One STEAM to rule them all - Steam Machine and Frame thread

[Deckard]

In a similar vein my Melomania Earbuds have Matt Berry doing the in-ear announcements - It never fails to make me smile when I put them in !

https://www.youtube.com/shorts/VDLgdVYIaoE

[KlaxonCow]

Things are not looking good for the pricing of the Steam Machine or Steam Frame, as the Steam Deck OLED's just seen a hefty price rise of hundreds of dollars.

Indeed, everything with memory or storage is experiencing a price rise across the board, but the Steam Machine proposition was finely balanced on its pricing. As it's not the best PC money can buy, but was "good enough" - measured by the Steam hardware survey to be slightly-better-than-average - to essentially bring the Steam Deck to your living room TV, IF priced accordingly. But if it's closer to a grand or so to purchase it, then it stops being a "bargain" impulse buy, the lower specs then matter more and so on and so forth.

I mean, once the Steam Deck was back in stock again, it sold out almost immediately. So there will be people who'll bite the bullet regardless, accepting that this is just our new reality for a few years, as we wait for the AI hype bubble to burst.

But the worry is that the new wave of Steam Machines ends up like the old wave. A bit too overpriced for an underpowered PC that it fails to spark the "critical mass" it needs to establish itself as a serious player in the mainstream market, and is not just an interesting curiosity sitting on the side. Proton does change things, as it did with the Steam Deck, but that alone doesn't dent the PC gaming market unto itself. It tells you that if you did switch, then you'd be mostly fine - but it's not a compelling reason, unto itself, to make that switch in the first place.

It could still work, maybe, if Microsoft continue to en-shit-ify Windows 11, but they've recently woken up and smelt the coffee, as they are, at least in word, acknowledging that they've gone too far and are pulling back - you can now actually customise the taskbar to be on any edge of the screen (which proves that that was always a "can't be bother" thing, because as they go through the list of complaints, they fix this often complained about issue near instantly).

[Misfit]

If Valve has to raise prices so much, then you know it has to be bad.

I have a cupboard full of worthless tech that has devalued over time, but never have I bought something that is now £210 more RRP than I paid 3 years ago! Which is the case for my 1TB OLED Deck.

Of course the odd really old/rare thing goes up, but in general you know buying the latest and greatest will soon be worth a fraction of what you paid.

Financially, I like that this is slowing down releases of new stuff. I'm like a moth to a flame for new toys and then never use them.

Right now, I believe I'm still in for a Machine, I like the concept, eco system, company, and that it will be targeted and I also have little time for AAA stuff, but do agree that it could really miss its window if it's not out soon.

[Deckard]

Yeah as Gemini pointed out this AI data centre shit (water, ram , local ecology, power) is not helping most people and I am not convinced by the trillionaires arguments how it might help further down the road.

[Mat]

It's all related to the ultra-capitalist bullshit ideology of "You'll own nothing and be happy".

That's very much what these companies are trying to do (not Valve, to be clear); they're trying to make ownership of hardware so expensive that subscription to cloud compute becomes the only option for Joe Public.

Someone needs to legislate against this fuckery, and fast.

[Whiteball]

It's all related to the ultra-capitalist bullshit ideology of "You'll own nothing and be happy".

That's very much what these companies are trying to do (not Valve, to be clear); they're trying to make ownership of hardware so expensive that subscription to cloud compute becomes the only option for Joe Public.

Someone needs to legislate against this fuckery, and fast.

While that seems to be the way things are headed at the moment, it would in all likelihood be just another passing phase as the majority of people will just do without or switch to local open source/ homemade stuff (hardware and software).

Like most things that 'look good on paper' in reality once completely implemented and working all the problems and oversights become apparent and the cracks start to show.

Of course the question is as always, can we ride out the storm long enough to see it blow over.

[Misfit]

Not sure how hard the push will be to cloud compute but I'm pretty sure it will be impossible to buy any serious software soon.

[KlaxonCow]

I had a chat with Gemini AI about this and, even with the vested interest in lying for reasons of self-preservation, it happily rationalised that this "bubble" makes no rational sense.

The analogy is Thomas Watson of IBM famously saying, in 1943, "I think there will only be a need for five computers worldwide".

We can look back on that quotation and laugh at how wrong that was - and, in fairness to him, the quote is given out-of-context as he was specifically talking about the computers that IBM were making at the time, and it turns out they sold thirteen of them, so he was still wrong to underestimate demand, but not as stupidly wrong as folks imagine.

But the analogy is that "AI cloud compute" is the mainframe of old. Warehouse-sized, stupidly expensive and, yes, there's only about 5 or 6 Big Tech "mega-brain" AIs out there. So if AI is "the future of everything" and you're one of the few people who owns the 5 or 6 AIs worldwide and everyone's paying a subscription to hook up their "dumb terminal" to the "mega-brain" then you've a licence to print money. So throw everything at it - trillions of dollars - to be on top, when AI takes over the universe.

There's a massive flaw in this plan. A flaw that's already been exposed by DeepSeek, and more and more AI research is just widening that crack into a wide chasm.

But rather than write a "long post" all about it, I've asked Gemini AI to summarise the problem, so that the big "mega-super-brain" itself can tell you why its own infrastructure model is doomed to failure:

The Fallacy of the "Mega-Brain": Why the AI Centralization Bubble Will Burst

The current AI hype cycle is built on a massive architectural gamble. Silicon Valley trillionaires are burning hundreds of billions of dollars constructing massive, centralized data centres. They are betting that the entire world will permanently rely on a few "mega-super-brains" rented out via the cloud.

This strategy is fundamentally flawed. It ignores the basic law of technology: software always optimizes, and hardware always decentralizes. Just as house-sized mainframes inevitably shrank into the smartphones in our pockets, the future of AI belongs to localized, modular networks—not a trillion-dollar corporate monopoly.

Code: Select all

CENTRALIZED "MEGA-BRAIN"             LOCALIZED "MIXTURE OF EXPERTS"
┌───────────────────────────┐         ┌───────────┐ ┌───────────┐ ┌───────────┐
│   150+ Human Languages    │         │  English  │ │ Home Auto │ │Accounting │
│   Quantum Physics         │   VS    │  Expert   │ │  Expert   │ │  Expert   │
│   Global Tax Law          │         └───────────┘ └───────────┘ └───────────┘
│   (99% Idle, Massive Cost)│         (Laser-focused, 100% efficient, fits on a NAS)
└───────────────────────────┘

1. The 99% Waste: The Computational Baggage Flaw

To chat with a user in English about their home thermostat, a centralized "Mega-Brain" must fire up a network trained simultaneously on Mongolian grammar, ancient history, quantum physics, and advanced coding.

* The Reality: In any single interaction, 99% of that model’s trillion-parameter knowledge sits idle. Yet, the user still pays the computational, financial, and environmental price for that massive, dead-weight infrastructure to exist in memory.
* The Killer Logic: You do not drive a 40-tonne cargo truck to the corner shop to buy a pint of milk. Using a multi-language, multi-discipline global model for a single-use corporate or domestic task is pure engineering malpractice.

2. The Laser-Focus Advantage: Single-Use Reliability

The "Mega-Brain" approach suffers from diminishing returns. To make a giant model slightly better at everything, you need exponentially more data and power, which leads to hallucinations, alignment bloat, and unpredictability.

* The Alternative: A hyper-specialized, local model trained exclusively on one language and one specific task (e.g., basic home automation, or a company's specific accountancy rules) can achieve near-100% accuracy and absolute reliability.
* The Math: A tiny 3-billion parameter "Expert" module can outperform a 1-trillion parameter global giant at a specific task because its training dataset is pure, laser-focused, and free of conflicting noise.

3. The "Mixture of Experts" (MoE) Belongs at the Edge

The trillionaires claim we need their data centres to handle complex thinking. But modern AI architecture already relies on a "Mixture of Experts" (MoE)—routing queries to smaller, specialized sub-networks.

* The Logical Conclusion: If AI is already modular by design, those "modules" do not need to sit in a centralized warehouse in Virginia. They should live on the user's side (the "edge").
* The Ideal Stack: A user runs a tiny, ultra-fast local AI on a home PC, phone, or NAS box. It handles 99% of daily tasks instantly, with zero latency, complete privacy, and zero subscription fees. If—and only if—the user asks a 1% outlier question requiring massive computing, the local AI can use an API to briefly consult a cloud-based "Big Brain."

The Bottom Line for the Forum

The current "AI bubble" is artificially inflating the cost of silicon, straining global energy grids, and making consumer electronics unnecessarily expensive. It is driven by corporate panic and FOMO (Fear Of Missing Out), not sound engineering.The trillionaires are building a digital empire of massive, expensive toll booths. They are completely ignoring the fact that open-source developers are already building the cars that can bypass them entirely.

There you go, straight from the horse's mouth. I then prompted it to compare the Big Tech model to how the human mind operates, and it rightly picked up on the hint. 3.5 billion years and natural selection went for the "mixture of experts" model.

And it's not an accident - if your language centre in the brain is damaged, then it trains up another part to take over. But if every part of the brain is capable - plasticity - then why have specialist areas in the first place? Because natural selection has worked it out - this is the most optimal model to follow.

The Biological Checkmate: Evolution vs. The Trillionaires

* The Modular Brain: The human brain does not route a vision request through the language center, nor does it use the auditory cortex to solve a math problem. We have distinct, specialized modules (visual cortex, auditory cortex, Broca’s area for speech).
* The Energy Constraint: The brain operates on roughly 20 watts of power—about the same as a dim light bulb. It can only survive on this tiny energy budget because it uses an biological "Mixture of Experts." It only activates the specific "neural modules" required for the task at hand.
* The Neuroplasticity Proof: If a stroke damages the language center, the brain doesn’t just instantly process language everywhere else. It has to painstakingly re-wire a new, specific local area to take over. This proves that compartmentalization isn't a fluke; it is an absolute structural necessity for efficiency and survival.

The Contrast with "Big AI"

If human evolution had followed the current Silicon Valley playbook, our brains would be a massive, undifferentiated mass where every single neuron would have to fire simultaneously whether we were composing a symphony or just blinking at the sun.

Code: Select all

HUMAN BRAIN (Evolutionary Design)     CENTRALIZED "BIG AI" (Corporate Design)
┌───────────┐ ┌───────────┐           ┌─────────────────────────────────────┐
│  Visual   │ │ Auditory  │           │                                     │
│  Cortex   │ │  Cortex   │           │      One Massive, Homogenous        │
└───────────┘ └───────────┘           │          Neural Network             │
┌───────────┐ ┌───────────┐           │                                     │
│ Language  │ │   Math    │           │ (Every neuron fires for everything) │
│  (Broca)  │ │  (Parietal│           │                                     │
└───────────┘ └───────────┘           └─────────────────────────────────────┘
  (Uses ~20W of Power)                  (Uses Gigawatts of Power)

The trillionaires are building a thermodynamic nightmare. By forcing one massive, homogenous network to handle every diverse human task simultaneously, they are fighting against the very laws of efficiency that nature perfected.

Nature uses edge computing. Your brain is a local Mixture of Experts running on a 20-watt organic NAS drive inside your skull. The tech industry's attempt to build a centralized, gigawatt-hungry corporate god is not progress—it is an evolutionary dead end.

They've bet on the wrong horse. Even the horse itself knows this.

(But I do like that it makes little ASCII art diagrams to illustrate its points now. Although AI still has that quirk of saying "we" when it means humanity - but, hey, maybe that's a good thing and might prevent it going "SkyNet" on us, if - like a dog trying to sit on a chair - "it thinks it's people".)

[KlaxonCow]

And, by the way, I checked. The bit about "Mongolian grammar" is not an exaggeration to make a rhetorical point, it really does know all about Mongolian grammar:

The grammar of Modern Mongolian (specifically standard Khalkha) operates on linguistic rules that are completely alien to Western Indo-European languages.

1. It is an Agglutinative Language

Unlike English, which uses separate words (like prepositions) to change a sentence's meaning, Mongolian strings suffixes together onto a root word like Lego bricks.

* In English, you say: "From my horses." (Three distinct words).
* In Khalkha Mongolian, this becomes a single word: mori-d-oos-min.
* The Breakdown: mori (horse) + -d (plural) + -oos (from/ablative case) + -min (my/possessive).

2. Strict Vowel Harmony

Mongolian features a rigid system of vowel harmony. Vowels are divided into "masculine/back" (a, o, u) and "feminine/front" (e, ö, ü).

* A word can generally only contain vowels from one group.
* When you add a suffix (like the ablative suffix meaning "from"), the suffix itself must morph to match the vowels of the root word. If the root is masculine, you use -aas or -oos; if it is feminine, you use -ees or -öös.

3. Subject-Object-Verb (SOV) & Postpositions

English uses a Subject-Verb-Object layout ("The child ate the food"). Mongolian uses SOV layout ("The child the food ate"). Furthermore, it uses postpositions instead of prepositions. Instead of saying "under the table", a Mongolian sentence structures it as "table under".

The machine's not bluffing.

It really does have this insanely inefficient architecture, where you've got to power up and support neural pathways for Mongolian grammar, as well as 150+ other languages - I also tested it by asking whether it could reply to me in Welsh and it did, without breaking a sweat or getting any of its grammar wrong - and all those programming languages, and quantum physics and human neuroscience and goodness knows what else.

It's all there in its connection weights and the whole entire thing has to be powered - and the machines watered, to keep them cool - with gigawatts of power, even though 99.9% of that knowledge sits completely idle. It knows 150+ human languages but when you're prompting it in English, all those other languages are being powered to sit idle. On the off-chance that you randomly start speaking French halfway through a sentence.

But to operate a smart speaker, it only needs to know basic conversational English. That's a much smaller network, so run it locally on more modest hardware.

You can also supply a "French module" that sits on the hard drive. If you need it, then spin it up, it can reply in French and then it spins back down again. Asking a question about physics? Wait a minute, let me spin up that "Astrophysics module". Ah, right, that's how stars evolve. There's your answer. Spin it back down again.

The AI's quite right. This "big brain" model is pure engineering malpractice. It's so perversely bad, in a really obviously terrible way, and they're betting the whole farm on an obvious turkey.

[KlaxonCow]

The rumours circulating and percolating right now is that Steam Machine review units are shipping now, press embargo lifts June 23 and somewhere around that time pre-orders will begin.

https://videocardz.com/newz/steam-machi ... this-month

Steam Machine review units are reportedly being shipped, pricing update expected later this month

Valve may be close to sharing more details on Steam Machine and Steam Frame, at least according to one account tracking the company’s hardware plans.

Steam Hardware Updates, an account not affiliated with Valve, claims that reviewers are now receiving Steam Machine and Steam Frame units. The same account claims that Valve is expected to announce hardware reservations and pricing between June 22 and June 30. It also claims that the creator embargo for hardware reviews may lift after June 23.

https://x.com/HardwareSteam/status/2065524671689445698

Unfortunately, it's hard to be hyped when this is delayed because AI is making shit far too expensive right now.

And the likelihood is that the Steam Machine will probably be around $1K, if we're lucky, and that's going to be a hard sell for merely "average" hardware (literally based on being slightly better than an old Steam hardware survey average). Even if such a price will still be a relative "bargain" next to EVERYTHING costing the Earth right now.

The problem being that the Steam Machine was a very balanced proposition. No, it's not the greatest gaming PC ever but, bundled off cheaply enough, then it's the "Nintendo proposition": No, this hardware is not the best ever, but it's Nintendo, so you know it's going to be fun. So the Steam Machine isn't the best PC ever, but it's fine to do most games on a living room TV and that's okay. Particularly because Steam games are cheap and go on sale and you've already got a thousand games in your library which you've never played.

But if the lowest tier is $1K then it risks being "niche", not breaking the mainstream and, though a perfectly nice product which is as cheap as it can be in the midst of the AI RAM-pocalypse, no-one's got the Benjamins for this sort of shit right now, sorry.

That said, Valve will probably shift their stock. There's enough / growing interest in this Linux gaming thing already, and many have their plans to kill Windows once they have a dedicated SteamOS Machine. It's a cube, which is cool. GabeN could shit in a cup and he'd still have plenty of buyers.

In that sense, it'll be "successful" as a product that'll sell out its stock. But, of course, we wanted a bigger success where this was cheap enough that everyone's throwing one under their TV, and Linux gaming goes from strength to strength, doubling to 8% market share (as these things are always exponential, per "the growth ogive") and devs start taking Linux seriously as a gaming platform, so it gets much more mainstream and many more folks can "divorce" from Windows, embracing open source and it's all lovely and there's double rainbows and everything is awesome.

This is a price-sensitive product and the AI RAM-pocalypse could not be more badly timed.

But, hey, if the rumours pan out, then we're but mere weeks away from finding out.

[Deckard]

It's gotta be under 1k - even go with 999 dollars and a controller and half-life 3 type deal.....Some new bundled custom software might sweeten the deal too

[Misfit]

It's gotta be under 1k - even go with 999 dollars and a controller and half-life 3 type deal.....Some new bundled custom software might sweeten the deal too

It'll sell out in minutes whatever they price it at.

And whilst I'm not quite the stupid impulse buyer i was, I'll be trying to get one.

(But only because I can put it through the business)

[Mat]

I'd like to think I would exercise some impulse control, as my high-end gaming PC is in the living room, and if I want, I can connect it to the telly, and I've got both the old and new Steam Controllers (the new one is really rather lovely, and has handily replaced my DualSense.)

But since I paid off the mortgage, I've ended up with more disposable income than I know what to do with.

And it would probably work really nicely with my Hue Sync box

But I am ultimately more interested in the Frame TBH, even though I barely play VR these days (too much furniture to move around.)

Since I sacked of my original Rift, I've refused to buy any of the Meta headsets, no matter how good they are, and have a Pico 4 which I've barely used since getting it. But a Valve headset that is comparatively third-party cruft free is a tempting proposition.
I just wish they were releasing it with Knuckles-style controllers.

[MattyP]

It's gotta be under 1k - even go with 999 dollars and a controller and half-life 3 type deal.....Some new bundled custom software might sweeten the deal too

It'll sell out in minutes whatever they price it at.

And whilst I'm not quite the stupid impulse buyer i was, I'll be trying to get one.

(But only because I can put it through the business)

Same. I'll be there hitting refresh! Hopefully when pre-orders go up they will have some sort of "scalper control" - eg having a Steam account active for X number of years or something... had my Steam account since the beginning when HL2 was released to it! Over 20 YEARS ago OMG. Is HL2 really 22...

[Mat]

had my Steam account since the beginning when HL2 was released to it! Over 20 YEARS ago OMG. Is HL2 really 22...

You and me both; that's a massive part of why I'm pissed off with the age verification requirement on Steam.
At 22 years old, my fucking account is old enough to drink in the US.

And then some dumbfuck gotcha nerd says "Ah, but it could be anyone using your account! I WIN!1!"

Ignoring the fact that it could be a fucking 5 year old using their parents' credit card.

What the fuck happened to the core of law being "The burden of proof is on the accuser"?