Memory Dethrones the GPU: The Week HBM's Chokepoint Redirected AI's Profit Pool

The Week's Decisive Shift: The Crown Moves From GPU to Memory

In the final week of September 2026, the U.S. semiconductor narrative quietly flipped. MarketWatch reported that Micron (MU) could overtake Nvidia (NVDA) as the biggest driver of S&P 500 profit growth, and analysts polled by Stocktwits and Yahoo Finance are modeling a 350% revenue jump for Micron's imminent earnings print. The tape already agrees: Micron cleared $1,080 on September 27. For three years, the throne of the AI narrative belonged to the GPU. This week's news said, from multiple angles at once, that the throne is moving to HBM.

The clinching signal came from the hardware side. Shattered.io reported that Nvidia's next-flagship Rubin Ultra has been forced to give up 33% of its memory capacity because of the HBM shortage — an admission from Nvidia itself that no matter how good the compute die is, the spec cannot be filled without the memory stacks. In the same week, Supermicro began shipping Vera Rubin NVL72 racks with 1,152 GPUs and 1.8MW inline cooling — the infrastructure is ready, but the memory that goes inside it is not. That contradiction is the picture of the week.

What the Numbers Say: Half of AI Capex Is Now Memory

One analyst warned this week that more than 50% of AI infrastructure capex will flow into memory semiconductors (finance.biggo.com). At the same time, the DRAM module market surged 59%, and shortages are now forecast to persist through 2028 (eu.36kr.com). Today's DDR5 16Gb spot price sits at $57.83, refusing to soften from all-time highs.

The implication is unambiguous. Of every $100 a hyperscaler spends, $50 ends up on the top line of memory companies (Micron, SK Hynix, Samsung) — which means the biggest beneficiary of the AI capex cycle is now the HBM maker, not the GPU maker. Businesskorea reported that both Samsung and SK Hynix are on track for record Q3 operating profits.

But the Narrative Has a Fracture

In the same week, the market interrogated the sustainability of that same story. On September 28, SK Hynix fell 6% — memory weakened broadly after Bloomberg reported that SK Hynix's Solidigm NAND subsidiary was being valued at $100 billion ahead of a potential IPO (24/7 Wall St.). Seeking Alpha argued that SK Hynix's massive capex ramp is putting the company in a tough spot. Both stories carry the same message: the risk that supply expansion outpaces demand is starting to price in.

Qz.com added that Micron's FQ1 2026 print may show that the demand benefit from Nvidia's new chip launches is being delayed. So a 350% revenue-jump expectation and a delay warning coexist in the same week — that is the watch-point for next week's Micron print.

The Policy Axis: SK Hynix in America and Intel's Capex

A second axis this week was the geopolitics of memory production inside the United States. Asia News Network and Yahoo Finance reported that SK Hynix is evaluating a U.S. NAND facility, driven by CHIPS Act incentives and geopolitical supply-chain diversification. Separately, Barchart wrote that SK Hynix is finding advanced-memory supply opportunities inside Intel's $100 billion U.S. chipmaking push.

Meanwhile, simplywall.st reported that Intel (INTC) is only supplying 50% of its CPU orders amid AI-driven demand, and Naver Premium Content flagged that CPU pricing is starting to trace the same 3-year cycle pattern memory ran three years ago. Bank of America raised its AMD price target to $720. The GPU narrative is diffusing into memory and CPU simultaneously.

PM View: Where Should You Stand?

The political economy of this week's narrative, restated:

1. The HBM chokepoint is real. Rubin Ultra losing 33% of its memory is an Nvidia-acknowledged supply constraint, not a rumor.
2. The value-chain crown is moving. With ~50% of AI capex flowing into memory, MarketWatch's observation that Micron could top the S&P 500 profit-growth board is not hyperbole.
3. But priced-in risk is now large. With Micron at $1,080 and Samsung/SK Hynix guiding record Q3 profits, the 6% SK Hynix drop is the market's first serious question mark on the speed of supply expansion.
4. The policy axis remains supportive. SK Hynix's U.S. NAND plant review, Intel's $100B capex, and China's potential re-opening of Nvidia chip sales (Reuters, Yahoo Finance) — all three vectors underpin memory demand.

From a positioning standpoint, this is the week the "memory dethroned GPU" narrative was confirmed — and also the week its first counter-narrative appeared. Next week's Micron earnings will decide which of the two the market weights.

Key Sources: - Micron could dethrone Nvidia as S&P 500's biggest profit growth driver (MarketWatch, 2026-09-28) - Rubin Ultra Loses 33% of Its Memory to HBM Shortage (shattered.io, 2026-09-28) - SK Hynix Falls 6% on $100B Solidigm IPO Valuation (24/7 Wall St., 2026-09-28) - Micron Earnings Preview: Analysts See Over 350% Revenue Jump (Yahoo Finance, 2026-09-28) - Supermicro Begins Shipping NVIDIA NVL72 Racks With 1.8MW Cooling (StorageReview, 2026-09-25) - plus 42 more