Stargate Builds Five More Data Centers (Power Grid Weeps)

Training GPT-5 apparently requires enough electricity to power entire countries. These models are power-hungry monsters.

The Numbers Are Completely Insane

7 gigawatts across multiple sites. For context, that's about 1.5x the power output of Hoover Dam, or roughly enough to power Switzerland. The hundreds of billions price tag is the largest private tech infrastructure project ever - basically building the computing equivalent of the Large Hadron Collider, except instead of finding the Higgs boson, we're trying to make chatbots slightly less stupid.

Reality check: Traditional data centers use maybe 50-100 MW. These Stargate facilities are planning facilities that are 10-20x larger. The cooling requirements alone are going to be ridiculous.

Modern AI training runs are completely different from normal computing. Instead of running diverse workloads efficiently, these facilities run thousands of GPUs at 100% utilization 24/7 for months. The power density is nuts - we're talking about racks that draw 80-100kW each.

Why Texas Gets All the Fun Data Centers

Texas has cheap electricity, business-friendly regulations, and politicians who don't freak out about AI. Plus the power grid can actually handle industrial-scale loads without collapsing.

The Milam County facility partnering with SB Energy makes sense - you need dedicated power infrastructure for this scale. You can't just plug into the local grid and hope for the best.

Infrastructure reality: Most regions literally can't support 7GW of new demand. The grid upgrades alone would take years and cost billions. Texas already has industrial-scale power infrastructure from oil/gas, so they can actually deliver the juice.

Oracle's involvement in Shackelford County is smart business - they get to sell cloud services while owning the underlying infrastructure. Vertical integration at enterprise scale.

The Construction Timeline Is Impossible (They'll Try Anyway)

Traditional data center construction takes 2-3 years from planning to operation. Stargate is claiming they'll have facilities operational "next year" in some locations.

I've worked on data center projects before - the "fast-build" timeline is horseshit. But here's what they're probably planning:

• Pre-fabricated modules instead of traditional construction
• Parallel permitting and construction (start building while permits are still processing)
• Simplified cooling designs optimized for speed, not efficiency
• Massive workforce mobilization and overtime costs

The Trump administration's fast-track permitting helps, but you still can't magic away physical construction time. Expect delays and cost overruns.

Power Grid Reality Check

OK, enough ranting about construction timelines, here's the actual physics problem: 7GW of new demand doesn't just appear without consequences. That's a significant fraction of many state's total generating capacity.

Grid impact challenges:

• Transmission infrastructure upgrades required
• Peak demand management (these facilities don't scale down)
• Backup power systems for facilities that can't afford downtime
• Grid stability issues when massive loads come online/offline

Texas ERCOT is probably the only grid in the US that can absorb this much new industrial demand without major upgrades. Even then, expect power prices to increase for everyone else.

Why This Actually Matters

AI training workloads are fundamentally different from traditional computing. You can't just rent more AWS instances and call it a day. Training frontier models requires thousands of GPUs in perfect synchronization, which means custom infrastructure.

The bottleneck shift: We used to be limited by algorithms and software. Now we're limited by how fast we can build data centers and get enough power to run them.

Companies building these facilities aren't just buying compute - they're securing competitive advantage. If you can't access this scale of infrastructure, you can't train competitive models. It's that simple.

Prediction: This infrastructure arms race will determine which companies survive the AI transition. Access to compute becomes more important than access to talent or capital.

In January, Stargate sounded ambitious. By September, it became clear they were actually being conservative. The massive price tag happened because they realized AI training needs way more compute than anyone originally estimated.

When Your Initial Plan Becomes The Small Version

January's announcement was already huge by normal standards. But then they ran the Abilene, Texas facility for a few months and realized they could use 10x more capacity immediately.

What happened: The first NVIDIA GB200 racks came online in June and immediately maxed out. Training runs that they thought would take months were queuing for availability. The bottleneck wasn't algorithms or funding - it was raw compute capacity.

They ran the Abilene facility for a few months and realized the original plan was cute but nowhere near big enough. Training runs were queuing for weeks. Turns out building AGI requires stupid amounts of compute.

Why These Partnerships Actually Make Sense

SoftBank has energy money and knows power infrastructure. Oracle knows enterprise infrastructure and has the cloud platform. OpenAI needs somewhere to plug in their H100s and train GPT-5. Simple as that.

The SoftBank angle: They've been investing in energy projects for years through SB Energy. When you need 7 gigawatts of power, you don't call random contractors - you call people who've actually built power infrastructure at scale.

Oracle's play: They want to be AWS for AI training. Getting OpenAI as an anchor tenant gives them credibility and guaranteed revenue while they build out the platform. Smart move.

OpenAI's problem: They can't train bigger models without more compute. Microsoft's Azure is helpful but not infinite. Having dedicated facilities means they can run training jobs 24/7 without competing for resources with Office 365 workloads.

Geographic Reality Check

Texas wins because:

• Deregulated energy market (cheap power)
• Lots of land that doesn't cost Manhattan prices
• Business-friendly regulations (minimal red tape)
• Existing energy infrastructure

Ohio (Lordstown):

• Old GM plant infrastructure they can repurpose
• Skilled manufacturing workforce already there
• Decent power grid connections
• Cheap real estate

Midwest expansion:

• Spread the risk across multiple power grids
• Redundancy in case one region has issues
• Different regulatory environments as backup plans

The Construction Reality

Building five massive data centers simultaneously is going to be a shitshow. The "fast-build" timeline for Milam County is optimistic at best. Expect:

• Cost overruns: The massive price tag will grow even bigger
• Timeline delays: "18 months" actually means 24-30 months in construction reality
• Supply chain nightmare: Good luck getting GB200 racks when everyone wants them
• Power grid stress: Local utilities are definitely not ready for this load

The modular construction approach makes sense - build standardized components off-site, then assemble on location. Less weather delays, more quality control, faster deployment when it works.

What This Actually Means

Whoever controls this infrastructure controls the future of AI development. If Stargate works, OpenAI gets a massive advantage over Google and Microsoft in the race to AGI. If it fails, they're stuck with whatever compute they can rent from cloud providers.

The 7GW power requirement is no joke - that's more than some small countries use. Local power grids will need serious upgrades to handle the load. Expect rolling blackouts and angry residents when their power bills spike to subsidize AI training.

Bottom line: This is the biggest infrastructure bet in AI history. It'll either make OpenAI the dominant AI company or bankrupt them trying. No middle ground.