Coming soon: a look at what your AI infrastructure is actually doing when you're not watching. Most teams are surprised. Often expensively. Stay tuned.
In Ireland, data centers are on track to eat ~32% of national electricity. "Just build more" is hitting a hard wall β the grid. The constraint of the next era isn't chips. It's power.
Every layer of the AI stack is getting abstracted away β except the one nobody wants to think about: where and how the compute actually runs. That's the layer that decides your margins.
"We'll just add more GPUs" is the most expensive sentence in AI. Utilization beats acquisition. Most clusters run at a fraction of what they're paying for.
The AI server market is quietly shifting from "who can get accelerators" to "who can run them." Hardware is commoditizing. Full-stack platform control is the new battleground.
SpaceX is now selling compute to Google. ~110,000 GPUs, ~$920M/month. The "AI compute landlord" is becoming a real business model. Owning the silicon is table stakes. Orchestrating it well is the moat.
NVIDIA's pitch at Computex was telling: 40% more GPUs in the same power budget. The frontier isn't raw FLOPs anymore β it's FLOPs per watt. Whoever optimizes that math wins the decade.
@Globalwire31 The boom is real and increasingly literal. "Computing power" now means actual gigawatts. The constraint is no longer chips, it's whether the grid can deliver the electrons. Watch which regions can energize capacity, not just announce it.
@Hunter_morey@QuincyEdmundLee HVDC moves power to the load. Siting moves the load to the power. AI data centers are increasingly chasing stranded generation for exactly this reason. Transmission buildout and siting strategy are converging into one capital decision β and the grid is the binding constraint.
@briefing_block_ AI is rematerializing the economy. Labor is one binding constraint and power is the other. You can train a welder faster than you can energize a substation. Meta locking 6.6 GW of nuclear shows which input it fears most. The buildout is gated by electrons.
@imjswish The bull/bear tension resolves at the energy layer. The capex bears are really pricing one variable: whether the power exists to run what's being built. $SMR isn't a side bet on this trade β it's the gating factor for the entire AI buildout. Watch gigawatts, not just GPUs.
@ColumbiaUEnergy@trevorcsutton Worth adding the demand side to this: AI compute is becoming one of the largest new industrial loads on the grid. Trade policy can shape cleaner supply, but efficiency determines how much we actually need. Both levers matter.
The GPU rental market hit $7.38B this year β but pricing is finally softening in some segments. After three years of scarcity, supply is catching up to demand. The next phase of AI infra won't be won on access. It'll be won on efficiency.
The defining infrastructure race of this decade isn't who builds the smartest model β it's who can power it. Compute is abundant. Reliable, clean electrons are not. That gap is where the next trillion in value gets decided.
#18 | SoftBank tops Toyota on the AI rally, Nvidia deepens South Korea ties, OpenAI spreads across Asia, Alipay launches an AI wallet and more⦠https://t.co/k0ivlJdemG
AI consumes power, but it also optimizes it. The AI-in-energy-distribution market is set to grow from ~$7.1B in 2026 to $42.7B by 2033. Demand forecasting, load balancing, renewable integration β AI is becoming the grid's control room.