Newsletter · · Ashutosh Agarwal

Samsung Out-Earns Nvidia as the Shovels Win the First Half - The AI Capex Tracker - Week of July 10, 2026

The AI Capex Tracker for the week of July 10, 2026. Samsung became the most profitable company on earth on the back of AI memory, the physical-layer suppliers crushed the Magnificent 7 in the first half, and the first cracks in AI demand appeared as Meta began renting out spare compute.

The AI Capex Tracker

Week of July 10, 2026: Samsung Out-Earns Nvidia as the Shovels Win the First Half

Issue: Friday, July 10, 2026

TL;DR

  • Samsung just became the most profitable company on earth, and it did it selling memory, not phones. Q2 profit hit $58.5B, edging past Nvidia's $53B, with roughly 94–96% of that coming from AI memory. Memory makers keep hiking prices (HBM up ~90% in Q1, another 50–60% in Q2, and a further ~20% teed up for Q3) because supply is short into 2030. (Limitless: An AI Podcast, Jul 9)
  • The picks-and-shovels crushed the spenders in the first half. Through six months, the S&P 500 was up ~9%, the other 493 names up over 14%, but the Magnificent 7 were up a paltry ~2%, with Microsoft, Meta and Tesla actually down. SanDisk led the whole index at +717%; the semiconductor index (SOX) rose ~74% while Nvidia rose just ~5%. (Stock Club, Jul 9; The Rundown, Jul 9)
  • The demand side finally flinched. Meta launched a cloud business to rent out its spare AI compute (the stock jumped ~10%, its best day in five months), which is exactly what a company does when it has more chips than it can use. At the same time, Palantir's Alex Karp went on CNBC warning that big companies increasingly doubt the return they're getting on AI. (The Twenty Minute VC, Jul 9)

What's New

The story this cycle isn't the chip designer, it's everyone downstream of it. Ranked by where the dollars and the risk actually sit:

1. Samsung out-earned Nvidia, and memory stopped behaving like a commodity. On Limitless: An AI Podcast (Jul 9), hosts Josh and Ejaz walked through how Samsung, a company most people still think of as a phone maker, just posted the most profit of any company in the world: $58.5B in Q2, ahead of consensus (~$55B) and ahead of Nvidia's $53B, with almost all of it from AI memory. That's up from $3.4B a year earlier, roughly $650M a day, or about $7,500 every second. The mechanism is a near-monopoly: SK Hynix owns about 60% of the market for high-bandwidth memory (HBM), the specialized memory stacked next to every AI chip, and the two Korean giants plus Micron are the only players at scale. So they can dictate price, HBM went up ~90% in Q1, another 50–60% in Q2, with a further ~20% planned for Q3, and it still isn't enough, because one gigabyte of HBM eats the factory capacity of about four gigabytes of ordinary memory. Gross margins tell the story: a grocery store keeps ~$3 of every $100, Apple ~$30, Samsung ~$52, and SK Hynix ~$72. New fab capacity doesn't arrive until ~2030, and the hosts peg demand at 3–5x supply over the next couple of years. (Limitless: An AI Podcast, Jul 9)

2. Storage is the second bottleneck, and SanDisk is the poster child. On Stock Club (Jul 9), the hosts ran the first-half scoreboard: the top ten S&P 500 performers were almost all physical-layer AI names, SanDisk +717%, Micron +274%, Western Digital +245%, Intel +230%, Seagate +222%, Dell +207%, Marvell +205%, Applied Materials +157%, Corning +153%, Flex +146%. SanDisk, which makes the solid-state drives that store the data AI models chew through, "literally cannot make them quick enough": a $42B backlog and 78% gross margins (up from 23% a year ago), with quarterly revenue jumping from $1.7B to $6B year on year and operating income going from $2M to $4.2B. All three HBM makers are sold out for 2026, and hyperscalers are "negotiating for access, not price." The catch, as the hosts flag: Amazon, Microsoft, Google and Meta do not want to keep paying 78% margins, so the pressure to find a workaround only grows. (Stock Club, Jul 9)

3. Meta blinked, it's now renting out its own compute. On The Twenty Minute VC (Jul 9), Harry Stebbings, Jason Lemkin and Rory dug into Meta launching "Meta Compute," a cloud business selling access to its AI infrastructure, hosted or raw GPUs by the hour, the same model as CoreWeave or Nebius. The market loved it: the stock rose ~10%, its biggest single-day gain in five months, while the neo-cloud incumbents (Nebius, CoreWeave) fell 10–15% on the new competition. But the read underneath is more double-edged: as one host put it, Meta and SpaceX have both now "buy a load of compute to build proprietary assets, fail to build those assets, and then decide instead to sell that compute to others." Renting out capacity is a great business, but it's also what you do when you have more than you can use, which cuts against the "infinite demand" story. (The Twenty Minute VC, Jul 9)

4. Nvidia is now the "cheap" AI stock, and it's financing its own customers to stay sold out. On The Rundown (Jul 9), Zaid Admani noted Nvidia has quietly shed about $1 trillion in market value in two months (down ~15% from its mid-May high) and now trades around 18x forward earnings, cheaper than the average S&P 500 stock ("freaking Hershey's has a higher forward PE"). It still owns 97% of the server-GPU market, and SpaceX's latest model, Grok 4.5, was trained on tens of thousands of Nvidia's newest GB300 chips; Wall Street's average target ($300) implies ~50% upside. The bear overhang: reports that its next-generation server racks are running a year behind schedule (Nvidia disputes this), and, per Mike Green on Full Signal (Jul 8), the worry that Nvidia is "manufacturing scarcity by financing its customers," the same vendor-financing that flattered Cisco's margins in 1999 before the write-downs came. On The Twenty Minute VC (Jul 9), the hosts described Nvidia's new "compute now, pay later" deals for neo-clouds: it books the hardware revenue upfront and hands the buyer put-back rights if they can't use the chips, legal under accounting rules, but "a derivative bet on keeping this thing going." (The Rundown, Jul 9; Full Signal, Jul 8; The Twenty Minute VC, Jul 9)

5. The customers are all trying to design their way off Nvidia. On AI Inside (Jul 9), Jason Howell and Jeff Jarvis covered two threads that matter for the chip designer's long-term moat: DeepSeek said it will build its own AI chip (as Nvidia's Jensen Huang has warned would happen if the US keeps restricting exports), and China is reportedly weighing walling off its best AI models from the rest of the world, a mirror image of US export controls. Separately, on The Twenty Minute VC (Jul 9), the hosts noted Anthropic has opened talks with Samsung to build its own chip. The pattern, every big buyer chasing custom silicon, is the slow-moving threat behind Nvidia's cheap multiple. (AI Inside, Jul 9; The Twenty Minute VC, Jul 9)

The Debate

This week the argument moved off the chip designer and onto Meta, a clean test of whether spending this much on AI is a smart bet or a bad one. In a genuine bull/bear debate on Pitch The PM (Jul 9), Avory CIO Sean Emory (long-time Meta owner) took the bull side against a hedge-fund host who thinks 2027 is the peak of hyperscaler capex.

Bull steel-man, the fundamentals are accelerating, and the build is funded. Emory's case: Meta is the second-fastest-growing name in the Magnificent 7 (behind Nvidia) with the highest gross margins, so "they've earned the right to invest." Advertisers are growing ~33%, the Meta AI app now sits inside the top 10 (just above Gmail), and, using outside real-time data, revenue is tracking 33–34% monthly growth versus consensus of ~27%, yet the stock trades at just ~22x earnings. The memory and storage world backs the "demand is real" view: HBM and SSDs are genuinely sold out into 2028–2030, and GE Vernova has a decade of turbine orders it physically cannot fill fast enough. (Pitch The PM, Jul 9; Limitless: An AI Podcast, Jul 9; Motley Fool Hidden Gems Investing, Jul 9)

Bear steel-man, the stock no longer trades on earnings, it trades on a leap of faith. The host's point on Pitch The PM is the one to sit with: Meta's stock now has a 90% correlation to 2027 free-cash-flow estimates, which have collapsed to roughly zero (maybe slightly negative) because Zuckerberg's AI bet, $60–80B a year, versus ~$20B for the old metaverse push, is eating the cash. Emory himself models capex eventually plateauing around $165B, still a year and a half to two years out. And the chatter of a possible Meta equity raise would, in his words, be read as a negative, a signal even more capex is coming.

The wider bear case got two named heavyweights. On RiskReversal Pod (Jul 8), Guy Adami and Dan Nathan read out Jim Chanos's new note ("The Second Derivative"), which argues the AI build-out is "a credit-driven real estate cycle hidden in tech clothing", funded like 2008, not 2000, with "manufactured demand" as hyperscaler capex loops back into money-losing labs like OpenAI and Anthropic as cloud credits, leaving Wall Street with concentrated credit exposure to pre-profit tenants. And on Full Signal (Jul 8), Mike Green tied it to market plumbing: passive 401(k) flows buying regardless of price, and triple-levered semiconductor ETFs (SOXL was down ~20% at one point) creating forced, self-reinforcing flows that can reverse violently. Meanwhile Alex Karp's CNBC warning, that enterprises increasingly question their AI ROI, went viral, and Palantir rose 9% on the day anyway. (Pitch The PM, Jul 9; RiskReversal Pod, Jul 8; Full Signal, Jul 8; The Twenty Minute VC, Jul 9)

As the host on Pitch The PM (Jul 9) put it: "The stock has 90% correlation to 2027 free cash flow revisions, which have essentially gotten to zero."

Sell signals to watch: any hyperscaler actually cutting forward capex (none has); more "excess compute" resale deals like Meta's cloud and SpaceX's rentals (the demand-blink tell); memory or storage gross margins rolling off their ~72–78% peak; Nvidia having to de-book prior revenue if a financed neo-cloud can't pay (its new put-back deals triggering); a Meta equity raise; and the rate of change of AI revenue growth flattening even as spend keeps climbing.

Stocks in Play

NVDA. Bull: still ~97% of the server-GPU market, trains everyone's frontier models (SpaceX's Grok 4.5 ran on tens of thousands of GB300s), and at ~18x forward it's cheaper than the average S&P 500 stock, with an average target near $300. Bear: ~$1T of value gone in two months, every major customer (DeepSeek, Anthropic) is designing its own chip, next-gen racks are reportedly a year behind (disputed), and margins are partly propped up by financing its own buyers. Next: Q2 earnings in August; June CPI and the start of Q2 season next week. (The Rundown, Jul 9; The Twenty Minute VC, Jul 9)

AVGO. Bull: the wave of hyperscalers building their own chips is arguably a bigger win for Broadcom (the custom-ASIC arms dealer) than for the companies doing it, Meta's own-silicon push included. Bear: if in-housing goes far enough, it eventually points at merchant ASICs too. Next: no fresh operator signal in this cycle's podcasts; watch custom-silicon design wins. (Motley Fool Hidden Gems Investing, Jul 9)

AMD. Bull: the Motley Fool Hidden Gems hosts noted Meta signed "the biggest AI deal in history" to buy AMD chips for its hardest AI jobs, a real merchant-GPU win. Bear: still no fresh roadmap color, and MI450X/Helios remains a show-me story on execution. Next: AMD Advancing AI Day, July 2026. (Motley Fool Hidden Gems Investing, Jul 9)

MSFT. Bull: balance sheet to fund anything, and it bounced off its late-March low of ~$350 back to ~$390. Bear: down year-to-date, the group's poster child for "spending a lot, showing little," and the recent software rally looks more like an oversold trade than a trend change. Next: FY26 Q4 capex, late July. (RiskReversal Pod, Jul 8; Stock Club, Jul 9)

GOOGL. Bull: one of only two Magnificent 7 names outperforming in the first half, with its own TPU chips and the ability to monetize compute directly on its own platform. Bear: the way its chips are now financed (a ~$35B private-credit fund raised to buy Google TPUs for customers) is Exhibit A in the "financialization" critique. Next: July capex guide. (Stock Club, Jul 9; RiskReversal Pod, Jul 8)

AMZN. Bull: up slightly in the first half, has its own Trainium silicon, and can fund the build almost at will. Bear: increasingly leaning on debt (a fresh ~$25B deal) to pay for capex that's growing faster than revenue. Next: late-July earnings. (RiskReversal Pod, Jul 8)

META. Bull: second-fastest-growing Magnificent 7 name behind Nvidia, highest gross margin, ~22x earnings, advertisers +33%, Meta AI now a top-10 app, and it just launched a cloud that the market rewarded with a ~10% pop. Bear: the stock now trades on 2027 free-cash-flow estimates that have gone to zero, capex is $60–80B a year headed toward a ~$165B plateau, a possible equity raise would be read as a negative, and renting out spare compute quietly undercuts the infinite-demand story. Next: late-July earnings; watch for any equity-raise signal. (Pitch The PM, Jul 9; Motley Fool Hidden Gems Investing, Jul 9; The Twenty Minute VC, Jul 9)

Read-Throughs

  • Power and grid, the constraint is energized megawatts, not announced ones. On The Data Center Frontier Show (Jul 9), AFCOM's Bill Klayman argued data-center operators are becoming "one of the largest unregulated utilities in the world" through behind-the-meter power (nuclear and gas turbines), and that the industry could draw upwards of 200 gigawatts by 2030, roughly "four Californias." His filter is the one that matters: "Announced megawatts are not the same as energized megawatts." On Switched On (Jul 9), BNEF analysts Natalie and Lara put Texas's large-load interconnection queue at close to 500 GW against an ERCOT peak of under 90 GW, with the new Senate Bill 6 requiring loads above 75 MW to prove they can flex, which pushes always-on data centers toward on-site generation. Most actionable trade: the turbine and grid-gear makers with real backlog, plus behind-the-meter gas/nuclear, over the paper-gigawatt announcers. (The Data Center Frontier Show, Jul 9; Switched On, Jul 9)
  • The clearest power name: GE Vernova. The Motley Fool Hidden Gems hosts made the case it's the best-positioned power stock in the AI build-out: a $163B backlog expected to hit $200B next year, an electrification unit that booked more data-center orders in Q1 than in all of 2025, and turbine production essentially sold out for close to a decade. It has ~400 GW of turbines running today and expects to add ~200 GW over five to seven years, and the real money is the decades of high-margin service that follows each install. The catch: most of the recent stock gain is valuation, so the thesis needs demand to hold for ~10 years. (Motley Fool Hidden Gems Investing, Jul 9)
  • Memory and storage, own the shovels, respect the cycle. Samsung, SK Hynix and Micron in HBM; SanDisk and Western Digital in storage, all sold out, all with record margins (Limitless; Stock Club). The bull view is that new supply can't arrive before ~2030, so pricing power persists. The bear view is that this has always been a boom-bust industry, the customers hate paying 72–78% margins, and price is a lagging signal at the top. Both can be true; size it knowing memory is where the whipsaw lands first. (Limitless: An AI Podcast, Jul 9; Stock Club, Jul 9)
  • Optics and networking. No fresh operator color on Coherent, Lumentum, Astera Labs or Credo this cycle, but the Stock Club first-half board is a useful read: Marvell +205% and Corning +153% (glass and fiber feeding data centers) both cracked the top ten, so the market is still paying up for anything physically wired into the build. (Stock Club, Jul 9)
  • Utilities and nuclear. The behind-the-meter shift keeps favoring incumbents with dispatchable and nuclear capacity, Dominion Energy's Stan Blackwell and Constellation are both slated to headline the coming data-center power discussions Klayman previewed. Regulated utilities that can actually energize load, plus nuclear operators, own the buildable pipeline. (The Data Center Frontier Show, Jul 9)

What Changed vs Last Issue

Yesterday (Jul 9, "Nvidia just gave back a trillion") the story was Nvidia's drawdown and money migrating to memory and power. Today the migration got its receipts, and the demand side finally showed a crack:

  • The rotation is now a first-half fact, not a forecast: the S&P 493 up ~14% and the Magnificent 7 up just ~2% (with Microsoft, Meta and Tesla down), SanDisk +717%, the SOX +74% vs Nvidia +5%. Last issue framed this as "MAG7 +1.1% YTD vs. S&P +10%"; the half-year print hardened it.
  • The memory winner has a name and a number: Samsung out-earned Nvidia ($58.5B vs $53B) to become the most profitable company in the world, a sharper headline than last issue's SK Hynix listing story. HBM pricing (+90% / +50–60% / +20% across Q1–Q3) and 72–78% margins quantify the "best pricing power of any sector" claim.
  • Nvidia's cheapness got a multiple: ~18x forward, below the S&P average, with a ~$300 average target, versus last issue's qualitative "cheapest since before the AI boom."
  • ERCOT's number moved again: the large-load queue is now ~500 GW (vs. last issue's ~445 GW), against sub-90 GW peak demand, with SB6's 75-MW flexibility rule now the gating mechanism.
  • The demand-blink signal we flagged to watch actually fired: last issue's sell-signal list included "'excess compute' resale deals multiplying." This cycle Meta launched a cloud to rent out its own compute (+10% on the day), the tell, not just the theory.
  • The circular-financing critique got heavyweight backers: last issue's Google and SpaceX example is now joined by Jim Chanos's "Second Derivative" note (a 2008-style credit read) and Mike Green's vendor-financing/passive-flow framing, plus Nvidia's own "compute now, pay later" deals.

Next prints: June CPI and the start of Q2 earnings season next week; late-July hyperscaler results (MSFT, AMZN, META) that test whether $60–190B-scale capex is finally buying visible revenue, and whether Meta signals an equity raise.