Technical Product Manager | AI Enthusiast | Making Kubernetes Work Smarter (CKA, CKAD) | Figuring out AI to build better Products | AI PM | PLGC™️ | PSC™️

I lead the Product Strategy for GPU-Powered AI Supercloud as a Group Product Manager.

My goal is to make advanced AI accessible, affordable, and scalable by building the cloud foundation that powers the next wave of AI innovation.

I’m passionate about shaping the future of AI infrastructure, bringing world-class GPU technology closer to startups, enterprises, and researchers.

At NeevCloud, I focus on scaling GPU Supercloud, so organizations have the compute power to train, fine-tune, and deploy AI models responsibly and efficiently.

I was also a
- Speaker: KubeCon Japan panel on Responsible AI Products
- Podcast Guest: LLM Inference Explained.

Why Kubernetes Gang Scheduling is Critical for LLM Workloads?

LLM training is fundamentally a Distributed workload.

It needs multiple GPU workers to start and communicate simultaneously.

But there is a  problem with Default K8s Scheduling

The standard scheduler places Pods one-by-one. 

If your job needs 10 GPU Pods but only 8 are available, 

The scheduler launches those 8 GPU Pods.


The Problem? 

Those 8 GPU Pods sit there,  

Consuming expensive GPU resources but performing zero computation.

The GPUs are reserved,  the Pods are running,  but no actual training is happening.

This is a massive cost inefficiency and a cluster deadlock.


The Solution (Gang Scheduling)

It enforces the "All-or-Nothing" rule. 

Gang Scheduling treats the LLM job workers as a single unit. 

It enforces a strict rule:

Do we have room for the entire team?

→ Yes: 
Launch all 10 Pods instantly. 
Training begins immediately.

→ No: 
Launch zero Pods. 
The job waits cleanly in a queue until the resources are fully available.


This Improves:
→ The utilization of high-value GPUs
→ Guarantees all distributed workers are ready for synchronized communication (NCCL)
→ Eliminates "zombie" jobs that clog up the cluster without doing any work.


Are you using Gang Scheduling today for any of your LLM workload?


-----------------------------------------------------------------
#kubernetes #k8s #scheduler #gpu #nvidia #h100 #a100 #utilization #docker #container #pod #cost #roi #infrastructure #productmanagement
#ai #genai #agents #prompt #aipm #pms #pm #data #llm #vllm #kubernetes #infrastructure #Amper #Hopper #Blackwell #gb200 #AIInfra #datacenter #aifactory

undefined undefined

Group Product Manager | AI PM (Platform/Infra) | LLM Inference | GPU Cloud @NeevCloud

I hit 8,000 followers on LinkedIn today!

And let me be honest

I’m not going to call this a “family,”

because I barely know less than 1% of the people here.

But that’s the beauty of it.

You follow me for the same reason I follow others,

To Learn, Grow, and Get Inspired.

And wow… 

I’ve learned so much from this network.

Every day, I try to give that value right back by sharing 

What I discover, 

What I try, 

And what I Fail at.

For the past two years, I’ve shown up consistently,

Putting out my learnings, thoughts, and experiments.

And hitting 8,000 today feels like fuel to keep going even harder.

Thank you for being here.

Let’s keep learning and leveling up together!

Stop thinking “Data Center.” Start thinking “AI Factory.”

We often talk about AI in the Cloud

But today I want to show you what it actually looks like on the ground.

At RackBank® (NeevCloud®’s parent company),  we’re building the critical infrastructure that will power the AI era. 

To run massive workloads like LLM Training, Inferencing and Generative AI, you need far more than just racks and space. 

The architectural view below gives you a glimpse into the engineering complexity behind a modern AI Factory.


1️⃣ The Production Floor (Main Building)
→ Inside are high-density GPU clusters (NVIDIA H100s/A100s/B200s). 
→ Each rack can draw 50–100 kW+
→ Nearly 10x a traditional rack
→ Reinforced floors to handle heavy liquid-cooled racks
→ This is where raw compute becomes intelligence.


2️⃣ The Lungs (White Cooling Units)
→ AI runs hot, extremely hot. 
→ Dedicated cooling towers (white units) to manage the thermal output of thousands of GPUs firing simultaneously.
→ These industrial cooling systems use advanced liquid and direct-to-chip cooling to reject heat loads that air cooling simply cannot handle.


3️⃣ The Heartbeat (Green Generator Units)
→ The green containers in front are heavy-duty backup generators 
→ AI training runs for weeks, sometimes months.
→ A millisecond of power loss can cost millions
→ These systems ensure uninterrupted continuity


4️⃣ The Nervous System (Fiber Interconnect)
→ Mostly invisible but essential
→ Miles of InfiniBand fiber interconnect thousands of GPUs 
→ So they operate as a unified supercomputer.


We’re not just building a place to store servers.

We’re building the physical foundation of the AI economy. 🇮🇳


Visit AI Factory: 
https://lnkd.in/dz4SZG7k

------------------------------------------------------
#ai #genai #GPU #Nvidia #utilization #agents #prompt #aipm #pms #pm #data #llm #vllm #kubernetes #infrastructure #Amper #Hopper #Blackwell #gb200 #AIInfra #datacenter #aifactory

What does it mean to build a 0 → 1  Product?

It means

Creating something new where nothing existed before.

No playbook.

No past data. 

Just a vision and a problem worth solving.

As a PM, 0 → 1 work is a different kind of challenge

→  You discover real problems that users haven’t voiced yet
→  You run fast experiments to learn what works
→  You align teams around a future that isn’t visible yet
→  You make decisions with limited data
→  You focus on proving Product-Market Fit (PMF), not just metrics

It’s messy. 

It’s uncertain. 

It’s risky.

But if you get it right, you create something that changes behavior, 

Unlocks new markets, and sets the foundation for the future.


A few Examples of 0-1 Products:

→ The first iPhone 📱
Smartphones as we know them didn’t exist before.

→ Airbnb 🏠
Renting homes to strangers online wasn’t normal before.

→ Uber 🚖
Every day ride-hailing through an app was new.


A simple way to think about it:

0 → 1   = Discover and Prove Value
1 → 10 = Scale and Optimize Value

Both are important.

But 0 → 1 is where the magic begins.

----------------------------------------------------------------
#ai #pm #aipm #productmanagement #productmanagers #aiml #gpu #zerotoone #data #genai

Our AI Factory is Coming...

AI has grown beyond what traditional data centers were built for.

Modern Blackwell GPUs demand far more power, cooling, and scale, and the infrastructure around them needs to evolve quickly.

Today, RackBank® ( NeevCloud's Parent Company) is taking that leap forward.

Meet the AI Factory.

A new kind of facility created specifically for the GB200 NVL72 generation,

Built for high-density racks, liquid cooling, and fast-to-install GPU clusters.

This isn’t an old setup modified to fit AI.

It’s an entirely new architecture designed from day one for accelerated computing.

What makes the AI Factory different?

→ High-Power Density: 
Supports racks up to 150kW without performance loss.

→ Liquid Cooling by Design: 
Direct chip cooling and CDUs targeting around 1.3 PUE.

→ Rapid Scaling:
Pre-built GPU modules that can grow to 16,000+ GPUs in a single hall.

→ AI-Focused Infrastructure:
Tailored for LLM training, large-scale inference, RAG workloads, and fully AI-driven companies.

This is the foundation for the next era of AI.

The AI Factory is on the way.

------------------------------------------------------
#ai #genai #GPU #Nvidia #utilization #agents #prompt #aipm #pms #pm #data #llm #vllm #kubernetes #infrastructure #Amper #Hopper #Blackwell #gb200 #AIInfra

RackBank AI Factory 

𝐓𝐡𝐞 𝐀𝐈 𝐫𝐞𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐡𝐚𝐬 𝐨𝐮𝐭𝐠𝐫𝐨𝐰𝐧 𝐭𝐫𝐚𝐝𝐢𝐭𝐢𝐨𝐧𝐚𝐥 #𝐝𝐚𝐭𝐚𝐜𝐞𝐧𝐭𝐞𝐫𝐬.
𝐁𝐥𝐚𝐜𝐤𝐰𝐞𝐥𝐥-𝐜𝐥𝐚𝐬𝐬 #GPUs are rewriting the rules of power, cooling, scale and the infrastructure behind them must evolve just as fast.

𝐓𝐨𝐝𝐚𝐲, 𝐰𝐞’𝐫𝐞 𝐨𝐩𝐞𝐧𝐢𝐧𝐠 𝐭𝐡𝐞 𝐝𝐨𝐨𝐫𝐬 𝐭𝐨 𝐭𝐡𝐚𝐭 𝐟𝐮𝐭𝐮𝐫𝐞.
𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐭𝐡𝐞 𝐀𝐈 𝐅𝐚𝐜𝐭𝐨𝐫𝐲.
A purpose-built environment engineered for the 𝐆𝐁𝟐𝟎𝟎𝐍𝐕𝐋𝟕𝟐 era, where 𝟏𝟓𝟎𝐤𝐖 racks, 𝐥𝐢𝐪𝐮𝐢𝐝-𝐜𝐨𝐨𝐥𝐞𝐝 architectures, and rapid-deploy 𝐆𝐏𝐔 pods become the new standard.

𝐓𝐡𝐢𝐬 𝐢𝐬𝐧’𝐭 𝐫𝐞𝐭𝐫𝐨𝐟𝐢𝐭𝐭𝐞𝐝 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞.
This is architecture designed from the ground up for accelerated #computing.

Inside the #AIFactory, every layer is built for what modern #AI demands:
𝟏. 𝐄𝐱𝐭𝐫𝐞𝐦𝐞 𝐃𝐞𝐧𝐬𝐢𝐭𝐲: Engineered for up to 150kW/rack, no throttling, no compromises.
𝟐. 𝐍𝐚𝐭𝐢𝐯𝐞 𝐋𝐢𝐪𝐮𝐢𝐝 𝐂𝐨𝐨𝐥𝐢𝐧𝐠: Direct-to-chip cooling loops and CDUs powering a near-perfect 1.3 PUE.
𝟑. 𝐖𝐚𝐫𝐩-𝐒𝐩𝐞𝐞𝐝 𝐃𝐞𝐩𝐥𝐨𝐲𝐦𝐞𝐧𝐭: Pre-fabricated GPU pods that scale from zero to 16,000+ GPUs per data hall.
𝟒. 𝐌𝐚𝐧𝐮𝐟𝐚𝐜𝐭𝐮𝐫𝐢𝐧𝐠 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞, 𝐍𝐨𝐭 𝐉𝐮𝐬𝐭 𝐇𝐨𝐬𝐭𝐢𝐧𝐠 𝐃𝐚𝐭𝐚: Designed for #LLM training, #hyperscale inference, RAG architectures, and AI-native enterprises.

𝐓𝐡𝐢𝐬 𝐢𝐬 𝐭𝐡𝐞 𝐛𝐚𝐜𝐤𝐛𝐨𝐧𝐞 𝐛𝐞𝐡𝐢𝐧𝐝 𝐭𝐡𝐞 𝐧𝐞𝐱𝐭 𝐝𝐞𝐜𝐚𝐝𝐞 𝐨𝐟 𝐀𝐈.
𝐓𝐡𝐞 𝐀𝐈 𝐅𝐚𝐜𝐭𝐨𝐫𝐲 𝐢𝐬 𝐜𝐨𝐦𝐢𝐧𝐠.

𝐑𝐞𝐬𝐞𝐫𝐯𝐞 𝐞𝐚𝐫𝐥𝐲 𝐚𝐜𝐜𝐞𝐬𝐬.
https://lnkd.in/dCGMYyJC

#AIDatacenter #BlackwellReady #GPUInfrastructure #LiquidCooling #LLMTraining #NextGenDC #PurposeBuiltForAI #RackBank

What Mixture-of-Experts (MoE) Means, and Why It's Hard to Scale Today?

MoE breaks a big AI model into many smaller “experts,” 

Each good at different things.

A router decides which experts should handle each token, 

So only part of the model runs at a time. 

This makes the model more efficient.


The problem:

Modern MoE models (like DeepSeek-R1 and Kimi K2) are so huge that the experts have to live on many different GPUs. 


This causes issues:

Each GPU needs more memory bandwidth to load the right experts weights for every token.

Communication becomes slow when the experts are spread across GPUs on different servers (anything beyond ~8 GPUs makes this worse).


What the NVIDIA GB200 NVL72 Changes?

The NVIDIA GB200 NVL72 is a full rack system with 72 Blackwell GPUs, 

All connected with very fast NVLink and sharing 30 TB of high-speed memory.

This setup lets MoE models spread their experts across all 72 GPUs with:

→ Less memory load on each GPU
→ Support for much longer context windows
→ More users running at the same time
→ Much faster communication between experts


------------------------------------------------------
#ai #genai #GPU #Nvidia #utilization #agents #prompt #aipm #pms #pm #productmanagement #productmanagers #data #llm #vllm #kubernetes #infrastructure #Amper #Hopper #Blackwell

taplio

Shamsher Ansari's Linkedin Analytics

Shamsher Ansari

Performance ScoreShamsher Ansari's LinkedIn Performance

Shamsher Ansari's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Shamsher Ansari's Linkedin Analytics

Shamsher Ansari

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Shamsher Ansari's LinkedIn Performance

Shamsher Ansari's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Shamsher Ansari's Linkedin Analytics

Performance ScoreShamsher Ansari's LinkedIn Performance

Shamsher Ansari's Best Posts (last 30 days)