Latest / Elon Musk Podcast / Anthropic takes control of SpaceX's Colossus supercomputer

Transcript

Anthropic takes control of SpaceX's Colossus supercomputer · Elon Musk Podcast

0:00
Anthropic just took total control of every single chip
0:03
inside Spacex's Colossus 1 supercomputer in Memphis,
0:07
securing hundreds of thousands of NVIDIA GPU's and hundreds of
0:11
megawatts of power that were originally built to train Elon
0:14
Musk's XAI models. Yeah, the size of this
0:17
acquisition completely rewires the power dynamics of the entire
0:20
tech sector. We are going to trace exactly
0:23
how a pretty catastrophic miscalculation and compute
0:28
demand forced a really bitter rivalry to turn into this
0:31
sprawling infrastructure alliance, right?
0:34
And you know how this sudden influx of hardware is unlocking
0:36
artificial intelligence that literally dreams between coding
0:40
sessions? So how does a company go from
0:42
being publicly labeled as evil by Elon Musk to renting his
0:45
entire supercomputer? And what does this bizarre
0:47
alliance actually mean for the physical limits of artificial
0:50
intelligence? Well, the origin, this whole
0:52
crisis really comes down to a forecasting error of just
0:55
historic proportions. Oh really?
0:57
Yeah, Entropic expected their growth to multiply by like a
1:00
factor of 10, but in reality, they experienced an 80 fold
1:03
surge in revenue and usage. Wait, 8080?
1:06
Yeah, and that completely exhausted their available
1:09
infrastructure, just pushing their servers to the absolute
1:12
breaking point. Wait, back up.
1:13
How does a major tech company underestimate their own growth
1:17
by that wide of a margin? I mean, forecasting A tenfold
1:20
increase is already incredibly aggressive for any enterprise
1:23
operation, right? Absolutely.
1:25
So hitting 80 times your expected growth sounds like a
1:29
fundamental misunderstanding of how your own customers are
1:32
actually using your product. That's exactly what it was.
1:35
It was a complete misunderstanding of how users
1:37
were interacting with the product.
1:39
We are looking at the shift from conversational artificial
1:42
intelligence to what the industry calls agentic
1:45
workflows. OK, agentic.
1:47
Right. Previously you interacted with a
1:48
chatbot in a very linear, predictable way, and you type a
1:52
question, the model processes your text and it generates an
1:56
answer, right? That is single transaction.
1:58
It consumes A finite, easily measurable amount of
2:01
computational power. You can build out a highly
2:04
accurate financial model for a data center based on like
2:08
millions of people asking single questions.
2:10
Because if I just ask for a recipe for banana bread, or you
2:14
know, I ask the AI to debug A5 line Python script, the
2:18
interaction is finished the moment the text appears on my
2:21
screen. Exactly.
2:22
The server can immediately allocate its resources to the
2:24
next person waiting in. Line right, but users stop
2:27
treating these systems like search engines.
2:30
They started deploying tools like Claude Code to run
2:33
continuous autonomous loops inside their software
2:36
repositories. You are no longer asking for a
2:39
single block of code. You are telling the agent here
2:42
is an entirely new feature I want in my application, Go read
2:46
my existing code base, understand how the database is
2:49
structured. Write the new code for the
2:51
feature. Run the software tests.
2:53
Wow. OK.
2:53
And. If those tests fail, read the
2:55
error logs, rewrite the code, and test it again until it
2:58
passes. So the software isn't just
3:00
answering a prompt, it's acting like an employee trying to solve
3:03
a puzzle over and over again without stopping.
3:05
And that changes the math completely.
3:08
An average developer is keeping these tools running constantly.
3:11
The compute consumption of an agentic loop is just
3:14
astronomical compared to a simple chat request.
3:17
Because it never really stops. Exactly, and understand why it
3:21
uses so much power. You have to look at how these
3:24
models process information through what are called.
3:27
Tokens, right? Tokens A.
3:29
Token is essentially a fragment of a word.
3:32
The word hamburger might be split into 3 tokens by the
3:35
system's tokenizer. When you ask a question, the
3:38
model reads the tokens in your prompt and then generates new
3:41
tokens for the answer. And every single token requires
3:44
the processors to perform thousands of mathematical
3:47
operations just to predict what the next fragment of a word
3:51
should be. Yes.
3:52
Now I apply that math to a genetic loop.
3:54
Every single time the agent writes a piece of code, runs a
3:56
test and fails, has to try again.
3:58
Yeah, but to try again, it has to reread the entire history of
4:01
what it just did. Oh, I see.
4:03
It has to ingest your entire code base, the original
4:06
instructions, the code it just wrote, the error log from, the
4:10
failed test, and its own internal reasoning about why the
4:13
test failed. That sounds massive.
4:15
It is. It has to process all of those
4:18
tokens again just to understand the context of the new attempt.
4:21
It consumes millions of tokens in a single afternoon.
4:25
So the original infrastructure was built for people asking
4:28
single questions, and suddenly millions of users are running
4:32
nonstop autonomous loops that consume computing power
4:35
exponentially. Yep, you have a server farm
4:38
expecting people to sip water from a fountain and instead
4:41
everyone showed up with industrial fire hoses.
4:43
That's a great way to put it. And that caused the product
4:45
experience to degrade severely. The infrastructure literally
4:49
could not handle the concurrent load right because the system is
4:52
trying to process billions of tokens simultaneously.
4:55
Paying users slammed into hard usage ceilings.
4:58
The servers simply did not have the processing cores available
5:01
to run the math. I mean, imagine you are a
5:03
developer. You have integrated this tool
5:06
completely into your daily workflow.
5:09
You rely on it to hit a critical project deadline and suddenly it
5:12
just stops working. Yeah, it's a nightmare.
5:14
You hit a button and get a spinning wheel, or worse, an
5:17
error message telling you to come back in three hours.
5:20
And it was happening constantly. Their accounts were actively
5:24
throttled, particularly during peak business hours when
5:28
everyone on the East Coast and West Coast of the United States
5:30
were working at the exact same time.
5:33
Enterprise clients, you know, the ones paying premium
5:35
corporate rates for guaranteed access, found their access
5:38
restricted just when they needed it most.
5:41
They were seeing time out errors and drop sessions right in the
5:45
middle of complex architectural tasks.
5:47
Which creates an existential threat to the company's
5:50
subscription model. If you are selling an autonomous
5:53
digital Co worker, it cannot arbitrarily decide to go on
5:57
strike at 2:00 in the afternoon because the server room is
6:00
running hot. The reliability is the entire
6:02
value proposition. If the tool is not dependable,
6:06
enterprise clients will just abandon it, regardless of how
6:08
intelligent the underlying model actually is.
6:11
Anthropic was facing a crisis where their own success was
6:15
basically destroying their product experience.
6:18
They needed computing power immediately, and they could not
6:20
wait years to build a new data center, right?
6:23
The only solution was to find an existing facility with the raw
6:27
physical hardware already installed and ready to go.
6:29
So to solve this crisis, Anthropic secured the Colossus
6:33
One facility in Memphis, right? And this is a supercomputer
6:36
built in a fraction of the usual construction time, and it is
6:39
housed inside a former Electrolux appliance factory.
6:43
Yeah, the physical reality of Colossus 1 is fascinating When
6:46
you examine the hardware involved, it operates with
6:48
hundreds of thousands of high end NVIDIA accelerators.
6:51
We are talking about H1 hundreds, H2 hundreds and the
6:54
absolute newest generation of chips designed specifically for
6:58
artificial intelligence workloads.
6:59
For anyone wondering why a standard computer processor
7:02
can't do this job, it comes down to how the chips handle tasks.
7:05
The CPU in your laptop has a few very powerful cores.
7:10
It is designed to execute a few complex tasks very quickly in
7:13
sequential order. Right one after the other.
7:15
But in NVIDIA GPU, a graphics processing unit has thousands of
7:19
smaller, less powerful cores. It is designed to do thousands
7:23
of simple mathematical calculations simultaneously.
7:26
Training an AI model or running in a gentic loop requires
7:30
millions of simultaneous matrix multiplications.
7:33
The GPU's are really the only hardware capable of that level
7:37
of parallel processing, and the power density required to run
7:41
hundreds of thousands of those specific accelerators in one
7:44
location is just staggering. An individual server rack fully
7:48
loaded with these chips draws exponentially more electricity
7:51
than a traditional cloud storage server rack.
7:53
Hold on, why use an abandoned appliance factory instead of
7:56
building a proper data center from scratch?
7:58
That's the crazy. Part I mean, if you were
8:00
deploying the most advanced sensitive computing hardware on
8:02
earth, why put it in an old manufacturing plant with a leaky
8:05
roof? Because of the utility grid
8:07
bottleneck, traditional data centers face interconnection
8:10
queues that are becoming impossibly long.
8:14
What do you mean by interconnection queues?
8:16
Well, if you want to build a new hyperscale facility today, you
8:19
have to ask the local utility divider for hundreds of
8:22
megawatts of power. OK, The utility cannot just flip
8:26
a switch. They have to conduct massive
8:28
environmental and engineering studies, physically upgrade high
8:32
voltage substations, and lay new transmission lines across miles
8:36
of public and private land. Oh wow.
8:39
That queue can last well over a decade just to get the permits
8:42
approved. So you simply cannot build fast
8:45
enough. If you rely on the standard
8:47
electrical grid, you are trapped by the bureaucracy and the
8:50
physical limitations of the local power company.
8:52
Exactly. And if your software is crashing
8:54
today, you don't have 10 years to wait for a substation.
8:57
By retrofitting an industrial site, the builders bypass the
9:01
grid interconnection delays almost entirely.
9:03
An appliance factory already has significant industrial power
9:06
infrastructure feeding into the property, but more importantly,
9:10
the builders brought their own power.
9:12
Yet this site relies on dozens of massive natural gas burning
9:17
turbines installed directly on the property.
9:19
They essentially built their own private fossil fuel power plant
9:23
just to run the servers independently of the Memphis
9:25
grid. They did, and power generation
9:27
is really only half the physical equation here.
9:29
Right. The heat generation rated by
9:31
hundreds of thousands of high end accelerators operating at
9:34
full capacity is immense. You cannot cool a dense
9:37
deployment of these chips with traditional air conditioning.
9:40
Just blowing air on them doesn't work.
9:42
No. Moving cold air over the racks
9:44
simply is not efficient enough to pull the thermal energy away
9:47
from the silicon before the chips physically melt.
9:50
You need liquid cooling. Which means pumping water
9:53
directly into the server racks, circulating it over specialized
9:56
cold plates attached directly to the processors, and piping the
10:00
hot water away to dissipate the heat.
10:02
Yes, water is vastly superior to air when it comes to thermal
10:06
conductivity. It absorbs heat much faster.
10:09
Exactly, but that requires an incredible amount of water
10:12
moving continuously through the system.
10:14
That is why the site features a custom wastewater treatment
10:17
plant. A whole treatment plant just for
10:19
the servers. Yeah, they process millions of
10:21
gallons of water just to handle the immense liquid cooling
10:24
requirements of the facility. They created an entirely
10:27
self-contained ecosystem of power generation and thermal
10:31
management inside an old factory shell.
10:33
And the consequence of plugging into this massive,
10:36
self-contained supercomputer was immediate for Anthropic users.
10:40
The rolling usage caps for Pro subscribers basically doubled
10:43
overnight. The artificial constraints
10:46
simply vanished. All of the peak hour throttling
10:49
that had plagued developers was entirely removed because the
10:52
back end processing power suddenly expanded to meet the
10:56
demand. And for enterprise developers,
10:57
the API input limits jumped by 1500%.
11:01
Which is huge. Think about what that actually
11:03
means for a software engineer. The context window?
11:06
The short term memory of the AI? The amount of text it can hold
11:09
in its brain at 1:00 time expanded so massively that they
11:13
can process giant code bases in a single prompt.
11:16
It alters the entire utility of the tool.
11:18
When you increase the input limit by that magnitude.
11:21
You move from processing individual files to analyzing
11:24
entire system architectures, right?
11:26
The agent can see your front end user interface, your back end
11:31
database routing, and your security protocols all at the
11:34
exact same time. It can spawn an error in the
11:37
database that is causing a visual glitch on the website
11:40
because it has the processing power to hold the entire
11:42
architecture in its active memory.
11:45
So for the everyday developer, the frustrating usage limit
11:48
reached screen essentially vanished overnight.
11:52
The tool changed from a novelty that might abandoned you at a
11:55
critical moment into a highly reliable Co worker.
11:58
It transitioned from a brittle experiment into robust
12:01
foundational infrastructure. But you know, the acquisition of
12:05
this specific facility reveals a very complex layer of corporate
12:09
strategy, particularly when you look at the personalities
12:11
involved here. Right, because Elon Musk
12:13
previously attacked Anthropic publicly.
12:15
Very publicly. He called the misanthropic and
12:17
heavily criticized their approach to safety and
12:20
alignment, suggesting they were designing models to be
12:22
politically biased or restrictive.
12:24
But he personally approved this massive lease because, in his
12:28
words, no one set off my evil detector during their meetings.
12:32
The change in rhetoric is striking, but honestly, it is
12:35
driven entirely by cold financial logic and the
12:39
structural change is happening within Musk's own organizations.
12:42
You have to look at the larger corporate board.
12:45
OK, XAI, his artificial intelligence company, merged
12:49
with SpaceX to form a new combined entity called Spacexai.
12:54
And they are preparing for an upcoming public offering.
12:56
And Wall Street analysts are valuing that combined entity at
13:00
well over a trillion dollars. When you are taking a company
13:03
public at a trillion dollar valuation, you need a flawless
13:07
equity narrative to present to institutional investors.
13:10
Colossus 1 was originally built to train XA is internal models
13:14
like Grok, but having a massive natural gas burning
13:18
supercomputer that only serves your internal research division
13:21
is a massive cost center. It's just burning money.
13:23
It burns cash constantly just to keep the servers running and the
13:26
cooling water flowing. But renting Colossus One out to
13:29
a paying, well funded enterprise customer like Anthropic changes
13:34
the financial profile completely.
13:36
It transforms a massive capital expenditure into a highly
13:39
lucrative operational revenue stream.
13:41
Exactly. It proved to investors that
13:44
SpaceX AI is not just a hardware manufacturer with an artificial
13:48
intelligence research wing, but a highly profitable cloud
13:51
infrastructure provider. Oh.
13:53
That makes sense it. Secures a stable, massive
13:55
revenue source for the upcoming roadshow.
13:57
And it does not actually hinder Musk's own artificial
14:00
intelligence development, right? His internal projects, like the
14:04
next generation of Grok, the computer vision models required
14:08
for autonomous driving, and the neural networks powering his
14:10
robotics initiatives, have already migrated to the even
14:14
larger Colossus 2 facility. Right.
14:16
They built a second, exponentially larger
14:19
supercomputer strictly for their own internal use, which freed up
14:23
the first facility to be monetized on the open market.
14:26
Wow. It's a brilliant piece of
14:27
balance sheet management. But, you know, the geopolitical
14:29
angle surrounding this infrastructure deal is just as
14:32
critical to understand. Right.
14:33
This brings in the United States government dynamic.
14:36
The Pentagon recently blacklisted Anthropic as a
14:38
supply chain risk. That blacklisting was a major
14:41
event in the tech sector. The Pentagon designated them as
14:44
a supply chain risk specifically because Anthropic refused to
14:49
remove safety restrictions for military operation.
14:52
Right. Really.
14:53
Yeah. The military wanted to integrate
14:55
the models into certain logistical and analytical
14:58
applications, and Anthropics leadership decided those
15:01
applications violated their internal usage policies
15:03
regarding potential harm and autonomous decision making in
15:07
combat scenarios. So they walked away from
15:09
massive, highly lucrative defense contracts to maintain
15:13
their internal safety standards, and as a result, the Department
15:17
of Defense effectively banned them from certain federal
15:20
integrations. They did, and counter
15:21
intuitively, it created a highly unique and valuable market
15:24
position for them. How so?
15:26
Well, if you look at the other major cloud providers, the
15:29
companies that typically rent out the computing power needed
15:32
to run these massive models, they are deeply entangled with
15:35
defense contracts. They provide the foundational
15:38
cloud infrastructure for global military logistics, intelligence
15:41
gathering and drone operations. But Anthropic is now relying
15:45
exclusively on SpaceX hardware for their primary computing
15:49
needs. And this provides them with a
15:50
form of sovereign compute because they are running on an
15:53
independent infrastructure stack completely separated from the
15:57
traditional massive cloud providers who are deeply
16:00
embedded with the Pentagon. They offer a very distinct
16:03
alternative to the corporate world.
16:06
This appeals heavily to enterprise clients who want to
16:08
completely distance their workflows from any military
16:11
applications. Exactly.
16:12
Imagine you are the chief compliance officer of a massive
16:16
European financial institution, a global healthcare conglomerate
16:20
or a multinational pharmaceutical corporation
16:22
operating in regions with strict data sovereignty and privacy
16:26
laws. Those executives might be
16:28
extremely hesitant to run their highly proprietary, regulated
16:31
corporate data through a cloud infrastructure system that is
16:35
intimately linked with United States military operations
16:38
simply due to the risk of espionage, regulatory blowback,
16:42
or, you know, public relations disasters.
16:45
Anthropic's isolated non military compute environment
16:48
opens up entirely new corporate markets for them globally.
16:51
So having this limitless hardware finally secured gave
16:55
Anthropic the physical freedom to explore software features
16:58
they simply could not afford to run when they were strapped for
17:01
compute. Yes, and this is where it gets
17:02
really interesting. They launched a feature called
17:05
Dreaming, which allows AI agents to review their own past
17:09
sessions and rewrite their memory stores while the system
17:12
is idle. This represents A fundamental
17:14
shift in how artificial intelligence retains and
17:16
processes information. Dreaming operates as an
17:20
asynchronous background job. When the agent is not actively
17:24
working on a coding task or answering a prompt for a user,
17:27
the servers do not just sit quiet right?
17:30
The agent initiates this reflection process on its own.
17:33
Wait, I need clarification on how this is actually different
17:35
from an AI just remembering what you told it.
17:37
If I tell a standard chat bot my name is John, it remembers my
17:41
name for the rest of the conversation.
17:43
Sure. How is dreaming functionally
17:45
different from standard memory? We have to distinguish between
17:48
in session memory and cross session consolidation.
17:52
A standard artificial intelligence is stateless across
17:55
sessions. Meaning what?
17:56
It starts every single interaction as a completely
17:59
blank slate, relying only on its original frozen training data.
18:04
It forgets its repeated mistakes every time you start a new task.
18:07
So if I open a new chat window and ask the agent to format a
18:11
SQL database and it makes a syntax error, I correct the
18:14
error and it finishes the job. But tomorrow when I open a brand
18:18
new window and ask it to format a similar database, it will make
18:22
the exact same syntax error all over again because the new
18:25
window has no idea the previous window ever existed.
18:28
Exactly because it has no cumulative operational
18:31
experience. Yeah, but a dreaming AI behaves
18:35
entirely differently. It curates its own persistent
18:38
memory file. As you interact with the agent
18:41
across weeks and months, the system writes continuous notes
18:44
to a persistent file about your preferences, your coding
18:47
environment, and the mistakes it has made.
18:49
But if you just let an AI write infinite sticky notes to itself,
18:52
eventually the desk is completely covered and the
18:54
system slows down because it has to read every single sticky note
18:57
before it takes an action. That's the danger, yeah.
18:59
The agent might learn a highly specific workaround for a
19:03
software bug on Monday, write a note about it, but by Friday
19:07
that bug is patched by the vendor and the workaround is
19:09
completely obsolete. And the persistent memory file
19:12
becomes bloated and highly contradictory.
19:16
The agent ends up wasting massive amounts of your
19:18
computing power, reading outdated instructions and
19:21
conflicting rules, which severely degrades its
19:24
performance and reasoning abilities.
19:26
That degradation is a phenomenon called context rot.
19:30
Dreaming is the cure for context rot.
19:33
O How does the asynchronous job actually clean U the sticky
19:37
notes? During the idle time, the agent
19:39
actively evaluates all of its stored notes against the
19:42
transcripts of its most recent experiences.
19:45
It scans dozens of past sessions.
19:47
Oh wow. It recognizes the pattern of its
19:49
own failures. It independently decides to
19:51
prune the obsolete workarounds, merge duplicate instructions,
19:55
and synthesize a clean, highly optimized memory store.
19:58
That's incredible. And it executes this entire
20:00
cleanup operation without a human ever explicitly telling it
20:03
what to remember or what to forget.
20:05
Anthropic also introduced a parallel feature alongside
20:08
Dreaming, where up to 20 specialized agents work
20:11
simultaneously on a single project.
20:13
And all twenty of those agents are feeding their individual
20:16
experiences back into this shared dreaming memory.
20:20
This concept is called multi agent orchestration.
20:23
You have a lead agent acting as a primary project manager and it
20:27
autonomously delegates tasks to highly specialized sub agents.
20:31
How does that work in practice? Well, you might have one agent
20:35
tasked exclusively with searching AIR logs, a second
20:38
agent tasked entirely with reading external API
20:40
documentation, and a third agent writing the actual Python code.
20:44
All at the same time. They operate in parallel
20:46
simultaneously, which drastically reduces the physical
20:49
time it takes to complete a highly complex software
20:52
architecture project. The pricing structure for this
20:54
orchestration is highly unique. Users pay a few cents per active
20:58
session hour, plus the standard token cost for the input and
21:01
output. You are effectively paying for
21:03
the infrastructure time, the active session hour, to keep the
21:07
autonomous loop open and running on the server, in addition to
21:10
the raw intelligence which is the tokens right, It reflects
21:14
the new physical reality that hosting an active looping agent
21:18
requires dedicated, sustained server resources rather than
21:23
just a momentary blip of processing power.
21:25
This alters how software is built from the ground up.
21:28
Agents no longer suffer from context rot where their
21:31
instructions become bloated and confused.
21:33
No they don't. They transform into self
21:36
improving entities that accumulate genuine operational
21:38
experience over countless tasks. A development team working with
21:42
a dreaming agent for six months will have a fundamentally
21:46
different tool than a team that just started using the exact
21:49
same model today. Because it's learned their
21:51
habits. The agent has actively adapted
21:54
to their specific coding standards, learned their
21:56
preferred architectural patterns, and deeply
21:59
internalized all the strange quirks of their legacy systems.
22:03
It evolves the artificial intelligence from a generic
22:06
digital assistant into a highly specialized, customized
22:09
colleague. But even with Colossus One, the
22:11
physical demands of these dreaming, looping agents are
22:14
still growing exponentially. The electrical grid on Earth
22:18
simply cannot sustain this trajectory.
22:20
It really can't. This leads to the ultimate
22:22
escape valve. The infrastructure agreement
22:25
includes A provision for Anthropic and SpaceX to jointly
22:29
explore building GW scale data centers in Earth's orbit.
22:32
The appeal of placing supercomputers in space is
22:35
driven by the absolute exhaustion of terrestrial
22:37
resources. The electrical grid on Earth is
22:40
entirely tapped out. We discussed the decade long
22:43
interconnection queues just to build a substation.
22:46
Furthermore, local communities are actively and aggressively
22:49
protesting the environmental impact of massive data centers.
22:53
Right. People do not want dozens of
22:55
natural gas turbines running constantly next to their
22:58
suburban neighborhoods, and they absolutely do not want millions
23:01
of gallons of their local municipal water supply
23:04
evaporating into the atmosphere just to cool server racks for a
23:08
tech company. Exactly.
23:10
Orbit offers an environment entirely free from those
23:12
terrestrial constraints. There are 0 zoning laws in low
23:15
Earth orbit. There are no local community
23:17
boards denying building permits or water usage rights.
23:21
Right. And most importantly, orbit
23:23
offers continuous, unimpeded solar energy.
23:26
A solar panel positioned in space can generate significantly
23:29
more power than the exact same panel on Earth because there is
23:34
no atmosphere, no clouds to filter the light, and depending
23:37
on the specific orbital path, there is no night cycle.
23:40
You have access to a permanent GW level power source.
23:43
The energy energy generation is flawless, but the concept hits a
23:48
massive physics wall the moment you try to manage the heat
23:52
generated by the servers. It really does.
23:54
This is a fundamental flaw in the spaces cold argument that
23:58
most people will naturally assume.
24:00
People assume that because the void of space is freezing,
24:03
cooling a massive data center in orbit would be incredibly easy.
24:07
You just open a vent and let the cold in.
24:09
That's what everyone thinks. But the vacuum of space is
24:11
actually a near perfect insulator because there is
24:14
absolutely no air to carry the heat away.
24:16
On Earth, we rely heavily on convection and conduction to
24:19
cool our electronics. We blow cold air over a hot
24:23
surface. The air physically absorbs the
24:25
heat energy, and the fan moves that hot air away from the
24:28
device. Or you pump cold water over a
24:31
hot surface, the water absorbs the heat and you pipe it away to
24:34
evaporate in the cooling tower. Convection requires a physical
24:39
medium like air or water to transfer the thermal energy from
24:43
the hot object to the surrounding environment.
24:45
Think of space not as a freezer, but as a giant vacuum sealed
24:49
thermos. If you put boiling hot coffee
24:51
inside a high quality thermos, it stays hot for hours because
24:55
the vacuum layer prevents the heat from escaping.
24:57
Perfect analogy. There is no air inside the
25:00
vacuum to pull the heat away from the liquid.
25:02
A data center in orbit is trapped inside a cosmic thermos.
25:06
If you put a hot computer chip in the vacuum of space and run a
25:08
calculation, it will simply bake itself until the silicon
25:11
physically melts because the thermal energy has nowhere to
25:14
go. The only physical mechanism
25:16
available to remove heat in a pure vacuum is through thermal
25:19
radiation. You have to literally radiate
25:21
the heat away as infrared light. Which is really hard.
25:24
Very hard. This entire process is governed
25:27
by the Stefan Boltzmann law, a principle of physics which
25:30
details exactly how energy radiates from a black body in
25:33
direct relation to its temperature and its surface
25:35
area. And radiating heat as infrared
25:38
light is incredibly inefficient compared to terrestrial liquid
25:41
cooling. To dissipate just a small
25:45
fraction of the heat generated by these advanced AI chips, a
25:48
single satellite requires a radiator surface area roughly
25:52
equal to four professional tennis courts.
25:55
You cannot just attach a small metallic fin to the side of the
25:58
satellite and expect the servers to survive.
26:01
The radiators have to be massive to provide enough physical
26:04
surface area for the infrared light to escape into the void.
26:07
So the hotter the chip, the better it radiates.
26:09
The harder the chip is allowed to get, the more efficiently it
26:12
radiates heat away. But these highly sensitive
26:14
silicon chips have strict thermal limits.
26:17
If they reach a certain temperature, the logic gates
26:19
fail and the chip is destroyed. So you have to keep them
26:22
somewhat cool. So you have to maintain a
26:24
relatively low operating temperature, which
26:27
mathematically dictates that the radiators have to be
26:29
exponentially larger. Scaling this mathematical
26:32
reality to the GW level data centers they're actually
26:35
proposing means deploying square kilometers of cooling fins into
26:39
orbit. We are talking about delicate
26:42
metallic structures the size of small cities floating in space,
26:45
existing solely to act as heat sinks for the processors.
26:49
The mechanical engineering complexity required to launch,
26:52
unfold, deploy and maintain square kilometers of fragile
26:56
cooling fins in an environment filled with micrometeoroids and
27:00
dangerous orbital debris is staggering.
27:02
Just one little piece of debris could ruin it.
27:04
A single piece of space junk traveling at orbital velocity
27:08
could sever A coolant line. You have to pump liquid coolant
27:11
through miles of delicate piping exposed to the harsh radiation
27:14
and vacuum of space, constantly circulating the fluid back to
27:18
the server core. This physical constraint
27:20
severely limits what orbital computing can actually achieve
27:23
in practice. It forces a complete shift in
27:25
strategy. It really does.
27:26
Space based data centers will likely be restricted to specific
27:30
inference tasks rather than replacing the massive, tightly
27:33
coupled terrestrial clusters needed for core AI training.
27:36
Training a Frontier artificial intelligence model from scratch
27:39
requires hundreds of thousands of chips communicating with each
27:42
other instantaneously, sharing data across fiber optic cables
27:47
with 0 latency. And you can't do that in space.
27:49
The latency requirements and the thermal management necessary for
27:52
a month long training run are virtually impossible to achieve
27:56
in an orbital environment with our current materials science
27:59
and launch capabilities, but inference, which means using an
28:03
already fully trained model to analyze new data, is much more
28:08
feasible. If an observation satellite is
28:10
taking high resolution multispectral photos of the
28:13
Earth's surface, instead of beaming terabytes of raw image
28:16
data down to a terrestrial ground station which consumes
28:19
massive amounts of bandwidth, the orbital data center can
28:22
analyze the images directly in space.
28:25
The model can identify the relevant information, like a
28:27
specific crop yield or a fleet of ships, and only beam down a
28:31
few megabytes of text summarizing the findings.
28:33
That solves a massive bandwidth bottleneck for satellite
28:36
communications. The Orbital data Center
28:39
effectively becomes an edge computing node, processing the
28:43
heavy data directly where it is generated.
28:45
It is a highly specialized, valuable application for
28:49
military and commercial intelligence, but it absolutely
28:52
does not solve the fundamental terrestrial power crisis for the
28:57
companies trying to train the next massive generation of
28:59
language models. The insatiable demand for
29:01
autonomous coding tools has pushed terrestrial power grids
29:04
to their absolute limits. This reality is forcing fierce
29:08
competitors to share hardware and driving engineers to look to
29:11
the vacuum of space just to kept the servers running.
29:15
If our software requires supercomputers the size of
29:17
appliance factories just to function today, how much of the
29:21
Earth's surface will we need to dedicated to cooling fans
29:23
tomorrow? If you're not subscribed yet,
29:25
take a second and hit follow on whatever app you're using.
29:27
It helps us keep making this. We appreciate you being here.
29:31
Also, check out our YouTube channel for more business and
29:33
tech updates. There's a link in the
29:34
description.