Notes on Local AI - Part 1 - Storage

17 Aug, 2026

Table of Contents

1. Introduction: One Year of Self-Hosting AI
2. The Motivation: Breaking Free from the Cloud
3. The True Cost of Ownership: Cloud vs. Local SSD
4. Upping the Game: Enter the NAS
5. The 12TB Showdown: Cloud vs. NAS
6. The AI Power User: Defining the Tiers
7. Tier 1: The Starter Storage Setup
8. Tier 2: The Fine-Tuner Storage Setup
9. Deflating the TCO with Local AI
10. The Final Pillar: Data Privacy

Introduction: One Year of Self-Hosting AI

As I approach my 1-year anniversary of self-hosting AI, I want to share some notes on how it all started out and where I am right now in this journey. I have been putting a lot of thought into this over the last year, and my pricing model going forward is going to be strictly value-based.

Before we get into what my setup is and if it actually is a good idea to go local, I want to talk a little bit about motivation.

The Motivation: Breaking Free from the Cloud

Self-hosting is not for everyone. It is a pain in the ass to maintain things. This started way before AI for me. Phones used to support microSD card slots which were relatively cheap to buy, swap out if they got full, but still keep old content like photos, videos, and audio. But then it became annoying to manage my digital life over the span of a decade.

Big providers like Google and Apple offer a lot of convenience features like categorizing, geotagging, running face recognition (even though I didn't ask for it) on my photos, and bringing up memories of photos I don't even remember or see anymore (which was nice). But then it comes with a cost I have to pay for the cloud. iPhone comes with iCloud, and Android has its own thing based on providers.

Okay, all this is fine until I realize, hey, I pay $9.99 per month for 2TB of storage. It adds up quickly to $600 for 5 years. Of course, my digital footprint is huge because I take photos and videos of all kinds of crap that I think is interesting (I will write in detail about how obsessive snap culture in my mind works later). It is going to surpass 2TB very soon (because I need it to be in 4k 120fps—because why not). A 2TB SSD is like $350. So why should I pay for a cloud? ROI seems cheaper if I just get an SSD. Right? Right? RIGHT?

Okay, so now I got a 2TB SSD because my brain is convinced I am saving money in 5 years. I cancel that cloud subscription and move all my stuff to the SSD. Voila! It felt good to cancel that 2TB subscription.

The True Cost of Ownership: Cloud vs. Local SSD

But wait, my phone is 128GB to 256GB buffered local storage, so I gotta do this sync every time I hit full storage. Let's say this happens every x days. The time taken for me to do a "connect phone, copy paste, wait for this to complete, delete all of the stuff on my phone" is y hours. Let's also put in the doom factor where the SSD can just wipe itself clean (Oh yeah, it's a thing, check out Sandisk) or a house burning down, or you know, just data corruption causing me to lose all my data as a probability factor on top of this.

Then there is the part where I get no more memories, no more grouped photos; it's just a dump of randomly named files that I probably am never going to look back at again. Let's call this the John Cena factor: "You won't see me (again)".

The real cost of ownership—not just numbers but also real quality of life impact included (I will call it TRCONJNBAEQOLII, or just TCO)—can be summarized by the following equation: (AI slop below)

--- AI SLOP ALERT STARTS !!! ---

Here is the mathematical equation for TRCONJNBAEQOLII (The Real Cost of Ownership Not Just Numbers But Also Real Quality of Life Impact Included), or simply TCO, formatted in Markdown.

Variables

• $S_{ssd}$: Storage capacity of the SSD (e.g., 2TB)
• $C_{ssd}$: Upfront monetary cost of the SSD
• $S_{phone}$: Buffered local storage of your phone (e.g., 128GB - 256GB)
• $D_{rate}$: Daily data accumulation rate (how fast you fill your phone)
• $x$: Frequency of syncs in days, where $x=\frac{S_{phone}}{D_{rate}}$
• $y$: Time taken per sync operation (connect, copy, wait, delete) in hours
• $V_{time}$: The monetary value you place on your own free time (per hour)
• $P_{doom}$: The Doom Factor (probability of hardware failure, house fire, or corruption)
• $V_{data}$: The priceless value of the data being lost
• $C_{cena}$: The John Cena Factor (the unquantifiable loss of curated memories, facial recognition grouping, and accessibility—because "you won't see them again")
• $N$: Number of years you plan to use this setup (e.g., 5 years)

The Equation

The total cost over $N$ years can be calculated as:

TCO = C_ssd + [ N × 365 × ( y × V_time / x ) ] + ( P_doom × V_data ) + C_cena

Breakdown of the TCO

1. $C_{ssd}$: The initial illusion of saving money.
2. $N·365·\left(\frac{y·V_{time}}{x}\right)$: The Manual Labor Tax. Since $x$ gets smaller as your phone storage $S_{phone}$ fills up or $D_{rate}$ increases, you spend more and more of your life ($y$ hours) acting as a human data pipeline.
3. $(P_{doom}·V_{data})$: The Anxiety Premium. The expected cost of complete data annihilation. Without the cloud's multi-region redundancy, this probability $P_{doom}$ is significantly higher.
4. $C_{cena}$: The Nostalgia Deficit. A constant quality-of-life penalty applied because a folder named DCIM/100FOLDER/IMG_4921 on an external drive does not spark joy like a notification saying "Look back at your trip to the mountains."

--- AI SLOP ALERT ENDS !!! ---

Okay, that was not so sloppy; it kind of captured most of it right. My handwritten equation is much more sophisticated.

Yeah no, that felt good to say, but it was a compulsive lie where my ego is refusing to accept the beautiful equation that the machine generated. (Note to self: Don't reflect immediately when writing; you can reflect and choose to not write it down).

Anyways, side quest rant completed, let's get back to plugging some numbers into the equation.

The Contender (Defending Champion): 2TB Cloud Subscription

• Cost: $9.99 * 12 + tax
• Manual labor: 0
• Doom risk: 0
• JC factor: 0
• TCO 5 years: $659.34 at a 10% tax :P (C'mon, you gotta be accurate, or what's the point? Penny wise, pound foolish).

The Challenger: 2TB SSD Local Setup

• Cost of hardware: $329.99 + tax
• Sync frequency: every 90 days
• Time per sync: 1 hour (I need to find the cable, plug it in, wait for 100GB+ transfer, verify it's copied, delete from phone)
• Value of my time: let's say $25 per hour (I like to heavily undervalue my personal time)
• Doom factor: 5% chance over 5 years (the probability I pour coffee on my drive, it mysteriously disappears, or you know, wipes itself clean)
• Value of my memories: $25,000 (the amount I would pay to get everything back if it gets lost)
• JC factor: $1,000 (can't do daily memories, searchability, seamless sharing with friends and family from smart reminders)
• Total SSD TCO (5 Years): $3,106.39 (Show your work next time pls!)

Okay, that quickly went from "Oh wow, I save $296.35" to "I am losing $2,453.65 in time, risk, and quality of life."

Gosh, this is painful. But wait, I am not gonna get beaten down that easily. Let's do a little more optimization at the risk of sunk cost.

Upping the Game: Enter the NAS

There are a few axes we can explore next from here, but my favorite one is "Why don't we throw more money at the problem and see if cost amortizes over 10 years" 🤪.

We can build a cloud, or wait, there should be solutions already available for this.

Discovers NAS which has 4 bays + 12TB slots

Okay, we are upping the problem statement. I've convinced myself my data footprint will be 12TB, say, in 10 years. Let's solve for this now :P

With the NAS comes new terms in the equation, like the fact that it's software I cannot add features to (closed system) a.k.a. I need to convince myself to not have the need in the future to query to see all photos of me wearing a blue t-shirt. And also, I can potentially get hacked, especially with all the AI models running wild right now.

--- AI SLOP STARTS ALERT !!! ---

Here is the updated mathematical equation specifically for the NAS TCO, taking into account that an off-the-shelf NAS has a built-in software ecosystem that is a walled garden that will age poorly compared to trillion-dollar tech giants, plus the modern realities of cybersecurity.

The NAS Variables

• $C_{nas}$: Total upfront cost of the NAS enclosure and hard drives.
• $N$: Number of years in operation (e.g., 10).
• $E_{power}$: Annual Electricity and Heat Tax (power consumption + HVAC offset).
• $T_{nuisance}$: Annual Physical Nuisance Tax (time spent cleaning dust, listening to fan noise, replacing drives).
• $P_{doom}$: Probability of physical hardware destruction (fire, flood, dropped drive).
• $P_{hack}$: Probability of a network breach, ransomware, or AI-driven exploit locking your system.
• $V_{data}$: The priceless value of the data being lost.
• $C_{stagnation}$: The Blue T-Shirt Deficit. The compounding penalty of using closed, proprietary NAS software that cannot be extended or updated at the speed of cloud AI, forcing you to suppress your desire for advanced search queries.

The NAS Equation

TCO_nas = C_nas + [ N × ( E_power + T_nuisance ) ] + [ ( P_doom + P_hack ) × V_data ] + C_stagnation

Breakdown of the New Penalties

1. ${\sum }_{i=1}^N(E_{power}+T_{nuisance})$: The physical reality of hosting a server in your living room for a decade. It generates heat, it makes noise, it uses power, and it collects dust.
2. $(P_{hack}·V_{data})$: The AI-Era Hacker Tax. Because your NAS is connected to your home network (and likely exposed to the web so you can auto-sync from your phone), you are now a target for automated ransomware bots. If they encrypt your drives, you lose everything.
3. $C_{stagnation}$: The FOMO Tax. Apple and Google will inevitably release AI that can instantly generate a video montage of "Every time I wore a blue t-shirt on a Tuesday." Your NAS software won't. You are paying a penalty by actively convincing yourself that you don't need those features to cope with your closed ecosystem.

--- AI SLOP ENDS ALERT !!! ---

Yeah, needed to bite my tongue with that one and sigh a relief at the machine's ability to just do that like it's nothing (feels insulting, but hey, it's transformers and data centers vs. my puny power-efficient brain).

The 12TB Showdown: Cloud vs. NAS

The Contender: 12TB Cloud Subscription (Over 10 years)

• Cost: $59.99 * 1.1 * 120 = $7,918.68
• Maintenance: 0
• Risk: 0
• FOMO new features: 0
• TCO: $7,918.68

Our New Heavyweight Challenger: NAS (10 Years, 12TB)

• Hardware cost: $748.68 + $429.99
• 10-year living room toll: ($50 electricity per year + $20 annoyance tax per year) * 10 = $700
• Anxiety premium: 2% chance of evaporation + 3% chance of getting hacked = 5% chance of $25k = $1,250
• FOMO cost: Meh, let's do $500
• Total NAS TCO (10 Years, 12TB): $3,746.54

So, the $8,000 is a luxury premium to never having to think about hardware versus doing a little deep dive to make a better financial decision that scales over time.

The AI Power User: Defining the Tiers

Now that we have established local storage is going to scale better over time just for a regular person's use case, let's 100x it and get into an AI power user mode.

Before we get into power user mode, let's do a little bit of tiering here. I will stick to simple terms here: Pro, Max, Enterprise—LOL, even as I was typing those out I had to laugh out loud and type that out too (geez, I really need to bring the monologue down).

Okay, back to tiers.

1. Tier 1: Starter The tier where you want to just get model weights from HF and run inference locally. Inference can be text, image, video, 3D models—whatever you are into.

2. Tier 2: The Fine-Tuner Now, as you start doing more of this, you realize it's a lot of crap the models generate—not your fault, just the open-source models are sufficiently stupid enough to require a higher level of human prompting game to get things done right now. This is when you realize the off-the-shelf weights are a good base, but you really need to fine-tune to your use case to actually get quality things done.

• AI Generated (sharing since it was fun to read): This is the exact moment you realize that prompt engineering is essentially just trying to hypnotize a very stubborn, hallucinating toddler. You can spend hours typing "Please, I am begging you, DO NOT output markdown, just give me the JSON," or you can just forcefully rewire its brain (fine-tune it) so it never makes that mistake again.
• Some use cases for fine-tuning are:
  • Can you just write like me?
  • I have a lot of handwritten code from the past decade because I build my own software; can you train on those and keep my taste and style?
  • Your brains are so big but you break on JSONs, let me train you to parse them without breaking.
  • I like my photos to look this way: camera angles, color grading, match my hand-drawing style (whatever floats your boat).
  • Audio: fine-tune on my own voice so you can talk like me and I don't have to do tiring interviews anymore, and I can participate in podcasts :D
  • The ultimate: Think like I do, decide like I do, be my clone, understand my personal commands—when I say "sit" you actually "stand," and when I say "stand" you actually don't and keep sitting but turn off my lights so I can sleep. (bruh?)

3. Final Tier: Autonomous

• You don't chat with them anymore.
• It's not a tool anymore; it's a workforce.
• Multi-agent swarms.
• Infinite loop worker.
  • AI written: You have a lightweight model running 24/7 that monitors financial news, Twitter, and Reddit. When it spots a trend, it wakes up your heavy, fine-tuned analysis model. The heavy model reads the data, writes a comprehensive market report, uses your voice-cloned TTS to record it as a podcast, and automatically uploads it to Spotify. You haven't touched the system in three months, but you have 10,000 subscribers.
• Act as your actual personal assistant—negotiate and send out meeting invites on behalf of you, clear out spam.
• Realize you don't have enough data to fine-tune, so generate synthetic data—breed smaller AI.

So we just went from hobby local AI to sentient infra just like that. But now, does that justify our TCO with NAS, or do we need to evolve our thinking and home storage even further?

What does AI SAY?

No. The NAS absolutely cannot survive Tier 3. It will die screaming.

Let me guess (and I guessed right). We probably need PCIe gen4/5 NVMe SSDs. Long-term memory needs vector databases for agents. NAS is just a dumb blob of data.

Tier 1: The Starter Storage Setup

Note: I am going to assume you have some way of running it and enough VRAM / memory to run these models. Not getting into GPUs/Memory yet. Probably another post. This is just storage.

A typical workflow would be to download model weights from HuggingFace (safetensors, GGUF). 10-15 favorite models for different use cases like diffusion, whisper, LLMs, and you may or may not want these to access your personal media.

You cannot use a NAS to load weights etc., because it's just slow. You can technically, because if you can keep the weights cached in VRAM and you are okay with a one-time load overhead.

If you are accessing personal media on your NAS, then that is also likely another separate post because I will have to get into vector databases, RAG, semantic search because AI weights you download cannot read your mind and don't know about you. So you need to give it a notebook to search about you to get things done. Although that can just be solved by an SSD that comes with your GPU or Mac (more on that later). For this storage discussion, I will anyways introduce an SSD + NAS setup that can solve this use case as well.

SSD + NAS Setup:

• Case 1: You don't care about integrating personal media and just want to get LLMs to chat, image gen, coding.
• Case 2: You want to also do cool stuff with personal media.
  • Generate a 60-second video montage of my dog running on snow from 2016 - 2022.
  • Go find all the food photos from lockdown and write down recipes based on what it looks like.
  • Give me a visual timeline of my fashion to see how my clothes and haircuts have changed over time in the last 5 years.

In both cases, SSD + NAS is a win since weights can live in the SSD either way and speed up loads. AI computers that claim to run inference usually come with one of these, so as long as you have one, you already have this.

TCO pretty much remains the same as before, except you now have a computer with an SSD (which is an unavoidable cost if you want to run local AI).

Tier 2: The Fine-Tuner Storage Setup

This is almost the same as above. But IS IT REALLY? What do we need to store more than the weights here? There are actually quite a few things:

• Datasets
• Checkpoints
• Can't do 4bit inference anymore. Get bf16.
• Optimizer states (haha - what is this? Welcome to training).
• LoRA adapters/training logs etc.

A bunch of these can go to the NAS and fit in well with the local storage architecture.

I will stop here because Tier 2 and Tier 3 are something I am still actively experimenting with. NAS can be off-the-shelf with closed software, or you can build your own with TrueNAS (that has more TCO because now you are a system admin).

Deflating the TCO with Local AI

To wrap up, I will mention briefly how local AI deflates most of the TCO:

• Eliminates manual labor:
  • You can let local models write scripts, sort photos, move/rename files, identify faces, tag photos, and a bunch of other cool things.
• The FOMO risk is gone:
  • You can just write your own custom feature. Don't have dark mode? Done in a day instead of waiting 2 years on your closed system.
• AI generates value (offsetting hardware cost):
  • That podcast doesn't take so long to generate and edit anymore.
  • Digital chores are gone.
  • You don't have to pay for a subscription for that app that does one thing that no other app does; now your local AI can.

The Final Pillar: Data Privacy

In this entire journey of canceling a cloud subscription to building a hybrid AI storage, the most important one I did not talk about is: Data Privacy.

Projecting this out to a year, the data is clear. Local storage is not just a box of your stuff; it is the foundation of your own autonomous infra.