What Are We Actually Consuming?

A vegetarian approach to AI, environmental cost, and the question of how much is enough.

By Brittany Martin, President of Chrysalis

For about five years, from 2012-2017, I was a vegetarian. I read Jonathan Safran Foer’s Eating Animals after a vegetarian friend prompted me to do so, and I quickly learned enough about the environmental impact of industrial meat production that I could no longer quite enjoy a burger the way I once did.

I still liked meat. In fact, I remember dreaming about burgers; to the point that I would wake up in a cold sweat thinking I had accidentally actually eaten a greasy delicious burger.

Because I still ate honey and eggs, and I wasn’t very discerning about my leather consumption, I wasn’t classified as vegan. Which means there were plenty of people who could point to the omelette on my plate at brunch and explain the inconsistency in my personal environmental strategy, and they would have been correct.

But I wasn’t trying to achieve ecological purity; I had simply looked at something I consumed, understood more about what it cost to produce than I had before, and decided I wanted to participate differently.

Which makes it at least a little strange that I now help build and lead an AI company, especially as I’ve learned more about the environmental costs of the technology we’re building with. (But then again, I’m no longer a vegetarian—it destroyed my health—so maybe it’s not that strange. I live in shades of gray rather than white and black)

Chrysalis is basically taking a vegetarian approach to AI. We’re not vegan, and we haven’t opted out.

A metaphor, as an amuse-bouche…

I’ve started to think about the AI landscape almost like a dietary spectrum.

Vegans are anti-AI. As much as possible, they’re opting out on principle.

Vegetarians are AI moderates. They use it, sometimes a lot, but they don’t think every problem needs AI simply because AI is available.

And then there are theomnivores: AI all day. If a meeting can have an AI notetaker, add one. If the form can become a chatbot, make it a chatbot. If an agent can automate the task, automate it. More AI is generally presumed to be better.

Green AI and local models are a slightly different question; they’re more like caring about where the food came from. You can be an omnivore who buys locally sourced meat or a vegetarian eating an avocado that traveled 4,000 miles wrapped in plastic.

There’s the question of how much we consume, and then there’s the question of how that consumption is produced.

We're not anti-AI. We're also not putting AI in everything.

Part of what changed my relationship with meat was learning to see the infrastructure behind the thing on my plate. A burger stopped being just a burger; there was land, water, feed, transportation, processing, and an entire industrial system I had mostly never thought about because I didn’t have to see it.

I think AI has started to feel similar for a lot of us after the initial boom of 2023. The little box on our screen can feel almost weightless, but underneath it are data centers, electricity, cooling systems, water, hardware, and enormous physical infrastructure. The exact environmental cost of one prompt is surprisingly difficult to calculate because it changes dramatically depending on the model, data center, energy source, cooling system, location, and even time of day. But “hard to calculate” does not mean “free.”

Globally, data centers accounted for about 1.5% of electricity demand in 2024, and the International Energy Agency projects their electricity consumption could more than double by 2030, with AI as a major driver of that growth (International Energy Agency [IEA], 2025).

And then… there’s the water dilemma. Berkeley Lab researchers found that the water use associated with data-center workloads can vary by more than 10,000-fold depending on server efficiency, the electricity grid, cooling systems, climate, utilization, and where the workload runs (Lei et al., 2025).

The environmental cost of AI is being counted, though we don’t yet know the exact cost. And the exact cost of your particular use of AI is much harder to turn into a simple number.

I’m less interested in finding the perfect meme-able statistic about how many bottles of water a ChatGPT conversation consumed and more interested in understanding what we're actually consuming when we use these tools... and then deciding when that consumption is worth it.

And for some people, the answer is no, full stop. I understand that.

For me, the answer isn’t no. I use AI nearly every day. We build technology that works with large language models. I have experienced the very real capacity these tools can create, particularly for small businesses and people doing work that would otherwise require substantially more time, money, or human labor. I also think increasing human capacity is not automatically a neutral good. I want to know what that additional capacity is being used for, what it costs to create, and whether the value on the other side actually justifies it.

But I also don’t think the alternative to abstinence has to be AI all day, everywhere, for everything.

Because once you recognize there is a material cost underneath every seemingly weightless interaction, another question becomes available:

What is the appropriate amount of AI for the actual problem?

Maybe the AI can just listen.

It can organize, track, help surface something for review, carry information forward, or prepare the person with actual competency to enter the conversation. It doesn’t always need to respond, coach, interpret, generate another artifact, or pretend to be wise.

That particular brand of restraint has become increasingly important to how I think about what we’re building.

Looking Upstream

Earlier this year, I took a course on community health that asked us to look upstream.

The basic idea is that when people repeatedly arrive downstream with the same problem, there’s only so much value in teaching each individual person how to manage what keeps floating toward them. Eventually somebody has to walk upstream and ask why it keeps ending up in the water.

Personal responsibility can become a very convenient place to stop looking.

Use less plastic, manage your stress, eat better, make responsible choices with AI.

Sure.

And... what systems are making the harmful thing cheap, available, profitable, ubiquitous, or difficult to abstain from in the first place?

The important question to come back to is: am I thinking upstream enough? Am I looking at individual behavior while stopping short of the economic and political incentives producing the conditions around it?

Because the upstream incentive in technology right now is mostly more: more usage, more automation, more compute, more context, more places where AI can be inserted into an interaction because putting AI into the product is itself treated as evidence of progress.

Chrysalis is not somehow magically standing outside that system. We are building a company inside the same market incentives pushing the rest of technology toward more usage, more compute, and more automation, which means our own incentives matter, too.

I don’t want our business model to depend on convincing people to use as much AI as possible. I want it to depend on whether we solve a problem that matters.

So our vegetarian approach can’t just be, “Everyone should use AI responsibly.” It has to show up in what we design, what we measure, and what we reward.

The question I want us to keep asking is: what is the appropriate amount of AI for the actual problem?

Not the smallest amount because less is always better, but enough to do something genuinely useful for humanity, with a reason for every additional thing we ask the technology, or the person using it, to carry.

Health is where this began.

The foundation of Chrysalis sits in my husband, Brent’s, nearly thirty years of living with chronic illness.

People with complicated and invisible illnesses rarely suffer from a complete absence of information. They often have an absurd amount of it, scattered across labs, medications, imaging, specialist notes, portal accounts, treatments, and years of experience noticing what happens inside a body no one else lives in.

What’s missing is often where those disparate pieces begin to form a pattern across time, and where the information most relevant to the next conversation can actually become visible.

The patient is, at this point, the only person responsible for carrying their whole story across systems that each hold another piece of it.

And healthcare in and of itself is not a system that is bereft of harm, and we can’t assume more sophisticated technology will produce better outcomes. In a case study I did this year on the NAACP’s Health AI Equity Initiative, one of the things that became clearest to me was that AI can reproduce existing inequity without anyone explicitly programming inequity into it; if you build a model from systems where access, diagnosis, treatment, and whose symptoms are believed are already uneven, the model can make those existing patterns faster and larger (NAACP, 2025; Watkins et al., 2025).

Transparency without agency is not empowering.

It’s good for a patient to know AI is being used.

It’s considerably different for the patient to have meaningful control over what information is being used, what moves into another interaction, what remains private, and whether the representation of their own story is accurate.

And all of this together is why we built Chrysalis the way we did: around context the individual controls, selective sharing rather than automatic access, and technology that can support the work without requiring the human with actual expertise to surrender their judgment to the machine.

And now we get to test it.

Last week, Chrysalis was selected as one of 15 teams advancing into Phase Two of the HHS TOPx Tech Sprint for AI and Invisible Illness, out of 171 Phase One submissions.

Which means we now get to test this philosophy in one of the places where we think it matters most, alongside patients, practitioners, researchers, and others who understand the problem from very different vantage points.

One of the lessons I took from studying community health is that people affected by a system shouldn’t only become sources of information after somebody else has already decided what the intervention will be. Decisions about what gets built, what gets measured, and what counts as success are themselves forms of power.

So before we decide what role this technology should play in chronic illness, we need to listen.

What are people carrying manually today?

What gets lost?

What do you wish a new practitioner understood without requiring you to reconstruct fifteen years of your life in a waiting room?

Where could technology genuinely reduce the burden?

And where would you really prefer that we keep the robot out of it?

I think that may be the simplest articulation of our vegetarian approach to AI.

We’re going to use AI, we’re going to build with it, and we’re going to keep asking what it costs, what it’s actually solving, who has power inside the system, and whether more AI is really the answer simply because more AI is available.

Sometimes it will be.

Sometimes I hope we have enough conviction to leave it off the menu.

xo, Brittany

P.S. As part of Phase Two, we’re looking to talk with people living with chronic and invisible illness. If that’s you and you’d be willing to tell us about how you currently carry, organize, remember, and share your health story—and what you would or absolutely would not want technology doing with it—I’d love to hear from you. Explore what we’re doing a bit more​here​.

References

Foer, J. S. (2009). Eating animals. Little, Brown and Company.

International Energy Agency. (2025). Energy and AI.​ https://www.iea.org/reports/energy-and-ai​

Lei, N., Lu, J., Shehabi, A., & Masanet, E. R. (2025). The water use of data center workloads: A review and assessment of key determinants. Resources, Conservation and Recycling, 219, 108310. https://doi.org/10.1016/j.resconrec.2025.108310

NAACP. (2025, December 11). NAACP calls for equity-first approach to AI in healthcare, issues governance framework to build healthier futures.​ https://naacp.org/articles/naacp-calls-equity-first-approach-ai-healthcare-issues-governance-framework-build​

Stafford, K. (2025, December 11). Exclusive: NAACP pressing for “equity-first” AI standards in medicine. Reuters.

Watkins, S. C., Kammer-Kerwick, M., Turner Lee, N., Rathnasingham, R., Wilds, T., & Woolston, C. (2025, November). Building a healthier future: Designing AI for health equity. NAACP.