tech-ai September 18, 2026 By Blake Colman

Google, Nvidia & Anthropic’s 100GW Power Grab Is an AI Land Rush

AI is not running out of chips. It is running out of permission to plug them in. Google, Nvidia and Anthropic think they can unlock 100GW without waiting for new power plants.

AI is not running out of chips. It is running out of permission to plug them in.

That is the uncomfortable truth behind Google, Nvidia and Anthropic joining a new alliance that says flexible data centres could unlock 100 gigawatts of capacity from America’s existing power grid. If they are right, the next winners in AI will not merely be the firms with the best models. They will be the ones that get power first. ([techcrunch.com](https://techcrunch.com/2026/09/17/google-nvidia-and-anthropic-want-emerald-ai-to-find-space-on-the-grid-for-more-data-centers/?utm_source=openai))

The real AI bottleneck has moved from silicon to electricity

On September 16, Google, Nvidia and grid-software startup Emerald AI launched the AI Energy Management Alliance, or AEMA. Anthropic, AES, Constellation, National Grid, NRG Energy and a collection of other power and technology players joined the group. Its pitch is simple: data centres should stop behaving like entitled industrial toddlers demanding full power every second of every day. ([axios.com](https://www.axios.com/2026/09/16/tech-giants-launch-flexible-power-coalition-data-centers?utm_source=openai))

Instead, they should flex.

When the grid is under pressure — a brutal summer afternoon, say, when households are hammering air conditioners — a data centre could pause noncritical AI training jobs, shift workloads to another location, draw from batteries or use local generation. In return, utilities get breathing room at the exact moments the grid is most stressed.

That idea is not revolutionary. Factories have participated in demand-response programs for decades. What is new is the prize on the table: AI data centres are arriving with enormous loads, enormous urgency and very little patience for a five-year queue. AEMA says new data-centre projects can face five to seven years or more to connect in key US markets. ([aema.ai](https://www.aema.ai/home?utm_source=openai))

No founder with a real business plan wants to tell customers, investors and staff: “Great news, we have the GPUs. Bad news, we can’t turn them on until 2032.”

That is why this matters. The alliance is not some worthy little climate initiative. It is an attempt to turn grid flexibility into a commercial weapon.

100 gigawatts is a bloody big claim

AEMA says making data centres moderately flexible could unlock up to 100GW of capacity from the existing US electricity system — which it compares with enough power for 100 million homes. That is a headline-sized number, so keep your scepticism switched on: it is an industry coalition’s estimate, not money already in the bank or capacity already wired into a rack. ([aema.ai](https://www.aema.ai/home?utm_source=openai))

Still, the underlying logic is solid.

Power grids are built to survive the worst few hours of the year. That means a lot of infrastructure sits underused for much of the time. If large power users agree to reduce demand during those peak windows, utilities may be able to connect more load without immediately building new generation, transmission lines and substations.

TechCrunch reported that simply capping grid use at 90% for a few hours at a time could free up 76GW of capacity, citing a Goldman Sachs study. That does not make the 100GW figure guaranteed. It does show why everyone in this alliance is suddenly interested in treating compute jobs as movable rather than sacred. ([techcrunch.com](https://techcrunch.com/2026/09/17/google-nvidia-and-anthropic-want-emerald-ai-to-find-space-on-the-grid-for-more-data-centers/?utm_source=openai))

Emerald AI has raised $150 million in a Series A at a $1.05 billion valuation to sell precisely this proposition: software that coordinates a data centre’s power use with the needs of the grid. Its software is designed to reduce consumption during periods of stress while limiting the hit to AI workloads. ([emeraldai.co](https://www.emeraldai.co/news?utm_source=openai))

That is the commercial point. A grid operator does not care whether your model-training run is exciting. It cares whether you can get out of the way when the system is strained.

Why Google, Nvidia and Anthropic are really here

Let’s not pretend this is charity.

Google wants more cloud capacity. Nvidia wants more of its systems deployed. Anthropic wants more compute to train and serve increasingly expensive models. Utilities want large new customers without being blamed for blackouts or punishing household bills. Emerald AI wants to become the software layer sitting between giant compute buyers and the grid.

Everyone has a motive. That is generally a better sign than a press release full of good intentions.

Nvidia’s involvement is particularly telling. The company has spent years selling the picks and shovels of the AI boom. But chips only produce revenue once they are deployed, powered and fed work. A delayed data centre is not merely a construction problem for Nvidia; it is a delayed order cycle.

Google has already developed demand-response capabilities for its own operations. Emerald AI and Nvidia are also working on a 100-megawatt power-flexible AI factory in Manassas, Virginia, according to SiliconANGLE. That is important because the industry has had enough slide decks. What matters now is whether flexible operation works at real scale without wrecking performance commitments to customers. ([siliconangle.com](https://siliconangle.com/2026/09/17/nvidia-google-and-emerald-ai-launch-flexible-data-center-consortium/?utm_source=openai))

Anthropic’s presence matters for a different reason. Frontier-model companies consume compute in lumpy bursts. Training runs, fine-tuning, data processing and inference do not all carry the same urgency. If a company can classify work properly, it can protect the customer-facing jobs and move the less time-sensitive stuff around.

That is not an electricity trick. It is operational competence.

The overlooked angle: this is a political licence to build

The biggest mistake investors and founders can make is treating this as an energy story only.

It is a permission story.

Communities are increasingly asking a fair question: why should households pay more, face reliability risks or live beside massive industrial facilities so an AI company can generate more tokens? The answer cannot be “because AI is important, mate.” That will not survive a town-hall meeting.

Flexible data centres give developers a better answer: we will pay to be interruptible, use batteries and on-site power intelligently, reduce demand when the grid needs help, and potentially allow more infrastructure to be built without forcing everyone else to carry the cost.

Whether every operator follows through is another matter. But politically, it is much stronger than demanding permanent, round-the-clock access to scarce electricity and calling anyone who objects anti-innovation.

AEMA’s own materials are unusually candid about the policy dimension. It intends to work on interconnection, transmission and large-load policy, and to push standards for how flexible data centres operate. ([aema.ai](https://www.aema.ai/home?utm_source=openai))

That means the standard-setting battle starts now.

The company whose operating model becomes the benchmark will enjoy an advantage. The company that treats flexibility as a bolt-on afterthought may discover that it has built an expensive warehouse full of hardware with no fast path to usable power.

Here is the contrarian bit: flexibility will not save bad AI economics

There is a temptation to hear “100GW” and assume the power problem is solved. It is not.

Flexible demand can make better use of existing capacity. It cannot magically produce transformers, transmission lines, skilled electricians, gas turbines, nuclear plants or long-duration storage. It also cannot turn an AI product with rubbish margins into a good business.

If your model requires relentless, high-cost compute to produce a feature customers barely value, shaving power use during peak hours will not rescue you. You still have a bad business wearing a clever energy hat.

There is another catch. Flexibility is easiest when workloads are genuinely movable. A batch job, a model-training run or internal data processing can often wait. A real-time customer product, financial workflow or critical enterprise system may have far less room to pause. The value will go to operators that design their products, service levels and technical architecture around that distinction from day one.

That is why I would not bet blindly on “AI infrastructure” as a category. I would back businesses that can prove three things: they have contracted power, they can use it efficiently, and they have customers willing to pay enough for the output.

Power availability is becoming a moat. Power flexibility may become the moat inside the moat.

What this means for you

If you are a founder, stop asking only, “What model should we use?” Ask, “What happens to this business when inference costs rise, capacity is constrained or our cloud provider cannot serve us where we need it?” Build an architecture that can route work, delay noncritical jobs and measure compute cost by customer and feature.

If you run operations, treat electricity and compute as procurement disciplines, not engineering trivia. Know which workloads must be instant, which can run overnight and which should not run at all. Most businesses have more wasteful AI activity than they care to admit.

If you are investing, look past GPU counts. The sharper questions are: Who has power contracts? Who has interconnection certainty? Who can flex load? Who owns software that makes expensive infrastructure work harder? The next round of AI wealth will not all sit with model makers.

And if you are a saver watching this circus from the sidelines, remember the old rule: when everybody chases the shiny object, look for the bottleneck. Right now, the bottleneck is not intelligence. It is power, permits and the operators capable of turning both into revenue.

That is far less sexy than a chatbot demo.

It is also where the money usually gets made.