Thread by @DisruptionHedge

The 3 Big Shifts Rewiring AI (a mega-🧵)

Shift #1: Sovereign AI... A brand new, deep pocketed customer:

People often compare today’s AI boom to the dot-com bubble, with Nvidia $NVDA as the Cisco Systems $CSCO of its day.

The big difference is that Cisco’s customers were debt-fueled dot-com startups with no revenue. Nvidia’s customers are the largest, most profitable companies in the world.

US big tech companies will spend over $300 billion building out their AI data centers this year. That was Phase 1.

Phase 2 is even bigger: Sovereign AI.

Sovereign AI means every country is building its own “AI factory.” AI models trained on local data, running on local compute, under local control.

If you believe, as I do, that AI will run healthcare systems, defense planning, and education in the coming decades… do you really want those systems sitting on foreign chips in foreign data centers?

Imagine running America’s nuclear fleet on Chinese AI. Exactly. It makes TikTok’s “spy app” look cute.

Governments are already writing billion-dollar checks to make sovereign AI possible:

--Europe is building out national supercomputers to give startups access to AI compute.

----The UK just switched on its most powerful supercomputer ever.

--India approved a 38,000 GPU national cluster to train models in 22 local languages.

--Saudi Arabia earmarked $100 billion to turn itself into a regional AI powerhouse.

Canada, Japan, and South Korea are all in, too.

Mark my words: Soon, governments will spend more money on AI than big tech.

Nvidia’s Jensen Huang put it best: “Every country will have an AI factory, just like every country has a telco.”

This is effectively a giant wealth transfer from the rest of the world to US AI companies that make the AI gear. And that’s a great thing for clued-in investors.

Shift #2: From training to inference: compute goes prime time

When people talk about AI infrastructure, they usually mean thousands of GPUs crunching away for weeks to train a giant model. That’s the “training” phase.

Training is like building the engine of a Ferrari. You do it a handful of times, and it’s costly and complex. But once it’s built, it’s ready to race.

The real money is in driving the Ferrari every day, aka “inference.” Every time you ask ChatGPT a question or have it generate an image, that’s inference. And it’s exploding because more and more people are using AI.

Last spring, Google’s $GOOGL Gemini models were processing 480 trillion tokens per month. That number has doubled to 980 trillion. ChatGPT will soon hit 1 billion monthly active users!

This changes everything about AI.

Training can be done in a remote desert data center using cheap power. Inference needs to sit closer to the user, run 24/7, and deliver answers in real time.

If you thought training AI was costly… oh boy, wait until you see what we need now. Inference compute now represents most of the cost of running advanced AI models.

Nvidia isn’t the big winner here. Walk into a data center, and alongside the GPUs you’ll see giant cooling fans… storage disks… memory chips… networking cables… and so on.

More “inference” means more heat… more energy consumed… more memory needed… and more data flowing through thick networking cables.

This shift changes where the money flows.

To move data faster between chips, these clusters need more connections per rack and faster optical links, driving a massive upgrade cycle in optics.

On Nvidia’s new GB200 racks, the optics bill alone can top $500,000. And demand for low-latency communication favors cutting-edge AI networking solutions like Ethernet.

The key takeaway is we’re shifting from huge, one-off training runs to continuous infrastructure spend. Inference creates sustained demand for AI gear. If the training era was Nvidia’s show, the inference era is where the broader ecosystem shines.

Get ready for a mad dash to overhaul AI data centers.

Shift #3: Data center spending is rotating to power, cooling, and networking

In 2006, Google built its first serious data center in Oregon. It cost about $600 million, which was a big deal at the time.

Fast-forward to 2025, and the bill for OpenAI’s “Stargate” project with Oracle Corp. $ORCL and SoftBank is coming in at $500 billion! Individual sites within this project will consume more power than a small city.

These aren’t your father’s data centers. One AI rack now guzzles the same amount of electricity as a dozen legacy racks. Meta Platforms $META even tore down a half-built campus in Utah to rebuild it from scratch for AI workloads.

The big change is that AI data centers must act as one giant computer, not a warehouse of single servers.

Training and inference require every GPU to talk to every other GPU at blistering speeds. That’s why companies are packing chips closer together and spending billions on optical networking gear.

The bottleneck is no longer chips. It’s power. A rack of servers stuffed with AI chips needs about 10X more power than a normal “cloud” server. Meta’s Hyperion cluster in Louisiana alone will consume roughly as much power as all of San Francisco.

Electricity = oil for AI.

AI’s huge power demands are also forcing the whole industry to shift to liquid cooling.

Old-school cloud centers were cooled like office buildings with giant air conditioners pushing cold air through rows of servers. But AI chips run so hot, air can’t carry the heat away fast enough to keep them from melting (literally).

That’s why liquid cooling is becoming to go-to. Water is 17X more effective than air at removing heat.

Google and Microsoft $MSFT are ripping out old air-cooled systems and retrofitting water loops right up against the chips. At Google’s Oklahoma campus, pipes literally snake through the racks, pumping chilled, treated water directly to the GPUs.

The easiest way to picture the AI buildout is to break data centers into three buckets of spending:

--The computers (~60% of the bill): This is the heart of the machine: GPUs, servers, racks, networking gear, and fiber optics. A few years ago, chips alone were 80% of the tab. Now, it’s closer to 50% as networking and optics balloon in importance.

--Power and cooling (~25% of the bill): Think of this as the “life support system.” It includes transformers, switchgear, and liquid-cooling plants. This used to be barely 10% of the total cost. Now it’s the fastest-growing expense line. GPUs run so hot that air can’t cool them anymore.

--The site (~15% of the bill): Finally, there’s the real estate, land, and building shells. Major line items in this category include long-term power contracts, grid upgrades, and on-site gas turbines.

The next trillion dollars of AI spend won’t be on chips alone…

GPUs made Phase 1 billionaires. The pipes and power will mint Phase 2 fortunes.

I find it ironic the sexiest technology on Earth—artificial intelligence—depends on boring inputs like electricity, thick cables, and cooling towers. That’s where the puck is going.