Newsletter · · Ashutosh Agarwal

The Custom-Silicon Fight Becomes a Financing Fight - Custom Silicon vs Nvidia - Week of September 14, 2026

Custom Silicon vs Nvidia for the week of September 7 to September 14, 2026. Podcast synthesis on how the AI-chip story pivoted from design wins to financing: Blackstone's ballooning Google-TPU cloud, warrant structures that may hand hyperscalers custom chips for free, a more skeptical second look at Broadcom's guide, and hyperscaler debt issuance now rivaling US Treasury flows.

Custom Silicon vs Nvidia

Week of September 14, 2026: The Custom-Silicon Fight Becomes a Financing Fight

Last week the podcasts argued about who builds the best chip. This week they argued about who's going to pay for all of them. After the flood of design-win news around the Nvidia-MediaTek deal and Broadcom's print, the feed went quiet on genuine new chip leaks and loud on a different question: the money. The dominant thread this week was financing and "digestion": how the buildout is being funded, whether the accounting is honest, and who actually captures the profits from all these custom chips. That's not a step down in importance. If you own any of these names, the financing plumbing is now the single biggest swing factor, and this week it got both louder and more specific.

There were still two genuinely new, book-relevant developments underneath the macro noise: Blackstone's TPU-chip cloud with Google is quietly ballooning from $5 billion toward $10–20 billion, and a clever new detail on who really captures the economics of custom chips, the hyperscalers, not the chip designers. More on both below.

A quick reminder for new readers: an ASIC, or "custom chip" (the industry calls it an XPU), is a chip a company designs for its own specific job (Google's TPU, Amazon's Trainium) instead of buying a general-purpose Nvidia GPU off the shelf. Firms like Broadcom and Marvell don't own those chips; they're the engineering shops that turn a hyperscaler's design into a real, manufacturable chip at TSMC. This letter tracks whether those custom chips are eating into Nvidia's dominance, and, increasingly, whether the whole boom can keep funding itself.

As always, the tags: [OPERATOR] = a company executive on the record; [REPORTER] = a journalist reporting; [ANALYST/PUNDIT] = an analyst, strategist, investor, or host opinion; "relayed" = a number reported second-hand on a podcast, worth confirming against filings.

TL;DR

  • The money question ate the chip question this week. With no fresh design-win leaks, the podcasts pivoted to financing and "digestion." The scariest number: hyperscalers are now issuing about as much long-dated debt as the entire U.S. government is, according to macro strategist Jens Nordvig, "a few private sector companies... managing to issue as much debt in the long end of the curve as the US government" (The Compound and Friends). [ANALYST/PUNDIT]

  • New and actionable: Blackstone's TPU cloud with Google is scaling up fast. The joint venture, now named Crux AI, a "neocloud" built entirely on Google's TPU chips, is moving from an initial $5 billion commitment to "multiples" of that, "more likely between $10 and $20 billion," per new reporting from The Information's Dakin Campbell (The Information's TITV). That's a big, fresh vote of confidence in TPU demand, financed off Google's own balance sheet. [REPORTER]

  • The best new insight of the week: hyperscalers may get their custom chips "basically for free." Google's deal with Marvell and Amazon's new deal with Qualcomm both use warrant structures, the hyperscaler gets equity in the chip designer if volume targets are hit. Run the math and "these hyperscalers are getting custom ASIC chips for basically free if all the targets are hit," which tells you exactly where the bargaining power sits (Vaughan Nelson Investment Management). Revenue growth for Broadcom/Marvell is real; value capture is another question. [ANALYST/PUNDIT]

  • The second-week read on Broadcom got more skeptical. The Circuit's Ben Bajarin and Jay Goldberg dug into the print: fiscal-27 custom revenue guided to $115B (below the ~$150B "bogey," but seen as a floor), $230B in fiscal-28; Broadcom finally naming MediaTek as a Google competitor after "basically denying competition existed" two quarters ago; and a gross-margin dip management waved off with "don't worry about gross margin, worry about operating margin" (The Circuit). [ANALYST/PUNDIT]

  • A sharp challenge to the "chips as collateral" story. Broadcom is helping finance its AI-lab customers and keeping it off its balance sheet, "using their chips as basically the backstop." But a custom XPU is built to run one company's workload, so if a customer fails, how fungible is the collateral really? "How can you say you're backing the financing with these chips when the fungibility of them is in question?" (Investing Experts). [ANALYST/PUNDIT]

  • Optics is the next bottleneck. As racks scale past a million connected chips, copper gives out and light takes over. Broadcom's Tomahawk 6 switch moves "more than 100 trillion bits... every single second"; Nvidia's new Rubin systems move 260 terabytes per second, driving a re-rating case for Lumentum, Ciena and Applied Optoelectronics (The MoneyFlows Show). [ANALYST/PUNDIT]

  • Quiet on the chips themselves. No fresh operator color this week on Marvell design wins, Amazon's Trainium 3, Microsoft's Maia, Meta's MTIA, Arm, or Alchip. Flagging honestly rather than stretching.

What's New

Ranked by what actually moves a book, the genuinely new, actionable items first, then the deeper re-reads of last week's prints, then the macro-financing debate that dominated the week.

1. Blackstone's Google-TPU cloud "Crux AI" is scaling from $5B toward $10–20B. [REPORTER, new reporting]

This is the most concrete new, thesis-relevant fact of the week, and it comes straight from fresh reporting. The Information's finance reporter Dakin Campbell laid out new details on the Blackstone-Google joint venture, "effectively a neocloud for TPU chips", that was originally announced at $5 billion (The Information's TITV). [REPORTER]

What's new: it has a name (Crux AI), a CEO ("this gentleman who was at Google for a couple decades... very familiar with Google's TPUs and how to employ them in a data center, neocloud environment"), and, crucially, a much bigger ambition. Campbell: "they're now looking to put multiples of $5 billion into the venture... multiples is not, you know, one and a half times the $5 billion. We're talking north of $10 billion, more likely between $10 and $20 billion."

Why it matters. This is the demand side of the TPU story getting funded at scale by an outside financier. It says a private-equity giant is willing to underwrite billions of dollars of Google silicon in a dedicated cloud, a real-money read-through for TPU volumes, and by extension for Broadcom and MediaTek, who build those chips. It also fits the broader theme of the week: the biggest AI infrastructure moves are increasingly financing structures, not product launches.

2. The warrant twist: hyperscalers may get custom chips "basically for free." [ANALYST/PUNDIT, new framing]

This is the most useful new idea of the week, and it reframes a question this letter has been circling for months, if custom-ASIC revenue is exploding, why isn't the value obviously flowing to Broadcom and Marvell? The team at Vaughan Nelson explained the mechanism (Vaughan Nelson Investment Management): [ANALYST/PUNDIT]

"There have been multiple announcements, Google partnering with Marvell, Amazon with Qualcomm, which we didn't even put in the paper that actually happened this past week... what the hyperscalers have decided to do is they say, well, we don't want to give all this money to these smaller companies... And so what the companies have come up with are these warrant structures where they'll get equity value in these companies if they meet certain volume levels. And when you run the math, what's crazy is effectively these hyperscalers are getting custom ASIC chips for basically free if all the targets are hit."

The punchline: "that right there shows who has the market power." Note also the new tidbit in there, an Amazon-Qualcomm custom-chip deal that "happened this past week." (Unconfirmed by an operator on the record; worth chasing against filings.)

Their broader argument is the classic Nvidia bull case, restated cleanly: Nvidia isn't a chip company anymore, it's an "architecture company", you get a GPU, a CPU, networking, and the CUDA software layer, and the software keeps squeezing free performance out of the hardware. Their example: a software update from "tile RT" (their phrasing) sped throughput "three to five X by just doing a software update." Custom ASICs, they argue, "can build a chip that works fast... but you don't have the developer ecosystem." And their favorite tell: even Elon Musk, who has deep custom-silicon experience, is "standardizing on NVIDIA's infrastructure" at SpaceX, and the neoclouds are building on standardized Nvidia. Their read: "scale is what wins in semis... not the best technology."

3. Broadcom, the skeptical second look. [ANALYST/PUNDIT, deeper re-read of last week's print]

Broadcom reported on September 2 (covered last week). This week the semiconductor specialists at The Circuit, Ben Bajarin and Jay Goldberg, gave it the harder, more skeptical dissection now that the dust has settled (The Circuit). [ANALYST/PUNDIT] The stock was down ~4% on the week. Their points:

The numbers, in context. The "bogey" for Broadcom's fiscal-27 custom business was around $150 billion; management guided $115 billion ("most consensus notes... has been like, that's probably the floor"), and $230 billion in fiscal-28. For scale, they peg the total custom-ASIC market (compute plus networking plus attach) at $400–430 billion in 2028.

MediaTek is now openly a rival at Google. "Two quarters ago, they were basically denying competition existed within Google. And here they are now saying, yes, there's competition from MediaTek, but we're delivering faster... better IP." Their take: MediaTek "is executing better than people expected," and "no slouch in the custom ASIC business."

Gross margin got waved off. "Gross margins were down and management's defense of this was not great... what they said was don't worry about gross margin, worry about operating margin." That matters because the whole bear case on Broadcom is that as customers gain their own IP, "why would you continue to pay Broadcom's margins? They have the highest margins of anybody in custom silicon."

A useful gigawatt frame. Using Bernstein-style math, they land on "10 gigawatts in fiscal 27, moving to 20 gigawatts in 28" across all Broadcom customers. On a dollars-per-gigawatt basis: Nvidia is talking "$40 to $50 billion per gigawatt going to $90 to $100 billion," while Broadcom is "$10 to $15 billion per gigawatt... now going to $25 to $30 billion." A big chunk of both, they suspect, is memory being lumped in.

Broadcom vs Nvidia, operating leverage: a tie. Goldberg built a spreadsheet comparing revenue growth to operating-income growth for both. "The result was a tie. They're both growing revenue very strongly, much faster than they're growing expenses." His conclusion on the whole "Broadcom vs Nvidia" framing: "if that's the case, then it's a tie."

Modeling custom ASICs is getting genuinely hard. The old shortcut, "so-and-so won TPUs, we think TPUs is this much volume, I'll give them all of it", is dead. Now it's "mixing and matching": one vendor wins the compute tile, another the I/O, another the networking, each at different margins (Ethernet attach < copper < optical). Their warning: "people are going to get this wrong... we're going to end up with 150% market share kind of calculations."

The other big Circuit theme, and it lines up with the week's financing thread, is that 2027 is peak constraint, 2028 is when relief starts. The industry "can only scale as its deepest bottleneck," and greenfield capacity ordered in late 2025/early 2026 doesn't arrive until 2028. Watch for potential upside surprises then, TSMC alone has "20 fabs" under construction across nodes, and Micron/Samsung/SK Hynix are all adding. The flip side (Goldberg, playing devil's advocate): more capacity in 2028 could finally push memory pricing down, memory units have been flat while revenue "soared," so it's all been price.

4. Nvidia's Hugging Face deal, dissected: it's about inference and developers, not chips. [ANALYST/PUNDIT, deeper re-read]

We flagged Nvidia's ~$12.9B Hugging Face acquisition last week; this week it got a full episode (Chip Stock Investor Podcast). [ANALYST/PUNDIT] The framing worth keeping:

It "would technically be the largest acquisition ever for NVIDIA if you exclude the Grok aquihire," and it's Nvidia's first real move into the "service space, the application layer." But there's "more than a small likelihood this one gets shot down", much of Hugging Face's team is in France, so expect heavy EU (plus U.S. and China) scrutiny to keep AI "truly open source."

The strategic logic is inference and developer funnel, not near-term revenue. Nvidia's initial hyperscaler GPU growth "has worn off... but now Blackwell in production, Vera Rubin is ramping... revenue has re-accelerated." Owning Hugging Face keeps Nvidia at the center of AI software as inference spreads into enterprises and the physical world. The comp is Microsoft buying GitHub in 2018, which preceded its OpenAI/Azure infrastructure boom.

A tidy aside on the custom-silicon overlap: the AI labs themselves are custom-chip buyers, "Anthropic's Claude using a lot of Amazon's custom silicon", so the "open vs closed model" and "GPU vs ASIC" fights aren't as clean as investors want them to be. Same episode notes Broadcom "reaccelerating its AI semiconductor revenue right now, thanks to... Google and the TPU, as well as others like OpenAI." [relayed]

Investing Experts' Joe Albano added the wry read: it's "ironic that [Nvidia] now purchased the one thing that one of its clients essentially broke into" (a reference to the recent open-model security saga), and called Nvidia's best-ever deal its old Mellanox networking buy, "everybody went, why do you want networking?... until they realized they're building the best and biggest racks in the world now because of that networking" (Investing Experts). [ANALYST/PUNDIT]

5. The financing debate got loud, and quantified. [ANALYST/PUNDIT, the week's dominant theme]

This is where most of the week's energy went. The through-line: AI capex has now outrun the hyperscalers' own cash flow, so the buildout increasingly depends on outside debt and equity, which makes the whole complex newly sensitive to the capital cycle. The specifics worth carrying:

Cameron Dawson (NewEdge Wealth) put hard numbers on the swing: as of summer 2025 analysts expected hyperscaler capex to be flat; it's now "likely... going to grow between 90 and a hundred percent for this year," with consensus at "30 to 40% growth for CapEx next year." Because semis are ~18% of the S&P 500, that delta drove the entire earnings boom. Her warning: these companies "used to be agnostic to the capital cycle because they were very self funded [and] are now very sensitive." The market "is not saying, okay, we won't fund you. The market is saying we'll fund you, but at a bigger price" (On The Tape with Danny Moses). [ANALYST/PUNDIT]

The accounting is getting creative, a late-cycle tell. Dawson (relayed on Two Quants and a Financial Planner) walked through it: Nvidia "extended its accounts receivables in order to help its customers," while the hyperscalers "extended... accounts payables, which flatters their free cash flow"; Microsoft "lowered its CapEx guidance because it reclassified financial and operating leases," moving spend out of operating cash flow; and a growing pile is "not even hitting the balance sheet... because it's all happening in these SPVs" that are "super esoteric and very murky." Strip out one-time gains and "50% growth is dropped down to about 30% growth" (Two Quants and a Financial Planner). Her line: "if things were as great as you say that they are, you should not have to be playing the games that you are playing." [ANALYST/PUNDIT]

The macro backdrop is the scary part. Macro strategist Jens Nordvig (Exante Data, a former #1-ranked currency strategist) called this "the most interesting macro moment of my lifetime." His headline stat: hyperscaler bond issuance at the long end of the curve is now roughly "equal to what the US Treasury has" in flow, and that's with the U.S. running a 6%-of-GDP deficit. Worse, the funding need is nonlinear: "next year is going to be dramatically more than this year." With the 10-year around 4.92% (last seen October 2023) and the 30-year at 2007-era highs, he's watching for a "non-linearity" that forces an "equity puke" (The Compound and Friends). [ANALYST/PUNDIT]

And the Fed decision lands this week. On Inside the ICE House, the guest pegged "about 60% odds of a rate hike for this month" (he personally doubts Warsh will hike). His point for our purposes: a 25bp move is noise next to "a trillion dollars of AI CapEx coming down the line for the next year" and "$5 trillion of projected spend" over five years (Inside the ICE House). [ANALYST/PUNDIT] A hike into a market this dependent on cheap capital would be a live catalyst, mark your calendar.

The Debate

The motion: do hyperscaler custom chips structurally take share and cap Nvidia's market and margins, or do CUDA, systems integration, and Nvidia's relentless pace keep merchant GPUs on top?

This week didn't move the silicon argument much (no new design-win data). But it added a genuinely new layer to the economics argument, which is arguably more important for where the profits land.

For the custom-silicon side (ASICs take share). The volume story is intact and growing: Broadcom guided fiscal-27 custom revenue to $115B (a floor) and $230B in fiscal-28, into a $400B+ market; Blackstone is underwriting $10–20B of a TPU-only cloud; and the AI labs (Anthropic on Amazon silicon, OpenAI on Broadcom) keep widening the buyer base. Nvidia's own margin dip and its willingness to co-opt XPUs (last week's MediaTek deal) both concede that custom chips are permanent.

For the merchant side (Nvidia keeps the pie). The week actually strengthened this side in a subtle way. First, the value-capture problem: warrant structures mean hyperscalers can get their ASICs "basically for free" if volumes hit, so custom-chip revenue can boom while custom-chip profit accrues to the buyer, not the designer (Vaughan Nelson). Second, the architecture-and-software moat: standardizing on Nvidia keeps paying performance dividends via CUDA updates, which is why even Musk's SpaceX standardizes on Nvidia. Third, operating leverage is a tie, Nvidia is compounding just as efficiently as Broadcom (The Circuit).

The most useful new frame this week came from Joe Albano's dissection of "chips as collateral." Broadcom is backstopping its labs' financing with the very XPUs it sells them, but "the whole point of an XPU is that it runs your specific workload," so if a customer collapses, the next buyer "may [take the chips] because... we'll take whatever compute we get. But they're going to be trying the next day to find a better solution because it's not going to be as efficient." In other words, the thing that makes a custom chip valuable (workload specificity) is exactly what makes it weak collateral. "I think it's a house of cards in the financing. And as long as nobody blows on it, it's going to be fine" (Investing Experts). [ANALYST/PUNDIT]

Where I come out this week: the silicon call is unchanged, own the volume growth on both sides (Broadcom/Marvell for the value-added networking and inference sockets, Nvidia for the fungible architecture it sells around every XPU). But the week sharpened the real risk, and it's not the chip fight. It's that the entire complex is now funding itself with debt at multi-decade-high yields, off-balance-sheet SPVs, warrant sweeteners, and collateral (custom chips) that may not be as fungible as the lenders assume. As Cameron Dawson put it: price leads data. The chips are selling; the question is whether the financing holds while everyone waits for the ROI that consensus has pushed out to 2028.

Stocks in Play

Every ticker with fresh podcast commentary this week, with bull, bear, and the next thing to watch. Note: most of this week's commentary is analyst/pundit opinion, not operator disclosure, tagged accordingly.

AVGO (Broadcom). [ANALYST/PUNDIT + relayed] Bull: fiscal-27 custom guide of $115B seen as a floor, $230B in fiscal-28; AI-semi revenue reaccelerating on Google TPU + OpenAI; largest gigawatt custom-compute footprint of anyone; operating leverage matches Nvidia's (The Circuit). Bear: gross margins down with a weak management defense; MediaTek now an acknowledged rival at Google; warrant structures may hand the economics to hyperscalers; off-balance-sheet financing backstopped by workload-specific chips of questionable fungibility (Investing Experts); one chart-watcher thinks it "might have been one of the first to top." Watch: whether operating margin holds as custom mix grows; any hard number on the fiscal-28 ramp; the networking/optical attach mix (higher margin than the ASIC itself).

NVDA (Nvidia). [ANALYST/PUNDIT] Bull: Blackwell in production, Vera Rubin ramping, revenue re-accelerated; the "architecture not a chip" moat (CUDA squeezing 3–5x throughput via software); SpaceX and the neoclouds standardizing on Nvidia; Rubin systems moving 260 TB/s (Vaughan Nelson, The MoneyFlows Show). Bear: Hugging Face deal faces real regulatory risk and won't add near-term revenue; Nvidia extending receivables to finance customers is a late-cycle tell; CDS has widened. Watch: Hugging Face regulatory path (France/EU); the January-quarter margin print; the Fed decision this week.

GOOGL (Alphabet). [REPORTER + ANALYST/PUNDIT] Bull: Blackstone's $10–20B "Crux AI" TPU cloud is a huge external vote of confidence in TPU demand; Google now sources across Broadcom, MediaTek and Marvell; warrant structures mean it captures outsized economics from its own chips (The Information's TITV). Bear: multi-sourcing adds execution risk; capex is increasingly debt-funded. Watch: Crux AI's final size and chip volume; TPU external-customer traction.

MRVL (Marvell). [ANALYST/PUNDIT] Bull: winning networking and attached content at Amazon and Google as programs proliferate (The Circuit). Bear: the Google deal reportedly carries a warrant structure that could gift much of the upside back to Google; no fresh design-win disclosure this week; "being very precise" in how it frames its Google opportunity, read: it likely doesn't have the high-margin compute tile. Watch: any hard disclosure separating compute-tile wins from networking attach.

ORCL (Oracle). [ANALYST/PUNDIT] Bull: most aggressive capex ramp of any hyperscaler, massive OpenAI deal, more "torque" to the AI trade (The Information's TITV). Bear: RBC's Rishi Jaluria (sector perform) "would rather own Microsoft"; Oracle "burning cash," dodged the free-cash-flow-breakeven question, and its "bring your own hardware" and prepayment moves look like "a band-aid solution"; AI gross margin only "30 to 40%." Watch: next month's analyst day for a real three-year roadmap and margin proof.

AMZN (Amazon) / QCOM (Qualcomm). [ANALYST/PUNDIT, relayed] New tidbit: Vaughan Nelson referenced an Amazon-Qualcomm custom-chip deal that "happened this past week," using the same warrant model. Unconfirmed by any operator on the record, chase it against filings before acting. Watch: confirmation and terms; any Trainium 3 / Project Rainier color (none this week).

LITE (Lumentum), CIEN (Ciena), AAOI (Applied Optoelectronics). [ANALYST/PUNDIT] The optics read-through, see Read-throughs below.

BE (Bloom Energy), RBRK (Rubrik), CRDO (Credo). [ANALYST/PUNDIT] Joe Albano's "beyond Nvidia" picks: Bloom Energy for behind-the-meter power (90-day SLA delivered in "50 to 60 days"); Rubrik for agentic-AI security/rollback; Credo he sees as earlier in its cycle with support "around 148" and a possible longer correction first (Investing Experts).

MSFT (Maia), META (MTIA), ARM, ALAB (Astera Labs), Alchip. [COVERAGE GAP] No dedicated podcast commentary on these programs this week. Flagging honestly rather than inventing.

Read-Throughs

Optics is the 2027 bottleneck, and the sellers are cheap after a sell-off. As racks connect more than a million chips, copper gives out. Broadcom's Tomahawk 6 moves "more than 100 trillion bits... every single second"; Nvidia's Rubin systems move 260 TB/s. The MoneyFlows Show made the case for three beaten-down optics names (The MoneyFlows Show): [ANALYST/PUNDIT]

Lumentum (LITE): lasers and modulators; last quarter revenue +109% to >$1B, operating margins 15%→37%, adjusted EPS $0.88→$3.23; guiding ~$1.25B next quarter. (Relayed: Deutsche Bank target $1,200, Evercore ISI $1,100.)

Ciena (CIEN): optical networking ("the traffic director"); revenue +37% to $1.67B, earnings +215%, FY guide raised to $6.4B; co-packaged optics ramping 2027–2028; TAM $25B→$50B by 2029.

Applied Optoelectronics (AAOI): transceivers; revenue +86% to $192M (fifth straight record quarter), 800G shipments more than doubled QoQ, demand for 800G/1.6T "exceed its production capacity through the middle of 2027"; largest U.S.-based AI transceiver maker.

Memory is the swing variable in both margin stories. The Circuit repeatedly noted that a lot of the eye-popping "dollars per gigawatt" math for both Nvidia and Broadcom is memory being lumped in, and that memory units have been flat while revenue soared, pure pricing. If the 2028 capacity wave (Micron's seven fabs, plus Samsung and SK Hynix) lands, memory pricing could reverse, a headwind to memory equities but a tailwind to accelerator margins (The Circuit). Albano separately flagged Micron DRAM/NAND "hitting peak headwinds."

TSMC and the buildout are the governor on everything. 2027 stays the peak-constraint year; TSMC has "20 fabs" under construction across nodes and packaging, but relief doesn't arrive until 2028. The whole complex "can only scale as its deepest bottleneck."

The financing plumbing is now the master dial. Hyperscaler long-end issuance rivaling U.S. Treasury flow; SPVs, warrant sweeteners, receivable/payable games, and workload-specific chips as loan collateral. This is the fault line under every name in this letter, and this week it moved from worry to hard numbers (The Compound and Friends, On The Tape with Danny Moses).

What Changed vs Last Week

A clear reversal of tempo, and a shift in what the smart money is worried about.

From a design-win flood to a financing debate. Last week's issue was built on hard chip news, the Nvidia-MediaTek deal, the Google TPU split, Broadcom's print. This week produced no comparable new silicon leak. Instead the conversation moved to how the whole buildout is being funded. That's the honest headline: a quieter week for chips, a much louder one for the money behind them.

The Broadcom read got more skeptical. Last week the coverage was about the guide-up (AI-semi to double in 2027, "line of sight" to double again in 2028). This week's second look zeroed in on the soft spots: the $115B guide sitting below the $150B bogey, the weak gross-margin defense, and, new, Broadcom finally admitting MediaTek is a real competitor at Google after denying it two quarters ago.

New: the value-capture question has a mechanism. Last week we noted Broadcom's margin compression as XPUs scale. This week we learned why the value may not flow to the designers at all: warrant structures (Google/Marvell, and a newly-referenced Amazon/Qualcomm) can hand hyperscalers their custom chips "basically for free" if volumes hit. That reframes the entire "who profits" debate.

New: TPU demand is getting financed at scale. The Blackstone-Google JV, now "Crux AI", is scaling from $5B toward $10–20B. A concrete new data point on TPU demand and a template for how this buildout gets funded off-balance-sheet.

The financing worry went from qualitative to quantified. Last week the bears (Bahnsen, Niles, Damodaran) raised the alarm in broad strokes. This week we got harder macro numbers, hyperscaler long-end issuance rivaling the U.S. Treasury (Nordvig), capex growth of 90–100% this year decelerating to 30–40% next (Dawson), and the specific accounting maneuvers (Nvidia receivables, Microsoft lease reclassification, SPVs).

The Fed catalyst is now imminent. Last week Dan Niles was watching for a possible hike; this week the market is at ~60% odds of a hike at the decision landing this week. That's the nearest-term macro catalyst for the whole complex.

The mis-dated "Mixture of Experts" glitch didn't appear this week. After three straight weeks of that future-dated CES/Rubin recap surfacing, it didn't turn up in this week's sweep, the streak paused. Nothing to exclude on that front.

What did not change: still no operator color on Microsoft's Maia, Meta's MTIA, Arm, Astera Labs, or Alchip, and now also no fresh Marvell design win or Amazon Trainium 3 detail. The silicon-specific leak machine took the week off; the financing debate did not.