AWS just doubled down on the bayou. $18 billion. Three campuses. One bet that AI compute demand is a bottomless pit.
Amazon Web Services announced an expansion of its Louisiana data center investment from $10 billion to $18 billion, adding a third campus to the two it disclosed in August 2024. The move isn't just a headline—it's a signal. I've been tracking hyperscaler capex since the 2017 ether rush, and this is the kind of commitment that reshapes the entire compute landscape.
Context: Why Now, Why Louisiana?
The original $10 billion plan targeted two campuses in northeastern Louisiana. The new $8 billion injection brings a third site, pushing total IT load to an estimated 300–500 megawatts. That's enough to host hundreds of thousands of GPUs—think 50,000+ H100 equivalents per campus. Louisiana offers cheap industrial power (6–7 cents/kWh vs. national average 11–12 cents) and a grid that isn't clogged like Northern Virginia's PJM market, where interconnection queues stretch years.
This isn't just about capacity. It's about escaping the Northern Virginia bottleneck. AWS, Microsoft, and Google have been fighting over power allocations in Loudoun County for years. Louisiana is a flanking move—secure cheaper, faster-to-market power before the competition wakes up.
Core: The Numbers That Matter
Let's break down what $18 billion buys in 2025:
- Power: 300–500 MW IT load. At 50 kW per rack (high-density AI clusters), that's 6,000–10,000 racks.
- Chips: If AWS deploys its own Trainium2 chips (claimed 30–40% cost savings vs. H100), one campus could host ~200,000 Trainium2 accelerators. Three campuses = 600,000+ chips. That's a massive self-owned compute fleet.
- Water: The Mississippi River provides abundant water for evaporative cooling. At 50+ kW per rack, liquid cooling is mandatory. Louisiana's water access is a hidden technical moat.
- Timeline: Construction starts likely in 2025–2026, with first capacity online by 2027. That's a 2–3 year lead over competitors who are still scouting sites.
I've audited data center supply chains for years. The bottleneck isn't GPUs—it's transformers and switchgear. Lead times for large power transformers are 18–24 months. AWS likely locked those orders months ago. This is chess, not checkers.
Contrarian: The Risk Nobody Is Talking About
Everyone celebrates the $18 billion. But here's the gritty reality: AWS is betting that AI demand grows at 40%+ CAGR for the next five years. If that growth stalls—say, due to AI winter or a shift to edge inference—these campuses become stranded assets. Depreciation on $18 billion over 15 years is $1.2 billion per year. If utilization drops below 60%, AWS's cloud margins take a hit.
And there's the energy angle. Louisiana's grid is ~70% natural gas. AWS pledged 100% renewable energy by 2025. To meet that, they'll need to buy renewable energy credits (RECs) from other states, eating into the cost advantage. I've seen this play out in Ohio—cheap power, but compliance costs erode the savings.
Finally, the biggest blind spot: self-chip dependency. Trainium2 is unproven at scale. NVIDIA's software ecosystem is still miles ahead. If AWS's own chips underperform, they'll have to retrofit with NVIDIA GPUs, blowing up the TCO model. That's a risk I'd flag for any institutional investor.
Takeaway: What to Watch Next
This isn't just a real estate story. It's a bet on the vertical integration of AI compute. AWS is going from "cloud provider" to "chip-to-datacenter AI utility." The key metric to track: Trainium adoption in Bedrock and SageMaker. If AWS's own chips power 60%+ of new AI workloads by 2027, the Louisiana campuses will be the foundation of a new infrastructure monopoly. If not, they'll be expensive white elephants.
Watch the Louisiana grid interconnection filings. If AWS files for 500 MW+ of new load, they're all-in. If they hedge with smaller increments, caution is the word.
We don't trade on hope. We trade on data. And the data says: AWS is chasing the white whale again, but this time the whale is AI compute, and the harpoon is $18 billion of Louisiana steel and silicon.
Hunting spreads while the market sleeps. This is how you position for the next cycle.