Revenge of the Rice Ball

Programming note: ARPU will be back on November 23 to look at the shifting state of the AI narrative.

⚡ Above the Fold: Nvidia Hangs Up the Leather Jacket

Before we get to this week's issue on the memory chip boom, we should probably address the leather jacket in the room.

On Thursday, Jensen Huang—arguably the most powerful executive of the AI era—resigned following revelations of two extramarital affairs, one with a suspected Chinese intelligence asset who works at Megaspeed, a Singapore-based Nvidia customer. Nvidia's stock had a momentary 14% panic attack at the opening bell before settling down, shaving another $180 billion off a valuation that has already lost a third of its peak since October.

Will the business stall? Almost certainly not, at least not soon. Nvidia's product roadmap for the next three years—the Blackwell ramps, the Vera Rubin architectures—is already well in motion. The software moat that actually traps the AI industry, CUDA, does not suddenly vanish because the CEO stepped down. And demand so radically outstrips supply that no hyperscaler is going to cancel a ten-billion-dollar GPU allocation over an HR scandal. Next week, Nvidia will likely report over $100 billion in quarterly revenue, and the cash will keep pouring in.

But Nvidia didn't just lose a CEO. Huang played half a dozen roles at once: technologist, salesman, marketer, lobbyist, dealmaker and the leather-jacketed face of the entire AI boom. He could command priority allocation from TSMC, lobby Washington on export controls, and then turn around and use Nvidia's balance sheet to bankroll a fleet of neoclouds like CoreWeave. He didn't just sell silicon; he actively managed the geopolitical and financial equilibrium of the entire hardware stack.

Nvidia's 75% gross margins weren't just protected by floating-point arithmetic; they were protected by Huang's singular ability to keep customers, suppliers, and governments aligned with his vision.

The irony is that the same web brought him down. The scandal centers on Megaspeed, a major Nvidia customer suspected of diverting chips to China. That combination of customer relationships and export-control politics is exactly what Huang handled personally.

Interim CEO Colette Kress and Nvidia's formidable engineering bench can comfortably ship chips and beat earnings targets for a few more quarters. But Nvidia might become something it has never been: a regular company, run in a regular way. For a business whose margins rested partly on one man's ability to keep customers, suppliers and governments in line, that may be the biggest change of all.

Now, on to the companies that out-earned America's tech giants last quarter...

Selling Shovels, Charging Gold

One useful way to measure the memory-chip business is in rice balls. When Japanese memory champion Elpida collapsed into bankruptcy in 2012, its president, Yukio Sakamoto, lamented that a DRAM chip had become so cheap it could "hardly buy you one rice ball."

That has always been the defining character of the memory trade. The cleanrooms cost astronomical sums to build, but the product they make can collapse in price whenever supply outruns demand. Because a multibillion-dollar fab is most economical running near full capacity, manufacturers overinvest in the good times, flood the market when demand cools, and bleed cash through the downturn until the weakest player goes under.

Last quarter, the rice balls got unimaginably expensive.

For the first time, five global memory suppliers (Samsung Electronics, SK Hynix, Micron Technology, Kioxia and Sandisk) logged more combined operating profit than five of America's largest technology companies, Nikkei Asia reported. Driven by a tenfold surge in some memory prices over eighteen months, the memory makers earned roughly $204 billion. Alphabet, Microsoft, Amazon, Meta and Apple managed roughly $160 billion.

That's a striking change of fortunes. Silicon Valley has spent two years promising that AI would generate an unprecedented wave of corporate profits. For one quarter, at least, the biggest earnings went not to the companies building AI platforms, but to the ones selling them memory.

And the most interesting part isn't how much money the memory makers earned. It's that they earned it while shipping barely more chips than before.

The 70% Commodity

In the first half of 2025, the memory industry was behaving much like its old self, with operating margins in the 10% range. Today, industry margins have approached 70%, with South Korea's SK Hynix delivering an extraordinary 80% operating margin last quarter.

Margins like that are not supposed to exist in a business built on interchangeable silicon and ferocious price competition.

The conventional explanation is HBM. Nvidia and other AI chip designers need vast amounts of high-bandwidth memory, which stacks layers of DRAM vertically beside the processor to keep accelerators fed with data. HBM is hard to make and consumes far more production capacity per bit than ordinary DRAM. With new fabs taking years to build, every wafer devoted to premium AI memory is a wafer not making conventional chips, and manufacturers have every incentive to choose the premium product.

But the more telling number is volume. According to Counterpoint Research, global DRAM revenue rose 369% year over year in the third quarter, to nearly $200 billion. Bit shipments, the industry's measure of how much memory capacity actually left the factories, rose only about 4% from the previous quarter. While the two figures cover different periods, the direction is clear: this is a pricing supercycle, not a volume cycle. Memory makers aren't producing five times as much DRAM as a year ago. They're charging vastly more for the chips they do make.

Relief is coming, but slowly. Memory makers are investing heavily, yet Counterpoint doesn't expect major new fabs to start production until 2028. Until then, buyers are competing for supply rather than negotiating on price. When missing the next generation of computing is an existential risk, a discount on memory falls a long way down the agenda.

The Memory Tax

It's tempting to see the memory shortage as a problem for the hyperscalers alone. They're the ones spending hundreds of billions on data centers. If they want AI infrastructure at any price, they can pay any price for the parts.

The trouble is that AI's appetite for memory no longer stops at the GPU. HBM revenue nearly doubled last quarter, but conventional DRAM grew even faster, Counterpoint Research found: DDR revenue rose about 500% and LPDDR about 400% from a year earlier.

Part of the reason is architecture. As AI systems take on more complex, agentic work, they need more CPUs to orchestrate tasks and move data, and CPUs run on conventional DRAM. Industry estimates suggest that the GPU-to-CPU ratio in cloud AI systems has narrowed from about 4–1 to 2-1, and is heading toward 1-1. In Nvidia's Vera Rubin racks, each CPU socket carries up to 1.5 terabytes of DRAM, nearly triple the previous generation.

So AI is now squeezing memory from both directions: GPUs need HBM, and a growing fleet of server CPUs needs conventional DRAM. Unfortunately, those chips compete for manufacturing capacity with the memory powering the rest of the technology economy: LPDDR in smartphones and laptops, and DDR in PCs.

For phone makers, the memory tax is already biting. According to industry estimates, soaring memory prices have raised the bill-of-materials cost of some entry-level smartphones by roughly 70% year over year. In premium handsets, DRAM has overtaken the main processor as the most expensive component. Global smartphone shipments fell 20% in the third quarter as higher costs pushed up retail prices and forced production cuts, and Qualcomm reported a 26% drop in handset-chip revenue.

Memory isn't the only culprit. Consumers are holding onto devices longer, and phone makers face other pressures. But the pain is sharpest at the cheap end. Apple and Samsung can absorb higher costs or pass them on to premium buyers. Budget brands competing on thin margins can only raise prices, cut memory, strip out other features or drop models altogether.

That is the less obvious redistribution inside the AI boom. American hyperscalers and AI labs are bidding for computing capacity. Memory makers, most of them in Asia, are capturing the profits from the resulting scarcity. And the bill is landing on everyone else, through pricier smartphones, skimpier PCs and fewer affordable devices.

The usual cure will arrive eventually. Higher profits attract new capacity, and new capacity brings prices down, a lesson memory makers have learned the hard way for decades. But this cycle's fabs are slower and costlier to build, and its products harder to make.

Fourteen years ago, a DRAM chip could hardly buy you a rice ball. At today's prices, a tray of them could probably buy the entire convenience store.

Signal Stack

The operating reality beneath the headlines.

• SoftBank’s Failed Bet on Gas Turbines for AI Drove $3.7bn Charge (Financial Times) – SoftBank ordered more than 13GW of turbine capacity—enough to power 11 million homes—partly as a speculative trade on future data centre demand that never materialised, and has now written off $2.9bn of the $9.5bn it paid in advance.

You received this message because you are subscribed to ARPU newsletter - you may unsubscribe anytime at the bottom of this email.

If a friend forwarded you this message, sign up here to get it in your inbox.