Newsletter · · Ashutosh Agarwal

Wall Street Pledges 500 Billion for AI Data Centers and the Credit Market Flinched - The AI Capex Tracker - Week of August 11, 2026

AI capex and semiconductors newsletter for the week of August 11, 2026. Jensen Huang convened Goldman, BlackRock, Blackstone, KKR, Apollo and Brookfield behind a 500 billion dollar consortium to finance AI data centers, and the credit market answered by pushing Nvidia's default insurance to its biggest jump in two weeks.

The AI Capex Tracker

Week of August 11, 2026: Wall Street Pledges 500 Billion for AI Data Centers and the Credit Market Flinched

TL;DR

  • The single biggest question hanging over this whole trade, who actually pays for the $700B-and-climbing buildout, just got an answer, live on television. Jensen Huang sat down with the heads of BlackRock, Blackstone, and a who's-who of private-capital firms to unveil a consortium aiming to raise $500 billion to finance AI data centers, with Larry Fink saying plainly, "we're going to have to raise trillions of dollars over the coming years." The pitch: "compute is now an asset class," financed like mortgages or aircraft. (Closing Bell, Aug 10)
  • The credit market's reaction was the opposite of relief. Instead of easing worries about financing risk, the half-trillion-dollar announcement pushed Nvidia's cost of default insurance up "the most in two weeks," with spreads across AI-linked companies widening, a market where "credit is more skeptical than equities." (NAB Morning Call, Aug 10) The clean bear frame for it: the bubble doesn't need AI to fail, it only needs "deployable supply and capital commitments to grow faster than monetizable demand," the exact gap Oracle just illustrated by swinging from +$26B to -$24B of free cash flow in a single year. (Breaking Analysis, Aug 10)
  • But the bulls now have hard numbers, not just faith. The three big clouds are sitting on more than $1.5 trillion of committed backlog (AWS $496B, Google Cloud $514B, Azure $678B), and Amazon says it breaks even on the servers in "two to three years," then runs them at "full profit" (Equity Mates, Aug 10). Cantor's Eric Johnston adds the tell the market wanted: Microsoft and Amazon cloud growth accelerated, margins are expanding, and "very soon revenue growth is going to exceed CapEx growth." (Bloomberg Surveillance, Aug 10)

What's new

For weeks the argument has been about demand (is it real? is it circular?) and then profits (are they even cash?). This cycle it moved to the plumbing underneath all of it: the money. Who supplies the hundreds of billions, on what terms, and what the debt market thinks of the answer. Ranked by where the dollars of P&L impact land hardest.

1. Wall Street just made "compute" a financeable asset class, a $500B consortium, announced with Jensen Huang at the table. Closing Bell, Aug 10, a rare on-camera roundtable with the people who are actually the money. This is the most important development of the cycle by a wide margin.

Nvidia's Jensen Huang convened the six biggest names in finance (Goldman Sachs, BlackRock, Blackstone, KKR, Apollo, and Brookfield) around a single idea: pool their firepower into a vehicle to fund AI data centers. Larry Fink (BlackRock) put the target on the table: "We need to raise $500 billion. Obviously, that's an unprecedented amount of money, but we're going to have to raise trillions of dollars over the coming years." The economics he cited: "each gigawatt costs $50 to $60 billion to build out… in the United States alone, we're going to need over 70 gigawatts of power to fuel this."

The framing is the point. Jonathan Gray of Blackstone described compute becoming "a financialized asset class," and Fink reached for a specific historical analogy: this is "the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s… the next future for financial engineering." John Stutts of Goldman drew the comparison to how the U.S. already finances "$700 billion a year in automotive or a couple trillion dollars a year in housing," arguing lenders will start underwriting the compute itself, the way an airline loan is backed by both the airline's credit and the plane. Think of it, one panelist said, like Nvidia setting up the equivalent of GM Financial for customers who "are not called Alphabet" and can't just raise tens of billions on their own.

Two caveats a PM should hold onto. First, per the release it's a memorandum of understanding, a target, not locked-in deals; "the demand is not the issue," Jensen said, "now we've got to hustle and get all of our agreements done." Second, and this is the tell, the players are "mostly private capital players," not the big balance-sheet banks, precisely because "a lot of the customers… are going to be cash flow negative or not have as good of credit in the beginning." One host's read: this looks "almost analogous to like an aircraft leasing service where it's a little bit off balance sheet." Why it moves the thesis: the funding wall the bears kept pointing at now has a named, industrial-scale answer, but it also confirms the buildout has outgrown the hyperscalers' own cash and is being pushed into off-balance-sheet, private-credit structures. For context, one host cited Morgan Stanley's estimate that the hyperscalers alone will spend about $3.5 trillion across 2026-28.

2. The credit market didn't take the $500B news as good news. NAB Morning Call, Aug 10, a bank markets desk read.

Here's the whipsaw. The same announcement that was meant to reduce financing worry did the reverse in the bond market: "a measure of [Nvidia's] credit risk jumped the most in two weeks after the reports of this huge half a trillion dollar funding package." The desk's broader observation is the one to file: "in credit markets… the credit spreads for a number of these companies have pushed quite a bit wider over the last couple of months," a "structural disconnect where credit is more skeptical than equities" even as the earnings themselves come in "really, really solid." When the market's reaction to more available financing is to price more risk into the biggest name in the trade, that is the signal both bulls and bears have said to watch.

3. The bear framework got sharper, and Oracle is the exhibit. Breaking Analysis, Aug 10, Dave Vellante and David Floyer, a technology-supply-chain analysis show. This is the most rigorous bear case of the cycle, and notably it isn't a doom call.

Their thesis, stated cleanly: "The bubble bursting does not require AI to fail. It only requires deployable supply and capital commitments to grow faster than monetizable demand." The mechanism is what they call two clocks. The short clock (GPUs, custom chips, memory, networking) moves in quarters and is what shows up in bullish earnings prints. The long clock (land, substations, grid interconnections, power, cooling) takes a decade. A data-center building "can be physically complete while it's still waiting for transformers and switchgear." Because the clocks are out of sync, "capital can be committed, hardware can be allocated, suppliers can report strong orders and revenue, but the capacity still has to be installed" before it earns a dollar.

Their worked example is Oracle, and the numbers are stark. OpenAI reportedly committed $300 billion over five years to buy compute from Oracle starting in 2027 (part of Stargate). On Oracle's side that shows up as capex of $55.7B in fiscal 2026, up 162%, guided as high as $95B gross next year, and remaining performance obligations of $638 billion, but Oracle "expects only about 12% of that RPO to convert into revenue during the next 12 months." The kicker: Oracle's free cash flow "moved from approximately plus $26 billion… to negative $24 billion," and its credit rating slid "to the edge of investment grade." Their read on the whole memory-driven revenue boom is the same discipline: Micron's DRAM bit shipments rose only "six percent" sequentially while price per bit rose "in the low 60s," and "scarcity is doing most of the work in the reported revenue line." The bubble test they'd watch: if memory prices normalize while bit demand keeps accelerating, that's healthy (a shift from a price-led to a volume-led cycle); if prices fall and bit growth also stalls, "surplus is arriving faster than useful demand can absorb it."

4. AMD delivered a clean beat, and got sold anyway. The lesson is about the price of admission. Telltales, Aug 10, a numbers-first markets show.

AMD's quarter was, by the hosts' account, "about as clean a print as semis produced": revenue $11.5B, up 52%, gross margin 54%, net income $2.3B, EPS $1.30, and a Q3 guide of "$13 billion at the midpoint, above consensus." Beat, beat, and beat, and the stock fell 7%. Why? "Not the quarter. The price of admission." AMD went into the print at "88 times trailing-free cash flow" (on $8.7B of trailing FCF). "At 88 times, you are not buying a beat. You are buying years of uninterrupted acceleration." And management told you the data-center step-up is "second half weighted", meaning the proof arrives after the multiple already has to hold. In the same week AMD bought Talos, a Canadian startup that etches AI model weights directly into silicon, which the hosts read as management "hedging its own roadmap… the opposite of what a multiple in the high 80s is underwriting." (More on Talos in the read-throughs.) This is the single best illustration of the cycle's core tension: the fundamentals are strong, but the valuations already price them in.

5. The bull rebuttal is real: backlog is committed, ROI is showing, and depreciation is a feature not a bug. Equity Mates, Aug 10, Bloomberg Surveillance, Aug 10, and The Enterprise AI Show, Aug 9.

The bulls stopped hand-waving this cycle. The hard numbers: hyperscaler capex went from about $100B in 2020 to roughly $700B this year, and "we wouldn't be surprised if next year's figure actually goes closer to 1.2 to 1.3 trillion." Underneath that sits more than $1.5 trillion of committed backlog (AWS $496B, Google Cloud $514B, Azure $678B) which "is committed demand, which means they are legally obliged to come through" (Equity Mates). Amazon raised its capex about $20B this year "because memory prices are up so much," but Andy Jassy's line was the bullish catalyst: they "break even on the data center… server component… in about two to three years," after which "it's full profit."

The ROI is starting to appear in the actual prints. Cantor Fitzgerald's Eric Johnston: yes, capex steps up to "over a trillion dollars next year from the big four," but "what they're starting to show is the ROI… cloud growth has accelerated… you're seeing margins expand and you're seeing rent prices for compute starting to move higher," so cash flows are "likely reducing the need for as much capital markets as people expected." Microsoft's free cash flow this quarter "came in higher than what the market was expecting" (Bloomberg Surveillance). And on the depreciation bomb the bears keep flagging, a direct rebuttal: a data center "is going to be making money for like 30 years," the servers inside are depreciated over three years but "run for five or six," so "in year four… we're always going to get two or three years of essentially pure margin", and even if OpenAI and Anthropic vanish, the GPU capacity "is going to be used by someone," the dark-fiber analogy from the dot-com bust (The Enterprise AI Show).

The debate

Steel-manning both sides of the ~$700-800-billion-in-2026 (and ~$1.2-1.3-trillion-in-2027) hyperscaler capex thesis. Last week the fight was "are the profits even real?" This week it moved to the layer beneath: the money to fund it now exists at industrial scale, and the credit market got more nervous the moment it did.

Bull: the demand is committed, the returns are finally visible, and the financing just got institutionalized. Backlog is the anchor: >$1.5T of legally committed cloud demand, growing, with Amazon spelling out a two-to-three-year break-even and then years of "pure margin" (Equity Mates, Aug 10). The ROI is showing up in reported cash flow: Microsoft's FCF beat, cloud growth reaccelerated, compute rents rising, and "revenue growth is going to exceed CapEx growth" soon (Bloomberg Surveillance, Aug 10). Supply signals confirm it: TSMC and ASML showed "no momentum slowdown," and every hyperscaler that reported took capex up, not down (Earn Your Leisure, Aug 10). And the new piece: the $500B consortium means the funding is no longer a scramble of one-off SPVs and bank loans but a "one-stop" financeable asset class backed by the deepest capital pools on earth (Closing Bell, Aug 10).

Bear: the spend has outrun internal cash, the timing gap is where the risk lives, and credit is flashing first. The very existence of a $500B private-capital vehicle is the bears' point made for them: the buildout is now too big for the hyperscalers' own balance sheets and is being pushed off-balance-sheet, into non-recourse structures, funded by cash-flow-negative customers (Closing Bell, Aug 10). Vellante's framework says the pop doesn't need failure, just for "deployable supply and capital commitments to grow faster than monetizable demand," and Oracle is the live case: +$26B to -$24B FCF, $638B of RPO of which only ~12% converts within a year, a credit rating on the edge of investment grade (Breaking Analysis, Aug 10). And the tell both camps named is turning: Nvidia's credit-default cost "jumped the most in two weeks" on the funding news, with spreads wider across the group, "credit is more skeptical than equities" (NAB Morning Call, Aug 10).

The synthesis. The two sides no longer disagree about the mechanics: everyone accepts the backlog is huge, the ROI is starting to show, and the financing is moving into opaque, off-balance-sheet, private-credit form. The disagreement is about sequencing: does third-party revenue arrive to fill the committed capacity before the long-clock costs (power, depreciation, debt service) and the credit market force discipline? The most useful fact of the week is that the market gave a partial answer in real time, offered a bigger, better-organized pool of money, and the price of insuring Nvidia's debt went up. Watch spreads, not headlines.

Sell signals to watch:

  • Nvidia's CDS and the first consortium deal terms. The $500B vehicle is a target, not funded. If Nvidia's default-insurance cost keeps climbing and the first actual deals price wide, the financing that underwrites the build gets more expensive in real time, the gauge both bulls and bears named.
  • The two-clock gap. Watch four memory variables together: bit shipments, average selling prices, package lead times, package yield (Breaking Analysis, Aug 10). Prices falling while bit growth stalls is the dangerous combination.
  • Backlog concentration disclosure. The $1.5T of cloud backlog is only bullish if it isn't two frontier labs. The clouds "won't disclose it" (Equity Mates, Aug 10); any hint of single-customer risk in those numbers is a sell.
  • A clean beat that gets sold. AMD's 7% drop on a triple-beat is the template: when great fundamentals can't hold high-80s FCF multiples, the group is discounting perfection (Telltales, Aug 10).
  • A leverage-driven encore. July's Situational Awareness blow-up (below) showed how fast a levered book unwinds. The private-credit structures the $500B consortium is built on are the next place a forced deleveraging could originate.

Stocks in play

NVDA. Bull: it is now the organizing hub of the entire financing solution, Jensen personally convened the six-firm, $500B consortium, and supply-chain checks (TSMC, ASML) show "no momentum slowdown" in demand (Closing Bell, Aug 10; Earn Your Leisure, Aug 10). Bear: it is also the center of the circularity, helping finance its own cash-flow-negative customers, and its own credit-default cost "jumped the most in two weeks" the moment the funding package hit (NAB Morning Call, Aug 10). Next catalyst: its quarter around Aug 27, still the demand read for the whole complex.

AVGO. Bull: the custom-silicon theme got a marquee validation: Anthropic reportedly confirmed its own AI chips built with Broadcom (and Samsung's 2nm), and Broadcom is one of the three names "building the highways" alongside Marvell and Astera Labs as connectivity becomes the next binding constraint (The Six Five, Aug 10). Bear: no fresh company-specific print this cycle, so it trades on the complex's beta into a jumpier credit backdrop, and it was one of the "short chips" legs of July's blown-up leverage book. Next catalyst: read-through from hyperscaler and Anthropic ASIC ramps; otherwise quiet in the podcasts this cycle.

AMD. Bull: a genuinely clean beat: revenue $11.5B (+52%), EPS $1.30, Q3 guide above consensus, plus the Talos acquisition adds ultra-fast inference silicon (weights burned into the chip, claimed up to ~10-100x a GPU on specific workloads) (Telltales, Aug 10; The Six Five, Aug 10). Bear: it went in at ~88x trailing free cash flow with the data-center step-up "second half weighted," fell 7% on the beat, and lost SpaceX to an Nvidia-exclusive commitment, the Talos deal itself reads as management "hedging its own roadmap" (Telltales, Aug 10). Next catalyst: whether the Talos silicon shows up in an actual customer deployment before the multiple has to be defended again; Nvidia's Aug 27 read-through.

MSFT. Bull: the ROI poster child this cycle: Azure growth reaccelerated, free cash flow beat, and management leaned toward the lower end of its capex range (a discipline signal); Cantor's Johnston says "revenue growth is going to exceed CapEx growth" soon (Bloomberg Surveillance, Aug 10; Earn Your Leisure, Aug 10). Bear: it carries the largest single cloud backlog ($678B) with undisclosed customer concentration, and remains the anchor tenant of the OpenAI circularity (Equity Mates, Aug 10). Next catalyst: whether reported revenue growth genuinely overtakes capex growth in the next print.

GOOGL. Bull: the loudest single-name bull call of the cycle: Cleo Capital's Sarah Kunst thinks Google will "ultimately be the AI winner" given its baked-in base (Android across ~3 billion devices, 4 billion-plus users monthly) and expects it to "power the vast majority of on-device AI"; a $514B Cloud backlog and a new Pixel/phone launch this week add to it (Bloomberg Surveillance, Aug 10). Bear: it still carries the depreciation and paper-gains critiques from last week, and the backlog-concentration disclosure gap applies to it too. Next catalyst: this week's phone/AI hardware launch; whether Cloud backlog converts to margin.

AMZN. Bull: the cleanest ROI story: AWS cloud growth accelerated, a $496B backlog described as legally committed demand, and Jassy's explicit two-to-three-year break-even then "full profit" math, which the market read as a real monetization path (Equity Mates, Aug 10; The Enterprise AI Show, Aug 9). Bear: it took capex up ~$20B this year specifically because memory prices spiked (a cost, not a demand, signal), and the backlog's customer concentration is undisclosed (Equity Mates, Aug 10). Next catalyst: evidence that AWS AI demand is broad third-party rather than a few frontier labs.

META. Bull: it took capex up alongside peers and the underlying business remains a cash engine, so any pivot to spending discipline (the 2022 playbook) is a live upside lever (Earn Your Leisure, Aug 10). Bear: it's the one big spender the bulls admit "still needs a little bit more proving to do" on returns, it's openly toying with a debt-versus-equity raise (a dilution or leverage overhang either way), and Zuckerberg is talking up a free, open-source model to give away to billions, great for reach, unclear for monetization (Earn Your Leisure, Aug 10; NAB Morning Call, Aug 10). Next catalyst: the financing decision (debt vs. equity) and any concrete monetization plan for its compute.

Read-throughs

  • Memory / networking / custom silicon, the buildout's constraint keeps moving, and that's where the durable money is. The specialist take of the cycle came from Patrick Moorhead and Daniel Newman: memory demand is running "5x, 8x, 10x" what the industry can supply, a shortfall caused by years of underinvestment when memory was in glut, and "it takes a good three to four years from shovel to high-volume manufacturing." The fixes are coming but slowly: Samsung's ZHBM (memory stacked directly on the accelerator) is "the most consequential" but "four or five years" out; HBM4/HBM5 are nearer. Two custom-silicon moves matter: Anthropic confirmed its own AI chips (Samsung 2nm, reportedly with Broadcom), and AMD bought Talos, which burns model weights into the chip for extreme inference efficiency. And as memory eases, "the network becomes a binding constraint", which is why "Nvidia and AMD build the engines, Astera Labs builds the highways along with Broadcom and Marvell." Astera beat on the top and bottom line with record growth (though the stock initially sold off on guidance) (The Six Five, Aug 10). The actionable read: connectivity (AVGO, MRVL, ALAB) is the next bottleneck to own; the Nvidia print around Aug 27 is the read-through for the whole chain.
  • Memory, the cautionary version: SanDisk shows why "cheap and growing" isn't a layup. SanDisk crashed from over $2,000 in July to around $1,200, even though the fundamentals didn't break: still guiding +300-400% year-over-year revenue, adjusted EPS ~$45 (vs. $39.25 just reported), no debt, buying back stock. The catch is that growth is "flatlining" sequentially as the company signs five-year fixed contracts with hyperscalers, trading the "wild ASPs" of a shortage for stability, and "memory companies… are still cyclical." New supply is also stirring (Yangtze Memory in China, an SK Hynix/Solidigm IPO), and the next-gen High Bandwidth Flash spec (co-released Aug 3 with SK Hynix, for Google and Tenstorrent) is an end-of-decade story, not a 2026 one (Chip Stock Investor, Aug 10). The read: memory pricing power is real now, but the market is forward-looking and will fade the cycle before earnings actually roll over.
  • Optics: the recipients are round-tripping violently, and the tell is imminent. Coherent ran +44% last week and gave back 14% on Monday; Lumentum reports "tomorrow night" (i.e. this week), a day ahead of others in the group, and "they sure did move together today to the downside." The desk's framing is pure risk management: "there's risk upon risk upon risk," these are AI names with "10% ATRs," and money is quietly rotating out of chips and into software (IGV +2.3%, Twilio breakout) (Stock Market Today with IBD, Aug 10). The read: use Lumentum's print as the first optics tell of the week; the software-over-hardware rotation is the positioning story to respect.
  • Positioning: the July blow-up got its full autopsy, and the lesson is about leverage, not the thesis. New detail on the Situational Awareness fund (Leopold Aschenbrenner): seeded in 2024 with ~$200-225M, up roughly 1,000% since inception and +439% in the first half of 2026, ~$45B of assets at the start of July, then down 67% in a month to about $10B when AI infrastructure stocks fell 35% and its 4x leverage cut both ways; the short book (software names like Adobe) failed as a hedge because those stocks rose. Citadel bought the public positions "at pennies on the dollar." The one asset that survived: a private Anthropic stake ($5B) that "can't be margin called." The investing takeaway the hosts landed on is a caution for the whole "own the shovels" trade: infrastructure suppliers (memory, power) win while they're the bottleneck, "but at some point the economics turn because that is no longer the bottleneck… and the market is forward-looking" (Wall Street Wildlife, Aug 9). The read: the group's swings are still as much about leverage and forced selling as fundamentals; the private-credit plumbing behind the new $500B consortium is where the next shock could start.
  • Power: a "clean data center" pitch, and a reminder power is the real gate. On the softer, thematic end, GridServe founder Toddington Harper (now CEO of Gigatons, based in Abu Dhabi) argued you can build AI data centers directly on solar-plus-batteries with liquid cooling, faster and cheaper than the current default of "plugging in new power stations running on gas," and promised deliveries in well under a year (Everything Electric, Aug 11). It's a private, early-stage story with no direct listed-name trade, but it underscores the structural point the bears and Vellante keep making: the binding, slow-moving constraint is power and grid interconnection, which is exactly why the $500B financing math starts from "$50-60B per gigawatt" and "70+ gigawatts."

What changed vs last issue

Last issue (Monday, Aug 10, "Damodaran: Big Tech's AI profits are a mirage.") was about the quality of the reported profits, the argument that ~85-87% of Amazon's and Google's net-income growth was unrealized paper gains, that a ~$300B depreciation backlog is looming, and that "net income is now a deadly number." This cycle's new episodes moved the debate one layer down the stack, from the profits to the funding, and, importantly, the bull side finally showed up with concrete numbers.

  • The funding question got a concrete, named answer. The running "who pays for the buildout?" thread now has a specific vehicle: a $500B consortium of the six biggest private-capital firms (Goldman, BlackRock, Blackstone, KKR, Apollo, Brookfield) convened by Nvidia, explicitly framing "compute as a financeable asset class" like 1970s mortgage-backed securities, at "$50-60B per gigawatt" for "70+ gigawatts." This is new and it's the biggest development since we started tracking the financing angle.
  • The credit market repriced against the good news. Where Friday and the weekend gave us Nvidia CDS "at a record" as a static fact, this cycle we watched it move in reaction to the funding announcement, "jumped the most in two weeks", the clearest real-time confirmation yet that credit is more skeptical than equity.
  • The depreciation debate got a genuine bull rebuttal. Yesterday's issue was mostly the bear case (deferred depreciation as a hidden bomb). This cycle a hardware-savvy show pushed back with the 30-year-asset / 3-year-depreciation / "years 4-6 are pure margin" math, plus the dark-fiber "it gets used by someone" argument, the first substantive answer to the depreciation charge.
  • Committed backlog got quantified. New, hard numbers: AWS $496B, Google Cloud $514B, Azure $678B, over $1.5T of legally committed cloud demand, plus Amazon's explicit 2-3 year server break-even. This reframes the bull case from "trust the demand" to "here is the contracted demand."
  • Oracle became the bears' cleanest exhibit. New figures: FCF +$26B to -$24B in one year, capex $55.7B (+162%) guided to $95B gross, RPO $638B with only ~12% converting within 12 months, credit rating "on the edge of investment grade", a live example of commitments arriving years before cash.
  • A clean beat got sold, with a valuation number attached. AMD: rev $11.5B (+52%), EPS $1.30, Q3 guide above consensus, stock -7%, at 88x trailing FCF, data-center step-up "second half weighted," plus the Talos acquisition as roadmap hedge.
  • Custom silicon and connectivity advanced. Anthropic confirmed its own chips (Samsung 2nm, reportedly Broadcom); AMD bought Talos (weights-in-silicon inference); Astera Labs posted record growth as "the highways" alongside Broadcom and Marvell; Samsung's ZHBM is the long-term memory fix (4-5 years out).
  • Fresh, concrete data points: consortium target $500B (an MOU, not funded), $50-60B/GW, 70+ GW US power, Morgan Stanley hyperscaler capex ~$3.5T (2026-28); cloud backlog AWS $496B / Google $514B / Azure $678B (>$1.5T); aggregate capex ~$100B (2020) to ~$700B (2026) to $1.2-1.3T (2027 est.); Amazon capex +$20B on memory prices, break-even 2-3 years; Oracle FCF +$26B to -$24B, capex $55.7B (+162%) to $95B gross guide, RPO $638B (~12% converts in 12 mo); OpenAI–Oracle commitment $300B/5yr from 2027; OpenAI total forward compute commitments $600-665B through 2030; WSTS 2026 semi market ~$1.5T (>half memory); Nvidia data-center rev $75B, Broadcom AI semi $10.8B, AMD data-center $5.8B (last quarter); AMD Q2 rev $11.5B (+52%), EPS $1.30, 88x trailing FCF, stock -7%; Micron DRAM bit shipments +6% vs ASP/bit +~64%; memory demand 5-10x supply; SanDisk $2,000 to $1,200, adj EPS guide ~$45 (vs $39.25); Situational Awareness +1,000% since inception / +439% H1 / $45B to $10B, -67% in July, Anthropic stake ~$5B uncallable; retail bull targets (opinion) NVDA $250-260, Micron $1,400, SanDisk ~$2,200.
  • Still open: whether the $500B consortium converts an MOU into funded deals and at what spreads; whether Nvidia's Aug 27 quarter shows demand it isn't itself financing; whether the $1.5T backlog is broad or frontier-lab-concentrated (undisclosed); whether memory prices normalize with accelerating bit growth (healthy) or with stalling demand (dangerous); and whether Meta chooses debt or equity to fund its spend.

Next catalysts to watch: Nvidia's quarter around Aug 27 (the demand read for the whole complex, now shadowed by the financing debate); the first actual deal terms out of the $500B consortium, plus Nvidia CDS and tech-bond spreads as the real-time financing gauge; Lumentum's optics print this week; SanDisk's and Western Digital's memory prints; and Google's phone/AI hardware launch this week.