The global AI infrastructure boom is driving unprecedented investment in AI data centers (AIDC). Across Southeast Asia and other rapidly growing digital markets, developers are securing large power capacities—often 50MW, 100MW or more—to support next-generation GPU clusters.
But for a billion-dollar AI infrastructure project, one question is becoming increasingly important:
How quickly can invested capital begin generating cash flow?
Traditional data center development focuses heavily on construction cost, often measured in CAPEX per MW. For AI factories filled with expensive GPUs, however, construction cost alone does not determine investment performance.
Time-to-Revenue and Capital Efficiency can be equally important.
Consider a hypothetical 48MW AIDC project with 16,384 NVIDIA B300 GPUs. Instead of constructing the entire facility before starting operations, the project can be divided into 16 standardized 3MW modular AI Pods.
This changes not only how the data center is built, but also how capital is deployed and recovered.
What Does a 48MW AI Data Center Look Like?
For this model:
48MW = 16 × 3MW Modular AI Pods
Each 3MW Pod contains approximately:
- 128 eight-GPU AI servers
- 1,024 B300 GPUs
- Dedicated power infrastructure
- High-speed networking
- Liquid cooling infrastructure
- Monitoring and control systems
The complete 48MW AI factory therefore contains:
2,048 AI servers + 16,384 B300 GPUs
Assuming each eight-GPU server together with its associated high-speed networking infrastructure costs approximately $700,000, total server and networking investment reaches approximately $1.434 billion.
If modular AIDC infrastructure costs approximately $208 million, total project CAPEX reaches approximately $1.642 billion.
A conventional facility benefiting from larger-scale civil construction might reduce infrastructure CAPEX by approximately 5%, resulting in total investment of around $1.631 billion.
Therefore, conventional construction appears approximately $10.4 million cheaper.
But that comparison ignores one of the most valuable assets in an AI data center:
time.
Every Month of Earlier GPU Operation Has Financial Value
Assume GPU computing capacity is monetized at:
$4.80 per GPU-hour
with average utilization of:
85%
Once all 16,384 GPUs are operational, monthly revenue could reach approximately $48.1 million.
After electricity, cooling, bandwidth, maintenance and other operating expenses—with assumptions including a PUE of 1.15 and electricity price of $0.07/kWh—the project could potentially generate approximately $44.8 million in monthly NOI.
At this scale, bringing computing capacity online even one month earlier can represent tens of millions of dollars in potential cash flow.
That makes construction speed a financial variable rather than simply an engineering KPI.
Traditional AIDC: Build Everything Before Monetization
A conventional hyperscale data center typically follows a sequential development process:
Design → Civil Construction → MEP Installation → Testing → Commissioning → Server Installation → Commercial Operation
Suppose an aggressively managed 48MW project can be completed within 12 months.
During months 1–12, significant capital is continuously invested in buildings, electrical infrastructure, cooling systems, servers and networking equipment.
However, there may be little or no GPU revenue.
Only after the entire facility is ready does the complete 16,384-GPU cluster enter commercial operation.
The problem is therefore not necessarily that conventional construction is too expensive.
The problem is that a very large amount of capital can remain non-revenue-generating during construction.
Modular AIDC: Build One, Operate One, Monetize One
Modularized AIDC changes the deployment philosophy.
Instead of treating 48MW as one enormous facility, it can be treated as:
16 × standardized 3MW AI computing units.
Each Pod can potentially be independently manufactured, integrated, tested, delivered, commissioned and placed into operation.
For example, Pod 1 could enter construction in Month 1, followed by Pod 2 in Month 2 and Pod 3 in Month 3.
Once Pod 1 is commissioned, its 1,024 GPUs can begin generating computing revenue while later Pods remain under construction.
The investment model changes from:
Build everything → Commission everything → Start earning
to:
Build → Commission → Earn → Expand
This creates what can be described as a Rolling CAPEX Model.
Rolling CAPEX Can Reduce Peak Funding Requirements
This distinction is particularly important for investors.
The conventional project may require peak funding approaching the entire $1.631 billion CAPEX before meaningful operating cash flow begins.
Under the modular scenario described above, early Pods generate cash while subsequent Pods are still being deployed.
Based on the model assumptions, this could reduce peak external funding requirements by approximately $244 million, or nearly 15%.
This does not mean modular AIDC reduces total project CAPEX by $244 million.
Instead, it means investors may not need to have the entire project capital committed simultaneously.
For infrastructure funds, private equity investors, data center developers and AI cloud operators, this difference can be extremely important.
Higher CAPEX Can Still Produce Better Returns
Interestingly, modular construction can cost slightly more.
In our model:
MetricTraditional AIDCModular AIDCTotal CAPEX$1.631B$1.642BPeak Funding$1.631B$1.387B5-Year Net Profit$518.6M$575.4M5-Year CAPEX ROI31.79%35.05%Peak Funding ROI31.79%41.49%Estimated Annualized IRR~12.0%~14.9%
The modular project requires approximately $10.4 million more CAPEX, yet the model indicates approximately $56.8 million more five-year net profit.
Why?
Because GPU capacity begins generating revenue earlier.
The objective is therefore not simply to minimize CAPEX.
It is to optimize the relationship between:
CAPEX + deployment speed + revenue timing + capital utilization.
The New AIDC Metrics: Beyond $/MW
Traditional data center development frequently uses $/MW as a key benchmark.
For AI infrastructure, that is no longer enough.
Three additional metrics should become central to project evaluation.
1. CAPEX per GPU
How much infrastructure investment is required to bring one GPU into commercial operation?
For AI factories, the GPU—not the building—is ultimately the revenue-generating asset.
2. Time-to-Revenue
How many months pass between the first dollar invested and the first GPU-hour sold?
Reducing this period can materially change project economics.
3. Capital Efficiency
How much operating cash flow can be generated relative to peak capital employed?
A modular project that begins producing revenue progressively can potentially outperform a lower-cost project that requires all capital to be deployed before monetization.
Standardized 3MW Pods Can Transform the AIDC Supply Chain
The deeper advantage of modular AIDC is not simply the use of containers.
The real advantage is standardization and repeatability.
Imagine defining one standardized product:
3MW AI Pod = 128 servers = 1,024 GPUs
The supporting infrastructure can be engineered around that repeatable architecture, including:
- MV/LV electrical distribution
- Switchgear and transformers
- UPS and backup power architecture
- High-density busway or power distribution
- Direct-to-chip liquid cooling
- CDU systems
- Heat rejection equipment
- Network infrastructure
- Fire detection and suppression
- Environmental monitoring
- DCIM and intelligent control systems
Once standardized, significant portions of the system can be factory manufactured and pre-integrated in parallel with site preparation.
This reduces dependency on sequential on-site construction.
Factory production also enables standardized quality control, repeatable testing procedures and faster replication across multiple projects and countries.
TLS Energy Provides Modularized AIDC Solutions
TLS Energy is developing modularized infrastructure solutions for next-generation AI data centers and AI factories.
Instead of viewing the data center as one large customized building, TLS Energy applies a product-oriented approach to AI infrastructure.
The objective is to create standardized, factory-integrated modular systems that can support high-density GPU computing environments.
A modular AIDC architecture can include GPU/IT modules, liquid-cooling modules, power distribution modules, MV/LV electrical modules and other supporting infrastructure, depending on project requirements.
For large projects, standardized modules can be manufactured and integrated in controlled factory environments while foundations, grid connections and other site works proceed simultaneously.
This parallel construction strategy can significantly shorten project schedules compared with purely sequential construction.
More importantly, the modular architecture allows developers to deploy capacity according to actual demand.
A project does not necessarily need to start with 48MW.
It could begin with:
3MW → 6MW → 12MW → 24MW → 48MW
Each stage can add standardized computing capacity as customer demand, power availability and financing develop.
This makes modularized AIDC particularly attractive for rapidly expanding AI cloud providers, hyperscalers, sovereign AI infrastructure projects and GPU-as-a-Service operators.
From Data Center Construction to AI Factory Manufacturing
The traditional data center industry thinks primarily in terms of construction.
The modular AIDC industry can increasingly think in terms of manufacturing.
That distinction is important.
Construction tends to be project-specific. Manufacturing focuses on standardization, repeatability, supply-chain optimization, factory testing and production capacity.
Once a 128-server / 1,024-GPU / 3MW Pod becomes a standardized product, suppliers can reserve components earlier, optimize production lines and shorten delivery schedules.
Instead of asking only:
“How cheaply can we build 48MW?”
AI infrastructure investors can ask:
“How quickly can we manufacture and activate the next 3MW of revenue-generating computing capacity?”
Conclusion: The Real Competition Is Capital Speed
The 48MW, 16,384-GPU example demonstrates an important principle.
Traditional construction may save approximately $10.4 million in infrastructure CAPEX.
But modular deployment could potentially reduce peak capital requirements by approximately $244 million, generate approximately $56.8 million more five-year profit, and increase estimated IRR from around 12.0% to 14.9% under the assumptions used in this model.
The numbers will naturally vary by GPU pricing, utilization, electricity cost, financing structure, PUE, construction schedule and computing rental rates.
But the underlying principle remains:
For AIDC, speed has financial value.
The future AI factory may therefore not be one enormous building constructed all at once.
It could be a network of standardized, independently deployable 3MW computing units, manufactured in factories and brought online progressively.
For TLS Energy, modularized AIDC represents this transition—from building data centers to manufacturing scalable AI infrastructure.
Because in the AI era, the winner may not be the company that builds the cheapest MW.
It may be the company that converts capital into revenue-generating GPU capacity the fastest.
