Key Takeaways
Jensen Huang’s 2026 AI manufacturing unit map spotlighted NVIDIA’s DSX buildout framework.Marvell gained 241% YTD; AI infrastructure corporations may even see heightened investor focus.NVIDIA initiatives 100 GW of AI factories by 2030, shifting consideration to ecosystem companions.
The next visitor publish comes from Ziven.io, a public markets intelligence platform delivering knowledge on corporations uncovered to bitcoin mining, synthetic intelligence, and crypto treasury methods. Initially revealed on June 18, 2026, by Cindy Feng.
Since Jensen Huang stood on the Computex stage and referred to as Marvell “the following trillion-dollar firm,” MRVL hasn’t seemed again. A inventory that traded between $50-$100 as lately as April now sits round $300, with an ATH round $316 and a acquire of roughly 241% YTD. One sentence from Jensen, and an organization re-rated by a quarter-trillion {dollars}.
Not shocking, a brand new train has begun: comb via the whole lot Jensen says, discover the following identify he’ll bless and get wealthy.
I perceive the impulse, however what’s clear from listening to Jensen’s complete keynote is that most individuals are watching the mistaken factor. Jensen didn’t simply drop a sizzling identify, he laid out a full map of how an AI manufacturing unit really will get constructed, layer by layer, firm by firm. That map is the half value realizing, as a result of it nonetheless works lengthy after the hype fades. I’m going to stroll you thru that particular slide, however first let’s begin with the half that confused lots of people.
RTX, DGX, DSX: employee, crew, manufacturing unit
Jensen break up NVIDIA’s manufacturers into three layers, every a much bigger unit than the final:
RTX is the GPU, the employee. The chip that does the precise computing. One pair of fingers. DGX is the system, the crew. Wire a pile of these chips right into a single machine and also you’ve bought a DGX. A crew appearing as one unit. DSX is the infrastructure, the manufacturing unit. The constructing these groups work in, plus the ability, cooling, community, and software program to maintain 1000’s of them working across the clock.
RTX and DGX you’ve most likely heard of. DSX is the brand new one, and it’s the one value understanding, as a result of it’s the place NVIDIA stops promoting you a chip and begins promoting you a solution to construct the whole plant.
What DSX really is
In Jensen’s phrases, DSX is “a blueprint, a reference design for constructing and working AI factories at most effectivity and profitability”.
In plainer phrases, it’s a recipe and a toolkit for booting up a gigawatt of compute and protecting it worthwhile. NVIDIA even named the toolkit’s components: a digital twin to design and check the entire manufacturing unit earlier than a single rack ships (DSXSim), an working system to run it as soon as it’s stay (DSX OS), and instruments to pack extra GPUs into the identical energy price range and flex with the grid (DSX Max LPS, DSX FLEX). The pitch is that 100 gigawatts of those factories come on-line earlier than the last decade is out, and that DSX-built ones run cheaper and lean on the grid extra gently.
That every one appears like one thing NVIDIA would promote you by itself. It’s really not the case.
No single firm can construct a complete AI manufacturing unit
A one-gigawatt AI manufacturing unit is now a $30-100 billion undertaking, in keeping with Jensen. At that scale it stops being a server room and turns into infrastructure on the order of a refinery or an influence station.
NVIDIA can’t construct that alone. It doesn’t pour concrete, run high-voltage strains, manufacture chillers, or negotiate with the native utility. And you’ll’t bolt these items on one after the other, as a result of the chips, racks, community, energy, and cooling all need to be designed collectively from day one. Each hour the manufacturing unit sits idle is income misplaced, so a construct this costly has to work the primary time.
Due to this fact NVIDIA did the wise factor: it revealed the blueprint and assembled a coalition of companions to cowl each layer it doesn’t do itself. That coalition has a reputation, the AI Manufacturing facility Ecosystem, and Jensen put the whole roster on a single slide. That slide is the map.
The map: who really builds an AI manufacturing unit
Most of these firm are non-public or listed abroad, however nonetheless a loads of U.S. listed ones. I made a desk to listing all publicly traded names from the map. The final column is my tough learn of how a lot of every enterprise actually rides on the AI build-out, as a result of being on the slide (may lean on advertising goal) and being moved by it are two very various things.
Please notice OTC or international listed names are excluded from the desk. If you’d like the whole CSV listing, simply drop me a message and I’ll ship it over. Additionally a couple of names are nonetheless non-public with upcoming IPOs, similar to Lambda (US), Nscale (UK), Firmus (Australia) and Yotta (India).
Necessary Observe
One has to comprehend {that a} emblem show tells you an organization is concerned however it doesn’t inform you whether or not the involvement is materials. For CoreWeave or Vertiv, AI-factory demand is actually the whole story. For Caterpillar or Nationwide Grid, it’s a sliver of a far greater enterprise that may barely transfer the inventory. The “Excessive” rows offer you torque and volatility in equal measure. The “Low” rowsgive you a steadier firm with solely a skinny thread tied to AI built-out commerce.
Last Ideas
Possibly one in all these names turns into the following Marvell, possibly none do. That’s not a name I could make from a slide, and chasing whichever emblem you hope Jensen blesses subsequent is nearer to a guessing sport than a method.
The sturdy worth right here is the map, plus a sharper query to take into it. For any identify on this chart, how a lot of its enterprise really rides on the AI build-out?How a lot pricing energy does its layer maintain? Pure-plays, diversified incumbents and commodity positively have completely different leverages and threat profiles.
Right here’s what doesn’t change: Each hyperscaler deal you’ll examine, each “X-gigawatt knowledge middle” headline, quietly is dependent upon this complete stack to occur. Somebody designs it, somebody builds it, somebody powers it, somebody cools it, somebody racks the servers, somebody runs it. This chart is the solid listing. Decide a layer that pursuits you and weigh its publicity in opposition to how a lot pricing energy it holds. That’s the place the true work begins. The map gained’t inform you what to purchase, however it’s a framework you’ll be able to confer with.
