An AI factory is a data centre built to manufacture intelligence, and NVIDIA used Computex 2026 to argue that every country now needs one of its own.
On 31 May 2026, at GTC Taipei, NVIDIA set out a buildout pitched at the scale of national infrastructure. The AI Cloud expansion reaches across six continents, with CoreWeave, Firmus and Nebius among the partners standing up large training and inference sites. In parallel, more than 500 Taiwan partners — TSMC, Foxconn, Pegatron, Quanta — are ramping production of Vera Rubin NVL72 systems, the rack-scale machines that will run the next wave of models. Jensen Huang, NVIDIA's chief executive, put the thesis in one sentence.
Every company and every country needs AI factory infrastructure.
— Jensen Huang, NVIDIA, GTC Taipei keynote, 31 May 2026 (https://blogs.nvidia.com/blog/ai-cloud-ecosystem/)
Huang has said versions of that line before. What changed at Computex is the supply chain lining up behind it — and the speed of the ramp.
What NVIDIA announced at Computex
NVIDIA frames the AI factory as a new class of plant: not a building that stores data, but one that takes in electricity and data and produces tokens — the raw output of trained models and running agents. The AI Cloud ecosystem expansion NVIDIA says is now worldwide, with CoreWeave, Firmus and Nebius building out facilities for training, inference and agentic workloads across six continents.
The Taiwan piece is the one to watch. NVIDIA says its 500-plus Taiwan partners are moving Vera Rubin NVL72 — the company's rack-scale Blackwell-successor platform — into volume production, while deploying accelerated computing inside their own factories. TSMC, Foxconn, Pegatron and Quanta are the names. Live coverage of the keynote tracked the through-line: the firms that assemble the world's AI hardware are also becoming the first heavy users of it.
Two numbers anchor the scale. NVIDIA puts the Taiwan partner base at more than 500 firms. The buildout spans six continents. Neither figure is a forecast — both describe orders already moving through fabs and assembly lines in mid-2026.
💡The unit of measure changed
An AI factory is measured in megawatts, not square metres. The constraint is no longer chips alone — it is power, land, water and the grid connection. That is why the story has moved from the lab to the ministry of energy.
The AI factory is the new sovereign asset
For two decades, national technology strategy meant broadband, spectrum and a stock exchange that could float a tech company. NVIDIA is arguing that the list now has a new top item: domestic AI-factory capacity. A country without it rents its intelligence from a country that has it.
Sovereign AI is the polite term. The blunt version: compute is becoming a measure of national standing, the way steel output once was. A government that cannot point to its own AI factories is a government negotiating from weakness — on price, on access, and on the terms under which its citizens' data trains someone else's model.
Who gets a factory — and who does not
Here is the uncomfortable part. A buildout across six continents sounds inclusive until you ask which sites land where. The economics favour cheap power, cool climates, political stability and existing fibre. That map does not run through most of Africa, and it barely runs through the global South at all.
I have written about the Digital Berlin Conference — the way platform power is quietly redrawing the continent's digital borders. AI-factory geography is the next chapter of that story. If the factories cluster in Taiwan, the US Gulf Coast, the Gulf states and northern Europe, then the places that host the compute set the terms for the places that do not.
This is where the language of Emergent Intelligence — the dignity-first frame I use for what the industry calls AI — stops being abstract. If intelligence is becoming infrastructure, then access to it is a question of justice, not just procurement. A continent that imports all of its intelligence inherits all of the constraints that come with it.
What it means for boards and policymakers
For a board, the Computex message is a planning input, not a press release. Compute pricing, availability and location now sit on the risk register beside currency and supply chain. The question "where does our intelligence run, and who controls that site?" belongs in the next strategy review.
For policymakers, the move sharpens two debates already underway. One is the sovereign-compute scramble in Europe, where SoftBank has just committed €75bn to French data centres. The other is the tightening of US chip-export controls, which decides who is allowed to build a frontier-grade factory at all. NVIDIA sells the shovels; governments increasingly decide who may dig.
And it raises the recursion question NVIDIA itself put on stage: the same firms now run AI inside their fabs to build the chips that run the factories. I have written about that loop separately — AI is starting to run the chip fab.
Frequently Asked Questions
These are the questions boards, founders and policy teams have been asking since NVIDIA's Computex 2026 keynote. Short answers follow, drawn from NVIDIA's own announcements and the live keynote coverage.
What is an AI factory?
In short, an AI factory is a data centre purpose-built to produce machine intelligence — taking in electricity and data and turning out trained models and running agents. The answer, simply put, is that the unit of value is the token, not the stored file. The key is the constraint: research and industry data show the limiting input is now power and grid capacity, not chips alone.
How does NVIDIA's Computex buildout actually work?
NVIDIA accepts orders from cloud builders and national operators, then its partners assemble Vera Rubin NVL72 racks into operational sites. According to NVIDIA, the AI Cloud ecosystem now spans six continents with CoreWeave, Firmus and Nebius, while more than 500 Taiwan partners ramp production. Data from the keynote shows the assemblers — TSMC, Foxconn, Pegatron, Quanta — are also early users of the systems they build.
Why is sovereign AI infrastructure suddenly the headline?
Earlier AI strategy treated compute as a commodity bought from a hyperscaler. According to NVIDIA, that framing is over: a country without domestic AI-factory capacity rents its intelligence from one that has it. The evidence is the capital — hundreds of billions of dollars committing to land, power and racks in 2026, which reveals that governments now treat compute as strategic, not discretionary.
Who benefits from the AI-factory buildout, and who is left out?
The buildout favours regions with cheap power, cool climates, political stability and existing fibre — Taiwan, the US, the Gulf, northern Europe. In other words, the analysis shows most of Africa and much of the global South sit outside the map. That gap is the dignity question underneath the engineering: who hosts the intelligence sets the terms for who merely consumes it.
What are the real risks of the AI-factory model?
Analysis of the buildout reveals three durable risks. First, concentration: a single dominant supplier of the factory itself, which is why the non-NVIDIA compute bet exists. Second, power and water strain on host grids, measured in gigawatts. Third, geopolitical exclusion, where nations without factories lose their bargaining power over price, access and data. Evidence from the 2026 capital surge shows each risk is structural, not temporary.
NVIDIA sold a vision at Computex that is hard to argue with on the engineering and easy to worry about on the politics. Intelligence is becoming infrastructure. The factories that make it are becoming national assets. The open question is whether the map of those factories ends up looking like the old map of industrial power — or a fairer one. That choice is not NVIDIA's to make. It belongs to the governments now deciding whether to build, rent, or be left out.
Related on humphreytheodore.com:
