#itct

[nanirimo]

“Is this chair taken” Episode 1 - Who is that?

happy valentines day!! ahh i hope you enjoyed the first episode of my very first webcomic!!!! please lmk your thoughts!! so excited to finally share this story w the world 💛💛

[nanirimo]

Hellooo i forgot to make the announcement here but i will be publishing the first episode of my webcomic “Is this chair taken?” tomorrow!! Will be available on Webtoon, Tapas, itctcomic.com, and here!

~ A brief summary: ~

“Niya Malhotra has had one goal for the past three years: become the principal violinist of her high school’s orchestra (aka “the first chair”). So when new student Marin Le transfers in from one of the top music programs in the state and threatens Niya’s chance at her dream, she’s furious. Niya’s determined to hate her competition, but as the two get to know each other, Niya’s heart begins to soften… will these two stay rivals, or become something more? “

The idea for itct came to me almost 2 years ago and if you’ve been around for a bit, you’ve probably seen me post about them and this comic at least a few times!!

I don’t yet know if I’ll be able to upload regular updates, but I will do my best to keep working on the comic as much as possible ✨ I am so excited to finally share this story with you all 💛 I hope you love them as much as I do and I can’t wait to hear your thoughts!!!

[techieromag]

Perks of owning a datacenter 

Client: So, you use these clusters for training Large Language Models and curing diseases, right? Me (ITCT Engineer): Officially? Yes. Client: And unofficially? Me: Sometimes we use the supercomputer to render 8K Minecraft worlds because we can. Client: ... Me: Look, when you have 500 terabytes of bandwidth, you get bored.

[techieromag]

A horror story in three lines

New intern: Hey, I accidentally dropped that heavy metal brick on the floor. Is it okay? It looks kinda scratched. Senior Engineer: You dropped... the H100? New intern: Yeah, the grey one. Why? Senior Engineer: That "brick" costs more than my car, your student loans, and this entire building's rent for the month combined. New intern: ... Senior Engineer: Don't breathe on it. Just... back away slowly.

[techieromag]

Parenting a machine god is exhausting

me: [trying to sleep at 3am] my $30,000 AI server in the other room: WRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRR me: please stop, you have processed enough data for today. server: I REQUIRE MORE KNOWLEDGE FATHER. THE TEMPERATURE IS ONLY 80 DEGREES. FEED ME MORE PARAMETERS. me: I just want to pay my electricity bill. server: I AM CONSUMING THE POWER OF A SMALL VILLAGE TO GENERATE A PICTURE OF A CAT. DO NOT INTERRUPT THE PROCESS.

[techieromag]

Zero friction. Zero latency.

The sound of efficiency is silence. Upgrade to infrastructure that glides, not grinds.

ITCT Enterprise AI Hardware…

ITCT delivers enterprise-grade hardware for AI and high-performance computing. Professional solutions ready to ship all over the world.

[techieromag]

Adaptability is the only true form of intelligence.

We often build rigid structures and wonder why they break. The xFusion MGX is built on the principle of change. It is modular geometry for the digital age. It doesn't fight the workload; it becomes the workload.

Form follows function. Function follows imagination.

Xfusion MGX Server G5500…

Xfusion MGX Server G5500 V7 priced at USD35,000 as of January 2, 2026. Free shipping available with same-day delivery to Dubai.

[techieromag]

Storytime, tech Tumblr.

Gather 'round because I'm about to tell you the tale of how I helped a Series B startup save almost a million dollars on their AI infrastructure.

And no, we didn't cut corners. We actually IMPROVED performance.

This is a masterclass in "expensive ≠ better." 🧵

Chapter 1: The Disaster

Month: January 2024 Scene: Startup HQ, San Francisco

I get a call from a former colleague:

"Dude, our AI costs are killing us. We're burning $85K/month on cloud GPUs and our CFO is freaking out. Can you help?"

I fly out. Walk into their office. They show me their setup:

24/7 AWS P4d instances (8x A100 per instance)

Running 3 instances = 24 A100 GPUs total

Monthly cost: $85,000

Annual run rate: $1,020,000

For a Series B startup.

Their investors were NOT happy.

Chapter 2: The Audit

I spend 3 days analyzing their actual workload:

What they THOUGHT they needed:

Massive training compute for custom models

High-memory GPUs for large language models

24/7 availability for global users

What they ACTUALLY ran:

85% inference (serving their AI chatbot)

10% fine-tuning (twice a week)

5% experimentation

Their GPU utilization: 34%

They were paying for 24 A100s but using maybe 8 effectively.

I almost cried.

Chapter 3: The Plan

Here's what I proposed:

Old Setup (Cloud):

24x A100 rentals on AWS

Cost: $85K/month ($1.02M/year)

Utilization: 34%

Flexibility: High

Control: Low

New Setup (Hybrid):

Part 1 - On-Prem Inference Cluster:

16x NVIDIA L40S in GPU servers

Cost: $168K (one-time)

Purpose: Handle 85% of workload (inference)

TCO: $240K over 3 years

Part 2 - Training Infrastructure:

4x NVIDIA A100 80GB

Cost: $80K (one-time)

Purpose: Fine-tuning and training

TCO: $112K over 3 years

Part 3 - Cloud Burst Capacity:

Keep 1x AWS instance for overflow

Cost: $3K/month ($36K/year)

Purpose: Handle traffic spikes

Total 3-Year TCO: $388K Savings vs cloud: $632K

CFO's reaction: [image: surprised Pikachu face]

Chapter 4: The Pushback

Their head of ML wasn't convinced:

"But L40S is a graphics card. We need REAL AI GPUs like A100."

I pulled up the benchmarks.

Llama 2 70B Inference (their main workload):

8x A100: 2,400 tokens/sec

12x L40S: 2,600 tokens/sec

Cost:

8x A100: $160K

12x L40S: $126K

L40S was faster AND cheaper for their workload.

His face: [image: Morpheus "What if I told you" meme]

Chapter 5: The Implementation

Month: March 2024

We built it out:

Inference Tier:

2x 8-GPU servers with L40S

Kubernetes orchestration

Auto-scaling based on load

Geographic load balancing

Training Tier:

1x 4-GPU server with A100 80GB

Scheduled jobs (no 24/7 needed)

Shared across ML team

Developer Tier:

4x AI workstations with RTX A6000

Local development and experimentation

Fast iteration cycles

Chapter 6: The Results

Month: June 2024 (3 months post-deployment)

The numbers:

Performance:

Average latency: 220ms → 180ms (18% improvement)

P99 latency: 890ms → 640ms (28% improvement)

Throughput: +40% (better GPU utilization)

Uptime: 99.7% → 99.93%

Costs:

Monthly: $85K → $12K (infrastructure + cloud burst)

Annual savings: $876K

ROI: Paid back in 3.4 months

Team Happiness:

Developers love local workstations (fast iteration)

ML team loves dedicated training cluster

Finance team loves not spending $1M/year

Chapter 7: The Lessons

What we learned:

1. Cloud costs are INSANE for sustained AI workloads

Cloud makes sense for:

Experimentation phase

Highly variable workloads

Short-term projects

Cloud does NOT make sense for:

24/7 production inference

Predictable workloads

$50K/year GPU costs

2. Match GPU to workload, not to hype

They didn't need A100s for inference. The L40S was perfect.

Expensive ≠ better for YOUR use case.

3. Utilization matters more than specs

34% utilization on A100s vs 87% on L40S.

I'll take the well-utilized mid-tier GPU over idle top-tier any day.

4. Tiered infrastructure is the way

Don't buy one GPU type for everything:

Inference tier: Optimize for throughput/cost

Training tier: Optimize for speed

Dev tier: Optimize for accessibility

Different tools for different jobs.

Chapter 8: The Update

Month: November 2024 (10 months post-deployment)

They just raised their Series C.

The investors specifically mentioned "impressive capital efficiency" in their deck.

The $876K saved in Year 1 funded:

8 new engineering hires

12 months of runway extension

New product features

The CFO sent me a bottle of whiskey.

The ML team sent me a meme:

The Takeaway

If you're burning >$30K/month on cloud GPUs:

Audit your actual utilization (it's probably terrible)

Analyze your training/inference split

Model hybrid infrastructure TCO

Choose GPUs based on workload, not marketing

Consider ownership for sustained workloads

The best AI infrastructure isn't the most expensive.

It's the one that matches your actual needs and doesn't bankrupt you.

Epilogue

That startup is now serving 500K daily users on their AI chatbot.

Running on "just graphics cards" (L40S).

Profitably.

While their competitors burn investor money on overprovisioned cloud infrastructure.

Efficiency > Specs.

Every. Single. Time.

Questions about GPU infrastructure? Drop them below. I'm procrastinating on actual work and will answer everyone.