Last updated: June 2026
Key Takeaways
- AI's appetite for power is now large enough to strain electrical grids. RAND projects that global AI datacenter demand could approach California's entire power capacity by 2027, though the estimates carry a wide margin of uncertainty.
- That constraint already reaches your house. The same chip-fabrication capacity feeding AI accelerators is the reason memory and router prices climbed sharply in 2026.
- The durable hedge is not betting on a national champion. It is owning more of your own stack. Open-weight models on your own hardware are not a replacement for frontier cloud AI, but they are insurance against infrastructure you do not control.
A chart from the RAND Corporation has been circulating again, and the number on it is hard to forget: by 2030, the data centers built to run AI could demand more than 300 gigawatts of power, roughly four times the entire generating capacity of California. The viral version of this story frames it as a contest between nations, a question of who builds the most compute the fastest. That framing misses the more useful point.
The binding limits on AI are not flags. They are physics: how fast a grid can deliver power, how long an interconnection queue runs, and how much chip-fabrication capacity exists in the world. Those limits are already reaching past the data center and into your home network, your next hardware purchase, and the price of the cloud services you use. That is the part worth understanding, because it changes what a sensible person should do about it.
The chart everyone is sharing, and what it actually says
The figure comes from a 2025 RAND report by Konstantin Pilz, Yusuf Mahmood, and Lennart Heim. It extrapolates two trends, the growth in AI chip supply and the growth in training compute, to estimate how much power AI data centers will need. In the high-growth scenario, global demand climbs from about 21 GW in 2025 to 68 GW by 2027 and 327 GW by 2030.
The reference lines are what make the chart land. Ten gigawatts of added demand in 2025 alone is more than the total power capacity of the state of Utah. The 68 GW projected for 2027 is close to California's entire 2022 capacity of 86 GW. By 2030, a single category of computing is drawn towering over the largest state grids in the country.
One honest caveat belongs here, and the RAND authors make it themselves: this is an extrapolation, not a forecast. Other credible estimates, from SemiAnalysis, Goldman Sachs, and McKinsey, land lower, and the spread between them is wide. The point is not that 327 GW is certain. It is that even the conservative estimates describe a level of demand the existing grid was never built to absorb.
| Year | Projected AI datacenter power demand | For comparison |
|---|---|---|
| 2025 | About 21 GW total, with roughly 10 GW added this year alone | More than Utah's total power capacity (~9 GW) |
| 2027 | About 68 GW | Close to California's total capacity (~86 GW) |
| 2030 | About 327 GW | Roughly 4x California's total capacity |
RAND (2025) high-growth estimate based on AI chip supply. Other estimates run lower. State figures are 2022 totals. Projections, not forecasts.
Why this is a physics problem, not a flag problem
The viral reading of the chart is geopolitical: whichever country hosts the most compute wins. But the constraints in the data are not about who owns the model. They are about power generation, the queue to connect a new load to the grid, and the global supply of advanced chips. None of those respect a border.
The grid is the tightest of the three. Goldman Sachs projects that US datacenter power demand will more than double, from 31 GW in 2025 to 66 GW in 2027, pushing data centers from about 4 percent of peak summer demand to more than 8 percent. Its analysts warn that some regions, short on planned generation, may have to turn future data centers away. A separate RAND grid study estimates the United States can add only about 82 GW of net new capacity by 2030, and notes that a grid-connection request in a hub like Virginia can take four to seven years to clear.
This is why every major economy is suddenly building generation at a pace not seen in decades, from gas turbines and nuclear restarts to small modular reactors, and, in China's case, 36 reactors under construction and a record 315 GW of new solar installed in 2025. Read that not as a scoreboard but as a measure of the problem: the compute the industry wants needs more electricity than anyone has lying around.
The part that reaches your house
Here is the mechanism that turns a datacenter story into a you story. AI accelerators rely on high-bandwidth memory, a stacked form of DRAM that consumes roughly four times the wafer area per gigabyte of the standard memory in your PC, phone, and router. Fabs serve their datacenter contracts first, so the consumer supply is whatever is left over, and in 2026 the leftover is thin.
The result is visible on any price tracker. A 32GB DDR5 kit that sold for $80 to $120 a year ago was logged by Tom's Hardware at $374.97 on June 3, 2026, with DRAM prices up over 171 percent year-over-year. In February 2026, Micron wound down its consumer-facing Crucial brand to concentrate on enterprise AI customers. We walk through the full picture in our RAM shortage explainer.
It does not stop at memory sticks. The same chips sit inside routers and gateways, and vendors including Cisco, HPE, and Dell have said they are passing the higher component costs straight to buyers. We traced exactly how the AI boom is showing up on networking price tags in this breakdown of rising router prices.
Then there is the channel most people never think about: the cloud itself. If every AI query you make is rented from a hyperscaler, that provider's power and compute constraints quietly become your pricing, your rate limits, and your place in line. That is the understated version of a principle this site keeps returning to: whoever controls the infrastructure controls the experience.
Open versus closed is the real axis, and "open" is not a country
The thread that made this chart go viral folds open-source AI into a single nation, as if "open weights" and "China" were the same thing. They are not. Open-weight models are genuinely multi-origin: Google's Gemma 4 ships under a clean Apache 2.0 license and runs on hardware as modest as a Raspberry Pi, alongside Meta's Llama, France's Mistral, Microsoft's Phi, and, yes, Chinese labs' Qwen, DeepSeek, and GLM.
DeepSeek is the case the thread leans on, and it deserves a precise telling. The lab released its V4 model in April 2026 with open weights, and Huawei announced same-day support on its Ascend chips. Whether the largest version was actually trained end to end on Chinese silicon is contested, with reporting suggesting the flagship leaned on Nvidia hardware while only a smaller variant trained on Huawei. If you want the hands-on view of what it takes to run that model yourself, we covered it in our DeepSeek V4-Flash hardware reality check.
The structural point survives whatever the training details turn out to be. Open weights are the reader's leverage regardless of who produced them, because a model file can be downloaded, inspected, and kept. A closed cloud model can change its price, its terms, or its availability tomorrow. A file on your own disk cannot be revoked. Even the policy lever the thread debates is a moving target: after banning the sale of Nvidia's H200 chips to China, the United States approved limited sales under a new rule in January 2026, and those sales then stalled. Hardware policy shifts with the political wind; the weights on your drive do not.
What owning your stack actually buys you, and what it does not
Start with the honest limit, because the angle depends on it. You will not run frontier models at home. GLM-5.1 needs roughly eight H200-class GPUs to serve, and DeepSeek's largest V4 is a datacenter animal. If you need the most capable reasoning available every single day, keep paying for it. Local AI is a hedge, not a one-for-one replacement.
What you can run on real hardware has improved dramatically. Gemma 4 scales down to a Raspberry Pi; capable models in the 24 to 32GB class run well on a used RTX 3090 or a unified-memory mini PC; and sparse mixture-of-experts models fit comfortably on an ordinary 16 to 32GB machine. Our guide to the best local models by VRAM maps which model fits which machine, and our local AI hardware guide covers the buying decisions by budget.
What that setup actually buys is the thing the cloud cannot sell you: privacy, because nothing leaves your network; cost stability, because there is no per-query meter; and independence from a supply chain you cannot see. For the documents and ideas that made you want private AI in the first place, that is the entire point. It is the same logic as building a knowledge base you own outright rather than renting one.
A pragmatic playbook for the next two years
- Do not panic-buy hardware into the worst memory prices in years. Size your purchase to the model you will actually run, not the one you might someday want.
- Start free. Install Ollama, pull Gemma 4, and see what the machine you already own can do before spending anything.
- If you do buy, match the box to the job: a mini PC for always-on chat and document work, a used 24GB GPU for faster local inference, or unified memory for the largest models that will fit on one machine.
- Harden it. A local model with a web interface is a server on your network. Keep it bound to localhost, run it inside a container, put it on its own VLAN, and monitor outbound DNS so nothing phones home unnoticed.
- Stay hybrid. Use the cloud for frontier work when you genuinely need it, and keep a local default for anything you would not want logged, reviewed, or repriced later.
Frequently Asked Questions
Will AI really run out of power?
ai-power-demand-vs-states-2026.pngNot in a literal switch-off sense, but the projections describe demand growing faster than the grid can add supply, which is why analysts warn that some regions may have to turn data centers away. Whether the constraint bites hard or gets engineered around with new generation and efficiency gains is genuinely uncertain. For a consumer, the structural risk of depending on that infrastructure holds either way.
Does the AI boom actually make my RAM and router more expensive?
Yes, and the link is direct. AI accelerators use high-bandwidth memory that consumes far more fabrication capacity per gigabyte than standard memory, so chipmakers fill datacenter orders first and consumer supply tightens. DRAM prices rose more than 171 percent year-over-year heading into 2026, and router makers have passed those costs on, as we cover in our router price analysis.
Is "open source AI" a China thing?
No. Open-weight models come from many places, including Google, Meta, Mistral in France, and Microsoft, alongside Chinese labs. The meaningful distinction is open versus closed, not one country versus another.
Can I run something like DeepSeek V4 or GLM-5.1 at home?
Mostly no. The flagship versions of these models are datacenter-class and will not fit on consumer hardware. The encouraging part is that the open models which do fit a normal machine have improved sharply, and our model-by-VRAM guide shows what actually runs at each memory tier.
What should I actually buy for local AI right now?
Start with nothing: try Ollama on the computer you already own. If you outgrow it, match the hardware to the models you run rather than future-proofing into today's inflated memory prices. Our hardware guide and mini PC guide break down the options by budget.
Will my cloud AI get more expensive or rate-limited?
It is plausible. When the providers you rent from are themselves constrained on power and compute, those constraints can surface as higher prices, tighter rate limits, or waitlists. That possibility is the core of the argument for keeping a local option open, even if the cloud stays your default for the hardest tasks.
