TWIML Briefing Room

HPE and NVIDIA on the AI Factory: Data In, Intelligence Out

Thierry Piennar & Kaushik Shirhatti· CTO of AI & HPC at HPE; VP of AI Factory at NVIDIA at Hewlett Packard Enterprise & NVIDIA
·~15 min·English·TWIML
AI InfrastructureInferenceGPUBusiness Strategy
TL;DR

HPE's Thierry Piennar and NVIDIA's Kaushik Shirhatti reframe the 'AI factory' from a rebranded GPU cluster into an operating model — where the decisive work is data quality, organizational unity, and intentional prioritization, not the silicon.

01The Definition

Data In, Intelligence Out

An AI factory isn't a GPU cluster with better branding — it's a five-layer system where data and energy go in and tokens, intelligence, and business outcomes come out.

at a basic fundamental level, what we think about an AI factory, doesn't matter if it's big or small, is when you put data in, you put energy in, and boom, like magic, intelligence, tokens, business outcomes come out.

— Kaushik Shirhatti, NVIDIA
Key Insight
The factory metaphor is really a claim about economics: if intelligence is a manufactured output, then whoever controls the two inputs — data and energy — controls the margins. That's why the rest of the conversation keeps circling back to power and data rather than model architecture.
02The Shift

Training Was the Story, Inference Is the Business

Inference has moved to the forefront as the driver, and the enterprises running it have to optimize the whole request-to-token path — where the metric that matters is power-to-token, not raw speed.

You have to look at inference from a cost optimization perspective, power to token, and then also from a, let's say, architectural strategy perspective.

— Thierry Piennar, HPE
Key Insight
Pricing inference as "power to token" quietly reprices the whole stack. It makes the memory system — KV cache, CXL — and intelligent routing like Dynamo as strategically important as the GPU itself, because what makes enterprise inference pencil out is a cheaper token, not a faster one.
03The Bottleneck

Garbage In, Garbage Out

The hard part of an AI factory isn't the chips — it's the data: knowing where it lives, curating it, and protecting it, even as enterprise AI ROI has climbed from about 5% to 30%.

AI is really not meaningful without data that is good data, right? garbage in, garbage out type of thing.

— Thierry Piennar, HPE
Key Insight
Calling data "your context" updates the old garbage-in-garbage-out cliche for the LLM era. The first obstacle often isn't modeling at all but simply knowing where an enterprise's data lives, what shape it's in, and how to protect it before any factory can turn it into value.
04The Discipline

Pick Four Projects and Go Deep

The enterprises that win don't chase a hundred AI ideas — they pick four or five, go deep end-to-end, and reimagine the process instead of merely optimizing it.

But really being intentional and saying, "I'm gonna pick four or five projects, and we are going to go deep," and actually go see it end to end, like the workflow, measure the productivity, how is it actually being useful.

— Kaushik Shirhatti, NVIDIA
Key Insight
The "optimize versus reimagine" distinction is the real payload. Optimizing a broken process just makes a bad workflow faster; reimagining it is what makes the ROI "so obvious" that adoption spreads on its own, no executive mandate required.
05The Org Change

AI Broke the Silos

You can't build an AI factory one team at a time — AI forced storage, compute, software, data, and ML teams to stop operating in silos and plan as one.

Those still exist, but you cannot operate in a silo to go build an AI factory.

— Kaushik Shirhatti, NVIDIA
Key Insight
It's an unusual admission from infrastructure vendors: the biggest blocker to an AI factory is organizational, not technical. Their enterprise reference designs are as much a political tool for forcing one shared plan as they are a hardware bill of materials.
06Sovereignty

Sovereign AI Is Strategic Autonomy

Sovereign AI isn't about owning the whole stack — it's a growth flywheel where local models, local context, and homegrown talent turn strategic autonomy into economic value.

Sovereign AI, at the end of the day, really is about strategic autonomy, right? It is about a nation or an entity, it can be at the enterprise level as well, right?

— Thierry Piennar, HPE
Key Insight
The tell is "you never control the full stack." Sovereignty is redefined here from self-sufficiency to leverage — picking the few layers (language, context, talent) where a country can compound value — which is how even a Canadian telco can grow into a national AI provider.
07The Forward Look

You Won't Budget for AI Factories

In a few years nobody will budget for 'AI factories' any more than they budget for email servers — the on-prem-versus-cloud debate will simply look quaint.

just like we don't ask whether we need email servers, we're not gonna ask, do we need AI factories? Should we put AI factories in our budget next year?

— Kaushik Shirhatti, NVIDIA
Key Insight
The email-server analogy is a bet on invisibility: the sign that AI factories have won isn't more hype, it's when they disappear into the budget as ordinary infrastructure — and by then the frontier will already have moved on to physical AI.