AI-native energy companies are emerging as a "new species" in Silicon Valley
Since entering 2026, the anxiety in the tech circle has been spreading from models, codes, and chips to the deeper layers of the AI industrial chain. The industry has started to discuss a more fundamental question: Is there enough electricity to support all this AI computing power?
At the just - concluded NVIDIA GTC Taipei conference on June 1st, Jensen Huang introduced NVIDIA DSX, the third - generation MGX cabinet architecture, and an 800VDC power supply solution. He redesigned computing, networking, storage, power, cooling, and control systems as a whole, aiming to redefine the "power system engineering" inside AI factories.
What NVIDIA wants to do is to optimize the system efficiency inside AI factories, coordinate computing, networking, power, and cooling, and increase the token output per watt of electricity.
Meanwhile, the power supply outside AI factories is becoming a new bottleneck: Where to build data centers, where there is still available connection capacity, whether projects can be connected to the grid as soon as possible, and how to adjust the load according to grid pressure after completion.
Against this background and driven by these needs, a group of "AI - native energy companies" have gradually emerged.
Interestingly, although they are energy companies, they don't build power plants or lay power lines. They only rely on codes and algorithms to try to redefine the flow, price, and rhythm of electricity.
The capital market has also started to re - price such companies.
In May 2026, Sutter Hill Ventures, a Silicon Valley venture capital firm that was an early investor in NVIDIA, co - led a $64 million Series A financing round for a company named "GridCARE" with well - known investor John Doerr.
GridCARE applies AI to power connection and energy dispatching, helping AI factories quickly identify available power resources, complete connection planning, and participate in subsequent load dispatching.
In the past, the imagination space of energy technology companies mainly came from new energy, energy storage, batteries, and grid equipment. But after the explosion of AI computing power demand, those who can help data centers find, connect to, and make good use of electricity faster may become a key link in the AI infrastructure chain.
Companies similar to GridCARE have also started to emerge in areas such as Silicon Valley in the United States.
Emerald AI, a company headquartered in Washington, USA, has raised approximately $68 million in total over 16 months. Behind it stand NVIDIA NVentures, Energy Impact Partners, and power giants such as Eaton, Siemens, and GE Vernova. Jeff Dean and Fei - Fei Li also participated in the investment as individuals.
In May 2026, Shatterdome Energy, founded by entrepreneur Amann Shariff with a quantitative trading background, also completed a $3.5 million Pre - Seed round of financing.
Most of these companies are targeting the most restricted part of AI infrastructure: "finding electricity" in the power grid, determining where there is still available capacity and where connection can be made faster to shorten the grid - connection queuing period; adjusting computing power tasks and using electricity during off - peak hours when the grid is under stress; and conducting real - time power trading and dispatching for new energy, energy storage, and large industrial users using AI.
The rise of these companies has important reference significance for the industry.
In the energy competition of the AI era, it's not just about building more power plants and laying more lines, but also about how to more efficiently organize new energy, energy storage, the power grid, and computing power loads. In the future, those who can find, connect to, and dispatch electricity faster may gain a more favorable position in the competition of AI infrastructure.
Beyond chips and computing power, electricity is becoming the new speed limit for AI systems. And the power industry itself is being rewritten by AI.
The picture is generated by AI
01 Power Anxiety in the AI Era: It's Not a Lack of Electricity, but "Available Electricity"
On the surface, the power anxiety in the AI era is a "lack of electricity," but in essence, it's a "lack of available electricity." Many power resources do exist, but they haven't been fully identified, dispatched, and delivered.
The report "Roadmap: AI Data Center Stack" released by Bessemer Venture Partners, a Silicon Valley venture capital firm, in May 2026 presented a set of figures: As of early 2026, 190 gigawatts of hyperscale data center projects were announced globally, but only 12 gigawatts were actually put into operation, 21 gigawatts were under construction, and the remaining 148 gigawatts were still on paper. More than a quarter of the projects planned to be launched in 2025 were stuck in the power supply and licensing stages.
A research report released by Stanford University in December 2025 also pointed out that the utilization rate of the US power grid is less than one - third most of the time. GridCARE, a smart grid operation and maintenance enterprise, provided more specific figures: Even in areas with the most severe power shortages, the actual utilization rate of the power grid is less than 32%. There is no shortage of electricity; the shortage lies in the ability to deliver it.
Amit Narayan, the co - founder and CEO of GridCARE, named this phenomenon the "Time - to - Energize Crisis," referring to the time lag of several years between power demand and actual power supply. A large amount of existing grid capacity cannot be utilized due to the limitations of traditional dispatching and grid - connection processes.
He once described the current situation by saying, "The current AI frenzy has gotten out of control to the extent that people think sending chips into space might be faster than finding electricity on Earth."
There are huge business opportunities hidden behind this bottleneck. According to GridCARE's calculations, every 1 gigawatt of power connected to the grid ahead of schedule can unlock $25 billion in value.
Sutter Hill Ventures, the lead investor, was one of the early investors in NVIDIA and can be said to have participated in the rise of the "computing era" from start to finish. Vic Miller, the managing director of this firm, publicly stated, "A year ago, few people talked about power as a bottleneck for AI. But today, it has become an insurmountable hurdle for the entire industry."
John Doerr, a follow - on investor, was also an early investor in Amazon and Google. When explaining his investment logic, he simply said, "GridCARE provides affordable and sustainable energy by releasing the idle power in the grid we have built."
GridCARE has launched a "power acceleration" software. Its core technology uses AI to simulate and analyze billions of operating states of the power grid in real - time, including line congestion, power outage risks, weather changes, and demand fluctuations. Then it identifies the idle power and sends it to where it is needed.
Currently, this model has achieved its first successful case. GridCARE is collaborating with Portland General Electric to release more than 400 megawatts of grid capacity in Hillsboro, Oregon, which is sufficient to support the connection of six data centers. Among them, the first batch of 80 megawatts is expected to be put into use in 2026.
02 From Finding Electricity to Dispatching Electricity: Teaching AI Factories to "Use Electricity During Off - Peak Hours"
GridCARE focuses on the power grid side, trying to dig out more available connection capacity from the existing power transmission and distribution system.
There are also energy startups that focus on the software layer but have completely different entry points.
A company named "Emerald AI" is exploring turning AI data centers into dispatchable grid assets, enabling data centers to adjust their power consumption rhythm according to the grid situation. For example, when the grid is under high pressure, some AI tasks can be temporarily slowed down, postponed, or transferred to other regions for operation. After the grid pressure eases, they can resume higher - load operations.
The underlying logic here is that AI factories don't need to operate at full capacity all the time. Model training tasks can be paused and then resumed, and batch inference tasks can be moved to other areas. As long as data centers can actively reduce power consumption according to grid instructions, the grid pressure will be much reduced, and there is no need to invest in building new lines for peak loads.
Emerald AI has launched a product called the "Conductor" platform, which is like installing a "flexible and adaptable" brain for data centers.
Its function is similar to an intelligent valve installed between the power grid and data centers. When the grid is under stress, the platform receives a signal and immediately reduces the power consumption of facilities while ensuring that key AI tasks running on NVIDIA GPUs are not affected.
At COMPUTEX Taipei, Emerald AI announced a cooperation with NVIDIA and Silicon Valley Power, launching the first commercial multi - megawatt project in Silicon Valley.
The starting point of this project is the "Flexible Load Interconnection Program" promoted by Silicon Valley Power. The core of this program is actually to solve the problem of the long queuing time for data centers to connect to the grid.
Shivaram commented on this, "Silicon Valley Power's 'Flexible Load Interconnection Program' has proven that the regulatory path is feasible. NVIDIA's DSX OS, DSX Flex, and our Conductor platform have brought this technical solution to the commercial scale."
03 From Single - Point Dispatching to Platform - Based: The AI - Upgraded "Virtual Power Plant"
Compared with GridCARE and Emerald AI, the ambition of the AI energy company Grid AI seems even greater.
Grid AI wants to use a unified AI platform to connect all dispersed power resources, from the air - conditioners in a single household to the backup power supply of an AI data center, all under dispatch.
They have divided this idea into three levels for implementation.
The first category is ordinary households and small businesses. AI automatically manages devices such as air - conditioners, electric vehicles, and batteries in the background, helping users consume more electricity when the electricity price is low and less when the price is high or the grid is under stress.
The second category is commercial and utility scenarios, where assets such as energy storage, electric vehicle fleets, and distributed power sources are unified for dispatch and participate in the power market trading.
The third category is AI data centers and large industrial parks. By coordinating power generation, energy storage, and load, these high - energy - consuming facilities can use more stable and cheaper electricity.
To some extent, Grid AI is essentially building an "AI - version virtual power plant." Traditional virtual power plants gather many "small power sources, small batteries, and small loads" to relieve the grid pressure. Grid AI expands the boundaries of virtual power plants to AI data centers and large industrial parks, creating an AI energy dispatch platform that covers households, commercial, utility, and hyperscale power - consumption scenarios.
In addition to optimizing the power grid and load, AI has also started to enter the trading section of the power market.
The US AI energy trading service provider Shatterdome Energy positions itself as the "financial infrastructure layer" in the energy world.
In the past, a rooftop solar panel, a wind turbine, and a set of energy - storage batteries were just scattered power - generation equipment. But in Shatterdome Energy's system, they can be packaged into a tradable energy asset. The platform decides when to sell electricity, when to store it, and when to use trading tools to hedge price risks based on electricity price fluctuations, weather changes, power - generation forecasts, and market demand.
Shatterdome Energy's AI tools focus on the subtle signals in the power market that are difficult for human traders to detect in a timely manner. For example, when a certain line suddenly becomes congested, the power - generation speed in a certain area cannot keep up with the demand, or the electricity price at a certain node is about to experience abnormal fluctuations. Algorithms can make judgments as soon as these changes occur and complete transactions faster than humans.
As the proportion of new energy increases, the power market is becoming more and more difficult to predict: Weather can affect the output of wind and solar power, data centers can suddenly increase the load, and local grid congestion can also cause rapid price differentiation in different regions. For power companies, inaccurate predictions and slow dispatching can directly lead to fines and trading losses.
After the entry of AI, energy trading has started to become more like a high - frequency game. In addition to helping enterprises "save electricity costs," it also needs to help power companies more accurately predict supply and demand, respond more quickly to price changes, and reduce losses caused by misjudgments.
A survey by the technology service company Digiqt in September 2025 showed that AI traders are rapidly penetrating the energy market. They have brought real changes: A medium - sized power company used to lose 50,000 to 150,000 euros per month in imbalance fines due to inaccurate predictions. After connecting to AI, this part of the loss has been reduced by 15% to 30%.
04 "Flexible Load": A New Solution to the Problem of Connecting AI Factories to the Power Grid
Start - ups have told many stories, but what are the real effects? Can AI data centers really "obey the grid's instructions"?
In March 2026, an experiment provided an answer.
The UK National Grid, NVIDIA, Emerald AI, and the Electric Power Research Institute (EPRI) jointly conducted a test: After the grid sent a signal, within about one minute, the data center at the London site reduced its power consumption by one - third. More importantly, the AI tasks running on NVIDIA GPUs did not interrupt.
Another test lasted for ten hours. The data center kept its power consumption at about 10% for a long time, and the workload was not affected either.
These two results show that AI data centers are not just "rigid objects" that always consume electricity at full capacity. They can also act as an adjustable load and give way when the grid is under stress.
If operators can prove that they can actively reduce the load when the grid is under stress, the grid does not need to expand capacity according to the theoretical maximum. In this way, the construction pressure on the grid can be reduced, and the waiting time for data centers to connect to the grid can also be shortened.
This is the significance of the London experiment: Although it is a preliminary experiment, it has proven that at least on the side of AI data centers, "flexible response" is a verifiable ability.
05 Conclusion: Software is Redefining the Power Layer
Whether it's GridCARE's "seam - filling" dispatching in the crowded power grid, Emerald AI teaching data centers to use electricity during off - peak hours, or Shatterdome Energy participating in power trading with algorithms, they all point to the same trend: In the AI era, we not only need more electricity but also need to use and dispatch it more effectively.
These AI - native energy companies haven't built a single power station or erected a single high - voltage line. But the software layer they have created is becoming an important part of the power grid system.
This also echoes the "AI five - layer cake" framework previously proposed by Jensen Huang: Energy is at the bottom layer, followed by chips, infrastructure, models, and applications. Without continuous, stable, and dispatchable power, even the most powerful chips and models cannot truly run.
This may be a profound transformation in the AI era: The huge power grid, born in the industrial era, is being reassembled by lines of code.
Ultimately, those who have smarter algorithms will hold the key to driving AI civilization.
This article is from "Tencent Technology", author: Helen Li, editor: Qingyang Xu, published by 36Kr with authorization.