Newsletter · · Ashutosh Agarwal

SpaceX Picks Nvidia and the Bill for the Buildout Arrives - AI Accelerators: GPUs, Custom Silicon & Optics - Week of August 10, 2026

For the week of August 3 to 10, 2026, SpaceX used its first public earnings night to commit its AI data centers to Nvidia's Vera Rubin rather than AMD, Lisa Su pushed back on the record, and the harder question of who actually pays for the buildout moved to the center of the debate.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of August 10, 2026: SpaceX Picks Nvidia and the Bill for the Buildout Arrives

GPUs, Custom Silicon & Optics, a twice-weekly read of what the smartest podcasts are saying about AI hardware. Issue 014 · Monday, August 10, 2026 · covering episodes published August 3–10.

Two things happened this week that, taken together, tell you where the AI hardware story is heading.

The first was loud and easy to see. SpaceX reported earnings for the first time as a public company, and on the same night Elon Musk said the quiet part out loud: when SpaceX builds its own AI data centers, it is going all-in on Nvidia, not AMD. The Nvidia chip he named, Vera Rubin, the generation after today's Blackwell, will sit at the heart of a new SpaceX platform called StarMind. AMD's stock, which had been up 7% on the day riding a roaring chip rally, dropped about 7% the moment those words landed.

The second was quieter, and it matters more. Across a dozen different podcasts this week, the conversation shifted away from "which chip wins" and toward a blunter question: who is actually paying for all of this, and with whose money? The hyperscalers, the handful of giant cloud companies building these data centers, are now spending more cash than they generate. So they've started borrowing. They've started renting instead of buying. And the bond market, which tends to notice trouble before the stock market does, has started to twitch.

This issue is about both, the marquee event and the deeper current running underneath it. As always, everything below is sourced to a specific podcast episode, with a link, and I keep the operators (the people actually running these companies) separate from the pundits (investors and commentators talking their book). Earnings figures quoted on podcasts are relayed, not read from filings, I flag that where it matters.

TL;DR

  • SpaceX blessed Nvidia and snubbed AMD. In its first-ever earnings report, SpaceX said its future in-house data-center platform, StarMind, will run on Nvidia's Vera Rubin. AMD fell ~7% after hours on the news despite a strong quarter. It's the single clearest customer signal in the GPU race this week.

  • SpaceX also gate-crashed the capex club. Q2 revenue ~$7.8B (up 92% year on year), a brand-new AI rental business ("Elon Web Services") that tripled in a quarter to ~$2.6B, and capital spending of $18.4B, six times a year ago. The stock fell ~13% anyway. Being a new, cash-hungry entrant is now a liability, not a badge of ambition.

  • AMD's CEO fought back on the record. Lisa Su gave an exclusive interview insisting AMD's data-center business will grow "well over 100%" into 2027, that she has customer visibility "12, 24, 36 months in advance," and that losing SpaceX is simply "the right compute for the right workload." AMD is up 142% this year; Nvidia is up 13%.

  • The financing story is the real story. Goldman's credit desk now expects roughly a third of hyperscaler capex to be debt-funded this year, rising to ~35% in 2027. There's ~$1.65 trillion of data-center and chip commitments sitting off the balance sheet as leases. The bill is coming due, and the bond market, especially around Oracle, is starting to price it.

  • The bears got louder and more specific. New analysis this week claims 70–74% of Microsoft's and Amazon's AI revenue traces back to just two customers, OpenAI and Anthropic, companies the hyperscalers themselves are funding. Critics call it circular. Bulls like Ed Yardeni counter that the earnings are real and the buyers "wish they had more compute today."

  • Power and permits are now the hard ceiling. Texas has ~900 gigawatts of proposed data-center projects against ~90 gigawatts of total grid capacity, and the governor just hit pause. Expect on-site gas turbines, rural builds, and even home-sized GPU boxes.

  • Nvidia may quietly shrink the memory in its next chip because it can't secure enough of it, a small headline with big implications.

  • Still dark: pure optics operators (Coherent, Lumentum, Astera, Credo) got almost no airtime again, and there was little fresh on memory, TSMC packaging, or Marvell.

1. The lead: SpaceX picks Nvidia, and AMD's CEO answers back

Start with the moment that moved the market.

On Saxo Market Call (August 5), host John J. Hardy laid out exactly what happened and why AMD investors flinched:

"SpaceX CEO Elon Musk announced selecting NVIDIA's Vera Rubin chips (the generation beyond Blackwell) as the core of SpaceX's future StarMind AI data center platform, which pressured AMD stock down 7% after hours despite the SOX semiconductor index up 6.5% on the day."

(Saxo Market Call) In plain terms: the whole chip sector had a monster day, and AMD went along for the ride, right up until Musk named Nvidia as his exclusive design partner. Then it dropped.

Why does one customer's choice matter so much? Because AMD's stock has run on the story that it is becoming Nvidia's credible second source. Every marquee AI builder that commits to AMD strengthens that story; every one that walks past it weakens it. SpaceX is about to become one of the largest data-center builders on the planet, so its "Nvidia only" is a real dent.

AMD's Lisa Su, on the record, pushing back

Here's what made this week different from the last several issues: AMD's CEO didn't let the story sit. Lisa Su went on Squawk on the Street the very next morning (August 5) for an exclusive interview, and she was direct.

On the snub, she refused to concede an inch of technical ground:

"You need the right compute for the right workload. And from our standpoint, we have extremely strong roadmap as we think about Helios and going forward to our MI500 series."

On growth, she went further than Wall Street's models, and pointedly corrected the analysts who had marked her business down:

"We do see our data center business growing well over 100% as we go into 2027. And those are just very large numbers when you think about the base that we're on top of."

The word that matters there is "well." She was signaling that analysts who penciled in "over 100%" were still too low. And she explained where the confidence comes from, the data-center buildout works on long lead times, so customers now have to commit years ahead:

"These data center buildouts are really long-term buildouts. So we have to plan with our customers 12, 24, 36 months in advance… we have much, much stronger customer visibility in this time frame."

She sized the opportunity at more than most people dare to. In this interview she framed the overall computing and adaptive-computing market, not just AI chips, as going "to over $2 trillion as we go through 2030." (Elsewhere she and others have quoted a figure closer to $1.4 trillion specifically for AI accelerators; I flag the distinction because the two numbers get blurred on air.) And she reminded everyone who's actually winning the stock market this year:

Host Jim Cramer: "You're up 142 percent this year. NVIDIA is up 13 percent."

(Squawk on the Street) That's the paradox in one line. AMD is the year's better stock and Nvidia is the year's better business, and that gap is exactly why AMD sells off on any wobble. As the Elon Musk Podcast (August 5) framed the valuation, AMD holds roughly 4.5% of the data-center GPU market against Nvidia's ~95%, yet trades near 50 times forward earnings versus Nvidia's ~20, so the stock "had already priced in perfection" (Elon Musk Podcast). When you're priced for perfection, a single lost customer is enough to knock you down.

The quarter itself was fine: which was the problem

AMD's actual numbers, relayed on Closing Bell (August 4), were good: record earnings of $1.66 a share, revenue of $11.5 billion up 50%, data-center revenue of $6.7 billion (now 58% of the company), and a next-quarter guide of ~$13 billion against a Street estimate near $12.5 billion. The stock fell 8–10% anyway. Gross margin stuck at 56%, flat, because the expensive memory and the new Helios rack systems weigh on it. Bernstein's Stacey Rasgon (a pundit; he rates AMD outperform with a $600 target) shrugged off the sell-off:

"I think it looks fine. Like, there's nothing wrong with it… I still think the big AI ramp with Helios really is a Q4 story… if it's big, everybody should be doing great."

(Closing Bell) His point, the one AMD bulls cling to, is that the real Helios revenue is a late-2026 and 2027 event, so this quarter was never going to show it.

Why it matters for a book: the AMD long thesis is intact but back-end-loaded, and it now has a visible scoreboard. Watch three things Su all but dared the market to track, the Helios ramp timing (Q4 into 2027), the Anthropic deployment (up to 2 gigawatts), and whether that "well over 100%" data-center growth shows up. Nvidia, meanwhile, just picked up the single most aggressive new builder in the industry as a Vera Rubin customer.

2. SpaceX joins the club: and gets punished for it

The other half of the SpaceX story is that it is now, out of nowhere, one of the biggest spenders in AI infrastructure. And the market's reaction tells you the mood has flipped.

The clearest walk-through came on All-In (August 8), where the hosts went through SpaceX's first report as a public company (numbers relayed, not from filings):

"Q2 results were spectacular… $7.8 billion in revenue, up 92% year over year… AI revenue, Elon Web Services, more than tripled quarter over quarter to $2.6 billion… CapEx was up $18.4 billion in the quarter. That's 6x year over year… a run rate of about $75 billion."

Let that sink in: a company that essentially did not have an AI business nine months ago now has a ~$2.6 billion-a-quarter cloud-rental operation (renting spare compute to Anthropic and Google from its "Colossus" server farm) and is spending $18.4 billion a quarter to build more. Musk guided to $100 billion of annual recurring revenue by year-end and pulled his $1-trillion revenue target forward a full year, to 2030. For scale, Morgan Stanley's 2030 revenue estimate for SpaceX is $325 billion, so even the bank thinks big, and Musk is way past the bank.

And the stock fell about 13%. It's down roughly 40–50% from its June IPO peak, settling near a $1.4 trillion valuation after listing above $2 trillion. As one host put it, the pattern is familiar, "within six months of the IPO, almost all these tech stocks are down 50% peak to trough", but the reason is the tell. Investors are now nervous about surging AI capex even when it comes wrapped in a spectacular growth story. (All-In)

There's a second, subtler idea buried in that All-In discussion that's worth pulling out, because it explains a lot of this week's news. The hosts called it "channel conflict." Google, Microsoft, and now SpaceX all face the same tension: they can either use their scarce compute to build their own frontier AI models, or rent it out to customers like Anthropic. Renting is the safer, more profitable business, Microsoft's Satya Nadella cited a Morgan Stanley figure of "over a 30% return on invested capital in tokens as a service." So capital keeps flowing toward being an infrastructure landlord and away from model-building. That's the through-line connecting Google losing researchers, SpaceX renting to Anthropic, and everyone racing to pour concrete: the money wants to own the picks and shovels, not the gold.

For the hardware book, SpaceX matters in two concrete ways. It's a giant new Nvidia customer (StarMind on Vera Rubin), and it's a giant new power problem, which brings us to the two themes that dominated the week.

3. The real story: how the buildout gets paid for

Here's the shift that made this week feel different. For a year, the debate was demand ("is the AI compute real?"). This week, on podcast after podcast, it became financing ("even if it's real, how is it funded, and what happens when the cash runs low?").

The most sober version came from Exchanges, Goldman Sachs's own podcast (August 5), where a credit analyst named Amanda laid out the plumbing in numbers. Until now, hyperscalers built data centers out of their own enormous cash flow. That's ending, because they're now spending nearly everything they earn:

"Taking that $108 billion of issuance… in 2025, that actually represents 27% of CapEx from the hyperscalers. So far this year… we think that roughly 33% or one-third of CapEx will be debt financed in 2026… we're bracing for something like $250 billion of direct supply from the hyperscalers… And then we expect this to peak next year in 2027, where we expect 35% of CapEx to be debt financed."

Why the rising reliance on borrowing? Because, in her words, "CapEx and cash flow from operations are quickly approaching the same level, they're converging." Translation: the piggy bank is nearly tapped, so debt does the rest. She's not alarmed, she thinks the investment-grade bond market, private credit ($4.5 trillion of "dry powder"), overseas bond markets, and project finance ($300 billion of it expected in 2027) can absorb it. But she flagged the early cracks: "17 out of the 23 sort of data center JV deals are now trading wide to originated yields," and the single number she'd watch is the "double derivative" of capex, the point where spending stops accelerating. (Exchanges)

Then there's the debt you can't see on the balance sheet. On The Real Investment Show (August 6), the hosts worked through a figure making the rounds, roughly $1.65 trillion (a Nikkei estimate) of obligations that don't appear as "debt" because they're structured as long-term leases: leasing the data center building instead of owning it, and even leasing the Nvidia chips instead of buying them.

"These obligations are real. And they have to pay them regardless of whether they use the data center, regardless of whether they use the chips… you can think of it like a bond where they do owe these monthly payments. But they're not on the balance sheet."

Their honest conclusion was reassuring on four of five names and pointed on the fifth: fold those hidden leases in and Oracle's debt-to-equity ratio jumps from about 4x to 10x, while Google, Amazon, and Microsoft go from a comfortable 0.3–0.6x to around 1x (Meta to ~2x). The credit market agrees, four of the five hyperscalers' bond spreads are calm, and only Oracle's default-insurance cost is "shooting up into the atmosphere." Their advice was the oldest in the book: read the footnotes. (The Real Investment Show)

And Microsoft gave us a small masterclass in optics this week. On Big Technology (August 8), the hosts explained Microsoft's "clever trick": it reclassified roughly $15 billion from capital spending to operating spending, and extended the assumed useful life of a data center from 15 years to 25. That let Microsoft hold its spending outlook flat, looking disciplined next to Google and Amazon raising theirs, and the stock had, per the hosts, its "single best stock day of anyone in history." The catch, which one host nailed:

"GPUs are not going to last 25 years… on the low end, people say they'll last three years… And, of course, the chips are 50% of the data center build-out cost. So how do you put that on a 25-year scale?"

(Big Technology) The building may last 25 years; the Nvidia chips inside it, half the cost, will be swapped out in a handful. Stretching the depreciation clock flatters today's profits and pushes the reckoning down the road. This is exactly the "depreciation" concern that has hovered over the whole complex, now showing up in the accounting.

Why it matters: the financing turn is the most important new development in weeks. As long as spending was self-funded, "is it a bubble?" was an abstract argument. Now it's concrete, debt issuance, off-balance-sheet leases, stretched depreciation, and a bond market that's already singling out the weakest name (Oracle). The credit market is the tripwire to watch, and it's twitching.

4. The bear case gets specific: and the bull case answers

The financing angle put fresh ammunition in the hands of the skeptics, and this week they got granular.

The sharpest bear was Ed Zitron on Better Offline (two episodes, August 5 and August 7). His argument isn't "AI is useless", it's that the eye-popping cloud growth everyone is celebrating is, to a shocking degree, two customers spending money the hyperscalers themselves gave them. He cited a stack of Wall Street research to make it concrete:

"73% of all Amazon's AI revenues in both 2026 and 2027 will be Anthropic and OpenAI's compute spend" (Barclays' Ross Sandler). "27% of all Google Cloud revenues… for 2026 are estimated to be OpenAI and Anthropic… and then 44% in 2027" (UBS' Stephen Drew). And on Microsoft: "70% of Microsoft's AI revenues are from either its revenue share of OpenAI's money, or OpenAI spending money to rent out compute from Azure."

His "scandal" is the circularity: the same hyperscalers that book this revenue are also the investors keeping OpenAI and Anthropic alive. He pointed to Amazon quietly handing OpenAI a final $35 billion tranche this week, money that was supposedly contingent on an IPO or achieving AGI, neither of which happened:

"This is what you do when you have no other options, no other viable customers… Amazon kind of knew that they needed to pay OpenAI so that OpenAI could pay them."

(Better Offline, "The AI Demand Bubble" · Better Offline, "The Pale Horsemen Arrive") It's a one-sided, deliberately provocative case, and those revenue-concentration figures are analyst estimates, not disclosed numbers, so treat them as informed guesses. But the underlying worry is legitimate and now widely shared. On 20VC (August 6), the hosts framed the same dependency more calmly: if OpenAI and Anthropic fail to monetize, "70–80% of this compute" spend is at risk. (That episode also flagged the week's other shoe: Moonshot AI, the Chinese lab behind the Kimi models, reportedly raised $3.5 billion at a $35 billion valuation, the low-cost competitor isn't going away.) (20VC)

The most useful framing of the bear case, though, came from a bull: David Cahn of Sequoia on Big Technology (August 5). He's spent two years quantifying the "ROI question," and his math has gotten enormous:

"You basically have about $3 trillion that needs to get paid back just since ChatGPT", his running tally of cumulative capex ($200B in 2023, $600B in 2024, $850B in 2025, ~$1.5T in 2026). By his rule of thumb, ~$2 trillion of 2026–2027 capex implies ~$4 trillion of lifetime revenue needed to pay it back.

His verdict is stark and worth quoting because it captures the psychology of the moment:

"Nothing short of AGI will be enough to justify the investments now being proposed… we've reached a point in 2026 where not only is everybody in the game, but everyone is massively accelerating… people just don't care about the math anymore at all. It's all in, all the way."

(Big Technology)

On the other side, the calmest bull voice this week was Ed Yardeni on The Compound and Friends (August 7). His rebuttal to the "earnings bubble" claim, the idea that today's profits are fake because everyone is stockpiling chips they won't need, was simple:

"The vibes I get out of the hyperscalers, the managers wish they had more compute today because they could make a lot more money today if they had it."

He made the bullish observation of the week about valuation, too: the market's forward price-to-earnings ratio has actually fallen 8% over six months even as the index rose 11%, meaning stocks got cheaper on rising earnings, "because people don't believe it will continue." And he pointed to the memory makers as proof the earnings are real, not hoped-for: Western Digital just posted a $3.2 billion quarterly profit, up from $282 million a year earlier, selling out of supply at fat margins. But even Yardeni wouldn't wave away the new risk:

"This is not a very thrilling development on the debt side… usually the debt markets see the problems before the stock market does. And credit default swaps on Oracle, I think, are at like a multi-year high."

(The Compound and Friends) Notice that the bull and the bears now agree on the single indicator to watch: the bond market, and Oracle in particular.

One more data point for scale, from Chip Stock Investor (August 3): across Amazon, Microsoft, Google, Meta, Oracle, Tesla, and SpaceX, trailing-twelve-month capex is ~$580 billion, 2026 is tracking to ~$900 billion, and 2027 could reach $1.5 trillion, against combined cloud revenue of ~$400 billion. Several of these companies are now free-cash-flow negative. (Chip Stock Investor) That's the gap the debate is really about.

(A footnote worth flagging: the forced unwind of Leopold Aschenbrenner's "Situational Awareness" AI hedge fund, surfaced last issue as an unverified rumor, was discussed as a real liquidation event this week across Get the Check, 20VC, and others, and named as one trigger for the recent whipsaw. The exact fund size and losses still vary by source, so I'm not putting a number on it.)

5. Power is the ceiling now: and it's getting political

If financing is the money constraint, electricity is the physical one, and this week the power story turned into a political one.

The headline number came from Telltales (August 5), and it's staggering:

"Data center developers are proposing 900 gigawatts of projects in Texas. All of Texas' power capacity is about 90 gigawatts, so it's 10 times the amount of current capacity."

Texas Governor Abbott sent regulators a letter that, parsed, amounts to a roughly 12-month pause while each project gets audited, mirroring New York's one-year hold and Virginia's quiet "no more." The hosts' conclusion is the read-through that matters: since developers all want compute now (not in 2029), and you can't build grid connections fast enough, the only path is on-site generation, "the only way it can be satisfied is with on-site turbines. You can't build combined cycle fast enough, and turbines are completely sold out." (Telltales)

That squares with what WTR Small-Cap Spotlight (August 4) reported: about 90% of operating data centers are urban and grid-powered, but roughly 70% of planned projects are heading to rural areas with their own generation, to dodge grid interconnection queues that now average 2,100 days, up 60% since 2017. The example: a Microsoft-Chevron project ("Kilby") planning a 2.7-gigawatt gas plant in the Texas Permian (WTR Small-Cap Spotlight). Morgan Stanley's Thoughts on the Market (August 4) framed the gap as a ~38-gigawatt shortfall through 2028, with behind-the-meter power offered as a concession to wary communities (Thoughts on the Market). Open Circuit (August 7) captured the politics, the backlash is less about clean energy than about local trust and who pays, with regulators experimenting with "cost-causation" tariffs so data centers foot their own grid bills (Open Circuit).

And this week surfaced two constraints below even power. On Catalyst with Shayle Kann (August 6), the bottleneck was labor, specifically a shortage of skilled electricians to actually wire these buildings (Catalyst). And in one of the more creative operator ideas of the week, on Inevitable (August 4), SPAN CEO Arch Rao unveiled "Expra", a plan to put small, liquid-cooled GPU boxes in ordinary homes, tapping the ~19 kilowatts of spare capacity sitting unused in most 200-amp houses. SPAN has raised nearly $500 million and is deploying early units with Nvidia (Inevitable). It sounds exotic, but it's the same logic as SpaceX planning data centers in space: when the grid says no, you find power somewhere else.

Why it matters: power availability, not chip supply, is increasingly the thing that gates how fast this buildout can actually happen, which is bullish for gas turbines, on-site generation, electrical equipment, and the picks-and-shovels of the grid, and a genuine timing risk for the compute demand everyone is modeling.

6. Custom silicon and a memory warning

A few developments worth a book's attention on the chip roadmap itself.

Everyone now wants their own chip. On The AI Podcast (August 5), the host summarized the state of play: Anthropic is building an in-house silicon team and scouting Samsung as a manufacturing partner, which "ends Anthropic's status as basically the last major AI lab without a public silicon program." OpenAI already shipped its Broadcom-designed inference chip, "Jalapeno," in June; Google runs its own TPUs; Meta has its MTIA accelerators. The motive is blunt, get out from under Nvidia's allocation and pricing. But it's a long game: custom silicon takes 18–24 months and hundreds of millions of dollars per version, with a roughly two-year payback (The AI Podcast).

A real challenger operator spoke. On Eye On A.I. (August 6), d-Matrix CEO Sid Sheth made the most interesting non-Nvidia case of the week, that inference is splitting into tiers, and there's a premium tier where speed is worth paying up for:

"I don't want that level of interactivity, in which case I'll pay $2 for a million tokens. But if I want that high level of interactivity, I'll pay $20 for a million tokens. And people are willing to pay that $20."

His pitch is that chips which put compute and memory together deliver "an order of magnitude" more memory bandwidth than a standard Nvidia GPU for spitting out tokens fast, the "premium token economy." d-Matrix bought a company called Giga.io to build full racks and is partnering with Supermicro to manufacture them (Eye On A.I.). Fellow startup Etched got a name-check on Summation (August 4), reportedly "a billion in orders" and "shipping their racks this summer" (Summation). The challengers are shipping, not just pitching.

And the memory warning. The most consequential small headline came from The AI Daily Brief (August 7): Nvidia is reportedly considering cutting the memory on some versions of its next flagship chip, Rubin Ultra, not because demand is soft, but because it may not be able to buy enough high-bandwidth memory to build them. Nvidia's hardware chief Andrew Bell insisted the opposite in mid-July,

"We were in front of the memory problem, so it's not going to hold us back anytime soon. The pricing, of course, is a problem for the whole world."

, but as one analyst quoted put it, "This isn't weak AI demand. This is memory rationing at the top of the food chain." (The AI Daily Brief) If even Nvidia has to design down to fit the memory it can source, the memory shortage we've tracked for months is now shaping the flagship product itself, bullish for the memory makers (SK Hynix, Micron, Samsung), and a subtle cap on how fast models can keep growing.

7. The negative space

Per house rules, what didn't show up is a signal too.

  • Pure optics stayed dark again. No operator commentary from Coherent, Lumentum, Astera Labs, or Credo, only a pundit stock pick of Coherent and Corning on The MoneyFlows Show (August 8) (The MoneyFlows Show). This is now a multi-week silence; either these names didn't report in the window or the podcasts simply aren't discussing them. I'd rather tell you that plainly than manufacture a signal from nothing.

  • Arista's blowout happened on August 4 (earnings of $1.02 vs an $0.88 estimate, and a strong guide) but fell inside last issue's window; there was nothing fresh this week beyond the after-hours pop.

  • TSMC and advanced packaging got only a passing mention (a $100 billion Arizona expansion, profit up 77%), no fresh CoWoS or substrate angle.

  • Marvell and Broadcom were quiet on the custom-silicon front; Broadcom surfaced mainly as the maker of OpenAI's "Jalapeno" and, oddly, as a short position in the blown-up Situational Awareness fund. Meta's MTIA was name-checked only.

What changed since Issue 013 (August 6)

  • The lead moved from AMD's print to SpaceX's. Last issue was "AMD beats and still bleeds." This week AMD is a continuation story, but with a fresh, on-the-record CEO rebuttal, while the new marquee event is SpaceX's first earnings and its decision to build on Nvidia's Vera Rubin, not AMD.

  • The frame shifted from equities to credit. Issue 013 was about how the stock market sorted the hyperscalers (Microsoft and Amazon rewarded, Meta punished). This week the conversation moved to financing, debt issuance, ~$1.65 trillion of off-balance-sheet leases, Microsoft's capex-to-opex reclassification, and Oracle's widening credit spreads. The tripwire is now the bond market.

  • New bubble voices. Chanos and Gavin Baker carried last issue; this week it was David Cahn (Sequoia, the ~$3-trillion/AGI-or-bust bear-in-bull's-clothing) and Ed Yardeni (the "they wish they had more compute" bull). The circular-financing critique also got far more specific, with named analyst estimates on OpenAI/Anthropic revenue concentration.

  • A new challenger operator and a memory warning. d-Matrix's Sid Sheth gave the first fresh non-Nvidia operator voice in a while ("premium token economy"), and Nvidia's reported plan to shrink Rubin's memory is a new, concrete sign the memory shortage is now shaping flagship silicon.

  • Power escalated from forecast to politics. Last issue's ~38-gigawatt shortfall became this issue's Texas pause (900 GW proposed vs 90 GW of grid), an electrician shortage, and home-sized GPU boxes.

  • Still dark: pure optics operators (now multi-week), TSMC packaging, Marvell, and Meta's MTIA.