NVIDIA suddenly looks cheap in standard valuation-ratio terms. Smart-money investors may now be dumping Nvidia for Micron and friends, as the real constraint in the AI build‑out is no longer just GPU TFLOPs that you can program with CUDA…
It is worth pausing to remember that in 2018 and again in 2022, NVIDIA was seen as in serious trouble:
First, the market for computer gaming and NVIDIA’s edge in serving that market seemed to be in trouble.
And then NVIDIA was rescued from that by crypto.
Then the market for crypto mining and NVIDIA’s edge in serving that market seemed to be in trouble.
And NVIDIA was rescued by AI—and much, much more than rescued: Jason Huang of NVIDIA as the genius-impresario-entrepreneur of our epoch, master of the most valuable company in the universe!
But the worms are at least exploring their options.
All of the companies that had the money to pay and did not have the cybernetic managerial bandwidth to avoid paying the NVIDIA tax in the enormous run-up since 2022 are now under various forms of pressure to find some way to deal with their ongoing investments in Modern Advanced Machine Learning installations other than to simply write ever larger and larger checks to NVIDIA.
Hence NVIDIA is, as my Bloomberg stream flagged a couple of days ago, now cheaper in a price-earnings sense than Hershey:
Jeran Wittenstein & Rainier Harris: NVIDIA’s $1 Trillion Slide Sends Valuation [Ratio Down] to Pre-AI Boom Levels <https://www.bloomberg.com/news/articles/2026-07-08/nvidia-s-1-trillion-slide-sends-valuation-to-pre-ai-boom-levels>: ‘The chip[-designer’s]… graphics processing units… still dominate…. [But] investors [are] rejigger[ing] the AI trade by ditching NVIDIA in favor of competing semiconductor manufacturers, particularly those in the memory market. The selloff has NVIDIA… trading at 18 times earnings projected over the next 12 months…. The last time the shares were this inexpensive [in valuation-ratio terms] was early 2019…. NVIDIA’s shrinking valuation isn’t the result of a deteriorating outlook. On the contrary, Wall Street analysts have been raising their profit estimates for the coming quarters. Instead, the selloff shows how much the AI trade is shifting to… memory and storage stocks like Micron Technology Inc.. [and to] NVIDIA[’s chip-design] rivals such as Advanced Micro Devices Inc. and Intel Corp. have seen their share prices double or even triple this year….
NVIDIA is expected to deliver the fourth-fastest revenue growth in the the S&P 500 this year, but it’s still cheaper than about half of the stocks in the index, including candy maker Hershey Co. and the utility Dominion Energy Inc…
Brad DeLong back again: There is a delicate dance in today’s AI build-out
between:
GPU power,
memory bandwidth,
memory capacity,
CPU orchestration,
the power bill,
an efficient software stack,
an easily workable software stack.
Whatever is in shortest and inelastic supply right now gets the lion’s share of the money that is paid in the mammoth AI infrastructure build-out. For most of the build-out so far, that has been NVIDIA. It’s combination of GPU power, the fact that CUDA is a remarkably efficient software stack, and the fact that CUDA is an easily workable software stack have been triple aces in its hand. But right now it is memory capacity, and, perhaps soon, power that look to be scarcest and hence most valuable over the next year.
As I understand the situation (and I may not), the smart (or at least less-dumb) money right now is converging on three main potential bottleneck stories with respect to the next five years. This is what I think they are thinking. And they may be right; they may be wrong:
The punchline is: memory (bandwidth + capacity) and power, with a rising role for CPUs, are where people are actually betting bottlenecks—and rents—will live. The rest are being downgraded as likely to shift from “binding constraint” to “important complement” for the medium term:
Over the next five years, “smart money” is converging on three main bottleneck stories—and largely downgrading the others from “binding constraint” to “important complement.”
GPU power: The more technical and buy‑side notes are hinting that the real future choke points are the things wrapped around the GPU die, not the arithmetic units themselves. It is CoWoS packaging at TSMC that is the binding limits on H100/H200/Blackwell build‑out, not NVIDIA’s ability to design. Since the hyperscalers have already forward‑bought all 2026–27 NVIDIA capacity in TSMC fabs; that pushes everyone else into neoclouds and secondary markets. Thus the “GPU power” scarcity story morphs into near term (’26–’27): constrained availability as existing fabs and packaging run hot; and medium term: intense competition from AMD, Intel, and in‑house accelerators as more capacity actually lands.
Memory bandwidth: Here the evidence is compelling, with everyone in the choir singing the song that AI has forced a deliberate, structural reallocation of wafers and clean‑room capacity from “plain DRAM/NAND” toward HBM and high‑end DDR5. Extreme tightness “through 2027 and beyond” is the base case, not the tail risk. icf.: Tom’s Hardware. Each AI server eats 10–20× the DRAM and HBM of a conventional server. Hyperscalers are locking in multi‑year supply, sometimes through 2029–2030. Every wafer that becomes HBM stacks is one less for “boring” DIMMs or SSD controllers. So shortages propagate. And firms are deliberately underbuilding generic DRAM because they remember the last two cycles that bankrupted half the industry. So memory bandwidth is the acute AI bottleneck as you can pile up GPU dies but if you can’t stack enough high‑yield HBM3e nearby, you don’t have an installation.
Memory capacity: t+This is the chronic constraint that oozes into PCs, consoles, phones, NICs, routers, storage arrays. “RAMageddon” journalistic pieces and OEM pricing behavior line up with the analyst notes: DRAM is in a multi‑year scarcity-and-high‑price regime. Cf.:Business Insider. Hence I do not have my M5UltraMacStudio. What replaces the NVIDIA tax may well be a memory tax levied by the three cooperating oligopolists.
CPU orchestration: Until recently, everyone treated CPUs as boring plumbing behind the GPU show. Now the scenario smart money is shifting to is: (a) training era: one fat CPU per 4–8 GPUs, mostly doing data input, orchestration, host‑side kernels. CPU not the constraint; (b) agentic inference era: lots of smaller, interactive, multi‑tenant, tool‑using workloads, where CPUs run vector DBs, routers, retrieval, business logic, and sometimes non‑GPU ML. Now CPU demand scales more nearly 1:1 with GPUs. If you need one very beefy CPU per GPU, Intel/AMD/ARM land some of the rents previously captured by NVIDIA alone. And, of course, that increases total power and memory pressure, because more CPUs mean more DRAM sockets and more watts per rack.
The power bill: What is the physical ability to deliver tens of gigawatts of reliable power to clusters, fast enough? Data centers at 300 MW–1 GW are now routine in planning documents; AI‑only campuses in the multiple‑GW range are on the boards. But transmission build‑out, interconnection queues, large transformers, switchgear, and gas pipelines all have 3–7 year lead times; many regions are already turning away projects or forcing on‑site generation. A lot of serious energy‑economics work (and industry self‑serving white papers) converge on roughly 75–100 GW of new U.S. generation by 2030 just for AI/data‑center load. That implies a double‑digit percent increase in U.S. natural gas-electricity production plus big chunks of new renewables. And, just maybe, some nuclear? Cheap kWh in principle matter less than “can you actually get 500 MW of firm capacity onto this site by 2028?” Permitting, rate structures, and local backlash become central to AI deployment. And grid hardware, transformers, and even high‑voltage cable become weird, old‑economy bottlenecks in the AI story.
An easily workable software stack: But even if CUDA is no longer the obvious path to run your workload, the much larger number of programmers to have CUDA experience still a semi-decisive factor giving NVIDIA a big edge as people plan their workloads? I don’t pretend to have a guess.
But do look for GPU+CUDA to take somewhat of a punch-drinking refreshment break, and for shifts: HBM vendors, grid builders, and—oddly enough—gas turbine manufacturers and Xeon designers to lead the dance. Hence: NVIDIA cheaper in a valuation-ratio sense than Hershey.
If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…
Now all we need is an AI business model big enough to support all those data centers.
The coding market is one of the few success stories, but it is only so big, and we are already seeing its limits. If nothing else, software producers are more likely than others to take AI in house using tailored systems on local hardware.
The advertising market does not need peak AI to short circuit web search to cut out information providers. The scientific market faces both Trump era headwinds and reliance on users likely to develop their own systems for in house use.
An investment on the order of a trillion dollars needs a return on the order of a hundred billion dollars to provide suitable returns. Is the idea that each American will be spending $30 a month for an AI assistant? Is there some big industry out there that could benefit from AI spending, in aggregate, on an order of ten billion dollars a month? The current all-in investment pattern would make sense if AI were a sure thing.
Good analysis.
Now all we need is an AI business model big enough to support all those data centers.
The coding market is one of the few success stories, but it is only so big, and we are already seeing its limits. If nothing else, software producers are more likely than others to take AI in house using tailored systems on local hardware.
The advertising market does not need peak AI to short circuit web search to cut out information providers. The scientific market faces both Trump era headwinds and reliance on users likely to develop their own systems for in house use.
An investment on the order of a trillion dollars needs a return on the order of a hundred billion dollars to provide suitable returns. Is the idea that each American will be spending $30 a month for an AI assistant? Is there some big industry out there that could benefit from AI spending, in aggregate, on an order of ten billion dollars a month? The current all-in investment pattern would make sense if AI were a sure thing.
Nice Analysis...Thx