In my previous article, I wrote about Own your AI stack.
My core argument was that models can be rented, but your data, context, workflows, and orchestration should remain under your control as much as possible. What truly accumulates over time — and can even compound in value — is not which model you subscribe to today, but whether your own data and workflows can gradually become part of your own intelligence.
Recently, after watching discussions from Y Combinator around Personal AGI and the AI stack, I kept thinking about another question:
If we really want to own our own intelligence, does that mean every country and every company needs to build a massive AI data center of its own?
My answer is: not necessarily.
In fact, I believe that for the vast majority of countries, enterprises, SMEs, and even individuals, the answer is no.
Not Every Country Needs to Build an Aircraft Carrier
I am also involved in consulting work related to AIDC+ and AI infrastructure, so I am not arguing against building AI data centers.
Quite the opposite. I believe AI infrastructure will become increasingly important.
But there is an interesting pattern in the market today.
As soon as people start talking about AI infrastructure, the conversation quickly moves toward data centers measured in tens of megawatts, hundreds of megawatts, or even gigawatts, along with tens or hundreds of thousands of GPUs.
It sometimes feels as if entering the AI era means everyone should start building a mega-scale AIDC.
That reminds me of a simple analogy:
Not every country needs to build an aircraft carrier, let alone an entire carrier strike group.
Aircraft carriers are obviously extremely powerful.
But the first question you should ask is:
What is the mission?
If your goal is to train the next generation of frontier models and compete directly with the world’s largest AI companies, then yes — you will need enormous amounts of compute, power, networking, cooling, and capital.
That is an aircraft-carrier-scale competition.
But if your goal is not to train the next GPT, and instead you simply want to:
keep your own data under your control;
allow Agents to securely access internal enterprise data;
run inference locally;
build your own knowledge base, RAG, memory, and workflows;
and gradually develop your own enterprise intelligence;
then the infrastructure you actually need may be very different.
Most Organizations May Not Need Mega-Scale AIDC
For most enterprises, I believe a more practical direction may be:
Enterprise-scale AIDC, Edge AIDC, or even In-Rack or In-Row AI infrastructure.
It does not necessarily need to be a new standalone data center building.
It may simply be one rack, or a few racks, inside an existing enterprise data center.
Those racks could contain GPU servers designed for AI inference, storage, networking, and the company’s own data layer, Agent Runtime, APIs, MCP, and orchestration.
They could run quantized open-weight models locally, while calling external frontier model APIs when more advanced capabilities are required.
In other words, it does not have to be a closed system.
It can be a Hybrid AI Infrastructure.
Sensitive data, local data, and enterprise knowledge can remain on-premises.
Tasks that require large-scale compute or frontier intelligence can be routed to cloud APIs.
Models can be replaced.
Compute can be expanded.
But the most important assets — the data, context, and workflows — remain under your control.
This is essentially the idea of Own your AI stack from my previous article, extended one layer deeper into infrastructure.
What You Need to Own Is Not Necessarily the Largest Compute
I think there is an important distinction here.
Own your AI stack does not mean Own every GPU.
You do not need to own the largest model in the world.
You also do not need to own the largest GPU cluster.
What you really need to own are the layers that create long-term competitive advantage:
Data, Context, Identity, Knowledge, Workflow, Agent Runtime, and Orchestration.
Compute itself can be hybrid.
Some can be on-premises.
Some can be at the edge.
Some can be in a private cloud.
Some can simply be rented from a public cloud or accessed through frontier model APIs.
A company does not need to build its own power plant in order to claim ownership of its information systems.
The real question has never been whether everything is self-built.
The real question is:
Who controls the most important assets and the decision-making layer?
The Value of Edge AIDC Is Bringing Data and Intelligence Closer Together
Why am I becoming increasingly interested in enterprise-scale and edge AIDC?
Because the most valuable thing inside many companies has never been the model.
It is the data.
Transaction records in ERP systems.
Customer information in CRM systems.
Production data from factories.
Documents accumulated over ten years.
Email, knowledge bases, SOPs, and operational information generated every day.
A lot of this data is not suitable for being sent entirely to an external cloud.
Sometimes the reason is privacy.
Sometimes compliance.
Sometimes latency.
And in many cases, the reason is even simpler:
The data is already here.
If the data is already on-premises, then moving part of the intelligence closer to the data is a very natural direction.
That is also how I think about Edge AI.
Not because cloud infrastructure is unimportant, but because not every dataset should be moved to the cloud just to use AI.
Sometimes the more sensible approach is:
Bring AI to the data.
From Cloud AI to Hybrid Intelligence
For that reason, I believe the future AI architecture for many enterprises will not be purely cloud-based, and it will not be fully self-hosted either.
It will be hybrid.
It may look something like this:
Local / Edge Compute → Private Data → Knowledge → Tools / API → Agent → Frontier Model API
Simple tasks, internal knowledge retrieval, RAG, embeddings, and some Agent workloads can be completed locally.
When stronger reasoning or specific multimodal capabilities are required, the system can call external models.
Today that might be OpenAI.
Tomorrow it might be Anthropic, Google, DeepSeek, Qwen, or another model that does not yet exist.
The model can change.
Your data and intelligence architecture do not need to change with it.
That is why I keep emphasizing:
Own the Data and Orchestration. Rent the Model.
And now I would add one more sentence:
Own enough Compute to keep your Intelligence under control.
Not the most compute.
Just enough compute.
This May Actually Be Taiwan’s Real Opportunity
From this perspective, I think this direction is especially relevant to Taiwan.
Taiwan’s long-standing strength has never been limited to cloud services.
It lies in the broader ICT supply chain.
Servers, storage, networking, edge computing, embedded systems, ODM/OEM, system integration, semiconductors, and the broader hardware ecosystem.
If the future of AI infrastructure has only one answer:
Who can build the largest Mega AIDC?
then it becomes an extremely capital-intensive game.
Only a very small number of players will ultimately be able to participate.
But if the next market also includes:
Enterprise AI Infrastructure,
Private AI,
Edge AIDC,
In-Rack AI,
In-Row AI,
AI Appliances,
and Hybrid AI Infrastructure designed for different industries,
then the game changes completely.
This market requires more than GPUs.
It requires servers, storage, networking, cooling, power management, security, software stacks, model runtimes, Agents, MCP, data governance, and system integration.
Put all of those pieces together, and I believe this looks much closer to what Taiwan is already good at.
Instead of Asking “How Do We Build an AI Data Center?”, Ask “What Problem Are We Solving?”
So while everyone is talking about AIDC, perhaps we should return to a more basic question:
Why do we need this AIDC in the first place?
Is it for frontier model training?
Is it to provide GPU cloud services?
Is it for sovereign AI?
Or is the real goal simply to let an enterprise keep its own data, workflows, and Agents inside infrastructure that it controls?
Different missions should naturally lead to different infrastructure designs.
Aircraft carriers have their role.
But not every mission requires deploying an aircraft carrier.
Sometimes one rack, an edge AI infrastructure stack, and access to cloud frontier models may already be enough to accomplish what you actually need.
Conclusion: Not the Biggest AI, but Your Own Intelligence
I increasingly believe that the most important question in the AI era is not:
“How many GPUs do I have?”
It is:
“Can I turn my own data into my own intelligence?”
Mega-scale AIDC will certainly continue to exist, and it is indispensable for frontier model training, hyperscalers, and large AI service providers.
But for the vast majority of enterprises, SMEs, and potentially even future personal AI infrastructure, I believe another path may be more important:
Smaller. Local. Hybrid. Ownable.
Keep your data, context, and workflows on infrastructure you control.
Run suitable open-weight models locally.
Rent the world’s best frontier intelligence when necessary.
Then use Agents and orchestration to combine those capabilities into intelligence that is genuinely your own.
Not everyone needs an aircraft carrier.
But every individual, every company, and perhaps every country should think carefully about how much AI infrastructure they need to own in order to keep their most important intelligence under their own control.
That may be the next layer of the Own your AI stack discussion — and, in my view, one of the most important infrastructure questions we should be asking now.