Newsletter · · Ashutosh Agarwal

Big Tech's AI Now Leans on Two Loss-Making Startups - The AI Capex Tracker - Week of August 12, 2026

AI capex and semiconductors newsletter for August 12, 2026. Steve Eisman and Ed Zitron put a number on the concentration risk, roughly 70 percent of hyperscaler AI revenue coming from OpenAI and Anthropic, while the depreciation gap got measured at 434 billion spent against 149 billion booked, and Texas halted its fast-track data-center interconnection queue.

The AI Capex Tracker

Week of August 12, 2026: Big Tech's AI Now Leans on Two Loss-Making Startups

Issue: Wednesday, August 12, 2026

TL;DR

  • The bear case stopped being a vibe and became a number this cycle, and it's the same number from very different mouths. Roughly 70% of the AI revenue at Microsoft, Amazon, Google and Oracle comes from just two customers: OpenAI and Anthropic, two private companies that lose billions a year. "Big Short" investor Steve Eisman called that the "Achilles heel of this whole story" on CNBC (CNBC's Fast Money, Aug 11); AI researcher Ed Zitron, citing UBS, Barclays and Wells Fargo, put hard sub-numbers on it across three separate shows (Thoughtful Money, Aug 11).
  • The accounting gap the bears keep pointing at just got measured. The four biggest hyperscalers bought about $434 billion of property and equipment in the year to March 2026 but have booked only $149 billion of depreciation so far, and their combined free cash flow is set to fall more than 91% this year even as reported net income rises about 25% (Many Happy Returns, Aug 12). Cash is going out the door now; the profit is on paper.
  • The bulls have a real answer, and it's specific: the money is being spent because the returns are showing up. T. Rowe Price's $8bn tech PM Dom Rizzo says this is "a lot like the 1998 sell-off," a buying opportunity mid-cycle, with hyperscaler capex heading from ~$800B this year to $1.5–1.6 trillion in 2027 precisely because AWS, Azure and Google Cloud are re-accelerating (37% / 43% / 82%) at ~50% incremental margins and roughly 30% returns on the build (Excess Returns, Aug 11).

What's new

A note on the window: this is the Wednesday issue, sweeping roughly the last 24 hours. Yesterday's issue leaned on podcasts published Aug 9–10 (the day the $500B financing consortium was unveiled on TV). Everything below is genuinely newer, published Aug 11 and Aug 12, and the story moved in a clear direction. For a week the argument climbed the stack: first "is the demand real?", then "are the profits even cash?", then "who funds the buildout?" This cycle it landed on the sharpest question of all: who is actually buying the compute, and can they pay? The answer, from bulls and bears alike, keeps coming back to two names: OpenAI and Anthropic. Ranked by where the dollars of P&L impact land hardest.

1. The concentration risk yesterday's issue called "undisclosed" now has a number, about 70%, and it came from a short-seller on live TV. CNBC's Fast Money, Aug 11, Steve Eisman, the "Big Short" investor and former Neuberger Berman PM. This is the most important development of the cycle, because it turns a soft worry into a measurable one.

On the $500B financing consortium, Eisman was almost bored: "It's not a new asset class. It's a securitization of something that produces revenue. Period… People securitize songs. So if you can securitize songs, you can securitize an AI data center that, God willing, produces revenue." In other words, don't lose sleep over the plumbing.

What he would lose sleep over is the customer list. "Something like 70% of all AI revenue for Microsoft, Amazon, Google, Oracle is from Anthropic and OpenAI, period. Just those two. And then if you dig down even further, it's 70% of AI revenue. It's 25 to 35 percent of their cloud revenue. So the futures of these massive companies, in a sense, are a bet that OpenAI [and] Anthropic is going to succeed. Private companies whose financials, until they go public, are hidden to almost everybody… we know that they lose billions like it's water."

The trigger he's watching: "the Chinese open-weight models are much cheaper. And if they start really taking a lot of market share, and it sounds like, from what I'm hearing, that they're starting to, you could have a big price war. And then we have a problem. That, I think, is the Achilles heel of the story." He was careful on timing, the short-seller's eternal enemy: "that story could happen. But it could be a year from now… and it may not happen at all." Why it moves the thesis: the single biggest unknown in the bull case, is the $1.5T of cloud backlog broad demand or a couple of frontier labs, just got answered in the bears' favor, by someone with no ideological axe to grind about AI itself.

2. Ed Zitron carpet-bombed the bull case across three shows in 24 hours, and the sub-numbers behind the 70% are worse than the headline. Thoughtful Money, Aug 11, Tom Bilyeu's Impact Theory, Aug 11, and Better Offline, Aug 12. Zitron is a full-throated AI bear, so treat the tone as opinion; the underlying figures are sourced to UBS, Barclays and Wells Fargo.

The forensic version of Eisman's 70%: per UBS, 27% of Google Cloud's revenue this year is OpenAI and Anthropic, rising to over 48% next year, more than $124 billion. Per Barclays, AWS is 13% this year, 18% next. And on Microsoft, using his own reporting on OpenAI: "69% of the year-over-year growth of Microsoft Intelligent Cloud segment [in calendar 2025] was actually from OpenAI. Without that, it would have only grown 8% year-over-year, which is barely beating inflation" (Tom Bilyeu, Aug 11).

His mechanism, stripped of the profanity: the future growth of the hyperscalers is contingent on two companies' ability to keep spending more every year, "over $200 billion in compute in the next year. Probably $300, $400 billion in the future," while losing tens of billions and needing "to become the size of Microsoft and Google… and profitable" for the math to work, which he thinks is "impossible." OpenAI, he noted, "had to raise $217 billion in the first six months of this year" to spend roughly $100B on compute. And even if the labs somehow win, "we have still overbuilt capacity by about 10 times" (Thoughtful Money, Aug 11).

On the $500B consortium itself, Zitron was blunt that it is the tell, not the cure: "this is absolutely circular financing. Nvidia is helping companies raise debt, offering a guarantee on that debt… all so that Nvidia can be paid using the proceeds from that debt which Nvidia is guaranteeing. Its customers are companies that can quite literally not afford to buy its products." He read the deal as memoranda of understanding "not worth the paper they're printed on," and drew the Stargate parallel: the $500B announced with fanfare in January 2025 that turned into a $30B Nvidia investment in OpenAI and "zero gigawatts of data center capacity." His punchline is a set of consensus revenue estimates he finds absurd: Nvidia at ~$216B in FY2026, with the Street modeling $393.7B (FY27), $565.7B (FY28), $694B (FY29), over $1.6 trillion of GPUs sold by January 2029, with 90%+ of revenue from data-center hardware (Better Offline, Aug 12).

The most useful investing frame came from host Tom Bilyeu, who is not a bear: the risk isn't fraud, it's where it lands. The debt is being sliced into CLOs and mezzanine tranches and sold to "pension funds, life insurance," so "the banks aren't holding the risk, the public is." His verdict: "Not illegal, but whoa, is this risky?" (Tom Bilyeu, Aug 11).

3. The depreciation "bomb" got weighed: $434B spent, $149B booked, and free cash flow set to fall 91%. Many Happy Returns, Aug 12, Robin Wigglesworth, the FT columnist, on an even-handed markets show.

This is the cleanest quantification yet of the "profits are on a slower clock than the spending" argument. The four big hyperscalers "bought about $434 billion of property and equipment in the four quarters to March 2026… but so far, we've only had $149 billion of reported depreciation." Translation: a large slug of cost has been incurred but hasn't hit the income statement yet. Look at the cash instead and it's stark: hyperscaler free cash flow is "projected to fall more than 91% this year, even as their official net income… will rise by around 25%."

Alphabet is his exhibit: it "just reported… negative free cash flow. So almost $6 billion more went out of the door than came in. It raised around $70 billion in new equity. It suspended its buyback and it doubled its long-term debt. And it lifted its CapEx guidance to around $200 billion for this year." His broader warning to anyone leaning on cheap-looking multiples: Alphabet trades at ~18x trailing earnings, "but if you normalize those earnings and stripped out its stakes in SpaceX and Anthropic, it's more like 26 times." The trailing numbers look artificially cheap because the earnings are artificially high. Why it matters: it hands the bears a specific, model-able figure for the cash drain, and a reason to distrust the "it's not even expensive" bull retort.

4. The bull rebuttal is genuinely strong, and it comes from an operator who runs $8bn, not a pundit. Excess Returns, Aug 11, Dom Rizzo, portfolio manager of T. Rowe Price's global technology strategy.

Rizzo's framing is that this is a mid-cycle scare, not the top: "this feels a lot like the 1998 sell-off… which proved to be an incredible buying opportunity." His tell that it isn't a repeat of the dot-com bust: in 1998 the chip industry saw an 8% revenue decline; in 2026 it's a 64% revenue increase, yet the chip index fell a similar ~30–40% both times. He maps the ChatGPT era onto the Netscape-to-2000 timeline and lands us at "literally the halfway point."

The numbers underneath his call: hyperscaler capex grows about 75% in 2026 to roughly $800 billion, and he expects an acceleration to $1.5–1.6 trillion in 2027, even assuming memory prices calm down, "because the ROIC is just so attractive." His read on why the market rewarded Microsoft and Amazon but punished Meta is all about visible returns: Amazon broke out that its short-lifecycle assets (chips, networking) "break even in roughly two to three years and then last five to six years," with AWS accelerating to 37% at 50% incremental operating margins, Azure at 43% guiding to 45%, and Google Cloud growing 82% at 50% incremental margins, plus Satya Nadella publicly blessing a Morgan Stanley estimate of ~30% ROIC across the hyperscalers. Crucially, Rizzo names the same pitfall the bears do, "OpenAI is already something like 25% of Azure revenue… mid-30s of the backlog, probably even higher as a percentage of net new bookings," but thinks the risk is "a few years out at least." Notably, both sides now agree on the facts; they disagree only on the timing.

5. Texas just threw sand in the gears of the data-center land grab, a real, this-week regulatory catalyst. NPM Interconnections, Aug 11, Bracewell LLP energy partners Bryan Clark and Jared Berg, a specialist power-project legal show.

On August 3, Texas Governor Abbott directed the state's utility regulator (PUCT) and grid operator (ERCOT) to run a project-by-project audit of every data center trying to connect to the Texas grid, checking power usage, water usage, and local land-use rules, and in response ERCOT halted its "batch zero" process (the fast-track batch study meant to clear the flood of large-load interconnection requests). It applies to all data centers, not just the big ones (in Texas a "large load" is 75 megawatts or more). Tellingly, the biggest developers, Amazon, Google, Crusoe, QTS, came out in support of the audit, and the PUC chairman called an open meeting this week to hammer out the process.

The money at stake is concrete: developers put up about $50,000 per megawatt to enter batch zero, and the draft interconnection rule under Senate Bill 6 currently makes 80% of that deposit non-refundable if a project is abandoned before it's energized, though PUC staff have signaled that could drop to 20% non-refundable, with a final rule expected by end of September. The lawyers' plain read: this "is injecting a lot of uncertainty into projects that already have a lot of uncertainty," and some developers "may or may not have the unilateral authority" to stay in, because their deposits were financed with debt or equity that has its own triggers. Why it matters: Texas is the single most important build-out venue, and this is the first hard sign that the "power is the real gate" thesis now has a regulatory brake attached to it. Watch this week's PUC meeting.

The debate

Steel-manning both sides of the ~$800-billion-in-2026 (and ~$1.5-trillion-in-2027) hyperscaler capex thesis. Last week the fight was about the funding, who supplies the half-trillion. This week it narrowed to the thing the funding is supposed to serve: the demand, and specifically how much of it is two money-losing startups.

Bull: the returns are visibly showing up, the spend is disciplined where it needs to be, and the multiples aren't stretched on earnings. The clouds are re-accelerating and expanding margins at the same time: AWS to 37%, Azure to 43–45%, Google Cloud +82%, all around 50% incremental operating margins, with management pointing to ~30% ROIC on the build and 2–3 year payback on the short-life gear (Excess Returns, Aug 11). The asset itself is proving more durable than the accountants assume: the useful life of a GPU is now argued at 5–10 years, up from the 3–4 years people assumed a year ago, which stretches the payback and supports Jensen Huang's "GPUs are a productive utility asset" pitch (Morning Call, Aug 11). And even a self-described optimist macro strategist argues this is an "earnings bubble," not a "price bubble": Nvidia's earnings are growing faster than its stock, the NASDAQ-100 trades below its 10-year average P/E, so "it doesn't look like a price bubble" (Top Traders Unplugged, Aug 12).

Bear: the growth is dangerously concentrated, the cash is draining, and the financing is being pushed onto the public. The demand isn't broad, it's ~70% two unprofitable startups whose survival is the real bet (CNBC's Fast Money, Aug 11; Thoughtful Money, Aug 11). The cash has already turned: FCF set to fall 91% this year, $434B spent vs $149B depreciated, Alphabet cash-flow negative and issuing $70B of equity (Many Happy Returns, Aug 12). And the "compute as an asset class" pitch runs into an awkward law of computing, The Verge's Nilay Patel: "It is unclear to me why the price of that asset would ever go up… The price of compute goes down over time," as supply from TPUs, Trainium and new Nvidia parts floods in (The Exchange, Aug 11). Even the historical analogy cuts against the spenders: in the dot-com fiber boom Nortel, Juniper and Alcatel "massively invested… created a supply glut" and prices fell 90%, "the winners were not those who overinvested," they were the companies that used the cheap capacity later (Top Traders Unplugged, Aug 12).

The synthesis. The striking thing this cycle is that the bulls and bears now agree on the facts. Dom Rizzo (bull) and Steve Eisman (bear) cite the same ~25–35% cloud / ~70% AI-revenue dependence on OpenAI and Anthropic. They disagree only on two things: whether the labs can keep raising and spending long enough for third-party demand to arrive, and whether cheap Chinese open-weight models break the price umbrella first. The tell to watch, on the other side of the debate, was on The Exchange: RBC's Srini Pajuri made the strong bull case (GPU rental rates of "$5 to $11," Nvidia throwing off "more than $200 billion of free cash flow in the next 12 months," supply "very, very tight" for a couple of years), and Nilay Patel simply asked what would make the price of compute go up, and got no clean answer.

Sell signals to watch:

  • Chinese open-weight price war. Eisman's named Achilles heel. If Kimi, DeepSeek-class models take real enterprise share and force a token price war, the OpenAI/Anthropic revenue that ~70% of hyperscaler AI depends on compresses fast (CNBC's Fast Money, Aug 11).
  • The Anthropic IPO print. Anthropic is targeting a September/early-October IPO at a $965B valuation on a $47B annualized run rate, and is already having to reassure investors about Chinese competition, Trump-administration friction, and data-center backlash (The Rundown, Aug 11). A weak reception, or a delay, as OpenAI just did into 2027, is the first place the "labs can always raise" assumption gets tested.
  • The FCF-vs-net-income gap. If reported net income keeps rising ~25% while free cash flow falls ~91%, the market is being asked to pay for accounting profit, not cash. Watch whether depreciation starts catching up to the $434B of spend (Many Happy Returns, Aug 12).
  • Private-credit stress. Even a bullish strategist flagged that private-credit funds are showing "the highest loan defaults level since at least 2021," the same pipes the $500B consortium's non-recourse structures will run through (Morning Call, Aug 11).
  • The Texas audit outcome. If the PUC's SB6 rule lands at 80% non-refundable and the batch-zero pause drags, expect some Texas projects to be abandoned and the "power is the gate" trade to get a real-world stress test (NPM Interconnections, Aug 11).

Stocks in play

NVDA. Bull: it remains the demand hub and the cash engine. RBC pegs a ~$350B revenue run-rate this year and $200B+ of free cash flow over the next 12 months, GPU rental rates of "$5 to $11" imply attractive returns, and Nvidia's own memory access is best-in-class into the Rubin ramp (The Exchange, Aug 11). Bear: it is the center of the circularity, offering "residual value support" for up to 25% of financed deals so cash-flow-negative customers can buy its chips, and the consensus revenue path ($393.7B → $565.7B → $694B through FY29) is what the bears call "insane"; Eisman frames it as "a single point of failure" that would have to dip into its ~$50B/quarter cash flow if demand slips (Better Offline, Aug 12; CNBC's Fast Money, Aug 11). Next catalyst: its quarter in late August (around Aug 26–27), still the demand read for the entire complex.

AVGO. Bull: the structural case is that hyperscaler custom silicon (Broadcom-built ASICs, alongside Google's TPUs and Amazon's Trainium) is the credible alternative to Nvidia lock-in, and it's the counterweight the whole "escape Nvidia" trade rests on (The Exchange, Aug 11). Bear: no fresh company-specific print this cycle, it rides the complex's beta into a jumpier credit backdrop and the same lab-concentration risk as its hyperscaler customers. Next catalyst: read-through from Nvidia's late-August quarter and hyperscaler ASIC ramps; otherwise quiet on the tape this cycle.

AMD. Bull: it's now openly named as part of the competitive escape hatch from Nvidia (Radeon alongside Google's TPUs and Amazon's Trainium), the theme bulls point to as evidence the ecosystem is broadening (CNBC's Fast Money, Aug 11). Bear: no new print since last week's clean-beat-sold-anyway (‑7% on a triple beat at ~88x trailing free cash flow), and it carries the same valuation-perfection risk as the rest of the group. Next catalyst: whether its second-half-weighted data-center step-up shows up before the multiple has to be defended again; Nvidia's late-August read-through. Otherwise quiet on the tape.

MSFT. Bull: the ROI poster child. Azure re-accelerated to ~43% guiding to 45% at ~50% incremental margins, it's AAA-rated (higher than US Treasuries), and Satya Nadella publicly endorsed the ~30% hyperscaler ROIC math (Excess Returns, Aug 11). Bear: it is the most OpenAI-dependent of all, "69% of the year-over-year growth of Microsoft Intelligent Cloud [in 2025] was actually from OpenAI," and roughly 76% of its revenue growth through mid-2028 is projected to come from Intelligent Cloud, which is increasingly OpenAI's spend (Thoughtful Money, Aug 11). Next catalyst: any sign Azure growth is broadening beyond OpenAI; the Anthropic IPO as a read on frontier-lab health.

GOOGL. Bull: Google Cloud is the fastest-growing of the three big clouds, +82% at ~50% incremental margins, and management is spending into demonstrable returns (Excess Returns, Aug 11). Bear: it turned free-cash-flow negative for the first time since its IPO (~$6B outflow), raised ~$70B of equity, suspended its buyback, doubled long-term debt, and lifted capex guidance to ~$200B, and stripped of its SpaceX/Anthropic stakes it trades at ~26x, not the ~18x the screen shows; UBS sees 48% of Google Cloud revenue coming from OpenAI and Anthropic next year (Many Happy Returns, Aug 12; Tom Bilyeu, Aug 11). Next catalyst: whether Cloud growth converts to actual free cash flow rather than more equity raises.

AMZN. Bull: the cleanest monetization story. AWS re-accelerating to 37% at 50% incremental margins, with Jassy's explicit 2–3 year break-even then 5–6 years of useful life on the short-life gear; it's also constructively engaging Texas by supporting the ERCOT data-center audit rather than fighting it (Excess Returns, Aug 11; NPM Interconnections, Aug 11). Bear: Barclays sees AWS's OpenAI+Anthropic mix rising to 18% next year, and Amazon reportedly spent ~$54B of capex in a single quarter, cash out now, payback later (Tom Bilyeu, Aug 11). Next catalyst: evidence AWS AI demand is broad third-party, not a few labs.

META. Bull: it's spending to defend a core ad business that's still compounding in the mid-20s, and it's being singled out by policymakers as doing power procurement "the right way" via long-term off-take agreements that are unlocking nuclear and gas projects (Excess Returns, Aug 11; Thematic Investors, Aug 11). Bear: it's the name the market punished on capex because, unlike AWS/Azure/GCP, the spend is to "make the core business better," not to feed an obviously accelerating cloud line, so the ROI is the hardest to see, and Zitron flags Meta's growth is already slowing (Excess Returns, Aug 11; Thoughtful Money, Aug 11). Next catalyst: whether ad-revenue acceleration keeps outrunning the capex line into next print.

Read-throughs

  • Power and grid: the story shifted from "how do we build it" to "who's allowed to build it," and Texas is ground zero. The most actionable new development is regulatory: Texas halted its fast-track data-center interconnection queue pending an audit, with $50,000/MW deposits and an 80%-vs-20% refundability question hanging over developers (NPM Interconnections, Aug 11). Zoom out and the demand is staggering: the White House's own benchmark is 100,000 MW of new generation needed by 2030, "an entire additional Texas power grid," with about half from data centers, against a backdrop of the fastest US power-demand growth in 30–50 years (3–4% a year) (Thematic Investors, Aug 11). The actionable trade the guests laid out: large, established generators (gas, renewables, battery, nuclear) are the structural winners because "there's limited labor, limited equipment, limited land, limited permitting," new turbines are a 4–5 year wait, and the market stays tight into late 2028/2029. Contracted generation portfolios trade at just 7–11x free cash flow with 12–20 year contracts and mid-teens unlevered returns, the cheapest, most contracted way to own the buildout. And the new federal "fast-track" rule (curtailable data centers get interconnected in 60–90 days instead of 6–8 years) is the underappreciated accelerant to watch as FERC implements it.
  • The financing plumbing: "compute as an asset class" met its first serious intellectual pushback, and it's about physics, not sentiment. RBC's Srini Pajuri made the case that GPUs are a real, cash-generating asset (rental rates $5–11, older A100s still earning good rates well past their accounting depreciation). Nilay Patel's rebuttal is the one to sit with: an asset class needs its price to be able to rise, and "the price of compute goes down over time" as supply floods in and chips get more efficient (The Exchange, Aug 11). The GPU-lifecycle debate is the crux: extend useful life to 5–10 years and the financing works; hold it at 3–4 and it doesn't (Morning Call, Aug 11). The read: the durability of the depreciation assumption is the trade, it's the single variable that both makes the securitizations investment-grade and keeps hyperscaler net income above cash flow.
  • The credit market is the quiet tell: hyperscaler debt now rivals the banks, and one name sits a notch above junk. The big five hyperscalers are issuing $100–200B a year of debt, on track to match JPMorgan and Citi as an investment-grade sector; ratings run from Microsoft (AAA) down to Oracle (BBB-, "just a notch above junk"), with Oracle and CoreWeave the names whose spreads are drifting toward high-yield (Getting Credit, Aug 11; Top Traders Unplugged, Aug 12). The read: as long as the majors keep printing at investment-grade spreads it stays "manageable"; Oracle and CoreWeave are the canaries.
  • The Bitcoin-miner-to-neocloud pivot is becoming a real supply source, and a way to play the power bottleneck. Riot Platforms landed a 20-year, $9.1 billion compute deal with Anthropic for 191 MW at its Rockdale, Texas site (+15% on the day, +40% YTD), and Kiel Infrastructure (ex-BitFarms) is demolishing its Bitcoin facilities to rebuild as AI data centers, keeping only the thing that matters, the power purchase agreements (The Rundown, Aug 11; Morning Call, Aug 11). The read: former miners sitting on secured power and interconnects are quietly becoming the fastest new AI capacity, the power rights are the asset, not the GPUs.

What changed vs last issue

Last issue (Tuesday, Aug 11, "Wall Street pledges $500B for AI. Credit flinched.") was about the supply of money, the $500B financing consortium unveiled with Jensen Huang at the table, and a credit market that got more nervous the moment more financing appeared. This cycle the debate moved to the source of the demand that money is meant to serve, and the answer, named repeatedly, is two loss-making startups.

  • The "undisclosed backlog concentration" from yesterday now has a number: ~70%. Yesterday we flagged that the $1.5T of cloud backlog was only bullish if it wasn't two frontier labs, and the clouds "won't disclose it." This cycle both a short-seller (Eisman) and a bear-analyst (Zitron, sourcing UBS/Barclays/Wells Fargo) put the figure at ~70% of AI revenue / 25–35% of cloud revenue from OpenAI and Anthropic, with sub-splits: UBS 27%→48% for Google Cloud, Barclays 13%→18% for AWS, 69% of Microsoft Intelligent Cloud's 2025 growth from OpenAI alone. The biggest single change of the week.
  • The depreciation "bomb" got measured. New hard figures: $434B of property/equipment bought vs only $149B depreciation booked in the year to March 2026; free cash flow set to fall 91% this year while net income rises ~25%. Yesterday this was a qualitative worry; now it's a number.
  • Alphabet's cash turn got specific. New: negative free cash flow (~$6B outflow), ~$70B equity raise, buyback suspended, long-term debt doubled, capex guided to ~$200B, and a ~26x normalized multiple vs the ~18x trailing screen once SpaceX/Anthropic stakes are stripped out.
  • The $500B consortium got dissected rather than celebrated. New detail vs yesterday's announcement: the mechanism is GPUs-as-collateral with Nvidia backstopping up to 25% of individual deals; it's memoranda of understanding, not contracts, with "different vehicles rather than one large collaboration"; and the near-unanimous podcast read, even from non-bears, is that this is vendor/circular financing whose risk is being distributed to pension funds and 401(k)s via CLOs.
  • "Compute as an asset class" met its counter-argument. New: The Verge's Nilay Patel's point that compute prices structurally fall, so it's unclear why the asset would ever appreciate, the first substantive intellectual pushback on the framing that anchored yesterday's issue.
  • A genuinely new power catalyst: Texas hit pause. New: Governor Abbott's Aug 3 directive and ERCOT's halt of the batch-zero interconnection process, a project-by-project audit backed by Amazon/Google/Crusoe/QTS, $50,000/MW deposits at risk, SB6 refundability set to move from 80% to possibly 20% non-refundable, PUC meeting this week. Plus the White House's 100 GW-by-2030 benchmark and the federal fast-track (curtailable to 60–90 day interconnect).
  • Anthropic's IPO clock advanced. New: September/early-October target at a $965B valuation on a $47B run-rate, with investors pressing it on Chinese competition, Trump-administration friction, and data-center backlash, while OpenAI slips its own IPO to 2027. Plus the Anthropic and Riot $9.1B, 20-year, 191 MW Texas compute deal.
  • The bull case got an operator, not a pundit. New: T. Rowe's Dom Rizzo framing this as a "1998 buying opportunity," capex accelerating to $1.5–1.6T in 2027, and the AWS/Azure/GCP acceleration (37% / 43% / 82%) at ~50% incremental margins with ~30% ROIC, the most credible bull rebuttal to the week's bear barrage.