Google Funds 96 MW Nuclear Uprate, $900M Customer Payoff

Can Google squeeze 96 megawatts from two aging nuclear plants without raising your bill? That is the test in Georgia this morning, where Google agreed to underwrite more output from Hatch and Vogtle in a deal that could mean about $900 million in customer benefits. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Tuesday, September 22, 2026. Let's get into it.

Bloomberg reports Alibaba unveiled a new AI chip, the Zhenwu V900, alongside a target of over 20 GW of Alibaba Cloud capacity by 2032. Now, for you if you're new here, that 20 GW number describes the electricity appetite for a whole fleet of giant data centers, and we're treating it exactly as Bloomberg framed it, an announced target for 2032, not contracted or under construction, and not energized racks you could touch today. And that matters, because an announced gigawatt is not a built one, and who pays to power it, and with what chips, is still open. So what would make this real? We can't say yet from performance, because this one came to us as a gist without the full article fetched, so the specifics we can stand on are just the chip name and that over-20 GW by 2032 target. The skeptic's flag travels with it too, which is that the chip performance claim floating around this announcement is company-claimed, without an independent spec in what we saw today, and the target has no timeline breakdown or counterparty attached. My read is file it as intent, not capacity, and the thing that would move it is the supply timeline, cluster scale, and a site with power. When those show up, we'll chase the megawatts to the money.

Nvidia on the other hand gave builders something concrete to shop with. According to the Nvidia Newsroom blog posted September 21, the company launched Nvidia DSX Ready, a qualification program for partner products that meet applicable Nvidia DSX AI factory reference design requirements, which is just Nvidia's blueprint for how compute, power, cooling and software are supposed to fit together. Additional categories across infrastructure and software will be rolled out over time, in their words. Now, what does qualified actually mean? For cooling, they use a self-qualification suite. A qualified CDU may meet the relevant cooling criteria, for example, while the builder still evaluates how it will fit the planned facility. Reaction was quick on X. As @Tesla_Megapack put it on X, quote, "Megapack excels in off-grid and on-grid applications, supplying large data center loads, buffering oscillating demand, helping interconnection grid-compliance and strengthening the grid - all with less local fossil generation Megapack is @NVIDIAAIInfra DSX Ready" end quote. And as @LG_ensol put it on X, quote, "We're proud to launch alongside NVIDIA AI Infrastructure as an inaugural NVIDIA DSX Ready partner." end quote. So why does this matter to you? Because time to power, getting the transformers, the interconnection, the cooling right, is what slows data centers down, and a shared checklist could cut integration risk, if builders actually use it.

Staying on time to power, Utility Dive reports the Public Utility Commission of Texas voted Friday to adopt new large load interconnection standards, which are the rules for how giant users plug into the grid. I should be clear here, this one is single-source to Utility Dive in what we could fetch, and this show has not independently confirmed the docket text, so I'm giving you their reporting. Under what Utility Dive describes, the final rule set a flat $100,000 study fee for all large load customers, regardless of load size, and eliminated a March proposal for a non-refundable interconnection fee of $50,000 per MW of contracted peak demand. As @UtilityDive put it on X, "Texas PUC adopts softened rules on data center interconnection." Now, the other softening matters more for your queue watch. The final rule also softened the PUCT's proposed standards for when the interconnecting DSP or TSP "must notify ERCOT of the large load customer's non-utilized capacity" if the customer missed scheduled energization milestones, meaning the dates they promised to actually take power. And on money held as security, Utility Dive reported in its article that within 60 days of providing that notice, the grid company must apply the financial security to amounts owed, then return the balance, a shift from a proposal that would have kept most of it. The context Utility Dive gives is a state audit of ERCOT's 474 GW interconnection queue, approximately 90 percent of which is data centers. So the entry ticket just got cheaper, while the audit decides what's real.

Our main story today I am calling Georgia's pay-your-own-way nuclear test. Electric utility Georgia Power and tech giant Google said the companies have an agreement under which Google will support power uprates at two of the utility's nuclear power plants, according to POWER Magazine reporting on September 21. If you're new, an uprate is not a new reactor, it's squeezing more electricity from a reactor you already have by swapping turbines, pumps, motors and cooling gear so it can run hotter and push more steam to the generator. The deal would add some 96 MW of power generation capacity to Georgia's power grid, and it must be approved by the Georgia Public Service Commission, the state's utility regulator. That approval point is everything, because without it, this is a signed commercial handshake waiting on a regulatory yes. The mechanics, as POWER Magazine reported Georgia Power filed Monday with the state PSC, include a new nuclear uprate tariff structure, basically a special rate schedule, plus a request to approve an extended power uprate, an EPU, for Units 1 and 2 at Hatch, located in Baxley about 200 miles southeast of Atlanta. An EPU for Units 1 and 2 at Vogtle, located near Waynesboro, just more than 150 miles east of Atlanta, was approved by the PSC as part of Georgia Power's 2025 Integrated Resource Plan. So one plant still needs the uprate approval, the other already has that piece, and both need the new tariff to work. The money figure that stopped me cold was about $900 million in benefits for its customers over the life of the units, which the utility said on Monday the deal could generate.

Now, how does the money actually flow, and who gets what? Google, under a subscription-based program, will subscribe to the new NU-1 tariff and receive zero-emission credits, which represent the carbon-free attributes of nuclear energy associated with the power generated by the nuclear unit uprates. In plain English, Google pays through a dedicated subscription, and in return gets the paperwork that says this chunk of carbon-free power is theirs to count toward clean goals. Georgia Power, which serves 2.8-million customers in Georgia, said Hatch and Vogtle provide about a quarter of the electricity used in the state, and the plants are managed by Southern Nuclear, a sister company owned by Southern Co. that specializes in nuclear operations. The utility also spelled out that EPUs increase output through modifications to equipment such as turbines, pumps, motors, and cooling systems, enabling reactors to operate at higher licensed thermal power levels while continuing to utilize the existing reactor fuel design and plant infrastructure. Here's what the two sides said they think this proves. Aaron Mitchell, senior vice president of Strategic Growth for Georgia Power, said quote, "Many of our large customers have specific clean energy goals, and we continue to work with the Georgia PSC to create programs that not only help them meet their goals, but align with our Customer Protection Pledge, enhancing the reliability and resiliency of the power grid and creating savings for all customers." He went on, "We appreciate Google's continued leadership in this space and partnership in developing this program, which we see as the latest example of how large-load growth can benefit all electric customers." Lucia Tian, director, Advanced Energy Technologies, at Google, said quote, "Google data centers are long-term investments in the communities we call home, catalyzing local economic development and accelerating clean energy deployment. They also serve as a proof point for how we can unlock the significant opportunity to bring online new nuclear power through expanding the capacity of the existing nuclear fleet," end quote. She continued that this collaboration ensures digital infrastructure growth goes hand-in-hand with positive local impact, helping to deliver hundreds of millions of dollars in direct customer benefits and supporting a more affordable, clean, and reliable energy system for all Georgia residents. Here's my read: the hyperscaler underwrites firm megawatts from steel that's already in the ground, instead of asking households to socialize the bet. But does that $900 million survive contact with regulators? The caveat in the file is that benefits are projected over life of the units, not upfront cash, and the whole MW figure is Georgia Power's share only and pending approval. What would make this credible to me is the PSC order itself, the approved NU-1 terms, who holds stranded-cost risk if loads don't show, and an energization date, not just a subscription. And for context on Google's appetite, Google will invest at least €13 billion ($15.1 billion) in Finnish AI infrastructure over 2027 and 2028 and has signed a 22-year power purchase agreement with Fortum for up to 50% of the output from the Loviisa nuclear plant, Google's first nuclear power contract outside the United States. Time for the Hype Check. I give this a 7 out of 10 on substance, in my opinion, because 96 MW is contracted uprate capacity with a named payer and a named tariff, not a new-build fantasy, but it loses points until Georgia regulators say yes and we see metered megawatts. What I'm watching next is that PSC decision on NU-1 and Hatch, which will prove whether pay-your-own-way nuclear can clear in months instead of years.

If today's episode helped you see where your power bill meets the AI boom, follow the show wherever you listen so you don't miss tomorrow's PSC and queue fallout. This has been Concrete Compute, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care!

I also host Quickly Quantum: a daily quantum computing briefing you don't need a physics degree to follow. The breakthroughs, the funding rounds, and how much substance is really under each claim. Find it wherever you get your podcasts.