> ## Content Index
> Fetch the complete content index at: https://www.sotkn.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# What Is an AI Factory? Nvidia’s Name for a Data Center That Sells Tokens
- URL: https://www.sotkn.com/explained/ai-factory-explained/
- Published: 2026-11-15T13:23:23.000Z
- Updated: 2026-11-15T13:23:23.000Z
- Description: An AI factory is Nvidia’s name for an AI data center that is judged by what it produces: tokens. The chipmaker is paid when the racks ship; the factory’s owner is paid only when the tokens sell. Power caps how many it can make.
- Author: Sean
- Tags: Mechanism, #desk-explained

EXPLAINED · Mechanism · Last reviewed: 15 Nov 2026 · Next review: after Nvidia’s third-quarter fiscal 2027 results (late November 2026) and when Firmus’s Tasmania campus comes online

An AI factory is Nvidia’s name for an AI data center that is built and judged as a production line: electricity and data go in, and AI output, counted in tokens, comes out. The building is the same kind of hall as any AI data center; what changes is the scorecard, because once output is counted in tokens, the power supply sets the ceiling on revenue and the owner, not the chipmaker, carries the risk that the tokens do not sell.

**In short.** “AI factory” is Nvidia’s word for an AI data center whose product is tokens, the units of AI output. The hall, racks and cooling are the same as any AI data center; the difference is that output, not floor space, becomes the yardstick. Who gets paid: Nvidia and the equipment makers first, when racks ship (Nvidia’s data center revenue was $89.0 billion in the quarter to July 2026); the landlord when the lease starts; the factory’s owner last, only when tokens sell. The stuck step is power: the megawatts fix how many GPUs a site can run, so they cap how many tokens it can ever sell.

Why an investor cares: the phrase has become the standard way chipmakers, cloud companies and governments describe the AI build-out, and it quietly changes who carries the risk. A data center landlord with a 15-year lease is paid for space whether or not the AI inside earns anything. An AI factory owner is paid for output. This page builds on [what an AI data center is](https://www.sotkn.com/explained/what-is-an-ai-data-center/) and [what an AI token is](https://www.sotkn.com/explained/ai-tokens-explained/), and answers one question: what is an AI factory, and who gets paid when one is built?

## What is an AI factory?

Nvidia defines it as specialised computing infrastructure that handles the whole life of an AI model, from taking in data to training it, fine-tuning it and then serving it to users at high volume. In Nvidia’s words, the primary product is intelligence, “measured by token throughput.” Its solutions page goes further: AI factories are “purpose-built to manufacture intelligence at scale,” and they tie together five layers, which Nvidia lists as energy, chips, infrastructure, models and applications.

Strip away the marketing and three ideas are left.

- **It is one machine, not many.** A traditional data center hosts thousands of unrelated servers. An AI factory wires tens of thousands of GPUs together so they behave like a single computer, which is why it needs very fast networking and very dense racks. Our [GPU cluster explainer](https://www.sotkn.com/explained/gpu-cluster-explained/) and [rack-scale AI explainer](https://www.sotkn.com/explained/rack-scale-ai-explained/) cover how that works.
- **It has an output you can count.** The output is tokens: the chunks of text, code or image data a model reads and writes. Because tokens can be counted and priced, an AI factory can be judged like a plant: units out, per hour, per megawatt, per dollar spent.
- **Inference is the main job.** Training builds a model once; inference runs it every time someone asks a question. Nvidia’s glossary says an AI factory puts a strong emphasis on inference performance and energy efficiency, because that is where the tokens are sold. See [AI training vs inference](https://www.sotkn.com/explained/ai-training-vs-inference/).

So an AI factory is not a new kind of building. It is an AI data center described by what it makes rather than by how much space it has. That framing matters for money, because it shows who is paid for the machines and who is paid for what the machines produce.

## How is an AI factory different from a data center?

The building, cooling and power gear look much the same. The business does not. Our [hyperscale vs colocation explainer](https://www.sotkn.com/explained/hyperscale-vs-colocation-data-center/) covers the older models; the table puts the AI factory next to them.

| Question                        | Traditional data center                                                                             | AI factory                                                                                                                  |
| ------------------------------- | --------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------- |
| **What it sells**               | Space, power and cooling for many customers’ servers, or cloud services on general-purpose machines | AI output: training runs, inference, tokens, or GPU time rented to someone who sells them                                   |
| **How it is measured**          | Megawatts leased, racks filled, uptime                                                              | Tokens produced per second, per megawatt and per dollar of hardware                                                         |
| **What sits inside**            | Mixed servers, storage and network gear at far lower power per rack                                 | Dense GPU racks wired as one machine; at Firmus’s Tasmania design, about 164 kW of IT power per GB300 rack (our arithmetic) |
| **What limits output**          | Customer demand for space                                                                           | Power first, then demand and the price per token                                                                            |
| **Who carries the demand risk** | Tenants, through long leases                                                                        | The owner or operator, unless a big customer signs a long take-or-pay contract                                              |

*The Firmus rack figure is our arithmetic from its 84 MW critical IT load and about 36,800 GPUs. Full list under References.*

The last row is the one that matters most. In a colocation hall, the tenant takes the risk that its servers earn too little. In an AI factory run by the owner, that risk sits with the owner. It moves back to a customer only when that customer signs a long, fixed contract for the capacity, which is why the largest AI factory projects are announced alongside a named or unnamed anchor customer.

## How does an AI factory work?

Follow the electricity. Every step from the grid to a paying customer has a different owner, and each owner is paid at a different moment.

- **1\. Secure the power.** A grid connection, firmed supply and backup. The number of megawatts sets the maximum number of racks. Our [page on how much power an AI data center uses](https://www.sotkn.com/explained/how-much-power-does-an-ai-data-center-use/) and [contracted power explainer](https://www.sotkn.com/explained/contracted-power-explained/) show how that is counted.
- **2\. Build the hall.** Developers and builders put up the shell, power rooms and cooling, often with factory-built modules; see [who builds AI data centers](https://www.sotkn.com/explained/who-builds-ai-data-centers/) and [what a modular data center is](https://www.sotkn.com/explained/modular-data-center-explained/).
- **3\. Install the racks.** GPU systems arrive, are wired into one cluster and are tested before the customer accepts them.
- **4\. Produce tokens.** The operator runs training jobs and inference. How many tokens come out depends on the chips, the software and how busy the racks stay. Our [GPU utilisation explainer](https://www.sotkn.com/explained/gpu-utilization-explained/) covers that last point.
- **5\. Sell tokens.** Customers pay per token through an API, per GPU hour, or under a fixed contract for the whole cluster; see [AI tokens vs GPU hours](https://www.sotkn.com/explained/ai-tokens-vs-gpu-hours/).

![How an AI factory turns power into revenue in five steps: secure the power as the stuck step, build the hall, install the racks where Nvidia is paid, produce tokens, and sell tokens where the factory owner is paid last](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/ai-factory-five-steps.png)

Power in, tokens out; each hand is paid at a different time. NVIDIA, Firmus, YTL; Second Order.

## How does an AI factory make money?

By selling its output, in one of three ways: tokens sold directly to users through an API, capacity rented by the GPU hour, or a whole cluster leased to one large customer for years. The first earns the most per GPU when demand is strong and the least when it is weak; the last earns less but is far more certain.

Nvidia has put a number on the best case. In October 2025 it said a $5 million investment in a GB200 NVL72 system “can generate $75 million in token revenue,” a 15x return, using results from the independent SemiAnalysis InferenceMAX benchmark on the DeepSeek R1 model. It also said that, “for power-limited AI factories,” its Blackwell chips deliver 10 times the throughput per megawatt of the previous generation on one class of models.

Read that figure for what it is. It is revenue, not profit, so it ignores power, staff, financing and the wear on the chips. It is modelled at an assumed price per token, and token prices fall as models get cheaper to run. And it assumes the system is busy. Our view: it is a ceiling for a well-run, fully sold factory, not a forecast for any particular one. The two numbers that decide whether a real AI factory pays back are how full it stays and what each token sells for.

For how rental prices and chip lives feed into that, see [GPU rental prices](https://www.sotkn.com/explained/gpu-rental-price-explained/) and [GPU depreciation](https://www.sotkn.com/explained/gpu-depreciation-explained/).

## Who gets paid, and when?

The term “AI factory” makes the order of payment easy to see, and the order is the point.

- **Nvidia and the equipment makers, first.** Nvidia books revenue when it delivers systems. In the quarter to 26 July 2026, its data center revenue was $89,023 million, up 117% on a year earlier, and 92.5% of its total revenue (our arithmetic). Of that, $40,313 million came from what it calls AI clouds, industrial and enterprise customers, including AI start-ups and sovereign buyers, which is 45.3% of data center revenue (our arithmetic); the rest came from the hyperscalers. Its gross margin was 75.0%. See [who makes AI chips](https://www.sotkn.com/explained/who-makes-ai-chips/).
- **Nvidia’s own suppliers, booked years ahead.** Nvidia’s supply and capacity commitments, mostly for memory and manufacturing capacity, rose from $119 billion to $279 billion in that one quarter, about 2.3 times as much, and about 3.1 quarters of data center revenue at the latest rate (our arithmetic). That is money promised to chip, memory and packaging suppliers before any factory owner sells a token. See [what a purchase commitment is](https://www.sotkn.com/explained/purchase-commitment-explained/).
- **The landlord and builders, when the hall is handed over.** Rent starts at handover, whether or not the tokens sell. Our [data center lease explainer](https://www.sotkn.com/explained/data-center-lease-explained/) shows how.
- **The factory owner, last.** Whoever owns the GPUs, often a [neocloud](https://www.sotkn.com/explained/neocloud-explained/), a hyperscaler or a national champion, is paid only when output sells. It has usually borrowed to buy the chips, so it needs the factory full from day one.

![Bar chart of Nvidia commitments at 26 July 2026: supply and capacity commitments $279 billion, one quarter earlier $119 billion, data center revenue for the quarter $89.0 billion, and $36 billion of cloud capacity Nvidia committed to rent back from AI clouds](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/ai-factory-commitments.png)

Nvidia has lined up $279 billion of supply, about three quarters of data center sales. NVIDIA Form 10-Q; ratios are Second Order arithmetic.

There is one twist. Nvidia is not only a seller to AI factories; it is also a buyer from some of them. Its 10-Q says AI clouds buy its data center products while Nvidia commits to cloud service agreements with them, typically for six years, and those commitments totalled $36 billion at 26 July 2026\. The AI clouds can stop providing that capacity to Nvidia and sell it to others at better rates, and Nvidia may share in revenue from third-party customers. In plain terms, part of the demand that keeps some AI factories busy comes from the chipmaker that sold them the chips.

Our view: the AI factory label is accurate about the product and quiet about the risk. The maker of the machines is paid on delivery, with a 75% gross margin. The owner of the machines is paid on demand, with borrowed money. For how those purchases are funded, see [who finances GPU clusters](https://www.sotkn.com/explained/who-finances-gpu-clusters/) and [data center financing](https://www.sotkn.com/explained/data-center-financing-explained/).

## Where is the stuck step?

Power, and then demand. Neither is solved by buying more chips.

**Power sets the size.** Firmus’s Tasmania campus shows how tightly the two are linked. The company says the site has 84 MW of critical IT load, the power available to the computing equipment itself, and that it will host about 36,800 Nvidia GB300 GPUs. That is about 438 GPUs per megawatt, or about 511 racks of 72 GPUs, each drawing about 164 kW of IT power once networking and storage are shared out (our arithmetic). Add another megawatt and you can add roughly another 438 GPUs; without it, you cannot add any. That is why Nvidia sells its newer chips on throughput per megawatt: in a power-limited factory, more tokens per watt is the only way to grow output without new power.

**Demand sets the payback.** A finished, powered factory still needs buyers at a price that covers the chips before they age. The safest projects sign a large customer before the racks arrive. Firmus says its Melbourne deployment has a multi-billion dollar contract with a leading global technology company for about 18,400 GB300 GPUs, and that its build-out is backed by a US$10 billion debt facility led by Blackstone. Those contracts move the demand risk to the customer; without them, it stays with the owner. Our [take-or-pay GPU contracts explainer](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/) covers how those deals are written.

**Cash comes before tokens.** Because the chipmaker is paid first, the owner has to fund the chips long before output starts. Our Stack page on [what Firmus does](https://www.sotkn.com/stack/neoclouds/what-does-firmus-do/) shows what that looks like when the cash for chips is the tight step.

Our view: an AI factory’s capacity is decided at the substation, and its profit is decided in the market for tokens. The chips sit in between and are paid for first. When you read about a new AI factory, the two questions are: is the power already connected, and who has signed to buy the output?

## AI factories in Malaysia, Australia and Singapore

**Malaysia.** YTL Power completed an AI data centre with Nvidia at its 600 MW YTL Green Data Center Park in Kulai, Johor, in 2025, running liquid-cooled GB200 NVL72 systems, and says its YTL AI Cloud is operational alongside ILMU, a Malaysian-built language model. In August 2026 it agreed to buy about 145 acres at Sedenak Tech Park West, with an option on up to 400 more, for a new gigawatt-scale campus to sit beside Kulai, where it plans up to 1.2 GW. NEXTDC is also building in Kuala Lumpur. Power in Johor is the gate here too; see our [Malaysia Pulse on Johor’s grid connections](https://www.sotkn.com/markets/malaysia/malaysia-johor-data-centre-grid-connection-october-2026/) and our [Malaysia door](https://www.sotkn.com/markets/malaysia/).

**Australia.** Firmus calls Project Southgate the beginning of “Australia’s AI Factory era.” With CDC Data Centres and Nvidia, it has set out an A$4.5 billion first stage and a path to as much as A$73.3 billion and 1.6 GW of operating AI factories through 2028, about A$45.8 million per megawatt all-in if that full figure is reached (our arithmetic; the release does not split chips from buildings). It also says tokens surplus to local demand would be offered to allied markets, which treats tokens as an export. See [what NEXTDC does](https://www.sotkn.com/stack/ai-data-centers/what-does-nextdc-do/) and our [Australia door](https://www.sotkn.com/markets/australia/).

**Singapore.** Singapore is where many of the region’s buyers, investors and operators sit rather than where the largest new halls go up; Firmus lists offices in Australia and Singapore, and YTL pitches Johor partly on its proximity to Singapore. New data center capacity in Singapore itself is allocated by the government, so a large AI factory there is a question of winning an allocation, not just buying chips. See [what sovereign AI is](https://www.sotkn.com/explained/sovereign-ai-explained/) and our [Singapore door](https://www.sotkn.com/markets/singapore/).

## What would change this view?

- **Much faster grid connections.** If utilities in Johor, Tasmania or Texas cut connection times sharply, power stops being the cap and chip supply, or demand, becomes the binding step.
- **Token prices falling faster than costs.** If the price per token drops faster than each new chip generation cuts the cost of making one, AI factories that sell tokens directly earn less than their business plans assumed.
- **Long contracts becoming the norm.** If most AI factories are pre-sold to hyperscalers and model labs on long take-or-pay terms, the owner looks more like a landlord, and the demand risk moves to the buyer.
- **Nvidia’s rent-back shrinking.** If the $36 billion of cloud capacity Nvidia has committed to falls as AI clouds sell it to others at better rates, that would be a sign of strong outside demand; if Nvidia has to keep adding it, the opposite.

## How to read an AI factory headline

- **Megawatts or GPUs?** Check whether the number is power (MW of critical IT load) or chips, and whether the power is connected or only planned.
- **Planned or operating?** Firmus’s 1.6 GW is a plan through 2028; its Tasmania site was still being built in 2026.
- **Who buys the output?** A named or unnamed anchor customer on a multi-year contract is worth more than a pipeline.
- **Revenue, or modelled revenue?** Nvidia’s $75 million per $5 million system is a modelled example, not a result.
- **Which currency?** Firmus reports most figures in Australian dollars; Nvidia and its debt facility in US dollars.

## FAQ

**What is an AI factory in simple terms?**  
An AI data center described as a production line. Power and data go in, and AI output comes out, counted in tokens, the small chunks of text, code or image data that a model reads and writes. Nvidia popularised the term and defines the product as intelligence measured by token throughput.

**Is an AI factory different from an AI data center?**  
Physically, not much. Both are halls full of GPU racks with heavy power and cooling. The difference is the yardstick: a data center is usually sized and rented by megawatts or racks, while an AI factory is judged by how many tokens it produces per megawatt and per dollar. Our page on what an AI data center is covers the building itself.

**Who coined the term AI factory?**  
Nvidia popularised it to describe customers’ GPU data centers, and now publishes reference designs for gigawatt-scale AI factories. Other companies, including Firmus in Australia and NEXTDC, use the same language for their own sites.

**How does an AI factory make money?**  
By selling what the GPUs produce: tokens through an API, model training time, or GPU capacity rented to someone who does. Nvidia’s own worked example says a $5 million GB200 NVL72 system can generate $75 million of token revenue, but that is modelled revenue at an assumed token price, not profit, and it assumes the system stays busy.

**Who gets paid first when an AI factory is built?**  
The chipmaker and equipment makers, when the racks ship. Nvidia books revenue when it delivers systems; in the quarter to 26 July 2026 its data center revenue was $89.0 billion. The landlord is paid once a lease starts. The owner of the factory is paid last, only when the tokens or GPU hours sell.

**What limits how much an AI factory can produce?**  
Power, first. A factory cannot run more racks than its grid connection and backup can feed, which is why Nvidia sells its newer chips on throughput per megawatt. After power comes demand and price: tokens that are produced but not sold, or sold cheaply, do not pay back the chips.

**Are there AI factories in Asia?**  
Yes. YTL Power has an AI data centre with Nvidia GB200 systems at its Kulai park in Johor, Malaysia, and is adding a gigawatt-scale campus at Sedenak. Firmus is building Project Southgate in Tasmania and Melbourne, Australia. Singapore hosts headquarters and buyers more than large new halls, because new data center capacity there is allocated by the government.

**What is a token?**  
The unit an AI model reads and writes, roughly a word or part of a word for text. AI services are increasingly priced per million tokens, which is why AI factories are measured in them. See our explainer on AI tokens.

**Is this a list of stocks to buy or sell?**  
No. This page explains how the money and the constraints move when an AI factory is built. It does not give buy or sell calls. Not investment advice.

Related: [what is an AI data center](https://www.sotkn.com/explained/what-is-an-ai-data-center/); [AI inference](https://www.sotkn.com/explained/ai-inference-explained/); [GPU as a service](https://www.sotkn.com/explained/gpu-as-a-service-explained/); [PUE](https://www.sotkn.com/explained/pue-explained/); [where AI data centers are located](https://www.sotkn.com/explained/where-are-ai-data-centers-located/); [how long an AI data center takes to build](https://www.sotkn.com/explained/how-long-does-it-take-to-build-an-ai-data-center/); [the Stargate data center project](https://www.sotkn.com/explained/stargate-data-center-explained/); [Wire on Naver and Brookfield’s AI factory deal](https://www.sotkn.com/wire/naver-brookfield-9b-gpu-rental-ai-factory-wire/).

## Stack doors

Layer: [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Also: [Neoclouds](https://www.sotkn.com/stack/neoclouds/), [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/), [Semiconductors](https://www.sotkn.com/stack/semiconductors/). Markets: [US](https://www.sotkn.com/markets/us/), [Malaysia](https://www.sotkn.com/markets/malaysia/), [Singapore](https://www.sotkn.com/markets/singapore/), [Australia](https://www.sotkn.com/markets/australia/).

[Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when an AI factory is powered, a customer signs for its output or a chip order moves.

**How we check this**

Figures come from Nvidia’s quarterly report on Form 10-Q for the three months to 26 July 2026 (filed with the SEC on 26 August 2026), Nvidia’s AI factory glossary and solutions pages, Nvidia’s InferenceMAX blog post of 9 October 2025, Firmus’s releases of October 2025 and 2 March 2026 and its Southgate Tasmania project page, YTL’s releases of 31 October 2025 and 19 August 2026, and NEXTDC’s guide to AI factories. Nvidia’s token-revenue example is the company’s own modelled figure, not a reported result. Firmus figures are company statements about projects still being built. Firmus reports in Australian dollars except where it marks US dollars. Every per-megawatt figure, per-rack figure, multiple or ratio we derived is labelled “our arithmetic”. The featured photo shows server racks in NASA’s Discover supercomputer hall (NASA Goddard / Pat Izzo, public domain), graded by Second Order; it is not one of the AI factories named here. This page explains how the money and the constraints move. It does not give buy or sell calls. Not investment advice. Do your own research.

**Last reviewed:** 15 Nov 2026 · **Next review:** after Nvidia’s third-quarter fiscal 2027 results (late November 2026) and when Firmus’s Tasmania campus comes online

Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here.

[Our editorial standards →](https://www.sotkn.com/editorial-standards/)

## References

1. [NVIDIA — What Is an AI Factory? (glossary)](https://www.nvidia.com/en-us/glossary/ai-factory/?ref=sotkn.com) — an AI factory is specialised computing infrastructure that manages the AI life cycle from data ingestion to training, fine-tuning and high-volume inference; the primary product is intelligence, measured by token throughput; optimised for inference performance and energy efficiency
2. [NVIDIA — AI Factories solutions page](https://www.nvidia.com/en-us/solutions/ai-factories/?ref=sotkn.com) — AI factories are purpose-built to manufacture intelligence at scale and optimise token generation; they unify five layers: energy, chips, infrastructure, models and applications; DSX reference design for gigawatt-scale AI factories, built to maximise tokens per watt
3. [NVIDIA Blog — Blackwell InferenceMAX benchmark results (9 Oct 2025)](https://blogs.nvidia.com/blog/blackwell-inferencemax-benchmark-results/?ref=sotkn.com) — a $5 million investment in a GB200 NVL72 system can generate $75 million in DeepSeek R1 token revenue, a 15x return, based on the SemiAnalysis InferenceMAX v1 benchmark; for power-limited AI factories, Blackwell delivers 10x throughput per megawatt for mixture-of-experts models versus the previous generation
4. [NVIDIA — Form 10-Q for the quarter ended 26 Jul 2026 (filed 26 Aug 2026)](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm?ref=sotkn.com) — Data Center revenue $89,023 million, up 117%; Hyperscale $48,710 million; AI Clouds, Industrial and Enterprise $40,313 million, up 138%, driven by AI natives, enterprises and sovereign customers; gross margin 75.0%; supply and capacity commitments $279 billion, up from $119 billion a quarter earlier; cloud service agreement commitments to AI clouds $36 billion, typically six years
5. [Firmus — Multi-year agreement with a global hyperscale customer at Project Southgate (2 Mar 2026)](https://firmus.co/newsroom/firmus-signs-multi-year-agreement-with-global-hyperscale-customer-at-project-southgate?ref=sotkn.com) — multi-billion dollar contract for about 18,400 NVIDIA GB300 GPUs at the Melbourne facility; US$10 billion debt facility led by Blackstone; more than $300 million committed to a domestic supply chain able to support up to 1.5 GW of AI facility delivery a year; Tasmania campus to host about 36,800 GB300 GPUs, completion targeted late 2026
6. [Firmus — Firmus expands Project Southgate across Australia (Oct 2025)](https://firmus.co/newsroom/southgate-expansion?ref=sotkn.com) — alliance with CDC Data Centres and NVIDIA; A$4.5 billion initial investment with scale up to A$73.3 billion and 1.6 GW of operational AI factories through 2028; PUE 1.10; AI tokens surplus to domestic demand to be offered to allied markets
7. [Firmus — Southgate Tasmania project page](https://firmus.co/infrastructure/southgate?ref=sotkn.com) — Launceston, Tasmania; 84 MW critical IT load; firmed Tasmanian hydro, wind and solar; PUE below 1.10, primarily direct-to-chip cooling
8. [YTL — NVIDIA and YTL brief Prime Minister Anwar Ibrahim at APEC (31 Oct 2025)](https://ytlcommunity.com/shownews.asp?newsID=5583&ref=sotkn.com) — completion of YTL Power’s AI data centre facility with NVIDIA at the 600 MW YTL Green Data Center Park in Kulai, Johor, powered by liquid-cooled NVL72 Grace Blackwell (GB200) GPUs; YTL AI Cloud operational; ILMU large language model
9. [YTL Power and JLand Group — Gigawatt-scale campus at Sedenak Tech Park (19 Aug 2026)](https://ytlcommunity.com/shownews.asp?newsID=5664&ref=sotkn.com) — SIPP Power to acquire about 145 acres in Sedenak Tech Park West with an option on up to 400 more; complements the Kulai park with planned capacity of up to 1.2 GW
10. [NEXTDC — Building APAC’s AI factories: a strategic guide for hyperscalers](https://www.nextdc.com/blog/building-apacs-ai-factories-a-strategic-guide-for-hyperscalers?ref=sotkn.com) — AI factories treat the whole facility as one computing unit; NEXTDC’s new builds in Tokyo, New Zealand and Kuala Lumpur extend its platform for hyperscalers scaling GPU-as-a-service

S

**[Sean](https://www.sotkn.com/author/sean/)**

Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck.

Covers AI infrastructure, semiconductors, and second-order stack reads.

[Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)