The AI Desktop 'Throttling' Story Nobody Got Right

When John Carmack—yes, that John Carmack—tweeted that his NVIDIA DGX Spark was "maxing out at only 100 watts power draw, less than half of the rated 240 watts," tech media did what tech media does best: turned a data point into a crisis narrative.

"OG Doom Legend John Carmack Says His Nvidia DGX Spark Box Has Thermal Throttling Issues," one headline proclaimed. Except that's not what he said. And more importantly, it's not what's actually happening.

Alex Ziskind, a developer who reviews hardware, decided to find out what's really going on. He acquired four different Grace Blackwell-based AI desktops—the NVIDIA DGX Spark, Dell Pro Max GB10, ASUS Ascent GX10, and MSI Edge Expert—and subjected them to the kind of sustained stress testing that reveals what marketing specs can't hide.

What he found matters if you're considering dropping $3,000 to $5,000 on one of these machines. It also illustrates something about how we talk about hardware problems: we love simple stories, even when the truth is more nuanced.

Same Chip, Different Clothes

All four machines share the same Grace Blackwell chip—that's the GB in GB10—with 20 ARM cores and 128GB of unified memory. Enough to run substantial language models, vision models, even video generation locally. They all ship with DGX OS, essentially Ubuntu with AI tooling pre-installed.

But they diverge in ways that might matter. The ASUS weighs 220 grams more than the others (possibly better cooling hardware, Ziskind speculates). The Spark and Dell use magnetic back panels with hidden screws; the ASUS makes you work harder to access the SSD. Port labeling varies from nonexistent (Spark) to reasonably helpful (MSI). The Spark is the only one with a Gen 5 NVMe drive.

These are the differences you see. The differences you don't see are more interesting.

What Throttling Actually Means

Here's the vocabulary lesson nobody asked for but everyone needs: thermal throttling is when a system gets so hot that it reduces clock speeds to avoid damage. If your CPU is supposed to run at 3 GHz but hits 110°C, it might drop to 2 GHz or lower. That's throttling.

What Carmack observed—and what Ziskind measured across all four machines—isn't that. It's power capping, which is different and arguably more interesting.

Ziskind ran three test scenarios. First, a small model (Qwen 34B) for 45 minutes to reach heat soak—the point where the cooling system, heat sink, and chassis have absorbed enough heat that temperatures stabilize. Result: all four machines performed identically. GPU utilization around 96%, temperatures hovering at 80°C, power draw 140-160W total. Clock speeds remained consistent. No throttling.

"If we take a look at PP 4096, this is a prompt processing phase," Ziskind notes in his testing. "We're hitting like good tokens per second, 1,976, and it's about the same on all the machines... We're getting 61, 61, 61, and 61. Small model, not a big difference."

Second test: a bigger model (Nemotron 30B, unquantized). Still identical performance. Still no throttling. Temperatures rose slightly, power usage remained constant, but all four machines delivered the same token generation rates.

Third test: GPU burn, an open-source tool designed to maximize resource usage and stress the cooling system. This is where things got interesting.

The Software Ceiling

Under maximum stress, all four machines hit a software-imposed power cap at just under 100 watts for the GPU—the exact number Carmack reported. This isn't thermal throttling. It's the system saying "this far and no farther," regardless of thermal headroom.

The ASUS machine did something none of the others did: it hit both the power cap and registered thermal throttle events. Twice. Power draw dropped from 96W to 76W. But here's where it gets weird: temperatures were only 95°C when this happened, while the Dell had hit 99°C without triggering any throttle response.

Did this translate to performance degradation? Ziskind doesn't think so. "Just because the system detected it and it's showing it to us, does that mean that we see like a big dip in performance? I don't think so. I mean, look at the chart. If we zoom in here, yeah, it looks pretty much the same on the right as it does on the left."

The clock speeds remained stable across all machines even during these events. If thermal throttling were happening in the traditional sense—CPU speeds being knocked down to manage heat—you'd see it in the clock speed data. You don't.

The Actually Interesting Part

What this means: all four machines, despite different chassis designs and cooling solutions, are being artificially limited by software. The hardware could likely handle more—the 240W power adapter suggests NVIDIA knows this—but the firmware says no.

Which means two things. First, the breathless headlines about the Spark's thermal problems were misdiagnosing the issue. This isn't a cooling failure; it's a power management policy. Second, and more interesting: this could be fixed with a software update.

"In theory, next year, Nvidia could release a software update that would raise that cap and make our machines more performant," Ziskind notes.

Will they? Who knows. Maybe the current limits exist for warranty reasons, or battery life in portable scenarios nobody's actually using these for, or because pushing harder would create thermal management headaches across different OEM implementations. Or maybe NVIDIA just wants to preserve product segmentation—can't have the $4,000 box performing too close to the $10,000 one.

A similar test from Storage Review found comparable results across Gigabyte, Acer, ASUS, and Dell variants, with one exception: the Acer system peaked at just 76°C during heavy workloads while others climbed into the mid-to-upper 80s. Possibly better thermal design, possibly different firmware behavior, possibly measurement variance.

What This Means For Buyers

If you're shopping in this space, the takeaway isn't that these machines throttle or don't throttle. It's that they perform identically for actual AI workloads despite their physical differences, and they're all being held back by the same software ceiling.

The differences that matter are the ones you'll notice daily: port labeling, SSD accessibility, noise levels (the Spark is quietest, the MSI loudest), whether you value a front-mounted power button. The ASUS is hardest to service. The Spark has the fastest storage. The Dell looks most like the reference design because it basically is.

But the performance story? It's the same story on all of them, with the same artificial ending. Whether that ending changes depends on decisions NVIDIA hasn't announced and may never make.

In the meantime, we have clarity on what's actually happening inside these boxes when you push them hard. It's not the story the headlines told. It rarely is.

Marcus Chen-Ramirez is a senior technology correspondent for Buzzrag