AI Survived a 5% 10-Year Yield - Now Wall Street Watches Bottlenecks, Not GPUs
The Fed hiked again and Treasury yields topped 5%, but chipmakers barely flinched. The market is starting to ask who controls memory, power and capital, and who can fund it with their own cash.
The question the market changed this week
On September 16, the Fed raised its benchmark rate by 25 basis points to a range of 3.75% to 4.00%, its first hike in three years. The same day, the 10-year Treasury yield broke above 5% intraday. That combination is normally bad news for long-duration growth stocks.
Yet the market did not fall apart across the board. On September 18, the Nasdaq rose roughly 0.4% and held a positive weekly gain. Semiconductors were relatively strong. That single session says something bigger is shifting. Rates are no longer moving every growth stock in the same direction. Instead, the market is now asking how much cash flow and supply constraints can offset the shock to discount rates.
- 3.75% to 4.00% - the new fed funds target range after the September FOMC meeting
- 5%+ - the psychological line the 10-year Treasury yield crossed again this week
- $100+ - the price band where Brent and WTI have been sitting, a variable that pressures both rates and inflation at once
Bottleneck one: in HBM4, speed of scale matters more than exclusivity
SK hynix has already started mass production of HBM4, the core memory for Nvidia's Vera Rubin platform, and the company confirmed on its second-quarter earnings call that large-volume HBM4 shipments have begun. The market's attention is now moving one step further, to higher-stacked versions of the chip.
SK hynix is leading mass shipments of 12-high HBM4, the primary Vera Rubin SKU, while supply-chain trackers are already picking up early signs of initial-volume shipments of the next-generation, higher-capacity 16-high 48GB HBM4 part, evidence of a lead in high-stack technology. Whether that translates into full system-level mass supply still needs official confirmation, including TSMC's advanced-packaging yields and customer qualification.
One thing should be stated plainly: the HBM4 supply chain for Nvidia's Vera Rubin is an officially three-way race among SK hynix, Samsung Electronics and Micron. It is not a single-vendor lock. Still, supply-chain estimates and market research point to SK hynix holding a strong position in initial core allocations, while Samsung is rapidly expanding its share as it clears HBM4 qualification and improves yields.
Some supply-chain estimates suggest SK hynix and Samsung together could effectively lead more than 80% of Vera Rubin's HBM4 supply. That figure has not been confirmed by Nvidia and should not be treated as an official share. The numbers investors should actually watch are each company's real bit shipments, yields, average selling prices, and the mix between 12-high and 16-high parts.
The bigger picture is a memory-wide supply shortage. TrendForce has said DRAM supply growth is failing to keep pace with demand, and SK hynix management has warned that shortages could become severe by 2027. If HBM keeps absorbing wafer capacity and advanced-packaging capability, price pressure could spread into server DRAM, enterprise SSDs and commodity DRAM.
Bottleneck two: power may choke supply before GPUs do
Morgan Stanley's framing for 2026 is "the Politics of Energy." As data-center power demand runs into electricity-rate and community-cost fights, power has stopped being a technology problem and become a political, regulatory and permitting problem.
Morgan Stanley estimates US data centers will need about 68GW of power between 2026 and 2028, against roughly 30GW of grid capacity that is under construction or contracted. In some regions, grid interconnection can take five to seven years. That is why on-site generation, fuel cells, gas turbines and battery storage have moved from "backup options" to core infrastructure that buys time.
That is also why Bloom Energy is drawing attention. The company says it can deliver fuel-cell power for data centers in as little as 90 days, and it has worked with Oracle toward exactly that kind of 90-day on-site power target. That figure comes from Bloom itself and depends on project conditions. But what the market is pricing is not the technology alone, it is time to power, how fast electricity can actually be connected.
Bottleneck three: capital itself is now expensive
A 5% 10-year yield is not just a chart level. In businesses that require tens of billions of dollars, like data centers, power equipment and chip fabs, the cost of capital itself becomes a competitive factor.
In this environment, the gap widens even among stocks that all sell "AI growth." A hyperscaler that can fund capex out of massive operating cash flow, and a loss-making company that keeps needing outside financing, do not deserve the same valuation even at the same growth rate. As AI spreads, not every AI-linked company benefits. Companies where capital costs and depreciation grow faster than the revenue AI actually generates could come under real pressure.
That is also why Morgan Stanley has emphasized "AI Technology Diffusion" this year. Going forward, what matters is less about how much model performance improves and more about how fast real companies adopt AI and convert it into productivity. The earnings gap between companies that use AI well and those that do not is likely to widen into a stock-price gap.
SpaceX and the trap of flow versus fundamentals
This week, SpaceX offered a lesson in separating flow from fundamentals. Its addition to the Nasdaq 100 was estimated to require roughly $15.5 billion to $22.0 billion in passive buying.
But that money does not all arrive fresh on the morning of September 21. Most of it is typically executed at the September 18 closing rebalance, ahead of the new weighting taking effect. Even with large mechanical buying on one side, if the selling on the other side is bigger, the stock still does not go up.
| This week's theme | The headline story | What actually matters |
|---|---|---|
| HBM4 | Who is the No. 1 supplier | Qualification, yields, ASP, stack count, real shipment volume |
| AI power | Power demand is exploding | Time to power, grid interconnection, on-site power deals, fuel costs |
| SpaceX | Massive passive inflows | Increased float, lockup expirations, actual earnings and valuation |
| Big Tech | AI capex keeps rising | Free cash flow, depreciation, AI revenue conversion, cost of capital |
What individual investors should change
The most dangerous move in a week like this is either dumping all AI exposure because rates went up, or piling into semiconductors at the highs simply because they held up.
A more realistic approach is to split AI exposure inside a portfolio into three layers. The first is suppliers sitting closest to the actual bottleneck, names like Nvidia, Broadcom, TSMC and the memory makers. The second is platforms that can fund AI capex out of their own cash flow, like Microsoft, Alphabet, Amazon and Meta. The third is infrastructure that demand keeps requiring as AI investment grows, like power and networking.
By contrast, loss-making growth names whose reliance on outside financing is growing faster than their revenue, and thematic stocks whose commercialization has not yet been proven, could face steeper discount rates the higher yields go. Treating every "AI ETF" as interchangeable ignores that gap.
Insight Times Editorial Desk
