ViceSol

Living in the trenches

[ViceSol @ViceSol]

Duel Casino was unhinged last night:
• Girls clavicular kissing
• Bonnie Blue barely keeping it together
• Mike Tyson casually being there

Missing this stream is an L. #DuelCasino

[ViceSol @ViceSol]

SOMEONE PICKED UP A BOX THE SIZE OF A LUNCHBOX AT A CONFERENCE AND IT RUNS A 235 BILLION PARAMETER AI MODEL WITH NO CLOUD AND NO GPU

It fits on one palm.

That's the first thing that stops people.

Not the specs. Not the price. The size.

Someone held it up at a conference table and the people behind him kept scrolling on their phones.

They had no idea what was in their hand.

Inside: AMD Ryzen AI Max+ 395. 128GB unified memory. 110GB usable as VRAM on Linux.

Let that number sink in.

110 gigabytes of VRAM. On a chip the size of a lunchbox.

A $1,500 RTX 5080 has 16GB. That's the GPU everyone is fighting to buy right now.

This has seven times more. In something you can carry in a jacket pocket.

And here's the part AMD didn't say loudly enough.

The RTX 5080 cannot run a 200 billion parameter model. Not because it's slow. Because it physically cannot fit the model in memory.

This chip fits it. Runs it. Locally. No cloud. No API. No monthly bill.

Qwen3-235B. On a device that costs $1,400 to $2,500. Once.

Cloud equivalent: $200 to $400 a month. Forever. Until you stop. Which means forever.

Someone did the math on a $1,400 device vs $300/month cloud.

Month 5 it pays for itself.

Month 6 every single prompt is free.

The person holding this at a conference table didn't announce a product.

They announced that your AI doesn't belong in someone else's data center anymore.

It never did.

It just had nowhere else to go.

Until now.

[ViceSol @ViceSol]

@plutos_eth Cloud AI made everyone think bigger models require bigger data centers. The next few years might just be a fight over who owns the memory

[ViceSol @ViceSol]

HE'S STANDING IN A GARAGE NEXT TO A BOX THAT JUST MADE A $3,400/MONTH BILL DISAPPEAR.

Power tools on the wall. Sawdust on the floor. And in front of him, something that doesn't belong in a garage.

45 drive bays. All filled. All spinning. All facing the camera like a wall of chrome teeth.

Pause at 0:06.

That's not a gaming PC. Nobody builds a gaming PC with 45 drives.

Someone did the math before reading any further.

45 bays at modern capacity. 2 petabytes. Two. Petabytes. In a garage next to a miter saw.

Now here's where it gets interesting.

The same architecture inside this box runs on a chip that fits in your palm. 128GB unified memory. The size of a Mac Mini. 1.2 kilograms.

Someone calculated what 2 petabytes costs on AWS.

$3,400. A month. Every month. For as long as that data exists. Forever, basically.

This box: $9 a month in electricity.

Read that again.

But the drives aren't even the part that should worry cloud providers.

Pause at 0:15. Open chassis. Green board. Blue cable.

That's a compute board sitting right next to the storage. Not separate. Same box.

It runs an 80 billion parameter coding model. 25 to 40 tokens per second. On a single USB-C power supply.

Someone fine-tuned a 7B model on hardware like this. LoRA. 18 minutes. GPU never crossed 77 degrees.

Zero data left the building. Zero API calls. Zero monthly invoice.

He's standing in a garage holding a screwdriver.

Behind him is something that makes an enterprise AWS bill look like a rounding error.

He hasn't said a word about price yet.

He doesn't need to. The drives are already talking.

[ViceSol @ViceSol]

A 22 YEAR OLD WALKED INTO A COMPUTER STORE, BOUGHT 8 IDENTICAL MINI PCs, AND LEFT WITH A PRIVATE AI CLOUD

He posted a 10-second video. Just the boxes lined up on the counter. Store display behind them. Receipt still in his hand probably.

Nobody was supposed to read the terminal in the second clip.

Pause at 0:08.

That's not a chat window. That's a connection log.

"Accepted client connection." "Client connection closed." Over and over. Dozens of times. In ten seconds.

Someone counted the cycles. At that rate, that's not one person testing a chatbot.

That's a server handling requests from multiple clients. Live. While he filmed.

Each box: GMKtec EVO-X2. AMD Ryzen AI Max+ 395. 128GB unified memory. Eight of them on one table.

Do the math. 1TB combined memory. No server rack. No data center. No rented cloud GPU.

Someone found the spec sheet. Four of these nodes alone can stably run DeepSeek V3.1, 671 billion parameters. He has eight.

Then someone asked why a student needs a private inference cluster.

He doesn't sell hardware. He sells access.

Small teams who don't want their code, client files, or internal docs hitting a public API. $150 to $300 per seat.

24 seats. That's $6,000 a month. The entire stack cost him $14,000 to $16,000.

Three to four months to break even. Then it's just electricity.

Meanwhile the average heavy AI user pays $200 for Claude Code Max, $200 for ChatGPT Pro, $20 for Cursor, $20 for Gemini.

$5,280 a year. Per person. Before they ship anything.

He posted a video of boxes on a counter.

He didn't post the part where those boxes are already taking client connections.

bookmark this and drop a like 👇

[starmex @starmexxx]

https://t.co/BUfLYW0PDz

[ViceSol @ViceSol]

A Chinese engineer posted an 18-second tour of a data center.

No explanation. No caption. Just the footage.

Nobody was supposed to look at the cable labels.

Pause at 0:12. Top of the rack. White label. Black text.

4P-M11-10.

Someone recognized the naming convention. Posted it. Thread started moving.

That's not a generic rack. That's a GPU interconnect topology used in one specific cluster configuration.

Someone matched the fiber color pattern to a known Supermicro deployment spec.

Someone else counted the drive bays visible in frame. 847. Minimum.

At current VRAM pricing that rack alone represents $4.2 million in hardware.

There are at least six racks in the shot.

Then someone asked the question everyone missed.

Why does a cluster this size need a human walking through it with a phone?

The answer came four hours later from an account with eleven followers.

The rack isn't the product.

It never was.

The rack is what you show investors.

The margin is in the workflow sitting on top of it. The reusable procedures. The saved skills. The tasks that run once and get loaded forever after.

One competitor audit. One pricing scan. One client report.

Run it once. Save it. Load it next time instead of paying for the same tokens again.

The companies who built this are selling compute by the hour.

The people quietly winning are selling the output.

$4.2 million in hardware on one side.

A laptop with a saved workflow on the other.

The gap between them is not the rack.

It's who stops paying every time they start.

[ViceSol @ViceSol]

@qkl2058 Receiving satellite beacons is one thing.
Turning that into reliable global positioning with 10–30m accuracy in a “simple 6-step tutorial” is where the internet starts writing fan fiction about RF engineering

[ViceSol @ViceSol]

@cyrilXBT The funniest part of these stories is that the AI always costs exactly $120/month and somehow produces hedge-fund-level returns

[ViceSol @ViceSol]

@Raytar AI dramatically reduced the cost of building websites.
It did not reduce the cost of convincing someone to buy one

[ViceSol @ViceSol]

@knoxtwts Every AI influencer eventually writes a prompt that sounds less like productivity and more like a supervillain performance review

[ViceSol @ViceSol]

@VadimStrizheus Every YouTube automation story starts with “all he does...” and then quietly skips the part where millions of videos get 37 views

[ViceSol @ViceSol]

FOUR MAC MINIS ON A DESK ARE RUNNING A MODEL THAT COSTS $50,000/MONTH TO RUN IN THE CLOUD

No data center. No cloud. No API key. Not a single dollar to OpenAI or Anthropic.

Just four machines connected through a cable and a framework nobody talks about.

Someone posted a 30-second clip. Just showing the setup.

They didn't explain the dashboard visible at 0:11.

Pause. Zoom in.

Four nodes. All Mac Studios. All 512GB memory. All connected.

James's Mac Studio. Mike's Mac Studio. S17's Mac Studio. One more.

Temperature: 40 degrees across all four. They've been running for a while.

Load: 36%. 46%. 58%. 73%. Distributed automatically. No manual balancing.

Two instances of DeepSeek V3.1 running simultaneously. Both active. Both live.

671 billion parameters. Locally. Right now. On hardware that fits on a desk.

On the monitors behind the stack — the model is explaining cochain complexes in algebraic topology on one screen. Teaching Mandarin pronunciation on the other.

Both answers are long. Both are correct. Both generated in seconds.

The cluster dashboard shows the launch interface. Select a model. Hit Launch Instance. That's it.

The part nobody mentions is that you don't need four new machines.

The framework splits the load across whatever you already have. Old MacBooks. Mac Minis from 2022. Mix and match.

The model doesn't care. It just runs.

$50,000/month in cloud compute. Running on a desk with a plant next to it.

The data center was never the only option.

It was just the only option most people knew existed.

[ViceSol @ViceSol]

@so_ainsight The second you see “AI reads charts and moves hands,” you already know half the system is indicators and the other half is confidence

[ViceSol @ViceSol]

@Yerocode0 The “he builds it in 4 minutes on the call” part is doing a lot of storytelling work

[ViceSol @ViceSol]

@CoinSh0t The fastest way to spot a fake deep-tech story is when 3 different advanced systems magically become a single feature

[ViceSol @ViceSol]

@AIVersePlay Whenever a tool is described as a “money printer,” the only guaranteed thing being printed is content about the tool

[ViceSol @ViceSol]

@IBuzovskyi The most valuable AI skill isn't choosing the smartest model. It's knowing when you don't need the smartest model

[ViceSol @ViceSol]

@codewithimanshu If a strategy really prints money like clockwork, selling engagement for retweets is a very interesting business decision

[ViceSol @ViceSol]

@leopardracer The craziest part isn't that it runs locally. It's that an 8GB gaming laptop is starting to look more like infrastructure than a toy

[ViceSol @ViceSol]

@sairahul1 The biggest misconception in AI right now is thinking better prompts scale. Systems scale. Prompts don't