the group issuing tens of billions in bonds (~$75B in recent weeks; high-end estimates of up to $10T over several years), widening AI-bond spreads would be an early, market-wide warning. (Rebel Capitalist News, Jul 14; The Peter Schiff Show, Jul 15)

What Changed vs Last Issue

Yesterday's issue (Jul 14, "Eisman: buy the suppliers, not the AI giants") was about the abstract bear case getting scarier, Apollo's Torsten Slok telling Steve Eisman the economy now rides on AI capex, hyperscaler cash flow is heading toward zero, and the spenders may have no moat. Today, that abstraction turned into price action, and the two sides both sharpened:

  • The fear became a selloff. Yesterday it was a thesis; today the market actually sold the AI trade, led by memory. SK Hynix's record one-day drop and the Korea rout spilling into US chip names is the first hard, market-wide crack we've logged, and it hit the exact spot (HBM/memory) the bulls always said would show strain first. This is a real delta from a week of talk.

  • The bulls finally landed a heavyweight counter. Yesterday's bull material was second-hand (factor and credit lenses). Today Gavin Baker, a named investor who traded the 2000 crash, delivered the cleanest anti-bubble frame yet ("no dark GPUs," 40x vs 150-180x, rising ROIC), and Ben Pouladian showed this has been an earnings-driven move with Nvidia's multiple below its own history. Conviction that a break is imminent should actually go down on this, not up.

  • The bear case moved from the income statement to the balance sheet. Yesterday it was Slok's "free cash flow toward zero." Today that shows up as actual behavior: a "quarter-trillion-dollar onslaught of AI bonds," ~$75B issued in recent weeks by Nvidia, SpaceX and Amazon, even Nvidia borrowing. The debt stage is the classic late-cycle tell, and it's now visible.

  • Power got quantified again. Yesterday Quinbrook's "bring your own power." Today Robbie Orvis put a number on the turbine bottleneck (7-8 year GE wait, forcing old jet engines and diesel) and on the cost to consumers (~74% higher wholesale rates by 2035, ~7-year PJM queues), reinforcing the behind-the-meter/speed-to-power trade.

  • Deltas on the big numbers: 2026 hyperscaler capex still anchored at ~$750-755B (Hickey's ~$750B by the five biggest plus ~$100B outside is consistent with yesterday's ~$755B). New, sharp data points: Chinese models at ~46% usage share (up from ~11%); TSMC June sales +68% YoY (Q2 tracking ~+36%), reporting Thursday; ~$75B of fresh AI bonds. No hyperscaler has cut capex, Jason Ware's "flat guide" trigger has not been pulled.

  • On the Texas SB6 deadline: the $40,000/MW large-load queue deposit deadline flagged in prior issues falls today (Jul 15). It did not surface in any podcast this cycle, genuinely quiet on that specific item, so no fresh read. Worth checking the withdrawal numbers directly once ERCOT reports them; that remains the real-world test of how much of the Texas pipeline is buildable.

Next prints to watch: TSMC on Thursday (Jul 16) for the first hard data on whether chip demand is still accelerating; then the late-July hyperscaler earnings (MSFT, AMZN, META, GOOGL), and specifically whether any of them dares to guide capex flat, which is the trigger the whole market is now watching for.