Nvidia has a new problem, and it can’t fix it with a faster GPU. The company is now working on how data centers get and use electricity, because power, not silicon, is what’s capping how fast AI infrastructure can grow. That’s the thread The Information pulls on in a new report on Nvidia’s push to solve the data center power bottleneck.
The Information’s reporting frames this as a strategic shift: the chipmaker that made its fortune selling compute is being forced to think like a utility engineer. And that tells you a lot about where the AI buildout is heading over the next two to three years.
Why power became the constraint
The math got out of hand fast. A single rack of Nvidia’s Blackwell-era systems already draws well over 100 kilowatts. The company’s own roadmap points toward racks in the 600 kilowatt range by 2027 with the Rubin Ultra generation. Multiply that across a campus and you’re talking about gigawatts, the scale of a nuclear plant, for a single AI site.
The grid wasn’t built for this. Utility interconnection queues in the U.S. run three to seven years in many regions. Transformers are backordered. And AI training workloads have a nasty habit of spiking and dropping in sync, which stresses grid equipment in ways steady industrial loads never did.
So Nvidia can ship all the GPUs it wants. If there’s nowhere to plug them in, revenue stalls. That’s the real reason this matters now.
What Nvidia is actually doing
Based on Nvidia’s public announcements over the past year, its approach breaks into three layers:
- Rethinking rack power delivery. Nvidia is pushing the industry toward 800-volt DC distribution inside data centers, replacing the maze of AC-to-DC conversions that waste energy and eat floor space. Fewer conversion steps mean less heat, thinner copper, and more room for compute.
- Smoothing the spikes. Newer systems include energy storage and power-smoothing hardware at the rack level, so the grid sees a flatter demand curve even when thousands of GPUs go from idle to full tilt in a millisecond.
- Making data centers grid-flexible. Nvidia has backed Emerald AI, a startup whose software throttles AI workloads during peak grid stress. A demonstration in Phoenix with Oracle and the Electric Power Research Institute cut a data center’s draw by roughly 25% for three hours without breaking the jobs running on it.
The third one is the most interesting. If a data center can promise a utility it will dial down on demand, it can get connected years sooner. That’s a business-model unlock, not just an engineering fix.
What stands out here
Nvidia doesn’t make transformers or run power plants. But it’s inserting itself into the power conversation because it has to. Every quarter that a hyperscaler waits for a substation is a quarter of delayed GPU orders. Nvidia’s incentive is to make its hardware the easiest thing on the grid to accommodate.
This also widens Nvidia’s moat. If the 800V DC standard, the power-smoothing hardware, and the grid-flexibility software all come tuned for Nvidia racks, AMD and custom silicon from Google or Amazon face one more integration hurdle. Power efficiency becomes a competitive feature, the same way networking did with NVLink and InfiniBand.
There’s a regulatory angle too. Utilities and state regulators are getting nervous about AI campuses raising rates for everyone else. A vendor that shows up with demand-response tools is a much easier story to sell to a public utility commission than one that just asks for more megawatts.
The Future Cast: what to watch through 2028
Expect the following to play out:
- Power becomes the headline spec. Watts per token will matter as much as FLOPS. Vendors will market efficiency the way they now market memory bandwidth.
- On-site generation goes mainstream. Gas turbines, fuel cells, and eventually small modular reactors next to data centers, because waiting on the grid is no longer an option.
- Flexible load becomes a contract term. Utilities will start requiring AI campuses to commit to curtailment windows in exchange for faster interconnection.
- Power infrastructure companies get pulled into the AI trade. Transformer makers, DC power specialists, and grid software firms become part of the AI supply chain conversation.
Practical takeaways
If you’re building on AI infrastructure, ask your cloud provider about power headroom, not just GPU availability. Capacity promises that ignore the grid are promises that will slip. If you’re evaluating AI hardware, put energy per workload in the comparison. And if you’re an investor or operator, the companies solving the electricity problem may end up with better margins than the ones just consuming it.
The Information’s full report has more on Nvidia’s specific plans and who it’s working with. The short version: the next phase of the AI race runs on electrons, and Nvidia knows it.