Dan Niles — Why cash makes sense, even as the AI boom continues

0:00 Teaser and set-up

• The channel's cold open strings together later lines: the Fed should raise, 40% of households own neither stocks nor a home, and "at least another year of solid growth" in AI.
• Host Ty frames the episode: a September hike, "Fed hikes, high oil and gas prices, and a bond rout are torpedoes in the water," and a 10% midterm-quarter drawdown base rate.

1:47 Listen to what Warsh says — "65 months"

• Has Warsh flipped from dove to hawk? "I'm a big believer in listen to what they say and that tells you where the head is at at the time." People change their minds when they get more data.
• The quote that settles it: "we've had 65 months of sustained elevated inflation and the responsibility for that sits squarely with the central bank."

2:44 The calendar says September 16th

• Meetings on September 16th and October 28th; midterms November 3rd. He won't hike days before a vote "because he doesn't want to irk the ire of the White House" — so "you default to he's going to raise… on September 16th."
• It is global: Japan and the UK are dealing with elevated inflation and creeping inflation expectations too. His post on the speech was titled "don't fight the Fed."

4:39 The labor market is fine; the hike is for the 40%

• Pushback on U3 at 4.1%: "normal unemployment is closer to 6% and change. You're down near four. It's really a great economy."
• "If you're raising a quarter of a point, if that's going to damage the economy, we got a lot bigger problems."
• The distributional case: 40% of households own no stocks and no home, so 65 months of inflation with no asset offset is "killing them" — "those are the people you should care about the most."

6:40 Data centers: if a quarter point kills it, AI isn't real

• "If you do [think it kills it], then this AI thing is not real, and the people who've been calling for a bubble are 100% right. But I don't think that's the case."
• Pauses are healthy: Cisco had "a couple of 30% selloffs" on its way up ~4,000% from end-1994 to March 2000.
• A pause in data-center expansion forces thought over arms-race spending — Larry Page's "I'd rather go bankrupt than lose this race" mentality produces a "race to the bottom."

7:59 Straight-up markets and the Korean leverage wipe-out

• "I always view markets when they go straight up as very problematic" — Korean retail on leverage and levered ETFs on Samsung and SK Hynix got "wiped out" in the end-June to end-July sell-off after getting "incredibly greedy between March and June."
• So the months to the midterms are welcome: "throw a little bit of cold water on the trade for now."

8:56 Read Cisco's May 2001 release

• Bookings "went from up 70% year-over-year to down 30% year-over-year in just several months" — while the internet kept growing through 2001–2002.
• The size of the prior run sets the size of the fall: NASDAQ +86% in 1999 and +24% more into 2000, then −78% over 2½ years.
• Situational Awareness, "running at four times leverage," is today's high-profile example of who gets hurt in the collapse.

10:49 More than one hike? Follow Iran and the pocketbook

• Republican polling "is horrible," and "if gasoline prices are up a lot, the sitting party gets killed in the midterms."
• The precedent: Iran held the hostages 444 days and released them "just a couple of hours into President Ronald Reagan being sworn into office" after Carter lost 489–49.
• So "this Strait of Hormuz situation is going to continue to be a problem till at least the midterms are done" — elevated oil seeps into inflation, and the global data-center build adds resource pressure. "It wouldn't surprise me if there's more than one rate hike coming."

13:08 One soft payrolls print won't change it

• "You have 65 months of bad data. I don't think one month is going to do it."

14:44 The bond back-up is a debt problem first

• "You left out in my mind the most important part": a 6% US deficit, "the highest deficit outside of a major war," and $40T of debt against ~$33T GDP; Japan and the UK at multi-decade-high yields.
• GFC stimulus never backed off; COVID was "stimulus on steroids" and produced the fastest inflation since the 1970s.
• No fix without a crisis: politicians "only have one goal… to get reelected," voters got used to "free stuff," and the democratic socialists gaining seats is "not good for big business" or the stock market.

18:55 No S&P range — a midterm base rate instead

• From 1990 through 2025, end-July to November 9th: median peak-to-trough drawdown "10% in midterm election years, it's only 5% in non-midterm election years… double what it normally is."
• The sample includes the 1990 Gulf War, which he notes has "some similarities here."

20:50 The June 20th speed-bump call, and why the AI numbers still work

• He turned negative on June 20th on AI-bill control and open-weight models — "one heck of a speed bump from there to the end of July."
• Since end-May: cost per million tokens down more than 50%, tokens produced up more than 2.5x — "that's more than offset that drop."
• AWS, Azure and Google Cloud accelerated from 35% to 43% with operating margins up ~2 points; Anthropic reached (adjusted) profitability in the June quarter.

22:45 Debt funding is the scary part; the bubble still ends in bankruptcies

• "You've got Google hitting the debt markets and the equity markets. They went cash flow negative for the first time since they went public."
• Every life-changing technology — canals, railroads, radio, TV, internet — attracts overinvestment, so "when you ultimately start to plateau… you're going to have the bubble break," but "you can make a lot of money if that ultimate peak is a couple of years out."

25:11 China wins on energy; the US holds on models

• Energy: "the central government decides this is what we're going to do" — no community or politicians in the way, unlike the US data-center backlash into the midterms.
• Models: cut off from Nvidia, Chinese labs got good at low-cost models, but "Gemini Flash… is going to be pretty darn good in terms of low-cost token production," and Anthropic and OpenAI lead the frontier.
• The Ford vs Ferrari point: ~90% of use goes to low-end models; the top 5–10% still wants the best — "costs come down, volumes go up."

28:31 China will overtake Korea in memory

• "Yes. Because for one simple reason it is a national priority" — chip supply is "a national defense issue."
• Not leading edge, but Japan in the 1970s–80s and Korea in the 1990s weren't either. CXMT has listed, YMTC lists "shortly," and China's ~20% share of world PCs and smartphones gives a captive market.

30:54 Apple: lucky, not smart — and an upgrade cycle next year

• Apple's AI promised in 2024 is still US-only, English-only and beta; foldables arrive ~7 years after Samsung's.
• But licensing Gemini "for a couple of billion" instead of spending $200B "worked out very well," and a "product guy" as CEO is welcome since the stock's run has been multiple expansion.
• "I think you're going to get a huge upgrade cycle sometime next year when you do get this out of beta," with localized versions worldwide. "You can beat smart, but you can't beat lucky."

33:53 Where to shelter: money-market cash

• "Cash is always not a bad option. And by cash I mean you can put your cash in a money market fund and get some pretty good yields."
• Competition for equities is global: Japanese investors can finally earn a reasonable, safe yield at home — "that's the math that everybody's doing around the world."
• Like 2022, when stocks and bonds both fell, cash is viable again, and money markets "have a pretty solid yield no matter what country you're in."

34:50 An earnings bubble, not a multiple bubble

• Valuations are high relative to GDP or normalized earnings "because these companies are all over-earning to some degree."
• "Having memory companies having operating margins of 80% is not normal" — once China reaches its memory ambitions, "those margins will go back to below the average."

36:21 Humility: stocks overshoot both ways

• When he downgraded everything he covered in 2000 he "didn't think NASDAQ would go down 78%."
• Volkswagen quadrupled in two days in the GFC short squeeze; GameStop more recently — "but they can also go down way more than you ever imagined possible as well."

37:31 Seven months into agentic — not the peak

• Agentic "got formalized with OpenClaw on January 30th… it's hard for me to believe we've reached a fundamental peak when we're 7 months into people trying to figure out how to use agentic" — "at least another year of solid growth."
• The bear case he respects: token costs down 50% without matching usage growth, and data centers with nowhere to put the chips.

38:13 Broadcom's miss — and why not bonds

• Broadcom "reported after the close, but the numbers were nowhere near what people thought… that stock's going to go down probably tomorrow."
• "I don't pretend to be smarter than the market" — with a good chance of "a more than double normal drawdown," "the only safe thing right now is sitting in cash because part of the reason we're worried about the stock market is because bonds are selling off."