$500 Billion Bet: Are AI Factories the Next Internet Backbone or a Massive Bubble?

When a tech conglomerate and a chip company shake hands promising a $500 billion partnership, the first question worth asking isn't "how will this money be spent," but "who will pay for this AI factory before 2027." On the surface, the deal between SK Group and NVIDIA looks like an arms race for computing infrastructure, but behind it lies a business model experiment regarding the "industrialization of AI."

Short Term: Capital-Intensive "Resource Stacking" Logic

From a technical feasibility perspective, a 2-gigawatt (2GW) AI cloud computing center is not a fantasy. Currently, the world's largest hyperscale data centers, such as Google's and AWS's campuses, typically have power capacities between 500 megawatts and 1 gigawatt. 2GW implies needing output equivalent to a nuclear power plant, which is a national-level infrastructure project in any country.

Technical Breakdown:

Dimension	Estimated Value	Notes
Total Power	2 GW	Roughly equal to 2 large nuclear plants or 8 coal-fired plants
Typical GPU Power Consumption	700-1000W/unit (Vera Rubin estimated)	At 1.5U density, a single rack can reach 10-15kW
Computing Density	Approx. 2 million GPUs (if all Vera Rubin)	Actual deployment will be mixed, but the scale is staggering
Cooling Method	Liquid cooling (direct liquid or immersion)	Air cooling for 2GW is impossible; heat density is too high
Annual Electricity Cost	Approx. $1.5-2 billion	Based on $0.08/kWh, pure electricity cost

# Rough estimate: Annual electricity cost for a 2GW data center
power_mw = 2000
hours_per_year = 8760
utilization = 0.85  # Assuming 85% utilization
cost_per_kwh = 0.08  # USD
annual_cost = power_mw * 1000 * hours_per_year * utilization * cost_per_kwh
print(f"Annual electricity cost approx: ${annual_cost/1e9:.2f} billion")
# Output: Annual electricity cost approx: $1.193 billion

This figure already approaches the annual operational cost of the largest single data centers run by major cloud providers. The combined $500 billion from SK Group and NVIDIA clearly isn't just buying hardware—it includes land, power grids, cooling systems, and custom development of the entire AI software stack.

[!note] Key Conclusion: In the short term, this deal is a "bet" on the continuity of the CUDA ecosystem. By binding SK with Vera Rubin DSX (Distributed System Exchange architecture), NVIDIA essentially forces the client to buy out the future 5-10 years of computing growth path upfront rather than renting on demand. This puts immense pressure on SK's cash flow, but NVIDIA secures guaranteed orders.

Long Term: The "Commercial Reusability" Challenge of AI Factories

Implementation Difficulty Assessment:

• Technical level: 70% feasible (but dependent on Vera Rubin chips mass-producing on time, maturity of liquid cooling systems, and stability of power supply)
• Commercial level: 40% feasible (biggest risk is "who will rent this 2GW of computing power")

In the short term, SK can absorb some computing demand through the South Korean government and local enterprises (like Samsung, LG), but 2GW represents a significant share of global total AI computing power. According to IDC data, global AI server shipments in 2026 are approximately $200 billion. If a 2GW data center is fully loaded, its computing equivalent could account for 5%-10% of newly added global AI computing power. This means SK must find enough "AI factory customers"—not ordinary cloud users, but super-enterprises requiring long-term, exclusive, high-throughput training tasks (such as OpenAI, Google DeepMind, or future autonomous driving companies).

Long-Term Business Model Comparison:

Model	Traditional Cloud	AI Factory
Billing Method	On-demand/Hourly	Reserved contracts + Resource packages
Tenant Type	Thousands of small clients	Very few large clients
Utilization Risk	Lower (multi-tenant sharing)	Extremely High (vacancy equals loss)
Capital Recovery Period	3-5 years	5-8 years (assuming >80% utilization)

The essence of an AI factory is "computing wholesale," similar to large hydroelectric stations supplying direct power to aluminum smelters. SK needs to find a batch of clients willing to sign long-term "computing purchase agreements," akin to the "bandwidth wholesale" model used by internet companies. But the problem is: AI large model training demand fluctuates wildly. A company might need 100,000 GPUs within six months, then only 10,000 afterwards. Can SK's "dedicated rental" solution meet this elasticity?

[!abstract] My Judgment: The real winner in the SK-NVIDIA collaboration is NVIDIA. With this $500 billion order, it locks in production capacity and cash flow for the next two years while turning the concept of "AI factories" from PPT slides into financeable assets. For SK, this is a gamble—if AI demand continues to explode in 2027, they win; if the computing bubble bursts, they will be the biggest bag-holder.

Actionable Advice for Readers

If you are an enterprise CTO or CIO planning your own AI infrastructure, don't get swept up by news like this.

Short-Term Action List:

• Assess your own AI workload type: Is it training-heavy or inference-heavy? Training requires long cycles and high density; inference requires low latency and elasticity.
• Don't blindly build your own AI factory: Unless you have a $500 billion budget and government support like SK, prioritize hybrid cloud + dedicated computing pools.
• Monitor the ecosystem maturity of Vera Rubin: NVIDIA's chip iteration pace is accelerating, but the software stack...

Original link: https://www.ithome.com/0/981/787.htm