TRANSCRIPTEnglish

local AI's biggest problem... they solved it?

23m 36s5,191 words433 segmentsEnglish

What this video covers

WiCi introduces a wireless computing infrastructure appliance powered by an NVIDIA RTX 5090 that lets laptops, Macs, and mobile devices access desktop-class GPU power over local Wi-Fi. By utilizing a virtual driver and an aggressive latency-reducing protocol, the system seamlessly runs AAA games, 3D rendering, and local AI tasks without streaming from the cloud. Additionally, WiCi solves severe consumer hardware limitations by running massive Mixture of Experts (MoE) AI models directly from high-speed SSD storage using intelligent expert caching and task scheduling.

Key takeaways

  • WiCi packages an RTX 5090 and a mini PC into a standalone appliance that connects to client devices entirely over Wi-Fi.
  • A custom virtual driver convinces client operating systems that the remote RTX 5090 is installed directly inside the laptop or Mac.
  • The system solves Wi-Fi latency bottlenecks by bundling instructions, sending only changed frame data, and predicting responses, reducing bandwidth needs to approximately 400 Mbps.
  • Massive local LLMs like DeepSeek 4.1 Flash (750 billion parameters) run at full quality on just 32GB of VRAM by utilizing a 4TB SSD as secondary memory.
  • Storage bottlenecks are mitigated by re-chunking MoE model files for SSDs, caching frequently chosen experts in VRAM and system RAM, and distributing compute across the CPU and GPU.
  • A dedicated model scheduler parks inactive AI models on the SSD, dropping model-swapping times between chat, image, and video generation from roughly 18 seconds down to 4 seconds.
  • The hardware operates as a centralized 'family GPU', enabling multiple household devices to share a single powerful local accelerator.

Topics covered

Wireless GPU Architecture and Network Protocol
Explains how WiCi overcomes the massive latency differential between PCIe connections (microsecond response) and Wi-Fi (millisecond response) by eliminating chatty protocol overhead, compressing data, predicting answers, and simulating native local hardware through a specialized driver.
Running Trillion-Parameter AI on Consumer Storage
Breaks down how a 750-billion-parameter Mixture of Experts (MoE) model, which typically requires over 500GB of VRAM, is executed using an SSD by optimizing model alignment, reading in 64-chunk blocks, keeping active experts in high-speed tiers, and achieving practical inference speeds of 8 to 15 tokens per second.
GPU Multi-Tenancy and Model Scheduling
Addresses the inherent design flaw of consumer GPUs being unable to share resources across simultaneous AI applications by implementing an OS-level scheduler that parks and unparks memory states to an SSD, enabling quick context switching across multiple devices.
“You learn what a mixture of experts model is and why it's kind of awesome for local AI.”
▶ 13:56 · HEAR IT IN THE VIDEO

FULL TRANSCRIPT

0:00

This is the first Wi Fi GPU running an NVIDIA fifty ninety, and it's somehow

0:05

over Wi Fi showing up as this laptop's GPU.

0:07

I'm playing Cyberpunk on it right now.

0:09

Also, I'm running Blender on it, but that's not even the craziest thing it's

0:12

doing. Over here, have a Mac also using the GPU.

0:15

Not only is it generating images and videos very fast, all local AI, but it's

0:19

also talking to this, DeepSeek four dot one Flash, a 750,000,000,000 parameter

0:24

model running at full quality on a box with 32 gigabytes of VRAM.

0:27

This doesn't make any You would normally need over 500 gigabytes of memory to run

0:31

this. It's called WiCi, wireless computing infrastructure.

0:34

Get your coffee ready. I cannot wait to show you how it's doing all of this.

0:38

Now let's get this right out in front. This video is sponsored by WiCi, but don't

0:41

let that deter you. I heavily vet every sponsor and product I look at.

0:44

This one is crazy, which is why I said yes.

0:47

And here here we are. They are a new company, and they're doing something

0:50

pretty bold. And when they told me about this thing, claiming to be the first

0:53

wireless GPU, and it can run trillion parameter models, I had to try it.

0:57

Like, really? Does this thing work? And in this video, we will be trying everything,

1:00

video editing, gaming, and impossible local AI inference with image and video

1:05

generation. Like, it really doesn't make any sense what it's doing.

1:07

And when I got it, like, when it arrived, it was inside a suitcase inside a box full

1:11

of bubble wrap. That's how new this thing is.

1:13

And here it is, an early engineering model.

1:15

Inside this box is an NVIDIA RTX fifty ninety.

1:18

There's also a mini PC as well, and it's actually running OpenWRT for its

1:22

networking. That's how it shows up on my network.

1:23

Now if I get brave enough, I will undo this thing.

1:26

So if you're seeing that b roll now, I did it.

1:28

But I'm gonna save it for the end of the video so I don't break it before I make

1:30

this content. Now the big idea with this thing is local.

1:34

No cloud. Right? I'm already sold, but they've taken it a step further, a step

1:37

that I didn't think I wanted to take.

1:39

This isn't just one device with a crazy powerful GPU.

1:42

Like, I already have that in my studio.

1:44

Instead, they just have the crazy GPU that multiple devices can use.

1:48

Think about that. It's like a family GPU.

1:50

Kind of like how we used to have one family computer for everyone.

1:53

Yes. That was a thing. Comment below if you had one family computer.

1:56

So the idea is that if you have a laptop, a tablet, a phone, it can now use this

2:00

desktop class GPU like it's its own.

2:03

And it's not just throwing a job over to it.

2:04

It's actually using it. Like playing cyberpunk on this little bitty laptop

2:08

really hits it home. For proof, you can actually hear the jet engine over here

2:11

taking off. Oh, this thing's taking

2:16

off. And seeing that fifty ninety show up in the task manager like it belongs to

2:20

that laptop is so weird and kind of wonderful.

2:23

How are they doing this? Let's actually hit that first.

2:25

Because there are two things that were breaking my brain, and I have to figure it

2:28

out. How are they doing this over Wi Fi because Wi Fi is not the greatest

2:32

connection? And how is this laptop suddenly believing that it has a fifty

2:34

ninety? And we'll get to the whole trillion parameter model thing later.

2:37

Now this will honestly be the craziest part of this video.

2:39

I even built this visualizer that will show you exactly what's happening with the

2:43

model on that GPU. Now by way, have you hacked the YouTube algorithm today?

2:46

Let's make sure you do. Hit that like button, subscribe, notification bell,

2:49

comment. You gotta hack YouTube today, ethically, of course.

2:51

Okay. So let's talk about your GPU. Typically, GPU, your gaming GPU needs a

2:55

very, very fast way of talking to your PC, and it does.

2:59

Here's our fun little fifty ninety flying around.

3:01

Can you keep an eye on it? It's fast.

3:02

When you slot that sucker into your gaming PC...

3:04

This is gonna be a gaming PC right here.

3:06

You're putting that into the PCIe port.

3:08

This sucker's 512 gigabits per second, answering requests in under a microsecond.

3:12

Now let's get this gaming PC out of here. We're using a laptop now.

3:15

Now over Wi Fi, instead of PCIe, we're talking about one gigabit per second,

3:19

answering at around one point five milliseconds.

3:21

And that's under the best conditions because Wi Fi is kinda, you know, bleh,

3:25

and that is the technical term. Now don't miss this.

3:27

It's this right here. This round trip time, the time it takes to answer those

3:31

requests that becomes a huge problem.

3:33

And it actually might not seem that bad at first.

3:34

Right? Like, one point five milliseconds.

3:36

That's super fast. No. It's not.

3:38

If you compare that to a microsecond, I'll give you an analogy that help my brain

3:41

process this. If PCIe handed you a picture in one second, with Wi Fi, that

3:47

would take twenty five minutes.

3:49

That's And a big problem because the game and the GPU, they're pretty chatty.

3:52

They're sending instructions and data and answers to create a frame thousands of

3:55

times every second. If every one of those answers had to wait on Wi Fi, we'd be

3:59

playing with PowerPoint. But we're not.

4:02

Look at this. 92 frames per second over Wi Fi, and I'm rendering this in Blender.

4:07

And this is not remote desktop. It's running on this laptop, on this GPU, over

4:11

Wi Fi. How did Wi Fi solve it? Well, they created their own protocol.

4:15

The team at Wi Fi has a networking background so that tracks.

4:17

The first thing they had to do is convince this laptop that he has a real GPU, that

4:21

he's actually using this fifty ninety over here.

4:23

Like, this is your real GPU, buddy, even though you have no room for it.

4:26

They did that by installing a fake driver.

4:28

And what's crazy is cyberpunk, the game treats this GPU like it's right there.

4:32

Like, it's the real one. It does not know the difference.

4:34

Cool, but they still didn't solve the problem of goodness.

4:37

This right here is super slow, this communication.

4:39

So what they did was kinda genius.

4:41

They just told them to stop talking so much.

4:43

Like, literally, that's what they did. They figured out which messages actually

4:46

need answers and then stopped waiting on the rest.

4:49

For instructions, they bundled them, compressed them, and sent them without

4:53

waiting. So for data, it only sends what's changed and doesn't keep repeating itself.

4:57

For answers, it doesn't even wait for the question.

5:00

It knows what's going to be asked, so it just sends it.

5:02

This is the Wi Fi version of that meeting could have been an email.

5:04

And then the GPU sends compressed video back to the laptop.

5:07

These four things they did took what would actually need to be, like, eight gigabits

5:11

per second needed to now all they need is 400 megabits per second.

5:15

Proof is in the pudding. Right now, we are gaming at 92 frames per second.

5:18

Look at the Wi Fi right now. It's about 400 megabits per second going across.

5:22

And get this out of your brain. This ain't cloud gaming.

5:24

This also isn't remote desktop.

5:26

The game isn't being streamed to us from the cloud.

5:28

The game is still happening right there on the laptop, but it's using the GPU over

5:33

there, over Wi Fi, which is just the craziest thing to say.

5:35

Now what's funny is some of the team at WiCi actually used to work in cloud

5:38

gaming. I work in the cloud gaming team in Tencent before.

5:41

And it was around the time that game, Black Myth Wukong, think it's how you say

5:45

it, came out that they had this idea, the Weisai idea.

5:48

At a moment of Black Myth, we were thinking, can we still leave the front end

5:53

on the laptop but push back end to a remote place?

5:57

That's where we started. But then they discovered something kinda cool.

6:01

Okay. And later on, we found that AI has a way more scenarios than gaming.

6:05

K. This is the part that got me most excited when I first heard about WiCi.

6:08

Sure. The Wi Fi gaming is neat, but local AI, that's where things are happening.

6:12

That's where I'm most excited about everything.

6:14

And this is legit a new way to approach using big models with local AI.

6:20

So WiCi actually did two things, solved two big problems with local AI.

6:23

I think they might be a game changer.

6:25

Prove me wrong in the comments. Do it. And some of this might be news to you because

6:29

I'm not like a GPU expert. I don't study this stuff, but I did on this video.

6:33

By the way, these videos take a minute because when someone sends me a product

6:37

like this, I don't just go, neat.

6:39

Here are the talking points. Know, like, I actually get into it.

6:41

I learn the technology. I learn how it works so I can pass this along.

6:44

And when it's something that I think is interesting, I do wanna pass it along.

6:48

I get excited. So the two things that you actually solve or two problems that exist,

6:52

one you're already familiar with, but this first one you're probably not.

6:54

GPUs don't like to share. What?

6:56

We'll cover that here in a moment. And two, this is the obvious one.

6:59

Consumer GPUs can't run big models because they don't fit.

7:02

They don't have enough memory. And, yeah, they solve that.

7:04

So for the first one, GPUs don't like to share.

7:06

What does that mean? I mean, sure. This fifty ninety is great at running local AI.

7:11

Have you tried it? It's fast. Hold up.

7:13

Chuck from the future here. I'm gonna do a quick edit on this video for you.

7:16

Because right here, I started to talk about how the GPU doesn't share.

7:19

But I cannot wait for you to see the local AI stuff, so I'm gonna move that right

7:23

over to now. I just want you to see it.

7:26

And also, want you to see that visualizer I made.

7:28

It's so sick. It's like looking in the brain of DeepSeek.

7:31

Here you go. So, our premise is that you can't run large AI models at home.

7:36

Consumer GPUs like the one inside this box just don't have enough memory to run them,

7:40

yet this box can. How? Let's break it down.

7:43

Now first, local AI is in a fun space right now.

7:45

You can run a ton of amazing models on pretty much any hardware, and you can do

7:50

some cool things. But when we're talking about the smart ones, the frontier like

7:53

ones that get close to what you might experience with OPUS five dot five or GPT

7:58

six, and there are local models that are like that, like Kimi k three or GLM five

8:02

dot three or DeepSeek four dot one.

8:04

These local models do give you a level of intelligence that's pretty dang close to

8:07

these frontier models. And the best part is that it's unlimited, and it's free, and

8:11

it's yours. Sounds great. Right?

8:13

Yeah. Well, you can't do it. Normal people can't run this stuff.

8:17

It requires a lot of memory, like a data center level of memory.

8:21

And have you seen the price of memory? The model I wanna focus on right now is the

8:24

DeepSeek four dot one Flash. To run this sucker full quality, you need over 500

8:28

gigabytes of memory. There is nothing consumer level that can do that.

8:31

No GPU, no gaming PC that can support that much system memory.

8:35

Even the new m five Mac Studios with 512 gigabytes of RAM can't really run this

8:39

comfortably. So, how can this thing run it?

8:43

Because we're not using memory. We're using the SSD.

8:46

Because inside this box right now is a four terabyte SSD, and it's using it

8:52

like memory. Now by the numbers, this is a terrible idea.

8:56

It's such a bad idea. Why are they doing this?

8:58

Sure. SSDs, they give us more capacity.

9:01

You most likely can buy a four terabyte SSD.

9:04

How much are they right now? Field trip, four terabyte SSD.

9:07

A new egg is gonna cost us $500.

9:11

So capacity, we're getting that, but speed, we're losing.

9:14

And that's our main problem right now. And you need speed when working with AI.

9:18

Let me show you what I'm talking about and how bad it normally would be with an SSD

9:22

if we didn't have what WiCi was doing.

9:24

And what they're doing is pretty sick. So just for a comparison, like what we're

9:27

talking about, because there are different types of memory.

9:29

I keep throwing them Memory. VRAM, which is the RAM that you're gonna see inside

9:33

your GPU. When we're talking about the fifty ninety, we're gonna be around 1.8

9:38

terabytes per second. Wicked fast.

9:40

Standard system RAM, most modern stuff, DDR five, 90 gigabytes per second.

9:45

And then we get down to the SSD, 14 gigabytes per second.

9:48

So if VRAM hands you a page in one second, RAM will take twenty seconds, the SSD will

9:54

take two minutes. It's It's like that sloth from Zootopia.

9:57

How does Weisai deal with this? First, I want you to imagine the best scenario

10:01

possible. Imagine that somehow this 59 d has 550 gigabytes of

10:06

VRAM. Can you imagine? No. You can't, but we're gonna do it right now.

10:10

With that, we can run the full DeepSeek four dot one flash with 500 gigabytes of

10:15

required RAM. It also has 750,000,000,000 parameters.

10:19

Pretty smart dude. And in this dream situation, speed is our ally.

10:22

For every word or token that this DeepSeek LLM generates, it has to transfer all

10:27

750,000,000,000 parameters to the chip.

10:31

This equates to roughly about 508 gigabytes.

10:33

This transfer right here, that's 1.8 terabytes per second.

10:36

That's giving us 3.5 words or tokens per second.

GET MORE FROM YOUR TRANSCRIPTS

Sign up for free and unlock interactive viewer, AI summaries, translations, concept maps, and more. No credit card required.