When the AGI (Artificial General Intelligence) era truly arrives, tools will begin to think for themselves.
As ChatGPT upgrades to the 6.0 Astra version, its capabilities are shifting from "answering questions" to "end-to-end execution."
Jensen Huang stated that this version requires 100,000 NVLink 72 chips for training (previously stated as 300,000), and proclaimed, "The AGI era has arrived."
What does AGI mean? It is more like breaking the fourth wall, shattering the physical boundaries between humans and tools.
In the future, if intelligent agents like AGI enter human production and life on a large scale, the demand for computing power will no longer be just periodic expansion, but will transform into a long-term, rigid, and highly fluctuating load increment for the power system.
This means that large model training is like a marathon that cannot lose power, while inference services ebb and flow like tides. The competitive narrative of AI is also shifting from "whose model is smarter" to "who can obtain power in a shorter time, at a lower cost, and more stably."
The development direction of AGI is also very clear. End-to-end execution turns inference from "peaks and valleys" into a 7x24 resident load. What the power system needs is not "average power," but "committed capacity + fast response + low-carbon attributes + bankability."
This is precisely the combined battlefield of energy storage, HVDC (High Voltage Direct Current), smart microgrids, nuclear/geothermal/gas turbines, and green power PPAs (Power Purchase Agreements). In the future, AIDC (AI Data Center) grid connection will likely shift from "applying for capacity" to "submitting a verifiable flexibility solution."
AIDC Is Becoming a "Super Load" in the Power System
For traditional IDCs (Internet Data Centers), stability is the most important.
AIDC, however, adds a new requirement beyond stability: the power system must keep up with the rapid changes in computing power loads.
On the one hand, AI training clusters have massive instantaneous power demands.
On the other hand, as inference services increase, the computing power load will fluctuate significantly with user requests, agent tasks, and business peaks.
This poses new requirements for the power supply system. In the past, the UPS (Uninterruptible Power Supply) in data centers mostly undertook the task of "lasting a few minutes after a power outage." But in the future, the value of energy storage in AIDC may further expand to three levels.
First, it is a backup power source in the traditional sense.
Once the power grid fails, the battery must quickly take over the load to ensure that the GPU cluster does not suffer from interrupted computing tasks, data loss, or even hardware risks due to instantaneous power loss.
Second, it is a power fluctuation buffer.
When the AI computing load changes rapidly, energy storage can absorb part of the instantaneous power changes, reducing the impact of large-scale computing clusters on the power grid.
Third, it is an energy management asset.
When AIDC integrates with photovoltaics, wind power, PPAs, and the electricity market, energy storage can perform functions such as peak shaving and valley filling, demand response, ancillary services, and renewable energy integration.
This means that AIDC energy storage is transforming from "backup power equipment" into "power infrastructure."
The International Energy Agency (IEA) also predicts that by the mid-2030s, there will not be just one answer to the energy mix for meeting the new power demands of data centers. Multiple power sources such as new energy, natural gas, and nuclear power will play a role, while energy storage and grid expansion will become crucial supporting components.
Is the power system ready?
The construction speed of AI data centers is increasingly approaching or even exceeding the construction cycle of traditional power infrastructure.
Building a data center may only take a few years, but constructing transmission lines, substations, generator sets, and new grid infrastructure also requires long-term planning and approval.
The IEA has explicitly pointed out that the rapid expansion of data centers is encountering bottlenecks in transformers, gas turbines, and grid connections.
Since building stations on the ground is time-consuming and labor-intensive, Elon Musk has started looking to space for power, aiming to build data centers in space.
This directly solves a problem: building data centers closer to energy sources is equivalent to a space-based version of direct green power connection.
Compared to Musk's visionary ideas, real-world AIDCs care more about practical questions: Where is the power source? How much capacity can the grid provide? Is there energy storage? What is the proportion of new energy? Can a long-term PPA be obtained? Who will back it up when grid fluctuations occur?
This is also why the integration between AIDC and the energy industry will become increasingly deep in the future.
Energy Storage Ushers in New Business Opportunities
As AI upgrades to AGI, what AIDC needs is not simply "more power." It requires committed capacity + fast response + high reliability + low-carbon power + long-term bankability.
Energy storage sits exactly at the intersection of these demands.
Especially with the gradual development of high-voltage direct current architectures such as 800V HVDC, data center power systems are evolving from traditional UPS systems to architectures with higher power density and higher efficiency.
This also means that the evaluation criteria for batteries are changing.
In the past, energy storage batteries competed on capacity, cycle life, and cost.
AIDC batteries, however, need to further compete on C-rate, safety, consistency, response speed, lifespan, thermal management, BMS (Battery Management System), and system-level reliability.
Notably, AIDC is not a low-threshold market for unbranded players, but a market where leading cell manufacturers continue to widen the gap through safety, cycle life, consistency, and delivery capabilities.
Viewed in tiers, the first tier consists of "cell + system + scenario" all-rounders.
CATL is the most aggressive. While promoting 300+Ah sodium-ion batteries, 587Ah large cells, and TENER-class high-safety energy storage systems, it is also forming a closed loop of "energy storage cells, power electronics, energy dispatch, power supply and distribution, and computing power scenarios" through equity investments/cooperations with Damao Technology, Zhongheng Electric, and 21Vianet.
The second tier consists of dark horses in professional energy storage cells and systems.
On the cell side, it depends on product definition capabilities: Hithium quickly rose to the top with large cells, overseas capacity, and breakthroughs in North America; EVE Energy binds with leading integrators, leading the pace with large cells like 628Ah; BYD controls costs through vertical integration, Blade Batteries, and overseas projects; REPT Battero, CALB, Gotion High-Tech, Envision AESC, Pylon Tech, and Sunwoda are grabbing orders in different segmented price bands.
In addition to the lithium-ion battery system, sodium-ion batteries are also expected to get a share of this market with advantages such as "high C-rate and long-term cost." Hithium, EVE Energy, REPT Battero, and Envision AESC have all launched AIDC-specific sodium-ion batteries, while new forces like HiNa Battery, Highstar, and Zhongna Energy continue to follow up.
On the system side, it depends on the ability to "understand the machine room": Sungrow is strong in PCS (Power Conversion System) and global large-scale energy storage delivery; Kehua Data, Vertiv, and Zhongheng Electric are strong in UPS/HVDC/micro-modules; Narada Power, Sacred Sun, and Shoto Technology have existing channels in communication and data center backup power, as well as overseas operators and internet customers.
A deeper judgment is that behind AIDC energy storage, it is not just about simply selling containers, but understanding grid connection indicators, fire protection standards, SLAs (Service Level Agreements), carbon accounting, and dispatch interfaces.
Future orders will increasingly appear in a composite model of "capacity leasing + ancillary service sharing + green power trading commissions." Manufacturers that simply sell equipment will see their profits squeezed, while players with full-stack capabilities of "cell - system - dispatch - trading" will gain pricing power.
In the short term, HVDC renovation and backup power storage demand will see volume growth first, benefiting system-side manufacturers;
In the medium term, the increased penetration rate of large cells and sodium-ion batteries in AIDC scenarios will benefit cell manufacturers with product definition capabilities;
In the long term, the coupling of green power PPAs, virtual power plants, and the electricity spot market with computing power loads will give birth to a brand-new business model of "computing-power-electricity coordinated operation," and the asset and financial attributes of energy storage will be thoroughly activated.
Returning to the present, global computing centers are in an expansion period where "power is chasing computing power."
Overseas, hyperscale cloud service providers are competing to lock in power resources, and leading energy storage integrators are joining forces with power electronics giants and chip manufacturers to launch power-energy storage reference solutions adapted to the new generation of AI computing clusters;
Domestically, the "East Data West Computing" project, coupled with the accelerated construction of AIDC, means that the high fluctuation characteristics of computing power loads have turned energy storage from an "optional choice" into a "prerequisite for grid connection."
The hundreds of GWh-level long-term supply agreements recently signed by domestic battery companies and overseas leading integrators are the latest footnote to this trend.
The scale of a single order has exceeded the total shipment volume of leading battery factories in the past year. The signal it conveys is not about the victory or defeat of a single enterprise, but rather:
The procurement gate for global computing power and electricity infrastructure is opening. Relying on large cell technology, delivery capabilities, and cost advantages, China's energy storage industry chain has become an indispensable supply force in this construction boom.
From CATL's "cell + system + scenario" closed loop, to professional manufacturers like Hithium and EVE Energy binding leading customers with differentiated products. This confirms the previous judgment: the path for Chinese battery factories to enter the AIDC track is to bind with overseas leading integrators and leverage their platforms to go global and meet worldwide computing power and electricity infrastructure demands.
Facing the blue ocean of AIDC, the layout of Chinese battery enterprises is no longer about grabbing scattered orders, but a systematic overseas expansion with clear tiers and respective strategic positions.
Conclusion:
When the AGI era truly arrives, machines will begin to think for themselves, call tools by themselves, and execute tasks by themselves.
This means that the energy demand for AI will also change: training requires greater computing power, and inference requires more online computing power. Computing power demand will ultimately be mapped to the power system in some form.
AI determines the upper limit of computing power, while power determines the lower limit. What future AIDCs need is not just more batteries, but the ability to turn batteries into stable, fast, dispatchable, and verifiable power assets. For the energy storage industry, this is what truly deserves attention.