AI Supercluster Stocks: The Companies Building the DOE’s National Lab Network

“AI supercluster” is not a term you will find in a Nvidia earnings call. It is Marc Chaikin’s name, from his “The End of Elon?” presentation, for the Department of Energy’s plan to link its national laboratory supercomputers into one AI research platform. He calls the platform “American Atlas,” gives away one stock for free, and says his team has “pinpointed 17 companies that have signed deals” to build it. This article looks at the investable universe: who is actually supplying the DOE program, and which names fit Chaikin’s clues.

What the supercluster actually is

Strip away the branding and the program is the DOE Genesis Mission, a late-2025 executive order directing the Department of Energy to build what the promo quotes as an “integrated AI platform to harness Federal scientific datasets – the world’s largest collection of such datasets – to accelerate breakthroughs.”

Chaikin describes the hardware side in two layers. First, “nine brand-new AI micro clusters… all powered by FP64 hardware,” starting with Lux at Oak Ridge and including Discovery, Minerva, Janus, and Tara. Second, the conversion of the DOE’s existing fleet: “The Department of Energy already owns the most powerful fleet of FP64-powered computers on the planet… They number more than 50 in all.” He names Frontier, El Capitan, Aurora, Perlmutter, Polaris, Venado, Tuolumne, and Crossroads, all real machines at the labs he cites.

The distinguishing feature, and the reason it is investable, is precision. These machines run FP64 double-precision math for scientific simulation, which is a different hardware market than the FP8 and FP4 inference chips that dominate commercial AI. Chaikin puts it bluntly: “Colossus runs mostly on FP8 or even FP4. But this company’s new AI micro cluster will run a brand-new version of FP64 hardware.”

The free pick: AMD

Chaikin names Advanced Micro Devices outright. “It’s the first company on my American Atlas ‘buy’ list, Molly. I’m talking about Advanced Micro Devices — ticker symbol AMD.”

The case is straightforward. Lux and Discovery, announced in October 2025 by the DOE, Oak Ridge, AMD, HPE, and Oracle, are built on AMD Instinct accelerators. The promo highlights “the MI430X Instinct Accelerator,” which AMD describes as “engineered specifically for sovereign AI and scientific computing.” Chaikin adds, “At least $1 billion dollars are about to hit this company’s account as a result,” which matches AMD’s own public characterization of the partnership as a roughly $1 billion program.

There is a deeper reason AMD dominates this list. Frontier and El Capitan, the two most powerful DOE machines today, both run AMD CPUs and GPUs, because AMD has led Nvidia in FP64 throughput for several product generations. If the DOE networks its fleet, most of the compute is already AMD. Our full review discusses why this is a solid but not explosive thesis for a $300 billion company.

The teased networking pick

Chaikin’s first paid recommendation in The American Atlas Portfolio is described, not named. “The biggest challenge will be connecting all of these AI supercomputers together into one massive device, seamlessly. That’s where our first American Atlas recommendation comes in.” It is a company with “cutting-edge networking solutions… The unsexy stuff,” whose potential “the market hasn’t priced in.”

Candidates that fit:

Arista Networks (ANET). The dominant vendor of high-speed Ethernet switching for AI clusters and a Power Gauge favorite. Arista’s 800G platforms are exactly what a multi-site scientific network would deploy. The catch is that “the market hasn’t priced in” is hard to square with Arista’s valuation.

Ciena (CIEN). The likeliest fit for “connecting supercomputers seamlessly” across sites “bigger than the state of Texas.” Ciena builds the long-haul optical transport that links data centers over hundreds of miles, and it supplies ESnet, the DOE’s dedicated science network that already connects the 17 national labs. That is a direct tie to the program, and Ciena has been a quieter AI beneficiary than the switch vendors.

Credo Technology (CRDO). Active electrical cables and optical DSPs for inside-the-rack connectivity. Smaller, faster growing, and less covered, which fits “unsexy” and “hasn’t priced in.”

Cornelis Networks is the most technically relevant name (its Omni-Path fabric is specified for DOE systems), but it is private, so it cannot be the pick.

We lean Ciena based on the “connecting across sites” language, with Arista as the conventional alternative. Chaikin’s Power Gauge has to rate the pick Bullish, which is a constraint the paying subscriber can check.

The other names on the contract sheet

Hewlett Packard Enterprise (HPE). HPE’s Cray division is the system integrator for Frontier, El Capitan, Aurora, Lux, and Discovery. Its Slingshot interconnect is the fabric inside these machines. If there is a “supercluster” builder in the literal sense, it is HPE. The AI-supercomputer segment is a small piece of HPE’s total revenue, which limits the upside, but it is the most direct exposure to the buildout after AMD.

Oracle (ORCL). A named partner on Lux and Discovery, supplying cloud infrastructure and software layers. Oracle’s role is as the bridge between the lab machines and commercial AI workloads.

Intel (INTC). Aurora at Argonne runs Intel CPUs and GPUs. The promo separately cites the government’s $8.9 billion Intel stake as an example of what happens “when the government gets behind” a stock: “Intel jumped as much as 128%.”

Nvidia (NVDA). Present in the DOE fleet (Perlmutter, Venado, Polaris) and in the new Solstice and Equinox systems at Argonne. Chaikin’s framing puts Nvidia in the “legacy AI” column, but in practice Nvidia hardware is part of the supercluster too.

Vertiv (VRT), Eaton (ETN). Power and cooling for high-density compute. Not named, but standard suppliers to any national lab installation.

How to think about the theme

The DOE spends a few billion dollars a year on computing, which is a rounding error next to hyperscaler capex. What the supercluster theme offers is not size but durability and prestige: national lab contracts are long-lived, politically protected, and a signal to commercial buyers. AMD’s FP64 leadership is a genuine moat in this niche. Ciena’s ESnet relationship is a genuine moat in its niche.

Chaikin’s promo wraps that modest, real thesis in a “$248 trillion” and “1 trillion times more powerful” package. If you can hold both ideas at once, the branding and the underlying reality, the supercluster is a reasonable place to look. For the lab-level detail on the machines themselves, see Oak Ridge and Frontier Supercomputer Stocks. For the coined vocabulary, see What Is the American Atlas Portfolio?