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

Niles expects the Fed under Chair Warsh to raise rates at the September 16, 2026 meeting rather than on October 28, which falls just days before the November 3 midterms.

Warsh said the central bank is squarely responsible for 65 months of sustained elevated inflation. A hike on October 28, days before the midterms, would anger the White House, so September is the window. Japan and the UK are also hiking, which adds pressure on stocks. One soft payrolls or CPI report won't change it after 65 months of bad data.

Don't fight the Fed: take central bankers at their word about where their heads are, and don't bet against stocks' headwind when the Fed is raising.

Niles believes in listening to what officials say, because it shows their current thinking even if they said something else to get appointed. If the Fed is hiking, you don't position against it.

Long-running inflation hurts households that own no stocks or homes the most, because inflation also pushes up asset prices and only asset owners get that offset.

About 40% of US households own neither stocks nor a home. They pay more for heat, food, and gas without any gain in asset prices to make up for it.

The Fed should hike: the job market is strong, and if a quarter point damages the economy, the economy is far more fragile than anyone thinks.

Unemployment is near 4%, against what Niles calls a normal rate of 6% and change, so 'it's really a great economy.' The 40% of households with no assets are the ones the Fed should care about most, and inflation is hurting them.

A pause in data-center expansion is healthy, not a sign AI is failing. If a quarter-point move could kill the build-out, the bubble callers would be right, and Niles doesn't think they are.

Cisco had a couple of 30% selloffs on its way to a roughly 4,000% gain from late 1994 to March 2000. A pause lets companies rethink how much capacity they need instead of racing to top each other's spending, the 'rather go bankrupt than lose this race' mindset he attributes to Larry Page.

Markets that rise in a straight line are dangerous: greed and leverage build up, and the bigger the bubble, the bigger the collapse. Regular pauses prevent wipeouts.

Korean investors in leveraged Samsung and Hynix ETFs were wiped out in the selloff from late June to late July after a greedy March-to-June run. Cisco's bookings went from up 70% to down 30% year over year within months in 2001. NASDAQ rose 86% in 1999, then fell 78% over two and a half years.

Niles believes Iran will keep the Strait of Hormuz a problem until at least the November 3, 2026 midterms, to hurt the Republicans at the polls.

When oil and gas prices are up a lot, the party in power usually gets crushed in the midterms, because voters vote their pocketbooks. Iran held the US hostages for 444 days and freed them hours after Reagan was sworn in, once Carter had been routed.

It wouldn't surprise Niles if the Fed hikes more than once, because high oil prices and the data-center build-out keep pushing inflation up.

Oil staying high will likely seep into inflation. The global data-center build-out soaks up resources, which also pushes prices up. Inflation has been above the 2% target for 65 months. He says he'll adjust as the data comes in.

Soaring government debt is the main force behind rising long-term bond yields worldwide, and deficits won't be fixed until a crisis forces it.

The US runs a 6% deficit, the highest outside a major war, with about $40T of debt against roughly $33T of GDP. Japan's and the UK's yields are at multi-decade highs. Stimulus from the financial crisis and COVID was never pulled back, which led to the 2021-22 inflation. Politicians want to be reelected and voters have gotten used to 'free stuff.'

The US political shift toward promising free benefits, including democratic socialists winning seats, is bad for big business and the stock market.

Niles says their platforms openly call big business bad, and populist promises push deficits higher rather than lower.

From late July to November 9 in US midterm years, the market's median peak-to-trough drop has been 10%, double the 5% in other years, so Niles sees an unusually risky stretch into the 2026 midterms.

His data covers 1990 through 2025 and includes the 1990 Gulf War, which he says has similarities to today. These are medians, with outliers both ways. He declined to give an S&P 500 range.

For any technology: costs fall, more people can afford it, volume rises, and in a healthy market profits grow. Falling prices alone don't mean shrinking revenue.

Niles uses this to explain why the 50% drop in AI token prices is being more than made up by higher usage.

AI economics are improving: token volume and cloud profits are rising faster than token prices are falling.

Cost per million tokens fell more than 50% from late May, while tokens produced rose more than 2.5x. Combined growth at AWS, Azure, and Google Cloud sped up from 35% in the March quarter to 43% in the June quarter, with operating margins up 2 points. Anthropic says it reached adjusted profitability in the June quarter, which makes the build-out look less dependent on debt.

A world-changing technology draws winner-take-all overinvestment, so the bubble breaks and companies go bankrupt once growth slows. Before that peak, investors can still make a lot of money.

Like canals, railroads, radio, TV, and the internet, companies race with a 'rather go bankrupt than lose this race' mindset. When growth plateaus, the bubble breaks, but the peak may be a couple of years away.

China is clearly winning the energy race for AI, because its central government can put a nuclear reactor wherever it chooses.

In the US, data centers face local and political pushback ahead of the midterms, and politicians are moving from supporting projects to shelving or opposing them.

US AI companies won't fall far behind China's cheap models. Google's coming Gemini Flash should be very good at producing tokens cheaply.

Cut off from Nvidia's best chips, Chinese labs had to get smart about algorithms and cost. Meanwhile, Anthropic and OpenAI lead at the frontier.

AI use will split by tier: about 90% of tasks will go to cheap, lower-end models, and only the top 5-10% of uses will need expensive frontier models.

'A Ford works as well as a Ferrari for the corner store': companies that first used frontier models for jobs like summarizing email found it too expensive. This mix shift is behind the lower token prices and higher volume.

China will become a major maker of memory chips, including high-bandwidth memory, because chip self-sufficiency is a national defense priority.

Being cut off from US chips makes Chinese supply as important as a battleship. Its chips don't need to be leading edge, just as Japan's in the 70s-80s and Korea's in the 90s weren't at first. CXMT has gone public and YMTC will soon, which raises capital. China buys about 20% of the world's PCs and smartphones, giving it a built-in market for roughly 20% share.

The market has an earnings bubble more than a valuation-multiple bubble. Memory makers' 80% operating margins aren't normal and will fall below average once China ramps up.

Valuations look high against GDP and normalized earnings because companies are over-earning. Chinese memory expansion should push margins back down.

Apple's choice to license Google's Gemini instead of spending hundreds of billions on AI has worked, maybe more by luck than by plan, and it helped the stock this year.

Apple reportedly pays a couple of billion dollars to license Gemini while Google spends about $200B, so Apple isn't burning cash. Niles calls Google the best AI company across the whole ecosystem. The stock's run came mostly from a higher valuation multiple, not new products: Apple is late with AI Siri (US-only, English, beta) and with foldables, years after Samsung.

Niles expects a huge iPhone upgrade cycle sometime in 2027, once Apple's AI leaves beta, expands to other languages and countries, and a better product line arrives.

Apple's ecosystem is sticky and users find it hard to switch to Android, so late features still sell.

Niles would shelter in cash, held in money market funds, from now through the early November 2026 midterms, instead of stocks or bonds.

The midterm-year median drop is double the normal one, and valuations are high against GDP and normalized earnings. Money market yields are solid in every country, and even Japanese investors now get real yields at home. Bonds are no refuge, since their selloff is part of why stocks are at risk.

AI hasn't hit a fundamental peak: agentic AI took shape only about seven months ago (OpenClaw, January 30, 2026), so Niles expects at least another year of solid growth.

It's hard to believe the peak is here when people are only seven months into working out how to use agents. He admits smart bears point to token prices falling 50% since May and warn of trouble if usage stops growing.

Stay humble about extremes: stocks can rise and fall far more than you imagine and nobody knows where the peak is, so respect the market instead of trying to outsmart it.

Examples: Volkswagen quadrupled in two days in a 2008 short squeeze to become the world's most valuable company, and GameStop. When Niles downgraded all his stocks in 2000, he didn't expect NASDAQ to fall 78%.

Broadcom's results, reported after the close on September 4, 2026, were nowhere near expectations, and Niles expects the stock to fall the next day unless the conference call changes things.

He cites it as evidence for the bears' worry that data-center demand is stalling.