Transforming Compute

NVIDIA announced it won't break out gaming revenue as its own line item on the balance sheet anymore, instead lumping it into the Edge Compute category alongside robotics, game consoles, PCs, workstations, AI-RAN base stations, and automotive. To be fair, this makes sense: they own 73% of the PC Gaming discrete GPU market according to the Steam Hardware survey, and 95% of recent GPU sales.

They have decisively won that market, for now.

It also doesn't help that Graphics accounts for ~10% of their entire revenue under the traditional reporting structure; that's chump change when their compute and datacenter revenues are 9x that amount and making up the overwhelming bulk of their income. On paper, from a financial standpoint, this change makes sense and lets them bundle some less-than-stellar divisions (like automotive) alongside more lucrative ones (like gaming) to make the narrative sound better to outsiders.

In practice however, this move comes across as a gigantic middle finger to the personal compute category as a whole. NVIDIA has made it clear that gaming and personal compute simply isn't a core focus anymore, at least while they can keep multiplying their income through chasing fads like crypto mining and AI compute from large enterprise customers. For the rest of us plebeians, we can pay for an AI appliance and the power it consumes so we can then rent access back to its larger network - but not our own appliance - for a nominal monthly fee; I wish I was kidding.

The writing has been on the wall since the dawn of the public cloud era, but its messaging has grown impossibly loud in the past year as storage and memory costs exploded, personal compute became scarcer, and cloud compute became the default rather than the luxury: you will own nothing, and you will be happy.

This isn't my first experience in a transformation of compute itself and the consumer relationship to it. I was there building custom PCs from (literal) garbage at a time when you still had to do research on purchasing components such that you got an actual manual with it from a reputable manufacturer. The 2000s-2010s ability of being able to buy PC components in beautiful packaging and flush with accessories is fairly recent in the grand scheme of things, and a welcome transformation that let us customize compute into forms we needed, rather than the bland boxen churned out by Compaq and Gateway and Dell of the time. This made sense specifically because compute was so diverse and constantly changing and growing and evolving on a yearly basis. Hell, I was there for that fleeting moment when a high-end gaming rig would've had two GPUs, a physics accelerator, and a sound card, yet also pull less than 500W from the wall under load while struggling to play Crysis or UT3.

Fun times.

Point is, this is not the first time compute has been forced to undergo a transformation to survive. If anything, this recent transition feels a bit overdue: NVIDIA and AMD both have been shoehorning more raw power into the GPUs to eke out performance for a while because any substantial performance uplifts requires a fundamental reworking of underlying architecture or immense optimization by developers. The sole outlier has been Intel, who had the benefit of no legacy bullshit to support when dropping their Arc lineup - though their driver issues make playing older titles a crapshoot at times. The name of the game since roughly 2020 has been more power and more cores for more performance, something not exactly scalable in the face of climate change and a demand for more efficient resource usage.

Yet when you step foot outside the high-end PC gaming arena, suddenly you begin seeing the transformation of personal compute already in progress: single board computers aren't a joke anymore, but instead can be full-blown workstations and servers thanks to the efficiency of Linux and BSD operating systems. Apple's choice to dump x86 has paid dividends, yielding tiny boxes and chilly laptops that can outperform comparable x86 hardware while sipping a fraction of the electricity. In a lot of ways, it's a return to the 80s and 90s of compute: manufacturers building bespoke processors with the accelerators their customers demand, with either their own OS (Apple), an open source OS (Linux, BSD), or a commodity OS (Windows) to run the show. A return to weird form factors like keyboard compute modules and spatial compute and handhelds and VR headsets and adorable little compute cubes or NUCs.

Despite insane memory and storage prices rising into the stratosphere, personal compute has soldiered onward in its own way. If all you've been focusing on is bespoke gaming PCs, you've been missing the real transition towards smaller footprints and efficient deployments. It's an area NVIDIA has consistently been losing ground in for a while now, as gamers move to Linux or APU-based compute devices that AMD specializes in and NVIDIA has chosen to ignore. When your electric bill climbs due to AI data center builds nearby but you still want to play your Steam Library, a Steam Deck or Framework desktop is a far more appealing prospect than a gigantic(ly expensive) bespoke rig that can trip a circuit breaker with power consumption and still require DLSS for acceptable performance.

Personal compute isn't dying; it's changing.

If there's one thing I've learned during my career in technology, it's that no cycle lasts forever. I've seen RAM be insanely expensive during the P4/Doom 3 era, and be so cheap it basically came for free with the purchase of a bag of Doritos in the DDR/DDR2 era. I've built rigs where 1GB was a small fortune, and others where 64GB was a comparative bargain. I remember when 512MB GPUs were impossibly unaffordable, yet now have a 24GB 3090 chilling in an eensy SFF box.

Business has cycles, and technology, like it or not, is a business. This cycle is putting the squeeze on components, which means those who find success will be those who can squeeze out maximum performance from limited hardware; the indie scene is doing quite well for itself even as AAA collapses under its own inefficiencies and corporate mismanagement. If all you know how to do is build in times of excess, you will falter the second there's a drought.

I am very much a "frame puritan", in the sense I have an intense dislike of reconstruction, interpolation, and generative techniques and their associated artifacts. Knowing that the current crop of games requires DLSS or FSR on even the most cutting edge of hardware platforms for acceptable performance ironically makes cheaper, less-capable equipment more palatable to my tastes; if I'm going to be dealing with temporal ghosting artifacts and a blurred image due to upscaling reconstruction techniques anyway, then why the fuck should I spend $3k on a bespoke rig when a $700 Mac Mini or $1000 Steam Machine will perform just as admirably and not require Windows 11?

The same knowledge applies to the enshittifying trend of cloud compute as well. If I'm going to be paying $20 a month for a paltry 2TB of cloud storage and a bunch of AI slop that can be taken from me at any time by the vendor and with no means of recovery, then why not pay a little more for a 4TB Beestation that I own instead? Or go throw that money at PikaPods for a NextCloud deployment that's mine? Or at Wasabi for a bucket of S3-compatible storage?

This extends to modern software deployments as well. Docker Compose makes it easier than ever for anyone with basic technical literacy to deploy their own applications in a way that's quick, efficient, and (relatively) secure on even potato hardware. I run a Jellyfin instance and Owncast deployment with hardware transcoding, an RSS Aggregator, an RSS Bridge, two anonymous imageboards, IRC server daemon, and reverse proxy on an Intel N100 NUC - with enough compute left over to run that entire stack under load eight more times. Containers can and should be made more accessible to compute end users, as we can throw them onto older hardware and provide our own services for pennies on the dollar compared to hosted versions.

We're at this unique junction where those who can operate on constraints are in a position to thrive while everyone else flounders. The enshittification of subscription services and general fatigue with technology means folks are more receptive to putting in the work to save some cash or preserve some ownership. AI's theft of content for training has given hoards of consumers tired of ever-increasing subscription prices and dwindling content libraries a green light on sailing the high seas of piracy (since if it's okay for Anthropic and OpenAI and Meta to steal for personal profit, surely it's okay for John Q Public to pirate without seeking remuneration, right?).

So as much as it sucks to see the hobby and passion we love change in such fundamental ways, we should be looking at the opportunities these changes bring in terms of taking back control over our technological experiences and redefining our relationship to these companies in the first place. They're the ones dropping the ball in the name of personal profit and totalitarian control.

I say we take it, and build a future of compute based on independence and efficiency instead of gluttonous excess and insatiable greed.