Navigating the rapid integration of artificial intelligence requires addressing complex questions about business, education, national defense, and long-term human survival.

During a panel discussion at the Next Endeavor conference held in Mountain View, California this September—convened by Imagination in Action—Meta representatives Lisa Rivalin and Richa Mishra shared insights into how their organization is pioneering advancements in this critical sector.

The Dual Meaning of Sustainability

Rivalin opened the discussion by unpacking the dual layers defining sustainability within the context of artificial intelligence deployments.

There is a double meaning of sustainable here, meaning something that lasts long, and also reducing the impact on the environment, she noted. Then there is using AI to solve environmental problems, and sustainable AI, which is reducing the environmental footprint of the AI hardware itself.

Mishra introduced a compelling physical metaphor to illustrate the scale of modern infrastructure:

I want you to imagine a room of almost this size, which houses hundreds of thousands of metal bookshelves, which we call racks. They have flatboards, which are circuit boards housing all of these specialized GPUs, CPUs, you know, all of the chips that we’re either planning to buy, or we wish to have more of. And all of these are sitting in widespread data centers across the world, and they’re all connected within them, as we see, (with) the fiberoptics growing within these data centers. We have massive amounts of heat that is getting dissipated, getting extracted out of these places, and at the same time, there’s also massive amounts of power … across all of these sites. Our biggest problem that we see right now is the discipline across all of these operations to come together, and power, reliability, and carbon are three of the major pillars that are helping us essentially enable this layer of AI.

End users remain largely shielded from the intricacies of these massive operations.

As infrastructure engineers, our biggest success is being invisible, Mishra emphasized. And the only time you realize we’re there is when something fails.

Addressing the energy landscape, Rivalin pointed out that surging demand has brought these resource challenges to the forefront.

It is more visible now because of the accelerating demand, she explained. So there is an elephant in the room. If we need to build more data centers, we have more needs in terms of hardware.

Supply chain limitations compound the difficulty of acquiring necessary semiconductors.

Five years ago, we didn’t really have a good read on where exactly our carbon footprint was sitting, Rivalin remarked. We realized that semiconductors, because of the chemicals that are actually dissipated during the process and also the energy-intensive manufacturing, are what is causing us to be very, very carbon-intensive. So we’re working a lot with the semiconductor industry, and we actually have a call for action to everyone working in the semiconductor industry to think about that, and rework (things in) a way that we can be more sustainable.

A lot of agents are doing the work for humans, and humans are making decisions, Rivalin continued, detailing operational workflows. But before we actually had to reach this state, we had to go through a set of principles.

Key principles highlighted included rigorous verification, navigating domain ambiguity, and establishing clear context through specific practices:

  • Please keep your logic and your code outside of the template, outside of the agent, and point your agent to deterministic code.
  • Most of the time, you want a validation process and you want a human gate, and it’s not easy. We’ve seen a lot of examples where a lot of things are being automated and pushed into the fleet, and as you can imagine, at our scale, it can actually create a lot of issues.

Mishra later addressed the synergy between human domain experts and automated architectural components.

The biggest success happens only when developers are brightly coupled with the domain experts of any field, she stated, noting that component utilization rarely hits theoretical maximums. Sometimes, as we heard in the previous talk, we could be running at as low as 3% to 5% utilization. That is actually happening. Our job in production is to be able to surface some of these insights and be able to optimize, build modular solutions with the help of agents and software solutions, where we’re able to take advantage of power budgets on a component level.

Discussing electrical grids, she noted the delicate balance required to handle sudden computational spikes without disrupting regional stability.

Early Warnings and Paths to Success

Mishra also highlighted the operational threat posed by silent data corruption:

What essentially happens is a single chip is degrading faster than a neighboring chip that would be operating in a different condition. And all of these cases are going to be cascading failures for any large model that is going to run at scale. These are the kinds of things that we are trying to find spatial and temporal signals for.

Tribal knowledge accumulates organically through hardware design iterations, driving the need for better diagnostic accuracy to maintain fleet health and performance across generations.

Rivalin underscored the importance of cross-industry cooperation in establishing standardized carbon-tracking metrics:

We have many collaborations across the industry to create a new way to calculate hardware carbon footprint and standardize it across the industry. We’re also in contact with regulators and standards bodies to help us really define that. Then we have created a suite of software and agents to calculate the carbon footprint for every single component that we have on the fleet, including when we have to repair, swap components, etc.

By integrating carbon tracking directly into proprietary rack designs, Meta evaluates environmental efficiency from initial concept through end-of-life recycling, prioritizing component reuse to significantly lower overall emissions.

Mishra concluded with an invitation for broader academic engagement, noting that university partnerships remain vital as the tech industry navigates its next major computational frontier.