Newsletter · · Ashutosh Agarwal

Wall Street's Newest AI Fight Is Memory vs Optics - The AI Capex Tracker - Week of August 13, 2026

AI capex and semiconductors newsletter for the week of August 13, 2026. A fresh intramural rotation split the AI hardware trade as Nvidia's Rubin roadmap favors optics over memory, while Nvidia's credit-default-swap cost nearly doubled and a venture capitalist reframed the bear case around the data center as a depreciating utility.

The AI Capex Tracker

Week of August 13, 2026: Wall Street's Newest AI Fight Is Memory vs Optics

Issue: Thursday, August 13, 2026

TL;DR

  • A brand-new fight broke out inside the AI trade: is the next bottleneck memory or optics? Over the weekend, financial Twitter argued that memory chips (Micron, SK Hynix) are rolling over: Korean sellers are being forced out, and Nvidia is putting less memory on its next system, Rubin, and connecting the racks with light (optics) instead. Monday it hit the market: optics stocks opened up 4–5% and memory down 4–5%, then whipsawed so hard that optics name Coherent closed down 14% while memory bounced. The one sober read: optics has the cleaner multi-year story, but don't trade the narrative. (The AI Investor Podcast, Aug 12)
  • For the first time, the bond market (not the pundits) flinched on Nvidia. The cost to insure Nvidia's debt against default nearly doubled in a month, from about 0.40% to 0.78% a year, even though the company threw off roughly $50 billion of cash last quarter. And when Nvidia unveiled its $500 billion financing plan, Micron, normally the first stock to rip on any AI-spend headline, actually fell. When the usual winners stop responding to good news, that's the tell. (RiskReversal Pod, Aug 12)
  • The bear case got its most rigorous framework yet, from a venture capitalist, not a doomster. Paul Kedrosky argues a data center is nothing like an apartment building: it's a depreciating utility whose product (AI "tokens") gets 70–80% cheaper every year, needing "around 100 million-fold growth over the next six years" in usage just to stay even. His scariest point for anyone who owns an index fund: AI-linked names are now "something like 40-odd percent" of the S&P 500, so "you're in it whether you like it or not." (Big Technology Podcast, Aug 12)

What's new

A note on the window: this is the Thursday issue, sweeping roughly the last 24 hours. Yesterday's issue leaned on podcasts published Aug 11-12: the 70% customer-concentration bombshell (Eisman, Zitron), the depreciation math ($434B spent vs $149B booked), and T. Rowe's Dom Rizzo defending the spend. Everything below is genuinely newer (published Aug 12) and, crucially, the argument moved sideways into new territory: away from "who's buying the compute" and toward two fresh questions: which part of the supply chain wins from here (memory or optics?), and is the credit market finally starting to price the risk? Ranked by where the dollars of P&L land hardest for a book.

1. The memory-vs-optics rotation is the freshest, most tradable thing that happened this week. The AI Investor Podcast, Aug 12, hosts Eric and Austin, technology-focused investors. This is the one to internalize because it actually moved stocks in real time.

For 18 months, memory (Micron, SK Hynix) and optics (Coherent, Lumentum, Fabrinet) were the two best-performing corners of the AI hardware trade. Over the weekend, a widely followed analyst ("Jukin," who writes for the research shop Citrini) argued that the baton is passing from memory to optics, for three reasons. First, memory is being hit by forced selling out of Korea: "many people being margin called and money coming out of leverage ETFs attached to memory," plus fund investors still pulling money after the recent Situational Awareness hedge-fund blow-up. Second (and this is the structural one), "Nvidia is reducing the amount of memory they have on Rubin systems. And they're going to focus more on tying these racks together with optics. So it would be memory's loss is optics gains." Third, the consensus that "memory will peak in the next two years."

Then the tell: when the market opened Monday, "the optics stocks were all up 4% or 5%. Memory stocks were down 4% or 5%. And then within 15 minutes, the situation had completely flipped where optics stocks Coherent closed on Monday down 14%. And most of the memory stocks were up." Eric's takeaway is the useful one for a PM: don't trade the narrative: "it certainly seems like some people were waiting to pounce on people buying into these areas." But he flagged a genuine yellow light on memory: SanDisk said on its call it will merely hold gross margins at ~85%, and "historically, gross margin peaks have been the sign to get out of memory… if they're peaking, well, you're already normally too late." The one hedge against that old rule: memory makers are now signing long-term contracts, which caps the upside but may also break the historic bust pattern. Why it moves numbers: this is the first real intramural rotation within the winning trade, and the "less memory, more optics on Rubin" point is a durable, Nvidia-driven reason to prefer the optics names into 2027.

2. On optics specifically, Lumentum's call was about as bullish as these get. The AI Investor Podcast, Aug 12.

Lumentum fell 5% on the initial print, then ran (up 14% intraday Wednesday) once the call landed. The line that captures the scale of what's coming: "for one major hyperscaler, the network capacity connecting just two AI data center sites could double the total global backbone capacity they built over the entirety of the last decade." Management said visibility has sharpened (rebutting the earlier "co-packaged optics is delayed" fears), that its biggest customer, Nvidia, brings demand in the second half of 2027, with 800-volt products shipping in the back half of 2026, and that it is "running well ahead on almost every metric on revenue, on margin, on operating margin," with a new financial target due at an upcoming conference. It has even had to go back to a key supplier (AXT) to lock up more supply. The read: optics is the cleaner structural growth story right now, and Lumentum has the technology lead in exactly the pieces (800V, near-package optics) that Nvidia's roadmap needs.

3. The $500B Nvidia financing plan got its sharpest single-analyst autopsy, and the verdict is "it's just debt." Prof G Markets, Aug 12, Jay Goldberg, analyst at Seaport Global Securities. This matters because Goldberg is a specialist who walked through the actual mechanics rather than just the vibe.

Goldberg's history lesson: Nvidia has quietly been financing its own customers since 2022: first easy payment terms for the early "neoclouds" like CoreWeave, then outright backstops ("if you can't use it all, we'll buy some percentage of it"). It already carries "$30 billion of what they call compute service agreements" off the balance sheet. The problem the $500B plan is really trying to solve: lenders have never accepted the GPUs themselves as collateral (they wanted a Microsoft-style customer contract behind the loan), and now the hyperscalers are pulling back from signing those neocloud deals because "they want to use their balance sheet for their own purposes." So Jensen Huang is trying to convince Wall Street that "chips are an investable asset class": good enough collateral to lend against. Goldberg's blunt bottom line: "ultimately, it's debt… And I think debt is very pro-cyclical. When times are good, like they are now… this is going to amp that up considerably. The problem is when the cycle turns and the cycle always turns, this kind of thing will amplify the pain on the downside." He does give Nvidia real credit on one hotly debated point: older hardware "doesn't depreciate quite as quickly as the bears will say," because Nvidia keeps squeezing more tokens out of old systems through software. But he doesn't think "GPUs as collateral" will land with lenders, "which is ultimately why Nvidia is going to take on some form of obligation here," a guarantee he pegs (from Jensen's own blog post) at around 25% coverage. Why it matters: it reframes the $500B not as new demand but as Nvidia adding hidden, pro-cyclical leverage to its own P&L.

4. And the bond market started to agree: Nvidia's "insurance cost" nearly doubled. RiskReversal Pod, Aug 12, hosts Dan Nathan and Guy Adami, markets commentators/traders.

This is the first genuinely market-based crack, not another opinion. The price of a credit default swap on Nvidia (essentially an insurance policy against the company defaulting) "has nearly doubled the last month," from about 40 basis points to "about 77 and a half." In plain terms: it now costs roughly $77,500 a year to insure $10 million of Nvidia debt, up from ~$40,000, "for one of the best credits on the face of the earth," a company that "generated like $50 billion in free cash flow last quarter alone." Bond investors are getting "slightly uneasy about all the backstopping." The hosts relayed a wry new coinage for the $500B structure from CNBC's Ron Insana, "collateralized chip obligations" (the AI era's answer to the collateralized debt obligations of 2008), and noted Jim Chanos was circulating the same worry. The single cleanest equity tell in the whole episode: on the day of the bullish $500B roundtable, "Micron should be up on the day. Micron is not up on the day. And Micron is actually now down slightly." When the stock that used to rip on any AI-spend headline shrugs, "it seems to be falling on deaf ears": the point of diminishing marginal returns. Why it matters: this is the earliest hard evidence that the financing pile-up is starting to cost something, and it lands on Nvidia's own credit and on the memory names first.

5. Anthropic's $9 billion power deal reads as defensive hoarding, and the IPO race is now a land grab for capital. Elon Musk Podcast, Aug 12, a markets/tech explainer show.

Anthropic finalized a $9 billion deal with Riot Platforms for 191 megawatts of dedicated data-center power ("enough electricity to power a medium-sized city") at a site Riot originally built for Bitcoin mining. The episode's framing is the useful part: is writing a $9B check to a single power provider "a confident assertion of guaranteed future growth… or a defensive hoarding maneuver driven by a fear of missing out on raw compute capacity?" Their answer leans toward the latter: "tying up 191 megawatts for yourself also means actively denying that exact same electrical capacity to everyone else… an arms race where you buy every resource you can find simply so the other side cannot have it." Meanwhile the two frontier labs are running opposite capital playbooks: Anthropic is racing toward what "could be the largest initial public offering in history" (whoever lists first "gets to define the metrics for Wall Street"), while OpenAI just did "a $7 billion employee share buyback using their own balance sheet, keeping their valuation totally flat at $852 billion" and rejecting outside money from SoftBank and Thrive. The investor worries the show flagged are the now-familiar three: cheaper Chinese models undercutting price, U.S. export-control friction, and local "not in my backyard" resistance to the physical buildout. Why it matters: the Riot deal is another data point that frontier labs are buying power for strategic denial, not just measured demand, which is exactly the pattern that ends in a glut.

The debate

Steel-manning both sides of the ~$800-billion-in-2026 (heading toward ~$1.5-trillion-in-2027) hyperscaler capex thesis. Yesterday the fight was about demand concentration: how much of it is two loss-making labs. This cycle a venture capitalist reframed the whole bear case around the nature of the asset, and the credit market gave the bears their first real-world data point.

Bull: the demand is physical, visible, and supply is still short. The clearest bull evidence this cycle was in the plumbing, not the pundits: neocloud Nebius is "signing large deals at $20 to $25 billion per gigawatt" on long-term contracts, and reported alongside CoreWeave sent both stocks up double digits (Nebius +14%, CoreWeave +22%) (The AI Investor Podcast, Aug 12). Intel's surprise $15B stock sale to fund capacity was "massively oversubscribed… more than $100 billion in interest": when capital is offered to this build, "it has ravenous demand for it" (The AI Investor Podcast, Aug 12). On power, EPRI's Arshad Mansoor will bet a curry that the US adds 100 gigawatts of new demand by 2032 ("two and a half UKs"), a bet the hyperscalers wouldn't make if the returns weren't there (Cleaning Up, Aug 12). And even the $500B skeptics concede one point to Jensen: older GPUs are lasting and earning longer than the bears assumed (Prof G Markets, Aug 12).

Bear: the asset depreciates, the product deflates, and the debt is everywhere. Paul Kedrosky's framework is the most complete bear articulation to date, and worth quoting at length because it's a genuinely new lens (Big Technology Podcast, Aug 12):

"This is much more like a utility, a non-regulated utility who has continuing capital requirements which continually dilute the returns because you're having to raise more capital all the way down the path."

His three-part case: (1) duration mismatch: a data center looks like a long-lived asset but most of its guts (the GPUs) need "wholesale replacement… anywhere from a four to seven year period," some failing on an 18-month cycle if they were used hard for training; (2) a deflating product: "these tokens are among the most rapidly depreciating assets we've ever seen… falling 70% to 80% year over year… for at least the last four years," so the Jevons-paradox rescue ("cheaper tokens, so we use more") would need "around 100 million-fold growth over the next six years… Is it possible? Absolutely. Is it likely? No"; and (3) systemic spread: because AI-related names are now "something like 40-odd percent" of the S&P 500 and hyperscaler debt is a fast-growing share of both investment-grade and high-yield issuance, "mathematically, they must be holding this stuff," and it will "metastasize" GFC-style into PIMCO, then insurers, then European banks. Kedrosky won't short it himself ("I get more pleasure from sleeping well at night"), but disclosed he's been working with "two very large hedge funds who put on some fairly complicated positions… for close to a year now," with a roughly 18-month timeline.

The synthesis. The bull and bear now agree on more than they'd admit. Both sides accept that older GPUs last longer than the crude 3-year depreciation scare implied, that's a point for the bulls. But both also accept that the product price falls relentlessly and the financing is getting more circular, points for the bears. The real disagreement has narrowed to one variable: does token volume grow fast enough to outrun token price collapse before the financing has to be refinanced at a worse point in the cycle? The tell that the market is starting to lean bearish isn't a quote: it's Micron failing to rally on the $500B news and Nvidia's CDS doubling.

Sell signals to watch:

  • Memory gross-margin peak. SanDisk merely holding ~85% margins is, by the old memory-cycle rulebook, a top signal, unless long-term contracts have genuinely broken the pattern this time (The AI Investor Podcast, Aug 12).
  • Nvidia CDS and the winners going quiet. If Nvidia's default-insurance cost keeps climbing and Micron/DRAM names keep shrugging off bullish headlines, the equity market is telling you the marginal buyer is done (RiskReversal Pod, Aug 12).
  • A bond-market wobble, not a stock wobble. The Cleaning Up host's warning is the one to sit with: a stock-market correction here is near-certain and survivable, but "what happens when it's a bond market correction? That feels much more threatening," because the hyperscalers are "betting the whole balance sheet" (Cleaning Up, Aug 12).
  • Chinese open-weight price war. Still the live wire: cheaper Chinese models pressuring the labs that ~70% of hyperscaler AI revenue depends on (Elon Musk Podcast, Aug 12).
  • The Anthropic IPO print. Whether it lists near its targeted valuation (and sets the multiple for the whole complex) or slips like OpenAI's did (Elon Musk Podcast, Aug 12).

Stocks in play

NVDA. Bull: still the demand hub and cash engine (~$50B free cash flow last quarter), and it keeps extending the useful life of older systems through software, softening the depreciation attack (RiskReversal Pod, Aug 12; Prof G Markets, Aug 12). Bear: it is the center of the circularity: ~$30B of off-balance-sheet compute-service agreements, a ~25% backstop on the $500B plan ("ultimately, it's debt… pro-cyclical"), and now its own default-insurance cost has nearly doubled to ~0.78% (Prof G Markets, Aug 12; RiskReversal Pod, Aug 12). Next catalyst: its quarter in late August (~Aug 26-27): the demand read for the entire complex.

AVGO. Bull: custom silicon is the credible escape from Nvidia lock-in, and it just got a fresh data point: Microsoft is building its Maya 300 accelerator with Marvell as design partner, targeting "300,000 chips in 2027," proof the ASIC ecosystem is real and broadening (The AI Investor Podcast, Aug 12). Bear: no fresh Broadcom-specific print this cycle; it rides the complex's beta into a jumpier credit backdrop. Next catalyst: read-through from Nvidia's late-August quarter and hyperscaler ASIC ramps, otherwise quiet on the tape this cycle.

AMD. Bull/Bear: quiet on the tape this cycle, no fresh AMD-specific operator or analyst signal in the last 24 hours of podcasts. It carries the same valuation-perfection and complex-beta risk as the group. Next catalyst: whether its second-half-weighted data-center step-up shows before the multiple has to be defended again; Nvidia's late-August read-through.

MSFT. Bull: it's leaning into custom silicon (the Marvell-built Maya 300, 300k chips in 2027) to lower its Nvidia bill and scale harder than the market appreciates (The AI Investor Podcast, Aug 12). Bear: the OpenAI dependence is stark: per Zitron, "70% of Microsoft's AI revenue in fiscal year 26 was OpenAI," about $34B of AI revenue total, which implies Microsoft 365 Copilot is "a single digit billions" business against "$261 billion of CapEx Microsoft has spent since the beginning of 2022" (Squawk Pod, Aug 12). Next catalyst: any sign Copilot/Azure demand is broadening beyond OpenAI.

GOOGL. Bull: Gemini crossed "1 billion users" per Sundar Pichai, though much of that is Gmail integration, not standalone paid demand (Squawk Pod, Aug 12). Bear: the same concentration flagged yesterday, now with a forward number: UBS estimates "48% of Google Cloud's entire revenue next year will come from Anthropic and OpenAI… over $124 billion" (Squawk Pod, Aug 12). Next catalyst: whether Cloud growth converts to real free cash flow rather than more equity/debt.

AMZN. Bull/Bear: mostly quiet on the tape for fresh AWS-specific color this cycle; it remains the demand anchor most bulls cite, and its Trainium chips are part of the "escape Nvidia" ASIC story. The bear overhang is unchanged: how much of AWS AI growth is a few frontier labs. Next catalyst: evidence AWS AI demand is broad third-party, not two labs.

META. Bull: still spending to defend a core ad business compounding in the mid-20s. Bear: it's the poster child for the off-balance-sheet financing that makes the bears nervous: its $30 billion Louisiana data center is being built through a special-purpose vehicle ("Beignet") where the outside investor (Blue Owl) funds ~80% and Meta ~20% and leases the facility back, keeping the debt off Meta's own books; this against a hyperscaler bond-issuance surge from ~$20B (all of 2024) to ~$178B (2026 year-to-date) (The Zach Foust Show, Aug 12, retail commentary; the Beignet/Blue Owl structure itself is widely reported). Next catalyst: whether ad-revenue acceleration keeps outrunning the capex line into the next print.

Read-throughs

  • Memory vs optics: the actionable trade of the week, and it favors optics. The structural reason to prefer optics is Nvidia-driven: it's putting less memory on Rubin and connecting racks with light instead, so "memory's loss is optics gains" (The AI Investor Podcast, Aug 12). Lumentum (with Coherent and Fabrinet) has the technology lead into an 800V ramp (H2 2026) and Nvidia demand landing H2 2027; the demand math is almost absurd: one hyperscaler's link between two sites can double the entire global backbone it built over the last decade. On memory, respect the old rule (SanDisk holding ~85% margins = a classic peak signal) but weigh the new one (long-term contracts may cap upside and smooth the bust). The tradable posture: overweight optics, treat memory as a trim-into-strength rather than a chase, and don't get whipsawed by intraday narrative flips like Monday's.
  • Semicap equipment and neoclouds are the second-derivative winners. Chip-equipment names ripped this week (CamTech +19%, Anto/BE +28%) as the "build the factories" theme broadened, and Intel's oversubscribed $15B raise (>$100B of interest, likely upsized to ~$20B) shows capital is still ravenous for capacity (The AI Investor Podcast, Aug 12). Neoclouds Nebius (+14%) and CoreWeave (+22%) confirmed pricing at "$20 to $25 billion per gigawatt" on long contracts: bullish for compute pricing, but note these smaller names have the most fragile financing if the spot market ever reprices.
  • Power: demand is staggering, but so is the constraint list, and flexibility is the new unlock. EPRI's Arshad Mansoor pegs 100 GW of new US demand by ~2032 against a ~450 GW peak grid today ("two and a half UKs"), while only ~2.5 GW of AI compute was actually added last year; energy investor Jigar Shah's constraint-stacked math lands at ~34 GW by 2030 (Cleaning Up, Aug 12). The interesting operational shift: data centers used to demand 24/7 flat power; now "speed to power" is forcing flexibility: backup generators running renewable diesel (a tailwind for Caterpillar and Wärtsilä), batteries, and microgrids, so a project can energize in three years instead of five. At ~$60B per gigawatt, anything that shortens the interconnect wait is worth real money to the developers.
  • Credit is the quiet macro tell. Beyond Nvidia's doubling CDS, the mechanism the bears keep drawing is the GFC playbook: hyperscaler and neocloud debt packaged and sold on until it sits inside investment-grade funds, insurers and pension books: "collateralized chip obligations" (RiskReversal Pod, Aug 12; Big Technology Podcast, Aug 12). The read: as long as the majors keep printing at investment-grade spreads it stays "manageable"; the first names whose spreads drift toward junk are the canaries.

What changed vs last issue

Last issue (Wednesday, Aug 12, "Big Tech's AI now leans on two loss-making startups.") nailed the demand concentration: ~70% of hyperscaler AI revenue from OpenAI and Anthropic, the $434B-spent-vs-$149B-depreciated gap, free cash flow set to fall 91%, and T. Rowe's Dom Rizzo defending the spend as a "1998 buying opportunity." This cycle the debate moved off "who's buying" and onto which suppliers win from here, and whether the credit market is finally pricing the risk.

  • A whole new sub-trade appeared: memory vs optics. Yesterday this wasn't a topic. Today it's the most tradable development: a rotation call (Citrini's "Jukin") that memory is topping while optics wins, driven by Nvidia putting less memory on Rubin; Monday's violent intraday whipsaw (optics +4–5% then Coherent −14%; memory reversed up); SanDisk's ~85% margins as a classic peak signal; and a strongly bullish Lumentum call (Nvidia demand H2 2027, 800V H2 2026, "two sites = double the last decade of backbone").
  • The credit market gave the first hard number. New vs yesterday's qualitative worries: Nvidia's 5-year CDS nearly doubled in a month (~40 → ~77.5 bps) despite ~$50B/quarter FCF; Micron fell on the bullish $500B news day; "collateralized chip obligations" enters the lexicon.
  • The bear case got a rigorous framework, from a VC. New: Paul Kedrosky's data-center-as-depreciating-utility model: GPUs needing wholesale replacement every 4–7 years (some failing at 18 months), tokens deflating "70% to 80% year over year," needing "~100 million-fold growth over the next six years" to offset, and AI now "something like 40-odd percent" of the S&P 500 so index holders are "in it whether you like it or not," with GFC-style spread into IG/HY debt.
  • The $500B plan got dissected by a named analyst. New detail from Seaport's Jay Goldberg: ~$30B of off-balance-sheet compute-service agreements already, a ~25% Nvidia backstop, and the core problem: hyperscalers pulling back from signing the neocloud contracts that used to be the collateral, so Jensen is pitching "chips as an investable asset class" to fill the gap. Verdict: "ultimately, it's debt… pro-cyclical."
  • Anthropic's power deal re-read as a land grab; OpenAI's capital move contrasts. New: the $9B/191 MW Anthropic–Riot deal framed as "defensive hoarding… so the other side cannot have it"; OpenAI's $7B employee buyback at a flat $852B valuation, rejecting SoftBank/Thrive; the IPO race as a fight to set Wall Street's multiple first.
  • Meta's off-balance-sheet financing got specific. New: the "Beignet" SPV for Meta's ~$30B Louisiana data center (Blue Owl ~80% / Meta ~20%, leased back), and the hyperscaler bond-issuance surge from ~$20B (2024) to ~$178B (2026 YTD), plus capex from ~$18B/quarter (four hyperscalers, 2019) to ~$130B/quarter (Q1 2026).
  • Fresh, concrete data points: Nvidia CDS ~40 → ~77.5 bps (≈$40k → ≈$77.5k to insure $10M); Micron down on the $500B day; SanDisk holding ~85% gross margin; Nvidia reducing memory content on Rubin in favor of optics; Lumentum: Nvidia demand H2 2027, 800V H2 2026, "two sites double the decade of backbone"; Coherent closed −14% Monday intraday; Nebius +14% / CoreWeave +22%, deals at "$20–25B per gigawatt"; Intel $15B raise, >$100B interest, likely ~$20B; CamTech +19%, Anto/BE +28%; Microsoft Maya 300 ASIC with Marvell, "300,000 chips in 2027"; MSFT AI revenue ~$34B FY26 (~70% OpenAI), Copilot "single digit billions," $261B capex since 2022; UBS 48% of Google Cloud revenue from labs next year (>$124B); Gemini 1B users; Anthropic–Riot $9B / 191 MW; OpenAI $7B buyback at flat $852B; Kedrosky's GPU replacement 4–7 yrs / 18-mo MTBF, tokens −70–80%/yr, "~100 million-fold" token growth needed, AI ≈40% of S&P 500; Goldberg's ~$30B off-balance-sheet compute agreements, ~25% backstop; EPRI 100 GW new US demand by 2032 vs ~450 GW peak, ~2.5 GW added last year, Jigar Shah ~34 GW by 2030, ~$60B/GW; Meta "Beignet" SPV ~$30B (Blue Owl 80/20); hyperscaler bond issuance ~$20B (2024) → ~$178B (2026 YTD); hyperscaler capex ~$18B/qtr (2019) → ~$130B/qtr (Q1 2026); OpenAI planning ~$750B compute spend by end-2030 (WSJ, via Zitron).
  • Quiet vs yesterday: the Texas ERCOT/SB6 audit produced no fresh Aug-12 podcast signal (the specialist legal read is from Aug 11, already covered), so the this-week PUC/ERCOT meeting and the end-September refundability rule remain the pending catalysts, not new developments.
  • Still open: whether the memory-vs-optics rotation is a durable trend or a one-week Twitter squall; whether Nvidia's CDS keeps widening and drags the complex; whether the Anthropic IPO lands near target or slips; whether Kedrosky's 18-month clock is right; and whether Nvidia's late-August quarter shows demand it isn't itself financing.

Next catalysts to watch: Nvidia's quarter in late August (~Aug 26-27): the demand read for the whole complex, now shadowed by the concentration and credit debates; this week's PUC/ERCOT meeting on the Texas data-center audit and the end-of-September SB6 refundability rule; the Anthropic IPO as the first real test of frontier-lab financeability; and Nvidia's CDS and hyperscaler credit spreads as the earliest-warning gauge for the financing plumbing.