Loading...
Loading...

How does artificial intelligence become more sustainable? Richard chats with Darshna Shah about her experiences as Chief AI Officer at Elastic. Darshna discusses the resources required to run large language model experiments, the pressure this has put on the electrical grid, and more. The conversation turns to efficiency - and the idea that there is little incentive in the current land grab around LLMs to be more efficient - that the focus is on growing quickly. But as technologies mature, efficiency becomes a key competitive advantage!
Links
Recorded January 29, 2026
From RunAsRadio.com, you're listening to RunAsRadio.
The Internet Audio Talk Show for IT Professionals with Richard Campbell.
This is Brandon Wen announcing show 1028, Sustainable AI with guest Darshan Ashaw, recorded
Thursday, January 29th, 2026.
RunAsRadio is produced each week by Sound Thoughts LLC.
For more information, visit SoundThoughtsLC.com.
You can follow us on X at x.com slash RunAsRadio.
Hi, this is Richard Campbell.
Thanks for listening to RunAsRadio.
I'm here at NDC London.
It's kind of a nice day, but we're down at the Queenie Center, the conference center
here and enjoying the show.
I have an opportunity to sit down with Darshan Ashaw, who is a chief AI officer at Elastac
Cloud, and she leads a strategic development of AI accelerators across finance, legal
energy, sustainability, and security.
Darshan Ashaw also has a broad range in consultancy experience in which she developed and deployed
a wide range of machine learning models on Azure across a myriad of industries, which
has a PhD in neuroscience, specializing in neuronal network dynamics in drug-resistant
temporal lobe epilepsy, for which her pasture for data and AI stemmed.
That's a cool gig, so dealing with a particular class of epilepsy and things you could do
to help treat it?
Yeah, absolutely.
That remains one of the highlights of my career.
So I actually started out in psychology and cognitive neuroscience, and I was really
fascinated with helping people with things like depression, Alzheimer's disease, various
different rehabilitation programmes, and quite early on I got some experience with the
brain injury rehabilitation unit, or there were a charity called Headway, and that experience
was really eye-opening for me, it was really emotionally rewarding, but it lacked some
of the intellectual stimulation that I was looking for, and I got really kind of addicted
to looking into try and find causation as to how these different diseases developed over
time and how we could potentially treat them.
But a lot of the research that I was embedded in was very correlational.
So I was like, no, I want to make a difference.
I want to be able to kind of propel a lot of these innovative advances.
So I decided to go back to hard science, and I was looking at research particularly around
Alzheimer's disease, and also epilepsy.
So my PhD research was actually looking at mechanisms of drug-resistant epilepsy, as you pointed
out.
So I did that in Birmingham, and it was really fascinating from the perspective of, so we
were doing a lot of simulation work, a lot of animal work, but also working very closely
with the Birmingham Children's Hospital, and they had patients where they were having
like 10 to 20 seizures a day.
And the drugs are not working for them.
And the drugs are not working for them.
So the only way that they could get any kind of quality of life was to resect out the
brain tissue.
That's so scary.
Yeah.
I've even heard of like, severing the campus colosum, like separating the two parts of the
brain, because that's where the cross-firing is having with how you have this disassociated
mind.
Yeah.
And none of these things sound like good ideas, but you're just desperate to give this
child some quality of life.
Exactly.
And we obviously wanted to try and find some answers for them, as to what happened within
the neural tissue, and what could we kind of infer from a lot of kind of chemical experimentations.
So we would go into surgery.
We would have like a bucket of artificial cerebral spinal fluid that had been oxygenated.
We would take the tissue, obviously with consent, and take it back to the lab, where we would
slice the tissue and preserve it with a lot of different kind of chemicals that would
try and replicate what we have in human bodies.
So we're not trying to impact any more damage.
And then we would implant electrodes in that brain tissue and run various different
experimentations from wholesale patching, which was just looking at certain cells and
exactly the transfer of different ions between the cell, as you kind of activated it, with
different electrical pulses, to entire network activity.
So we would do things like try and manipulate the excitability of that tissue.
Right.
Because this was overexcited tissue, that's why I was exercise, this is the one that kept
misfiring.
Exactly.
And we wanted to find out exactly what happened to that network stability that was lost.
That we could actually infer why some of these drugs were not working.
So one of my other PhD lab partners was actually looking at cannabinoids and as the disease
kind of develops at which stage would certain drugs be the most effective to actually have an
impact, which was quite interesting when we look at kind of the developmental life cycle.
So in humans, typically a lot of epilepsy is either caused by kind of blunt force trauma,
so like your car accidents being hit over the head, et cetera, or by certain cancers.
And obviously the two are very, very different one, you're still kind of dealing with somewhat
normal tissue in the latter, you're kind of dealing with all different types of changes
to the neuronal tissue.
So it was really fascinating.
And I was actually, my PhD thesis was actually looking at how does long-term memory and the
repeated patterns of excitability in that neuronal tissue actually lead to that excitability
balance being mismanaged.
So yeah, really fascinating.
And machine learning came into play with this?
Yeah, so I always had a fascination with statistics.
I really liked a lot of the maths around how you would objectively and reliably say that
this is something that is statistically significant.
Right.
But when you start looking at neuronal network dynamics, that suddenly becomes very, very
complicated.
And in this case, you're talking about actual neurons, that's a neural net, not the synthetic
ones we build in software.
No, exactly.
It gives a moment here.
I think you're going to see a neural net about a neural net.
That's crazy.
Well, that's exactly where we were going really.
So I didn't know anything around a lot of the computational side of things that was going
on until I was doing my PhD.
And I was like, okay, so we were having issues where we were already, I wasn't involved
in cloud in any way or anything like that.
And obviously, you can imagine we're implanting multiple electrodes in this tissue.
So the amount of data that it would produce was so huge and we didn't have any cloud storage
at that time.
So we were basically doing things like mapping down when we would do certain manipulations
to the tissue and taking snapshots of this electrical activity within the experiment
at certain times and saying, okay, at this point, I added this particular drug.
This is how the activity changed.
So I started thinking I was like, okay, I was looking into ways that we could do this
better.
And that's when I started to get really, really fascinated in data science.
And at that time, I think it was over a decade ago now.
So it was being positioned as the kind of sexiest job of the 21st century.
Right.
This is the data scientist period.
Yeah, exactly.
So I was fascinated by it and I was like, what is this?
So I started looking at the time I was looking at R because we were using SPSS.
I was not a code or a tool.
So I was fascinated in the math side, had a very good foundation on statistics.
And I was like, okay, so how can I do this at scale, maybe in a bit more of an automated
way, with more robust capabilities of adding more data into the mix, that I can get an accurate
true picture of what's happening here.
So I started dabbling in a little bit of Python and R at the time.
And we were at a careers forever.
And my friend goes, oh, you've been talking about data science because I was obsessed.
I was running workshops with our lab around statistics and how we could do better data analysis
with the data that we had, as well as keep on top of the research and all of the lab work.
She's like, you're really fascinated.
I've come across this bootcamp.
Why don't you take a look?
Because I knew I didn't want to stay in academia.
Right.
I loved it.
The innovations were too incremental and it was also quite isolating.
I wanted the research that I was doing to have real world impact and I wanted to see
it fast.
Data analysts can get very esoteric, if you want, at this whole modeling, how many people
who survive on the Titanic, they're almost cliche.
Yeah, exactly.
So I went ahead and managed to get a place on this bootcamp, which was called Science
to Data Science.
And they were specifically taking academics and giving them real world exposure to how to
use machine learning.
So we had two or three weeks of boot camps on economics, coding, marketing, strategy,
all kinds of things.
And then the final three weeks, so five weeks, we were working with a finance company called
InvestTech.
Well, my particular group was anyway.
And we were asked to basically explore the asset managers and understand if there were
any patterns that could lead to better renewals of contracts and things like that.
So we took a whole unsupervised learning approach to the problem.
And within a few weeks, we went from analysing the data, cleaning it, modelling it and relaying
back the analysis to the business, which they could use.
And from that point, I was absolutely hooked.
I was like, this is the pace that I need.
I love that I'm working with such multi-faceted groups of people from business people, what
are very analytical people and very technical people.
So I was like, I want to do this.
And I didn't necessarily like, I loved neuroscience and I would have loved to have stayed in that
field.
But I was also willing to kind of broaden my horizons a little bit.
Sure.
And there was so much opportunity in that space.
Like, you were also doing machine learning before it was really that cool.
Yeah.
Like, it's not the madness that's going on now, but it was an emergent time.
The tools had gotten better and the generative AI models were really quite powerful.
We don't think much about the fact that today, you take a picture with your phone and
it can tell you what's in the picture, but that was impossible 15 years ago.
Yeah.
And now it's just built into every phone.
That's part of the generative AI models that have worked really well.
Yeah.
LLM's a little more challenging, but clearly we're in that part where the statistical models
matter.
You know, the error levels and the probabilities matter.
We seem to be hiding that more these days.
It's concerning.
Yeah.
I think it's astonishing the kind of record pace at which we've seen advances from machine
learning to generative AI.
And I think I was in a little bit of shock.
I think it's most people when GPT arrived on our doorstep, but I think the kind of pace
at which we've adopted that and a lot of the kind of assumptions we make around how it
should perform are kind of edging us back towards maybe we should look at some of the practices
we did have when we were doing machine learning around the quality of the data, like how much
confidence do we actually need and in the determinism of the outputs that we're having?
Sure.
Because I think that's a kind of key starting point.
Yeah.
And the fact that we've basically concealed that from the operator at this point to me is
someone who again was in that space before all this.
It's like, why don't I know what the confidence level is on this return?
Like that disturbs me.
Yeah.
And you can see it when you show maybe people that are consuming the models that you create
in the solutions.
When you show them that confidence score, it changes their entire interpretation.
For sure.
So now I'm telling my teams and I think obviously UI and AI being more and more closely
gathered together as we kind of consume it on this mass scale.
But just those little things around what you show people can really change the kind of
perspective around how much they're willing to trust it as well.
Which I was reading the, I believe it was a Harvard study about the less you knew about
these tools, the more you trusted them, that as soon as you understood them well, you
trusted them less.
Which basically, we are concealing important information.
We need to educate people.
They need to be careful with these tools and they're really not being given the things
they need to get to that point quickly.
It's crazy because again, I've been in this long enough to remember when people
use that excuse, well, the computer said so it must be true.
It's like, it's only as good as the data we put in it.
And now we're using these non-deterministic models to produce results, which are even
more questionable.
Yeah.
Exactly.
Oh.
Nice, definitely.
You're doing a talk around sustainable AI.
How does that, what do you even mean when that, that question that's presented to you?
What is sustainable about it?
So I think obviously the scale that we're using generative AI is pretty incredible.
So environmental impact?
Environmental impact.
So I've been interested in energy and sustainability for quite a while.
So I've been in machine learning and AI for the last decade.
And I would say a significant proportion of that has been working either with solar
farm asset managers trying to increase their yield using machine learning within ESG
intelligence providing.
So that was a big thing kind of around, probably just as the pandemic was starting around 2020,
we saw this big explosion of these really ambitious targets being set by a lot of enterprise
organizations around this is a critical challenge that we face need to decarbonize.
But I think one of the things that kind of frustrated me at that point was it's not realistic.
Right.
Like it's a very...
Folks called it greenwashing.
Yeah, exactly.
Exactly.
All of the news articles one after another were decarbonization, but it's kind of conflicted
with this marketing ploy around.
Right.
You're not actually...
If you actually look at the operations of the business, you've concealed all this information.
And then we saw a lot of court cases around people actually getting fined for greenwashing.
And I was like, I understand that Butter is obviously a key critical challenge for humanity
as a whole.
And I'll panic it.
We have to be realistic around how we're going to achieve this.
And if we then bring in the social and the governance side of things, I think it becomes
very messy.
Like our societies and our economic structures as a whole are based on capitalistic outputs.
So if it's not going to lead to profit and revenue, the kind of attention that you're
going to get from leaders of businesses is going to be very different.
Yeah.
So we are beholden to the shareholder value equation.
Unless you can paint a picture of shareholder value, you can have a tough time getting
results from it.
Exactly.
So I think that one of the things that has recently reignited my optimism for sustainability
is the narrative shift.
So now a lot of energy and decarbonization infrastructure is being funded by a lot of
AI demand and development.
It is an interesting handed-hand effect.
We've had this emergence of these new tools that demand an awful lot of compute.
Data centers were struggling for power before 22, like it was already an issue.
It's just exacerbated the whole issue, but these tech giants have enough money.
Exactly.
Power plants only cost so much.
I just think they're not realistic on the timeline.
It takes a few years to build a data center.
It takes even more years to build a power plant.
Yeah.
Not to be, but I don't know, obviously this is very circular in its economy.
We drive the demand of how we're consuming AI, which will then derive the demand of the
infrastructure upgrades that need doing in order to power data centers, etc.
There is that circular aspect to it.
But I think the fact that it's being tied to something that has economic value is a little
bit more reassuring because I feel like the attention and the focus that we will have
on it is slightly different to maybe the perspective we had before around, let's just move
towards decarbonisation because we need to, but it will have these effects.
So I think that positioning has given me a fresh perspective on things.
Whether it's realistic and how that's determined over time remains to be seen, I think.
Yeah, I feel like there's a lot of plans being presented right now that probably will
never come to pass because we are in a bubble and this rate of spend has to taper off at
some point.
Eventually people run out of money.
Yeah.
You know, these are very big companies on awful lot of money, but they sure are spending
it right now and they're not making as much as they're spending, like you don't think
you can keep doing that.
But if we learn a bunch of things along the way, I hate this end justifies the means part,
that it is pushing forward some thinking around how power could be produced.
Yeah.
Exactly.
For better or worse.
And we should take a break for some very important messages.
Reimagined what's possible for your career, your team, your business, and for work itself
at the Microsoft 365 Community Conference, it built 21st to 23rd, 2026 at the Low Sapphire
Falls Resort in Orlando.
Back with hundreds of Microsoft executives, subject matter experts and product makers
to unlock solutions to accelerate your AI transformation.
The Microsoft 365 Community Conference is your opportunity to spark inspiration, build
connections, and deepen your skills.
Feature speakers at the Microsoft 365 Community Conference include Jeff Teeper, President
of Microsoft Collaborative Apps and Platforms, Adam Harmads, Vice President of Product,
Caroline Agatima, Director of M365 Customer Advocacy, and Heather Cook, Principal PM Manager
of Global Community Evangelism for M365.
Let's just a run as radio can get a discount on registration by using the code run as at M365.com.
The Microsoft 365 Community Conference is at the Low Sapphire Falls Resort in Orlando,
April 21 to 23rd, 2026, go to M365.com to register.
We're back, it's run as radio amateur camel.
I'm sitting here with Darshnah Shah talking a little bit about sustainable AI.
Are there other angles besides just the power consumption that concerns you for sustainability?
I'm thinking more around kind of how we as developers are thinking about design of AI.
So at the minute, we are kind of pushing more and more towards performance, which I think
is obviously naturally where we would be.
And have been for decades, right?
Everybody wanted stuff to be faster.
Every time you could shave some time off anything, it was better.
Exactly.
And I think that's a great starting point, but if we look at kind of how we can make things
not just performance, well, actually, yes, just performance.
If we look at performance, a lot of the kind of design principles actually overlap
with how we can also create much more energy efficient solutions.
I mean, I guess there's a presumed correlation between if it's more
performance, it will consume less.
I mean, one would argue that LLM model turns it on its head where the most quote unquote
effective, we can argue that that truth LLMs are also the largest and most energy consuming.
Although I think deep seek may punch some holes in that like where it's never been true
before.
I don't know why they try and make that argument right now.
You should be more efficient deep seek specifically or well, these things sort of came on the
scene with this idea of you don't have to have that much compute to have a usable large
language model.
That's a mixture of expert time architecture.
Yeah.
And again, how much of this is true?
Like, how do you benchmark that versus that's how they positioned what they made?
I mean, it's certainly been an ongoing conversation for me about why do you think GPT-5 is better
than GPT-4?
Like, how do you measure that?
It's just because the numbers bigger or just because they've told us it's a better
version.
Oh, okay.
I see.
And then saying thing with a deep seek where it's like, oh, this model runs in a smaller footprint
and needs less GPU to run.
Yeah.
Is it better?
Like, what is better when you talk about these sort of non-deterministic models?
I mean, some tests, the thing that I find very challenging about the tests for, okay,
this does this test better than that one is these are already curve-footing tools.
So they'll get very good at running tests more efficiently.
It doesn't mean they're more useful in the field.
Yeah, I think the whole evaluation, um, very challenging.
Yeah, it's very challenging, um, also a big area of fascination for me.
Sure.
I mean, coming from the machine learning side of things where you had very statistical
outputs, um, to now trying to intermix that is quite interesting in how that's evolving.
Um, so part of the talk I gave here at NDC yesterday was covering some of that.
Like, how do we actually measure the, the usefulness of machine learning, sorry, of these
large language models?
And I really, I've tested a few different tools out there.
So there's rag-ass, deep evil.
I really like the comprehensive nature of the Azure evaluation SDK because they have
kind of brought a few different categories of evaluation metrics.
So you have your typical LLM as a judge type metrics.
Right.
So this is where LLM is basically marking their own homework to certain degree.
We give it the generated output.
I love that description, marking your own homework.
Yeah.
Yeah.
Um, so you, they will rate things at the generated output from zero to five, um, for things
like groundedness, coherence, fluency, relevance, et cetera.
And then we have, um, the NLP-based metrics.
So these are things like your F1 score, Rouge and Blarge scores.
And this is like, if you're from machine learning, you're really empathized with a lot
of the, the kind of precision at which these metrics can give you.
So F1 score ranging from zero to one, the closer to one you are, the better performing
it is.
Rouge and Blarge scores are looking specifically at the unigrams that are generated
from your LLM and how quickly that would match ground truth.
So you would need to put in a bit of effort into, okay, this is the request I put into
the LLM.
This is the answer I'm expecting, how close is it?
Right.
And it's literally looking at the word for word output.
Nice.
Semantics can get a little bit interesting there, because if we're looking at the unigram
specifically, words can change meanings of things very dramatically.
And that's not encaptured in those metrics.
Yeah.
And you're not expecting word-for-word repeatability out of LLM anyway.
Yeah.
And I think if anybody was looking at that kind of area of precision, we probably don't
want to be looking at generative AI.
And we probably, if you do still want to use generative AI for kind of unstructured
output, then the branding and the voice and the tone of the output is really important.
So you probably want to look more towards fine-tuning something.
Yeah.
And that's a whole different area within itself.
And you can go from one extreme to another in terms of complexity of approaches to fine-tuning.
Then obviously we've got the safety and the ethics and kind of metrics.
Which also comes back to the data sources and the quality and I think that's the biggest
concern when we're running these LLM's is they're pre-computed models that we really
don't know the sources of and for the most part, these organizations are not willing
to share.
Yeah.
Because they keep getting sued when they do.
Yeah.
And if you're using something like a rag type architecture, you want to test like, is there
any protected material that's being divulged maybe where you don't want to?
Because everybody has universal access or do we actually need to look at what permissions
does somebody who's asking something of this solution have?
What should we be servicing back to them?
Likewise if there's any kind of indirect attacks that could be propagated through certain
prompts, you want to kind of start testing all of that.
And moving on to that specifically, I think the way that we can kind of test solutions
is also evolving quite quickly.
We have these kind of standard metrics, but now there's a lot of simulations that this
idea that's been borrowed from robotics around how we could maybe use agents that maybe
have a bit more of a naive perspective around how the features were actually coded, etc.
The tools, etc.
And actually start testing that as maybe a user would and we can get a much holistic
kind of overview of how does this actually perform before we go ahead and scale it and
open it up to 100 people to start testing.
And all of this to get to the point of what is the footprint of this tool?
So now we can consider the sustainability effects that it's going to have.
Yeah, so I started looking into this because GitHub's Green Software Foundation have a whole
bunch of packages that they have suggested can be helpful and they most definitely can.
So things like code carbon, eco to AI is looking and I really like them because if you
were from a machine learning perspective, you can talk about the complexity of the algorithm,
the hardware required to train that model, the volume of data, etc.
But when we move to this model around, we're not training the model, we're consuming a model.
The metrics that we need are slightly different and some of them we don't actually control.
So we can measure things like latency, tokens used, throughput, etc.
But there are other things that we don't control.
So when that model is running on hardware, that we don't control and we don't have insight
into how it's being deployed, we have to rely on certain metrics that are divulged by
the cloud providers.
So things like power usage effectiveness, the water usage effectiveness, which are typical.
I'll call it carbon output per transaction, that kind of thing.
So there's a few different papers that I've started looking at when we start using these outputs.
What kind of metrics do we see?
I guess this is another area of debate.
It depends on who you are because if you're a sustainability officer,
you're interested in the entire life cycle from the hardware to training,
to deployment to use, right?
And of life cycle because you want to report on your scope one, two, and three.
But if you're a developer, you're probably only interested on the bits that are
controllable for you, which is around how it's being used, so the deployment.
And if we look at the scale of the emissions and water usage associated with different points
in that life cycle, a few different papers have highlighted, that's quite different, right?
So we automatically assume a lot of hardware is required for training and
keeping these models active, which is true.
So I think, particularly, I think this is from about 2023.
This paper was looking at, based on a 105 billion requests of GPT,
and they said it was around 1.4 million for the hardware, around 1.4 million tons of CO2,
around 6,229 tons for the AO training.
But an astonishing 45 million tons for consumption, so based on our use.
So the lowest energy and carbon emitting part of it,
ends up being the largest amount of consumption because we do so much more of it?
Exactly.
But you're always saying the training was the downside.
That was the expensive and consumptive part.
But now if you actually use these things, that's going to overwhelm.
Which I guess is in order, that does make sense to me.
Yeah, and it's still a problem.
Yeah, absolutely.
And I was looking at things like, okay, well, if I'm making a GPT request relative to
using TikTok and Netflix, how much is that?
And yes, there is a substantial difference there,
like making the GPT request is much lower in terms of the energy required.
But like you say, if we're thinking about the scale at which this is being adopted,
then those numbers are slightly disproportionate.
Yeah, become an issue.
Yeah, so I think that's where I'm kind of advocating for
using these design principles and implementing them within our blueprints a little bit
more fundamentally because it's not just it will improve our performance and profitability,
but it also have much better impact for the planet.
Right.
So I think that's where we're kind of moving.
That there were also, I think, fundamentally as a kind of broader construct,
the whole idea of sustainability within tech is not a good versus bad scenario.
It's really complex.
And I see a lot of good in terms of the investments in infrastructure,
a lot of the new technologies that are evolving.
So one of the recent interesting stories I was reading was around a spin-out out of UCL,
who managed to secure 10 million pounds in funding to use light instead of electricity
for data transmission.
So lights were already used for part of it anyway, but it needs the pulse electricity
to process it, and then it needs to be converted back into light.
Right.
To transmit.
Far from decay.
Exactly.
And they were basically saying if we take electricity out of the equation and use phototonics
to do the parts that electricity were doing,
could we just rely on light purely?
So they received another 17 million pounds in funding in order to kind of evolve
that and actually put it into practice.
So there were really interesting projects like that.
There was another one I was reading by somebody who had a PhD in genomics and was looking at
the topology and how we can use biodiversity to really improve the design of data centers as well,
which is quite interesting.
So there's a lot of cool advances going on that I think will obviously highlight some of the
the progresses we can make in this really complex challenge as well.
I have to wonder, I think we're going to continue to use the cloud to make models,
but there's really a push for local consumption.
And I think the footprints are going to get worse by doing that because it's sort of like the car,
you know, it's one thing to make the fuels.
I think to burn it in lots of different machines.
And when we switch to more of the electric model,
you only have one place to really deal with the pollution of making the electricity,
the electricity itself is less emitting.
And if we keep running in the cloud model, then we're paying for tokens,
but arguably those large data centers should be more efficient than if everybody's running their
own rack, i.e. 200s.
And however that power generated and that heat is mitigated and so forth, you know,
today large commercial data centers don't use air conditioning at all. They use
ambient heat and airflow to cool their machines.
So they consume less electricity that way, but they consume more water, where if we're building
out our own little racks for a compute load for a company, you're probably going to stick
an air conditioner on that and have a larger part of footprint as a consequence.
Because I kind of feel like we're going to end up with more local compute on a lot of these
things, but I'm also thinking from the sustainability point of view, that's going to actually be
harder to manage, absolutely. We don't have, obviously. I mean, there might be some kind of
generalized technologies that we could start adopting. But again, I think that we're already
seeing huge differences in companies that have capital to invest in making things more efficient
and performance versus the smaller companies. So I think that will continue to grow.
And the smaller companies are probably going to stick with the cloud because it's cost-effective
and it's all operating costs rather than capital expenditures. The bigger places,
perhaps, can have more choice. Maybe liquid cooling a bit of our solution, like I'm always pondering
that what those pieces are going to look like because they're pushed to local compute is big.
That also speaks to the data sovereignty issue where people don't want their data going into
the cloud or they want more control over it as well. So it's going to be a press there.
But they're also going to take on the sustainability issue. It's like if you're going to run the
workload, sustainability problem is yours. Yeah, absolutely. I'm kind of curious because you mentioned
there's a lot of money that is obviously being funneled into the development of data centers
and the upgrades into the power facilities that are needed to support data center running.
If the bubble was to burst and the money into the upgrades stopped flowing,
what do you think the net impact would be? Well, I think we're overspending right now,
and those additional costs are going to, at some point, the data center doesn't make sense.
It's too expensive. One of the things I did earlier this year with my sort of advice to
systems is this is a bad year to buy computers. It's better to buy extended warranties this year
because you keep machines around for five years and this is the year they'll be really expensive.
If you can wait a year until this eases off, it'll cost less to amortize over the period.
And I think this is the same thing that's going to happen to these data centers is you've created
such a demand. Everything is getting overpriced. These data centers don't make sense.
So I do think there's going to be a cooling off to some degree. I think they're deliberately
overprovisioning as much as they say demand is I. But stuff's also getting more efficient.
Like the latest reports out of OpenAI show they're spending less to provide the same service.
Like you would expect more optimization to happen. And I would argue we're not focusing on
optimization right now because we're in a bubble. Like the best thing that will come out of a bubble
is the bubble ending is going back and saying we need to be more efficient. We've been too busy
on the land graph and grow, grow, grow and bigger, bigger, bigger and not thinking about more efficient.
And the pressure to work with less will help us focus on efficiency.
That's really fascinating.
Yeah. I mean, it's, you could tell I've spent too much of my management to even think in terms of
if I cut your budget 10% and don't reduce your workload, I'm effectively forcing efficiency on you.
You will find a way to do this and be more efficient because you have less to spend.
That's essentially what we're talking about here is right now where the money is crazy flowing.
There's no incentive to be efficient. So we're having a worse sustainability outcome
because we're not putting the pressure on the right way except in the non-financial way.
It's unethical what you're doing with these data centers. Do better.
Well, like we already know that's a lesser extent than it's going to cost you a fortune not to be
efficient. And so, you know, one could make the argument. The best thing that could happen to
sustainability in AI is an end of the bubble. Take the money away. You have to be more efficient.
Right. What we don't know today at all, the thing that worries me the most is we don't know
actually what any of these workloads actually cost. Yeah. You know, I'm working with folks that have
multiple all you can eat agent accounts running huge amounts of their resources to do these
cool programming things that you're being wildly productive. Well, what does it actually cost?
Yeah. You know, a lot of the research I think that is
estimating that has so many assumptions baked into it because yeah.
No, and let's face it, the companies are incented. They're offering all you can eat for a reason.
They need more customers. The same reason we used to get unlimited bandwidth accounts and so forth.
They were all about trying to grab customers. It's part of the bubble.
So the bubble ending is going to have to end those things too. And we're going to get a realistic
accounting of this. I'm not going to say it'll completely self correct, but we will never really
correct this until we stop this bubble. That's such a brilliant perspective. So you think a lot of
the optimization will be driven once the bubble bursts. You started this conversation with
until you find the financial incentive. People tend not to do the right thing. The financial
incentive is financial constraint. So it gives me some optimism. It's going to be a rough ride.
There's no worries about it. There's a downside to going this fast is stopping is hard.
In some ways, I have an optimism just because I've seen this before. The web got way better after
the doc combo because we had to be efficient. We built better tools. We built better servers.
We got better at making websites. We thought about size and scale and so forth because we had to
pay for stuff. Right now where we're not paying for stuff, we're doing dumb things and being
inefficient. Yeah, we can afford to. I mean, I've been spending a lot of time with legal clients.
And the FOMO that is going on is astonishing. It's like several different practice groups,
all consuming, all indexing, like the same several different models from different providers,
indexing data over and over again. And it's like, well, you're all doing e-discovery. Why don't you
just come together and look at it holistically from one perspective. And they'll have disputes
within their own little tribes as well. So it's different. Yeah, my models better.
So I think that's quite funny. It's the only way to get them to actually play together to
constrain them. You only have a budget to do one. So now figure out how to do it together.
I think some of the kind of trends, I think it was the international energy agency that we're
looking at that. I don't know if you followed the story around Alpha Fold. Absolutely. Yeah, that was
brilliant. And the amazing transformation, well, even go back earlier to the different Go variants,
where the Alpha Go where it's just trained against itself was much more computationally efficient
than the original one that was trained on the historical games. Yeah. And so we literally are
developing more efficient training models as we're incentive to do so. And ultimately Alpha Fold
has, I would argue, done the most important thing in generative AI of the world. Those 200 million
protein fold solutions available public is like, it's my marquee for you cannot ignore this
technology. It is fundamentally change medicine forever. Yeah, absolutely. And when I was looking
into some of, I was like, great, this technology must have really accelerated our materials development.
Sure. But they were, the international energy agency were quoting things like only 0.01%
of new solar PV materials have been experimentally produced. And I'm like, we could use technology
like Alpha Fold to start increasing and accelerating the rate at which we're developing and designing
these materials. And then they went on to quote around only 1% of energy patterns are actually
using AI to further the kind of innovation around energy. And I think it goes back to this whole
thing around what you're talking about where we have constrained. Yeah. We'll actually start using
the technology we have at our disposal to make things more efficient. Well, and you know,
the administrators mantra has changed as good. You go first. Power generation is an inherently
conservative industry because it's not where people look for innovation. They expect reliability.
It's all that matters. They will inevitably be late movers. You have to prove this tech works
for them to put cycles into it because they, they, they, they work in single percentage points
of efficiency improvement. So they, they're not to place where you look for experimentation.
There are other areas in, in industry across that that where big returns matter and taking
chances produce pretty sufficient results. So it's one of the, I think the disconnects between
the concerns around power generation and the concerns around these latest technologies is
those thing, though the late technology stuff, the AI stuff is allowed to innovate because
the returns are so large. Yeah. But optimization and power plants is 1%. You can't afford to spend
money experimentally. You have to be very, very certain that you can make it improvement before
you're going to do it. It has to be well proven. There's anything we've learned about this tech.
It's everything but well proven so far. Yeah. Exactly. We can't, we don't have confidence in
the results. Yeah. So I'm not going to go too down on the power companies just because they're
conservative. They were designed to be. Yeah. That's what we want from them. But we are going to have
to have a meeting in the minds of some of this and some of them. Darshan, this is a really fun
conversation. Just who I think we've sort of explored a lot of problems in this space.
What are you thinking about next in terms of sustainability? I mean, I kind of feel like we're all
stuck until the bubble ends in some respect. I think for me, I'm focusing a lot on the performance.
I mean, like you say, a lot of confidence at the minute, the money's flowing. So it's not
too much of a problem. We can have 20 different sandboxes with lots of different experiments.
That's what's partoken. Yeah. That's not a bad place to focus. Exactly.
So I think I'm really invested in kind of understanding how we can tangibly and reliably
measure performance. That encapsulate not just statistical metrics, but also more of the domain
experts. Sort of that real world. Yeah. Yeah. If we can't measure, we can't make it better.
Exactly. Great place to focus. Thanks so much for coming on the show. Thank you so much for
having me, Rich. You bet. And we'll talk to you next time. I'm Ron Asradio.



