The AES Corporation is one of the largest suppliers of clean energy to the elite corner of the tech industry whose artificial intelligence ambitions are fueling load growth anxieties around the energy sector. To date, the company has more than seven gigawatts of renewables projects in operation and under construction for those customers.
U.S. chief commercial officer Kleber Costa said that that focus on hyperscaler customers dates back to the formation of AES Clean Energy, the company’s renewable energy development business.
When that business was created four years ago, there was an intentional focus on “large technology firms,” Costa said. And that focus impacted everything from the design of the organization to how the company structured and priced clean energy transactions.
In the last year, however, there have been a lot of changes in what AES’ largest tech customers are asking for, Costa said. The biggest one? Project timelines.
“The change we have seen here is that there’s a bigger appetite for projects that have earlier [commercial operation dates],” he said. “The other change is a bigger sense of urgency, particularly in the way some of these customers are approaching utilities, trying to partner with utilities in addition to renewable developers.”
The AI layer
Despite AES’ broad experience with the needs of these customers, there isn’t exactly a playbook for the structure or setup of a renewables project, Costa said.
Though the company spends a lot of time standardizing development and operation to target lower levelized cost of energy, he added, “we also spend a lot of time customizing the individual solutions for each customer hoping to capture a premium, because we are getting closer to something the customer values more than just a simple project.”
That said, renewables plus solar is the “clear winner” in the near-term, Costa said, at least until clean, firm sources like advanced nuclear and geothermal come into play.
It’s a model AES has been honing for the last five years. And it’s one that Costa said is ideally equipped to leverage an artificial intelligence layer; he pointed to a recent partnership with Amazon, on a solar and battery project in Southern California, as a “very successful” example.
“We oversize the storage and capture value within the solar hours,”he said.
The idea is that the system of four-hour duration batteries charge during the four cheapest hours in the CAISO market, and discharge during the four most expensive hours. But optimizing that set-up takes AI, Costa said.
Fluence, the solutions provider on the California project, estimates that the machine learning software will analyze billions of data points every year to determine when to buy, sell, and store energy.
Costa said AES is anticipating more projects that use this trio of technologies are coming. But, he warned, they face significant — and growing — challenges.
Predicting data center needs
Data center developers are facing “exactly the same issues” as everyone else in the energy sector at the moment, Costa said, citing interconnection and permitting challenges in particular.
And the difficulty in getting a project permitted or interconnected has meant that AES customers have had to be more flexible on where they’re built.
“I personally don’t see a lot of behind-the-meter, microgrid, isolated-from-the-grid projects,” Costa said. “But I do see a potential for more geographical flexibility in where they site new data centers.”
However, he is skeptical that some of the more creative data center solutions — like higher latency requirements — will scale: “Some would say that the latency requirements for the large language model training are not as stringent as for the traditional cloud computing business,” he said.
But at the end of the day, how flexible a data center can be will likely come down to price.
“Data center operators do not like to be flexible in the same way that if you ask a power plant operator, they don’t like to ramp their plants up and down,” Costa said.
Data center operators are looking to earn a return on their very high capex costs, he added. It’s probably cheaper and more efficient to run a data center 24/7, which will slow down the spread of dispatchable data centers. But that doesn’t mean that there aren’t certain contexts where flexibility will be necessary.
“For us to meet the hundreds of gigawatts of demand that we’re seeing in the next ten years, I think we’ll need a combination of all of the above,” Costa said. ”We will need a little bit of flexibility.”
