Dan Niles: Why Cash Makes Sense, Even as the AI Boom Continues

Take central bankers at their word: what a policymaker says publicly tells you where their head is at that moment, even if they said the opposite to get appointed.

Niles says people say one thing to get appointed, then change their mind as new data arrives; the Fed chair's speech citing 65 months of sustained elevated inflation and saying responsibility 'sits squarely with the central bank' is 'about as clear as you can have it'.

Dan Niles expects the Fed to hike at the September 16, 2026 meeting rather than wait.

The chair's speech flagged 65 months of sustained elevated inflation as the central bank's responsibility; the only other meeting, October 28, falls days before the November 3 midterms, so he 'doesn't want to irk the eye of the White House' — so you default to September.

When the Fed is raising rates, don't fight it — position with policy, not against it.

Niles says he posted 'don't fight the Fed' after seeing the chair's speech; he believes a hiking Fed puts a lot of pressure on the stock market.

The tightening is global, not just American, which compounds the pressure on equities.

Niles says Japan is going to raise as well, and the UK and many central banks are dealing with elevated inflation, partly because inflation expectations seem to be creeping higher.

Validator: anchor is 14 words (6-12 asked); link kept

The US labour market is genuinely strong, so weak-labour arguments against a hike don't hold up for Niles.

He says 'normal' unemployment is closer to 6% and change while the US is near four; you can always cut the numbers to find something that disagrees, but it's 'really a great economy'.

The Fed should hike because sustained inflation hits hardest the 40% of US households that own neither stocks nor a home.

Niles argues inflation also inflates stock and home prices, so asset owners are partly offset; the 40% without assets just face higher heating, food and gas costs, and 65 months of this is 'super painful' — 'those are the people you should care about the most'.

Validator: anchor is 13 words (6-12 asked); link kept

A quarter-point move cannot plausibly break the AI/data-centre buildout — if it could, the AI thesis was never real.

Niles' logic: if bulls believe a quarter point kills data-centre demand, then 'the people who've been calling for a bubble are 100% right'; he doesn't think that's the case, and says the same about the economy — if a quarter point damages it, we have much bigger problems.

Validator: anchor is 13 words (6-12 asked); link kept

Deep drawdowns inside a secular tech boom are normal and healthy, not evidence the trend is over.

Niles cites Cisco having a couple of 30% selloffs on its way to being up roughly 4,000% from the end of 1994 to its March 2000 peak; a pause in data-centre expansion is likewise 'good'.

AI capex is driven by an arms-race mentality that ratchets spending upward, so a capex pause is a healthy discipline, not a red flag.

Niles quotes Larry Page's 'I'd rather go bankrupt than lose this race'; without a pause you get one company expanding 80%, the next 90%, the next 100%, a race to the bottom, instead of asking whether new methods mean less square footage is needed.

Validator: anchor is 13 words (6-12 asked); link kept

Markets that go straight up are dangerous because they breed leveraged greed, and the damage lands on retail when the move reverses.

Niles says people and markets swing between fear and greed; Korean investors on leverage or in levered ETFs on Samsung and Hynix got wiped out in the sell-off from end-June to end-July after getting 'incredibly greedy' between March and June, helped by promoters who never mention the risks.

A technology can keep growing while the stocks and orders collapse — bookings can swing violently even when end demand never stops.

Niles points to Cisco's May 2001 earnings release, where bookings went from +70% year-over-year to -30% in just several months; the internet didn't stop growing in 2001-02, but because NASDAQ was up 86% in 1999 and another 24% to start 2000 it still fell 78% over two and a half years.

Niles believes the Strait of Hormuz / Iran problem persists at least until the November 2026 midterms are done.

High oil and gasoline prices historically get the sitting party killed in midterms, and Republican polling is already horrible; the 1979-81 precedent is Iran holding hostages 444 days despite sanctions and releasing them hours into Reagan's inauguration after Carter was destroyed in the polls (roughly 489 electoral votes to 49). Voters ultimately vote on whether they can feed the family and fill the car.

Validator: anchor is 13 words (6-12 asked); link kept

More than one rate hike would not surprise Niles, though he won't forecast a series.

If oil stays elevated it keeps seeping into inflation, and continued data-centre growth in the US and globally sucks up resources, adding upward price pressure; plus inflation has been above the 2% target for 65 months and is sustained. He prefers to watch the data and market reaction and adjust.

A single soft payrolls or benign CPI print before the September 16 meeting will not stop the hike.

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

The back-end yield problem is fundamentally about government debt, not just AI borrowing crowding out the long end.

Niles says the US runs a 6% deficit — the highest outside a major war — with $40 trillion of debt against roughly $33 trillion of GDP, and it's global: Japan's bond yields are the highest in multiple decades and the UK similar. Deficit-running began in the global financial crisis, was never unwound, then COVID was 'stimulus on steroids', producing the fastest inflation since the 1970s in 2021-22.

Deficits won't be fixed until a crisis forces it, because politicians in every country optimise for re-election rather than the national interest.

Niles argues voters got used to 'free stuff' during COVID and point to a fine stock market as proof it can continue, forgetting that it produced the fastest inflation in 40 years; politicians who promise free everything are the ones who win.

Validator: anchor is 13 words (6-12 asked); link kept

The rise of democratic socialists in the US is a negative for big business and therefore for the stock market.

Niles notes they are gaining many seats on a platform that explicitly says big business is bad, and the political winds point to deficits going up, not down.

Midterm-year seasonality implies roughly double the usual drawdown into early November — 10% versus 5%.

Niles' own study covering 1990 through the end of 2025, measuring end-July to November 9: median peak-to-trough drawdown is about 5% in normal years but 10% in midterm election years; there are outliers on both sides and the sample includes the 1990 Gulf War, which he sees some similarity to.

AI unit economics are improving, not deteriorating: token price collapse has been more than offset by volume, and the cloud businesses are accelerating with fatter margins.

Cost per million tokens is down more than 50% from end-May to now while tokens produced are up more than 2.5x; AWS, Azure and Google Cloud growth accelerated from 35% in the March quarter to 43% year-over-year in the June quarter, with operating margins expanding 2%; Anthropic says it reached (adjusted) profitability in the June quarter.

Validator: anchor is 14 words (6-12 asked); link kept

Falling unit costs driving rising volume and expanding profits is how technology markets are supposed to work — price deflation alone is not a bear signal.

Niles: costs come down, volumes go up because more people can afford it, and hopefully profits expand with it in a healthy market; he frames the 50% token-cost decline plus 2.5x volume growth plus improving profitability as exactly that pattern.

Funding AI capex with debt is the scary part of the buildout; independent AI-lab profitability is what makes it sustainable.

Niles notes Google has been hitting the debt and equity markets and went cash-flow negative for the first time since going public; against that, Anthropic reaching profitability 'has to make you feel better about the sustainability'.

If AI is genuinely a life-changing technology, a bubble break with bankruptcies is the inevitable endgame — but large money can still be made if the peak is years away.

Niles' chain: companies that believe they'd rather go bankrupt than lose the race will over-invest, so when growth rates finally plateau the bubble breaks and some go bust — same pattern as canals, railroads, radio, TV and the internet.

This is not the fundamental peak of AI — Niles expects at least another year of solid growth.

He dates the formalisation of agentic AI to around January 30 (which he associates with 'Open Claw'), only about seven months ago, and says it's hard to believe we've peaked when people are only seven months into figuring out how to use agentic AI; he acknowledges many smart people think the bubble breaks now.

Falsifier: If token costs keep falling at the recent 50% pace and token usage stops going up, 'we could have a real problem'.

China has a structural edge in AI energy because its central government can site power without local or electoral veto.

Niles: Beijing can decide to put a nuclear reactor somewhere and it happens — no community, no politicians; in the US, data centres face pushback precisely because midterms are coming and politicians flip from supporting to freezing or opposing projects to avoid being voted out.

US labs won't fall far behind China at the cheap end of the model market, so commoditised low-cost models aren't a US-specific threat.

Chinese labs got smart on algorithms because they were cut off from Nvidia's latest chips — necessity is the mother of invention — but Niles expects Google's Gemini Flash to be 'pretty darn good' at low-cost token production, while the US keeps the frontier lead with Anthropic and OpenAI.

Validator: anchor is 13 words (6-12 asked); link kept

Match the model tier to the task: cheap models for the bulk of workloads, frontier models only for the hardest few percent.

Niles' analogy: you don't take a Ferrari to the corner store for milk, a Ford works just as well; companies early on threw frontier models at things like summarising email, which is too expensive. He expects roughly 90% of usage at the low end and the top 5-10% of use cases still paying for the best models.

Niles is 100% confident China will build its own chip supply including high-bandwidth memory, and can take meaningful share without being leading edge.

It's a national defence priority after being cut off from US chips — 'as important as having a battleship or an aircraft carrier' — which was never true for Japan or Korea. Japan in the 1970s-80s and Korea in the 1990s also entered without leading-edge product. CXMT is already public and YMTC goes public shortly, raising more capital, and China consumes roughly 20% of the world's PCs and 20% of smartphones, so 20% share is achievable domestically with no need to export.

Memory makers' ~80% operating margins are abnormal and will fall below the long-run average once China reaches its memory ambitions.

Niles 'firmly believes' this; he frames it as the clearest example of companies over-earning today.

Global equity risk today is an earnings bubble more than a multiple bubble.

Niles says valuations are very high relative to GDP or to normalized earnings because companies are all over-earning to some degree — so the multiple looks less extreme only because the 'E' is inflated.

Validator: anchor is 15 words (6-12 asked); link kept

Niles' shelter for this window is cash in money market funds, held into the early-November US midterms.

Yields on money market funds are solid in most countries; bonds aren't the alternative because part of the reason to worry about stocks is that bonds are selling off; valuations are high globally and a more-than-double-normal drawdown is plausible. He calls cash 'the only safe thing right now' and notes it worked in 2022 when both stocks and bonds were losing. No size, entry or exit level given.

Falsifier: He says the picture would be different if valuations weren't as high as they are across the globe.

Validator: anchor is 14 words (6-12 asked); link kept

Money market yields are now a genuine global competitor to equities, which is itself part of why stocks are struggling.

Niles' example: a Japanese investor who arguably could not get a reasonable yield on Japanese bonds for three decades can now get one, safely — 'that's the math that everybody's doing around the world'.

Validator: anchor is 13 words (6-12 asked); link kept

Respect the market rather than trying to outsmart it: stocks overshoot in both directions far beyond what you can imagine, so never assume you know where the peak is.

Niles cites Volkswagen quadrupling in two days during the global financial crisis on a short squeeze to become the most valuable company in the world, and GameStop; and his own 2000 downgrade where he never expected NASDAQ to fall 78% peak to trough.

Apple's choice not to spend hundreds of billions on AI capex has worked out well, though Niles doubts it was planned.

Apple is licensing Gemini for the iPhone from Google for a rumoured couple of billion dollars while Google spends around $200 billion, so Apple isn't burning cash flow; the stock's strength this year came largely from multiple expansion rather than products, and Niles thinks Google is the best AI company across the whole ecosystem.

Validator: anchor is 13 words (6-12 asked); link kept

Niles expects a huge Apple upgrade cycle sometime in 2027 once the AI Siri leaves beta and goes multi-language.

The AI Siri arriving is US-only, English-only and in beta; the upgrade comes when it ships in Chinese and other native languages worldwide with a better product line. Apple is late — Samsung has had AI phones for years and introduced a foldable around 2019 — but ecosystem stickiness makes Apple users very hard to move to Android.

The September 16, 2026 FOMC meeting is the decision point Niles expects to deliver a hike.

He rules out the following meeting on October 28 because it is only days before the November 3 midterms.

Treat the window from now to the early-November US midterms as the high-risk stretch for equities.

Niles stacks Fed hikes, elevated oil from the Iran/Hormuz situation, selling-off bonds and high valuations, against a midterm-year median drawdown of 10%.

If token costs keep falling at the recent 50% pace while token usage stops rising, that is the signal of a real problem for the AI trade.

Niles presents this as the strongest bear argument alongside data centres that can't site their chips; he cites the >50% token cost decline since end-May as the starting point.

Validator: anchor is 13 words (6-12 asked); link kept

Apple's AI Siri is due next week (week of September 7, 2026) but only in the US, only in English, and in beta.

Apple promised an AI model back in 2024 and Niles says we're still waiting; the limited launch is why he sees the real upgrade cycle only in 2027.

Watch YMTC's upcoming IPO as the next capital-raising step in China's domestic chip push.

Niles notes CXMT has already gone public and YMTC goes public shortly, which will help them raise even more capital for the national chip effort.