这是用户在 2025-4-4 12:12 为 https://app.tegus.co/app/documents/interview/125140?q=+What+do+you+think+it+looks+like+in+2026+and+2... 保存的双语快照页面,由 沉浸式翻译 提供双语支持。了解如何保存?
Mentions
Annotations
Chief Architect at Microsoft
Chief Architect at Microsoft
Topics
Summary
The AI infrastructure industry is shifting towards a more cautious approach to building data centers, with a focus on profitability and leasing options. Core Scientific is facing challenges with project timelines and finding non-CoreWeave customers, while also exploring nuclear power options. The conversation between the Tegus Client and the expert covered topics such as the demand for compute data, the use of GPUs for AI, and the impact of market volatility on decision-making. Overall, the industry is moving towards smarter strategies and partnerships for growth and profitability.
Expert Details Compliance
Consultant
147 Other Interviews on Tegus

Chief Architect at Microsoft The expert can speak to the current market for HPC data centers, converting bitcoin mining operations into HPC compute data centers, and the CoreWeave/Core Scientific deal.

Interview
Tegus Client
00:00:00
Thank
you
for taking the time to chat again about
and
the space. Look, there's a lot of noise
in
the,
what
I call, the AI infrastructure ecosystem about people signing leases, people canceling leases, people changing footprint. There's just a lot of debate, I
think
out there. Maybe just to kick
it
off at a high level, how are
you
thinking about the AI infra world today? Has that view changed
in
the past two months?
Chief Architect at Microsoft
00:00:25
Still very bullish, very optimistic on
it
. When
it
comes down to
it
, there were a couple projects that we ended up stopping. Our planned data centers builds are very slightly down from
what
we thought.
It
wasn't that we outright just decided that there was not enough demand or anything
like
that.
It
was more so that the projects would not meet the deadlines
and
the schedule that had been communicated to us by the vendors.
With a couple of those,
like
one out
in
Ellendale, North Dakota was supposed to be between a one GB to a five GW site. As
it
turns out, the vendor that we were working with just flat out said, "Hey, we're having issues. We don't
think
we can build at that pace," things
like
that.
Given the outstanding timeline
and
the amount of money
and
just the investments we'd already made, we decided to cancel that project until the future, at some point, whether or not they could get the schedule together or whether or not we could find someone who could execute on
it
faster.
The infrastructure demand is still there. Over the last three months though, we've been very, how to say this, our CFO asked us to go on through
and
make sure that as we build these data centers that they'll definitely be profitable as soon as we light them up. There's just so many components to data centers.
You
get construction,
you
get the land,
you
've got transformers,
and
all the equipment,
and
HVAC
and
perimeter security
and
stuff
like
that.
You
have to run with all the labor challenges there
and
then PPAs
and
IPPs that we work with
and
stuff
like
that.
With Amy wanting things to be profitable, some of the projects that we've had to look at to say, "Unless we back this
in
, unless we move some of this equipment unless we pull this forward, we're not going to be able to have
it
lit up
and
be profitable as soon as we turn
it
on." Still bullish, still building at a very big pace.
I
think
that Satya even came out
and
said that from a building versus buying standpoint, we've been doing a lot of building. How to say this? For
like
that Ellendale one,
it
was going to be build-to-suit
and
then we just decided to walk away from
it
. I know AWS the next day came
in
and
said that they would be willing to go ahead
and
take that because they're looking to build too.
We're very good at building internally, but at the same time we also want to make sure that at the end of 10, 15 years, we're not holding on to just being a massive landlord too. Some of the projects we plan on building, we're seeing if we can actually just outsource that to construction companies
and
have them build
it
and
us lease from them so that we'll look at maybe a 10- or a 15-year lease on
it
.
Again, finding companies that can handle all aspects of
it
has been challenging. There are some that we work with who are very, very good at
it
. There are other ones, they might be good at construction, but they're not good at necessarily negotiating a power agreement or really good at understanding permitting
and
labor licensing
and
things
like
that too. There's a whole PR aspect of
it
. To answer your question, still good, along with Amy's edict is really just make sure that for these projects that we're not going to lose money on them.
Tegus Client
00:03:43
Cool. A couple follow-ups within that. First of all, just to confirm, that data center that
you
all walked away from, that was picked up by someone else very quickly?
Chief Architect at Microsoft
00:03:53
Yeah. AWS, two days later was
like
, "Hey, we'll take
it
." Applied Digital is who we were working with. That project's been under development for three years
and
we had to work with three different utility companies
in
the area to get power delivered there. We had to work with the governor, we had to work with just all these different parties.
Applied basically was
like
, "We don't
think
we can get this done at this timeline. We're going to need more time." We already had commitments from them to have
it
done by next year. They're saying they need three additional years. I was
like
, "No, this is way too long." We basically said, "Having already put three years into
it
,
and
you
're saying
you
need an additional three years
in
addition to the one year. That's eight years total time. Sorry, that's just not soon enough."
Tegus Client
00:04:43
Yeah. Just tough to
do
.
What
was the issue they had? Was
it
construction? Was
it
supply chain or power?
Chief Architect at Microsoft
00:04:49
I
think
the power we had helped to work with,
it
was more the construction
and
it
was them being willing to, I don't want to get too heavily into
it
, but let's just say from a risk management standpoint, because if
you
're going to build a five GW campus,
you
're going to have to put out billions
and
billions of dollars.
When
you
look at mitigating your risk on that,
in
the event that
decided we canceled or something
like
that,
it
's
like
, how much
do
you
want to have that capital outlay
and
then tie
it
up
in
court or something
like
that? I
think
a lot of
it
was just them saying, "Hey, we're not willing to build
it
unless
you
're willing to front us a bunch of money." I can't get too heavily
in
the details though.
Tegus Client
00:05:34
That's all good. The question is,
what
you
just said aligns with
what
Satya said
and
everyone said on the earnings calls. Is the rest of the industry doing the similar thing? Is
it
more that
you
guys move faster
and
now
you
're moving a little smarter while other people are still trying to move fast?
Chief Architect at Microsoft
00:05:50
I
think
the latter. We definitely have moved faster
in
the past. Now
it
's really,
it
's not that we put out too much capacity if anything,
and
we're still underserved. We definitely had a lack of capacity before for the heterogeneous compute
and
then for the AI compute. As a result,
and
it
's even said
in
earnings calls, we lost out on revenue.
Now, I've met with our global CTO beginning of February
and
I was asking him
like
, "Where are we at from a capacity standpoint?" He's
like
, "Actually, we basically saw ourselves as three to four months behind." As of beginning of February, he looked, "Yeah, we're definitely caught up. We're not going to have any more issues
in
terms of outages or not being able to recognize revenue."
I was
like
, "Does that mean we have a surplus?" He's
like
, "Oh, no. We definitely don't have a surplus. We've definitely built out enough facilities to where we're ahead of the demand curve. We're basically just keeping pace with
it
." Which is good.
Right now we rushed to keep to try to address the demand
and
now we're building at the same speed as the demand. Everything is just being more, as
you
put
it
, smart.
It
's more about, "All right, now that we definitely can validate the demand, let's light up that data center."
Tegus Client
00:07:06
What
about your peers? How are other people behaving?
Chief Architect at Microsoft
00:07:09
AWS has been playing catch-up. I will say that they've gotten really good at the power game. We have a whole team called Energy Strategy
and
they really focused with the
deal
and
looking at different sites that are near nuclear reactors or looking at negotiating PPAs or working,
like
what
we did with Three Mile Island. We find them talking to a lot of the same companies that we've been talking to
in
terms of dedicated power for data centers.
They've really done well with the power strategy
and
got better at that. Initially, they didn't want to
do
build-to-suit, but now they are doing more build-to-suit. I
think
they're a little also hesitant
like
we are get
in
the, being the owner. Eventually, we're talking about how we were going to own
like
900 or something data centers. Now
it
's
like
, "All right, we're still going to have 900 data centers, but we're not going to own all those."
I
think
that they're doing the same type thing of saying, "Let's see if we can find the, not the
, the Equinixes, but the CoreWeaves or the Core Scientifics."
and
Equinix are pretty big
in
like
Ireland
and
EU. There are definitely areas where we would rather start doing leasing rather than building
it
ourselves because we know the demand will be there. Just from a cost standpoint, there's a lot of upfront costs that we've been having to front to go through.
Tegus Client
00:08:35
Okay. Just a couple follows up.
You
mentioned the mindset shift a little bit more towards the leasing side versus the building-to-own side. If we were talking three months ago
and
you
said 900 data centers down the road, how much of those would be owned versus leased? Today,
what
do
you
think
that split
looks
like
if
you
had to look that far out
in
the future?
Chief Architect at Microsoft
00:08:57
At the beginning of this fiscal year, we had 312 data centers. I'd say that probably 90% of those were owned
and
the remaining 10%, I
think
we actually were leasing. Now we're moving to, trying to go to a 60/40 split of, "We'll keep building, but let's start shifting that to where we're definitely going to be leasing a lot more." They're building
and
/or leasing.
The type of data centers too,
and
it
's not just the AI ones, we got three separate tiers. The lowest tiers are the network accelerators, the middle tiers are the mobile data centers,
and
the biggest tier are the AI-focused data centers.
What
we're finding is that a lot of those network-focused data centers, those are a hit with customers.
They really
like
those. The fact that they can get on our backbone
and
transfer data much faster is great. The challenge
you
have with those are that a lot of those need to be
in
major metropolitan areas
and
we don't want to be landholders
in
major metro areas. We already have enough facilities
and
whatnot.
With those,
it
's who are companies out there that might have existing physical presence
in
major metro areas where we can say, "Hey, we want to lease facility space from
you
and
drop
in
high-speed networking equipment to get on
and
then all we
do
is bring
in
optical fiber to our network?" A lot of those companies that have metro-based up are being tapped for those.
It
's amazing because once
you
get customers into high-speed data connections, then they just start dumping everything
in
the cloud, which is great.
Tegus Client
00:10:33
Right.
What
do
you
think
it
looks
like
in
2026
and
2027
?
Chief Architect at Microsoft
00:10:38
The plan is to keep doing
it
until 2028.
The overarching goal was to have 900 data centers by 2028. As to whether or not we build those or we lease them, we're shifting more towards leasing, but
it
's still to have around 900 by 2028.
Tegus Client
00:10:56
Okay. Three hundred this year go to 400. That applies
like
another 100 to 120 something each year after that.
Chief Architect at Microsoft
00:11:04
Yeah. This is because originally back
in
2019, our goal was to
do
a 100 data centers
in
a decade. Back
in
2020, when the pandemic hit, we realized there just was not going to be enough data centers out there for the compute demand for AI. We went from a 100
in
a decade to a 100 per year. That number has been slowly going up. Originally we canceled some projects this year, but three of those projects we actually just came back
and
said we need to add them again. We basically uncanceled them.
Tegus Client
00:11:38
Okay. The point though is the number of data centers
you
're building per year isn't necessarily growing, but obviously, that's still a very big number for the next couple years.
Chief Architect at Microsoft
00:11:49
Yeah.
Tegus Client
00:11:49
Okay. There's also some noise around people canceling contracts with
. Any color
you
can add around that? I
think
and
both refuted that. I'm more just how
you
think
about
's doing today maybe is a better way to answer.
Chief Architect at Microsoft
00:12:05
Yeah. I remember I got a call two weeks ago about that.
It
was funny because that call was lined up
like
a week
in
advance
and
the guy jumped on. He was
like
, "Yeah,
you
told me that there's plenty of demand." I was
like
, "Yeah, I don't know who's saying this." I tried to track
it
down. We tried to track
it
down internally. We don't know who said that, but nothing's been canceled. If anything, the contracts haven't been canceled. They've been altered to where we're moving. I got to be careful how I phrase this, too.
There are three primary customers of compute data at
. We've got, obviously
Internal, so we've got our own product groups that run
like
Copilot for Office
and
Copilot for Defender
and
stuff
like
that. The second bucket are honestly our customers. The third bucket we have is OpenAI.
They run quite a lot of compute internally
in
our data centers for their models. Just because the amount of demand we've had from those three buckets, we have not been able to meet just on our data center.
We've had to go into
and
go into
heavily. Those terms
and
conditions
in
the original contract from, I
think
,
like
2022, we weren't able to bring one of those buckets over there. Now we are allowed to
do
that.
It
's definitely made
it
easier for us to leverage
. To a degree, the
investment they've got as well.
Tegus Client
00:13:29
Okay. If anything,
it
's more demand, is a better way to say
it
.
Chief Architect at Microsoft
00:13:33
There's plenty of demand, but they did not want a specific type of workload there from a liability standpoint.
It
had to
do
with liability is
what
it
comes down to. Now there's no issues with
it
. Now they're
like
, "Now we're fine with
it
."
Tegus Client
00:13:48
Cool. The other one is, I
think
Satya made comments about shifting spend incrementally from infrastructure to chips. Curious if
you
can expand on that a little bit more.
Chief Architect at Microsoft
00:13:58
Yeah. I would say definitely one of the things,
and
we foresaw this quite some time ago, is that the infrastructure, as
you
build
it
out there, there's definitely still a need for the GPUs. Part of me being a chief architect, I remember thinking to myself, there's going to be probably a company that gets out there
and
builds really good GPUs
and
that's going to be the one that everyone goes towards.
We had people back
in
2018 who foresaw that
and
said, "At some point, someone's going to build this rockstar GPU
and
the other ones won't be able to compete. Whoever's the first out of the gate is really going to capture the market." That became
. The challenge
you
have is that there are many, many different types of models out there.
You
've got your small language or large language models.
You
've also got your foundational models, some of which are vision-based.
You
've also got decision tree analysis models.
You
've got the more the math
and
scientific models,
like
did, where
it
's all numeric as opposed to using language
and
things
like
that.
We knew quite some time ago, with all these different models. Customers
and
I've seen this for the last three years, is that they go,
and
they're
like
, "We want the H100 from
."
It
's
like
, "Well,
you
don't even have large languages.
You
're using chemistry
and
stuff
like
that. Why would
you
want the H100?" "Well, we hear that's the best."
We built a technology called Smart Fabric, where
what
it
does is
it
's an abstraction layer.
You
send us your data,
you
send us your model,
you
send us basically
what
you
want to train
and
inference on. Behind the scenes,
what
we
do
is we go, "All right, even though
you
told me
you
want to run on the H100, this is a decision tree model. We're going to run that on
what
are called APUs, which are associated processing units. This is a cognitive one. This is using vision processing. We're going to run
it
on a VPU as opposed to a GPU."
We
do
it
behind the scenes. The reason we
do
that is because there's just not enough GPUs out there, if we were to run all those loads on all GPUs, then we'd have to be just basically turning away customers. We run
it
on other types
and
it
does two things.
Number one,
it
frees up the GPUs for the people who actually need them. Number two, the customer gets a lower bill at the end of the day. They usually come back
and
go, "Well,
you
told us this was going to be $30,000
and
it
's only $3,000? All right, well, given how cheap
it
is now, I'm going to run more." They end up spending more
and
more money with us as a result of that.
Getting back to your question on chips versus infrastructure, a lot of
what
you
're seeing is those models, the first ones that have come out, they've been trained on GPUs, or they've been trained on the appropriate processing unit. Now
it
's, "Well,
do
we want to keep inferencing on GPUs?" That's not really cost-effective. Maybe I could run
it
on something else that is more cost-effective,
like
an FPGA, or maybe have a custom ASIC built to run that model or something
like
that.
A lot of
what
we've been doing now are the build-outs of our own
and
, but also FPGAs. We've been using FPGAs for over a decade now. The nice part about those are that
you
can take pretty big robust models. An FPGA is basically
like
a huge emulator.
You
can run
it
.
You
basically create the model
in
there.
You
run
it
and
virtualize
it
,
and
then
you
can
do
the inferencing at a fraction of the cost. Where
it
costs,
it
's 1,250 W is how much a Blackwell chip runs. An FPGA, even the highest tier FPGA is run about 250 W.
It
's
like
a fifth of the cost of running from an energy standpoint, to run an FPGA.
Tegus Client
00:17:47
That make sense. Is your GPU spin slowing? Is
it
that
you
're just also adding this stuff on top to migrate things around?
Chief Architect at Microsoft
00:17:54
Also adding some stuff on top. We knew this was going to be coming.
It
's, "All right. Right now, everyone needs to
do
training
and
inferencing." Everyone needs to
do
training. That's why we bought so many GPUs. There's still not enough GPUs out there, I'll tell
you
that. At the same time, now that they've been switching over
and
going, "I can't keep paying for these GPUs,
what
else
do
you
have?" We'll say, "Yeah, we got FPGAs." We'd plan the strategy out for quite some time. Right now we actually work with several vendors to build custom FPGAs for us for this purpose.
Tegus Client
00:18:28
Got
it
.
and
the Bitcoin side, why
do
you
think
it
's taking so long, or really for
to find a non-CoreWeave customer?
Chief Architect at Microsoft
00:18:37
Why is
it
taking them so long to find? We have so much demand that we're just sending to
.
Tegus Client
00:18:43
For their ongoing purposes
and
concentration, they're trying really hard to find a non-
customer. Someone else, another hyperscaler. They talked about
it
being done by end of the year last year,
and
obviously,
it
still hasn't happened. I
think
the industry as a whole has not signed really any agreements with Bitcoin miners. I'm curious if
you
have any thoughts on that if there is anything
you
need to call out, or anything to highlight.
Chief Architect at Microsoft
00:19:10
I would say that when
you
look at those two different workloads, either AI versus Bitcoin, Bitcoin is just a volatility on
it
.
It
's not even just the technology alone, but at the beginning of last year, where we would had a bunch of
,
like
companies calling us, saying, "Hey, we've got this capacity, we've got these GPUs, could we convert them over to
do
AI?"
I remember going to several sites
in
Texas
and
some parts of Europe too,
and
visiting
and
talking to them
and
saying
like
, "Well, look,
you
don't have appropriate HVAC
in
here. This is incredibly noisy.
You
need liquid cooling,
you
need better perimeter security,"
and
things
like
that.
There's a lot of enhancements
you
have to
do
to make GPUs ready for AI, including high-speed networking. A lot of those Bitcoin rigs, as I said, they're just gigantic rigs. They're not doing a lot of data transfer within the data center. They're not redundant either too. Looking at them we're
like
, "Here's the checklist of stuff
you
would need to
do
."
Some of them decided, "Yeah, we're willing to
do
this, we're willing to convert over because we see AI as that's a long tail, that's a 15-, 20-year revenue stream versus Bitcoin."
It
's the more bitcoins that
you
find, the fewer there are available
and
the more costly they become.
It
's at some point that plank,
it
's a burning plank
and
you
're just going to basically put yourself out of business.
Some of them did decide to
do
that. I
think
with Trump coming
in
and
saying he really likes Bitcoin
and
he really embraced digital currencies
and
whatnot, some of those that had committed to converting over came back
and
said, "Actually, we
think
that bitcoins are going to be profitable again
and
we'll find new ways to make money off of them."
When was
it
like
a month ago? Several of them have come back to us
and
said, "Now I
think
we definitely want to convert them over to AI one." Several of them are out there,
like
, "
Do
I want to make the billions of dollars
in
investments to convert these over to AI data center?"
It
's not even necessarily billions, but
do
I want to raise the capital to convert these over
and
make them purely AI-focused or
do
I want to keep hedging my bet? Bitcoin keeps going up
and
up
and
then
you
see
it
take huge tumbles. I
think
a lot of them are trying to get out of the volatility game.
Tegus Client
00:21:34
Do
you
think
will be successful
in
finding a non-CoreWeave customer?
Chief Architect at Microsoft
00:21:39
Yeah. I
think
it
would be interesting if they work with
like
a Nebbius or someone
like
that who's got one data center out
in
Europe. I
think
if
were to partner with some of these other companies that have a lot of good funding behind them
and
be able to say, "Hey, we're willing to build data centers, as long as
you
can bring demand through your channel." The demand is out there
and
it
's going to continue to come with us. I
think
AWS, I know that they talk to
from time to time, but they would probably want a vehicle
like
to go through.
Tegus Client
00:22:15
Right. To be clear, I
think
they have the megawatts
like
they have the sites.
It
's not Greenfield versus some of the Bitcoin miners. I understand.
It
is Greenfield.
Chief Architect at Microsoft
00:22:25
Yeah. The power is the hardest thing to get, frankly. We've got companies that
do
land banking for us. We've got plenty of the ability to negotiate PPAs
and
stuff
like
that.
What
it
comes down to is how can
you
make sure that, especially for us, that power is clean, that
it
's sustainable, that
it
's renewable energy,
and
then guarantee the transmission to those sites?
I will say there are certain states
like
Ohio, where there's three major power companies
in
Ohio
and
in
Southern Ohio, when we wanted to build there, the power company said, "We're not going to make these tens of millions of dollars
in
investments
in
power infrastructure
and
then just have
you
build a data center there just for
you
. If
you
want us to make these investments, we want
you
to share
in
the build-out of them, but then we also want a cut of the revenue that's going over those regardless of not
you
're making money."
We went back
and
said, "
What
?" They're
like
, "Look, this is something we have to maintain those lines
and
so we want a share
in
the profits." I
think
a lot of those, the utility providers, they're getting to that point where they're being pressed to build out more infrastructure for data center providers. If
you
've got a Bitcoin miner that's already done the upfront work, the power infrastructure,
and
maybe already has a PPA
in
place.
I saw
, I
think
it
was three hours ago, I saw an article about how they had just negotiated down some of their energy rates, which gives them better profitability too. That's the thing is if
you
've got the shell
in
place
and
you
are able to work with a Clean Spark or RIS Energy, things
like
that,
and
have them bring down your overall energy cost, then
you
're definitely attractive to a hyperscaler.
Tegus Client
00:24:14
Okay. On the other Bitcoin miners, any thoughts on
like
someone
like
how Hut eight is doing
in
this market?
Chief Architect at Microsoft
00:24:20
Yeah, they're doing okay. Any specific location?
Tegus Client
00:24:23
A lot of these miners are thinking about converting. As far as ones
you
think
are farther along that curve, as far as finding customers professional, attracted to the hyperscaler customers or
kind of customers?
Chief Architect at Microsoft
00:24:36
They've got Hut eight. I
think
they've just bought
like
600 ac that they were going to
do
with 300 MW of power. Actually, when
you
look at
it
, that's a real smart investment.
It
's $2.5 billion. The good thing about that with 600 ac is that
you
can gradually expand on that. That's
what
we did out
in
Boydton, Virginia. We bought hundreds of acres. The nice part there is as we grow the demand, we can slowly just drop
in
more buildings
and
we got power coming to that campus
and
things
like
that too.
As long as
you
've got,
you
diversify your strategy. Some of those buildings will put GPUs
in
, some of them will put
in
FPGAs for imprinting. Some of them we put
in
HPC equipment for high-performance compute. I
think
Hut eight is actually positioned well.
Now them actually converting over to doing AI there, they've been doing more Bitcoin
and
high-performance computer
and
they've been slowly moving into doing more AI workloads.
You
got to be willing to put
in
the upfront CapEx to make that conversion. Once
you
do
,
you
're well-positioned to take advantage of demand going forward.
Tegus Client
00:25:43
A final one, just about nuclear. Any new thoughts on the nuclear side? How
you
think
about that as an option? Has that decision tree changed at all?
Chief Architect at Microsoft
00:25:52
That keeps coming. We're still pushing forward on that.
What
it
comes down to is
you
stand up a nuclear plant
and
that can basically give
you
1,200 MW of power for one nuclear reactor. We've been looking at Three Mile Island to see
what
we could
do
to maybe build another tower there
and
then to build a data center there.
That program has been going quite well, honestly. The SMR, the companies that we're working with, they're making a lot of progress there. I'm happy with that. I know
has been very happy with the progress that we're making on that.
It
still comes down to regulatory.
Some of
it
is
like
just, how to say this, there were people that we were working with that prior to two months ago were gainfully employed by the Department of Energy that decided that they no longer wanted to work for the government.
It
's
like
, "Hey, they got dogged."
It
's
like
, "All right, well,
what
do
we
do
now? Who
do
we talk to?" They're
like
, "Well, we currently don't have anyone
in
that position."
It
's
like
, "All right, well, how
do
we get the permits for building this stuff?"
They're
like
, "
You
're going to have to go wait for us to hire new people. Oh, by the way, there's a government hiring freeze. Sorry." Right now, we're doing a lot from a technology standpoint
and
advances there,
and
honestly,
it
's really impressive. The challenge is just
like
, who
do
we work with
in
the government? We can't just go out
and
build a nuclear reactor that's unpermitted
and
unlicensed.
Tegus Client
00:27:23
Cool. Thanks for your time
and
insight.