Latest / Elon Musk Podcast / Musk says Tesla's mega AI chip fab project to launch in seven days

Transcript

Musk says Tesla's mega AI chip fab project to launch in seven days · Elon Musk Podcast

0:00
Tesla is planning a $20 billion facility called the Terafab
0:03
Project to manufacture its own 2 nanometer AI chips, aiming to
0:08
produce up to 200 billion AI and memory chips annually.
0:13
I mean the sheer scale of that hardware target, it just
0:16
requires a complete recalibration of how we view
0:19
this company. We are looking at an objective
0:21
of 100,000 wafer starts per month.
0:25
And a wafer, just for context, is that large completely pure
0:28
silicon disk that they, you know, carve the individual chips
0:32
out of? Right exactly.
0:34
And 100,000 a month means they want to operate at the absolute
0:37
limits of global volume. It is one single organization
0:41
attempting to own every conceivable layer of artificial
0:44
intelligence. Yeah, they want to control the
0:45
physical silicon beneath the systems, the data processing
0:49
pipeline and the highest level software reasoning.
0:51
And they're attempting this massive vertical integration
0:54
while facing, frankly, immense skepticism from seasoned
0:57
semiconductor industry veterans. Which brings us to the core of
1:00
all this Can a company with absolutely no history and
1:03
semiconductor manufacturing successfully run the most
1:06
advanced chip factory on the planet?
1:08
To figure out if they can, we really have to look closely at
1:11
the specific hardware they intend to build.
1:13
The planned AI 5 chip is engineered to pack 40 to 50
1:18
times more compute performance than their current AI4
1:21
iteration. Wow. 40 to 50 times.
1:23
Yeah, and it's designed to hold 9 times more memory.
1:27
So we're talking about highly optimized inference hardware
1:30
intended to serve as the physical brain for the full self
1:33
driving software, the cyber cab robot taxi program and you know,
1:37
the optimist humanoid robot. So they're trying to build the
1:40
physical engine that makes their software completely autonomous,
1:44
operating without needing to constantly ping external data
1:47
centers. Designing an inference engine
1:49
like that is an incredible feat on its own, but transitioning
1:53
from designing those chips to like physically manufacturing
1:56
them is an entirely different discipline.
1:58
Oh, absolutely. Think of the difference between
1:59
writing a recipe and actually running a three Michelin star
2:02
restaurant. OK, writing the instructions for
2:04
a complex dish requires immense culinary skill.
2:07
You have to understand the flavor profiles, the chemistry
2:10
of the ingredients, the exact timing.
2:12
And you can test it in a private kitchen right until it is
2:15
absolutely perfect. Exactly.
2:17
The recipe is just the architecture.
2:20
But executing that dish, Florida State, thousands of times a
2:23
night involves managing a massive supply chain, the
2:26
intense heat of the kitchen, unpredictable human error and
2:30
just a million microscopic variables that the recipe never
2:33
explicitly covers. That's a great way to look at
2:36
it. Designing a chip is writing the
2:37
recipe. Building A2 nanometer Fab is
2:40
running the entire restaurant under the highest possible
2:43
pressure and. Controlling that incredibly
2:45
complex manufacturing process completely eliminates the supply
2:49
chain vulnerabilities that competitors face.
2:52
Like companies like Waymo and GM's Cruise are dependent on 3rd
2:56
party silicon. They literally have to wait in
2:58
line for their hardware. Yeah, by making the chips
3:00
internally, that vulnerability just vanishes.
3:03
If you own the factory, you control the pace of your own
3:05
innovation. It even opens up the possibility
3:08
of becoming a supplier to entirely different industries
3:10
down the line. But we have seen similar massive
3:13
manufacturing goals from them before, specifically with the
3:16
4680 battery cells. Right, the batteries.
3:19
The initial promise there was massive in house cell production
3:23
reaching 100 GW hours along with like a 56% cost reduction.
3:28
The stated goal is to deliver five times more energy, 6 times
3:31
more power and a 16% range increase.
3:35
All of this was supposed to enable a much cheaper $25,000
3:39
electric vehicle by just mastering the physical
3:42
production of the batteries. Well, Becca.
3:44
Are we saying they failed at battery manufacturing or just
3:47
missed the targets? They missed by a very wide
3:49
margin. Production only reached roughly
3:51
20 GW hours, which is a mere fraction of the original target.
3:55
The core issue was the dry electrode process, which proved
3:59
incredibly difficult to scale up.
4:01
How does that process actually work?
4:03
I mean, what makes it so hard? Well, traditional battery
4:05
manufacturing uses a wet process.
4:07
You mix solvents into a chemical slurry, coat the metal foils
4:11
with it, and then run those foils through these massive
4:13
ovens to Bake Off the liquid. And those ovens take up huge
4:17
amounts of factory floor space. OK, so the DRY process tries to
4:21
fix that. Right, the dry electrode method
4:24
attempts to skip the solvent and the ovens entirely, mixing dry
4:28
powders directly onto the metal foil.
4:30
It saves immense space and energy.
4:32
That sounds great in theory. In theory, yes, but getting
4:36
microscopic powder to stick uniformly to a metal sheet
4:39
without clumping or flaking off is exceptionally difficult.
4:43
Imagine trying to spread dry flour perfectly evenly across a
4:46
baking sheet without any water or butter to bind it.
4:49
Oh man, yeah, that sounds like a nightmare.
4:51
When they couldn't perfect it for both sides of the battery,
4:54
they had to rely on traditional suppliers for the cathode
4:57
materials. So the cheaper vehicle still
4:59
does not exist, and the batteries primarily went into
5:02
the cyber truck, which is struggled commercially.
5:04
Ultimately a major supplier, LNF Company, wrote down its deal by
5:08
99%. Wow.
5:10
So the difficulties in scaling battery manufacturing, a field
5:13
closely related to their core automotive business, lead
5:16
directly into the challenges of an entirely new discipline,
5:19
because fabricating silicon at A2 nanometer scale is governed
5:23
by a completely different set of physical laws.
5:25
Yeah, exactly. And look, they previously built
5:28
a highly capable chip design team, recruiting heavyweights
5:32
like legendary chip architect Jim Keller and Peter Bannon from
5:35
Apple's design team. That group successfully designed
5:38
the Hardware 3. Hardware 4 and Dojo training
5:41
chips. Which was a genuine achievement
5:44
in silicon architecture. It was.
5:45
However, much of that talent is now gone.
5:48
Key leaders left to launch their own startups.
5:51
The Dojo project lead, Ganesh Venkataramanan departed and took
5:55
about 20 Dojo team members to a new startup called Density AI.
5:59
Yeah, and eventually the entire Next Generation Dojo project was
6:03
killed, and Peter Bannon left as well.
6:06
The architectural minds who laid the groundwork for their current
6:08
silicon are largely absent from this new manufacturing push.
6:12
And there is a vast difference between design talent and
6:15
manufacturing talent. You don't just, you know,
6:17
reassign software engineers to run a fab.
6:20
You need process engineers who specialize in lithography,
6:23
etching and chemical mechanical planarization.
6:25
Which, for those of us who don't run chip factories, basically
6:28
means sanding down the silicon at a microscopic level so it is
6:31
perfectly flat. Right.
6:32
And then you have to operate extreme ultraviolet or EUV
6:36
equipment. And EV machines are incredibly
6:38
complex. They work by blasting
6:40
microscopic droplets of tin with lasers to create plasma.
6:44
That plasma emits light that is bounced off incredibly flat
6:47
mirrors to carve patterns smaller than a strand of human
6:51
DNA. It's mind blowing.
6:53
It really is, and that requires decades of institutional
6:56
knowledge that this company has never employed.
6:58
Which brings us to a major disconnect regarding the
7:01
approach to that manufacturing environment.
7:03
Elon Musk claimed he could eat a cheeseburger and smoke a cigar
7:06
inside A2 nanometer Fab because the wafers would be fully
7:09
contained. Right, I saw that.
7:11
The argument is that the semiconductor industry gets
7:14
clean rooms wrong by over engineering the spaces outside
7:17
the specialized machines. He believes that as long as the
7:20
wafers are transported in sealed pods, the surrounding factory
7:24
floor does not need traditional purification.
7:26
But that logic directly contradicts the fundamental
7:29
physics of semiconductor manufacturing based on the
7:33
source material. Modern leading edge fabs require
7:36
ISO Class 1 to 3 standards for a very specific reason.
7:40
In an ISO Class 1 clean room, the air is filtered hundreds of
7:44
times per hour. Yeah, the purity levels are
7:46
insane. Because human breath alone
7:48
introduces millions of contamination particles that
7:50
destroyed nanometer scale chips. We are talking about transistors
7:54
that are only a few atoms wide. If a single microscopic speck of
7:58
dust or you know, a crumb from that cheeseburger lands on A2
8:01
nanometer transistor, it ruins the entire chip.
8:04
Not to mention the cigar. Exactly.
8:06
Smoking a cigar in that environment would release
8:08
organic contamination that would permanently damage those
8:11
incredibly sensitive EUV mirrors used to direct the lasers.
8:14
The contamination risk is absolute, and the chemistry
8:17
inside a fab is highly volatile. You cannot simply isolate the
8:21
wafers and ignore the surrounding atmosphere when the
8:23
machines themselves must be opened, maintained and
8:26
calibrated by human workers. So making chips is hard, Really
8:29
hard, yeah. And this approach has drawn
8:31
explicit warnings from industry leaders.
8:34
NVIDIA CEO Jensen Huang explicitly stated that matching
8:38
the manufacturing capabilities of TSMC is virtually impossible.
8:42
TSMC spent decades and 10s of billions of dollars refining the
8:46
science and artistry of chip making.
8:48
Right. Intel has struggled for years to
8:50
regain its manufacturing edge despite having thousands of
8:53
experienced fab engineers and over $100 billion in
8:57
investments. Samsung's foundry business still
9:00
trails TSMC and yield rates at advanced nodes despite massive
9:04
investment that really limits the realistic expectations for
9:07
Terrafab's immediate success. So if the hardware manufacturing
9:10
is an uphill battle, they have to put massive pressure on their
9:13
software capabilities to close the gap with competitors.
9:16
Which is interesting because industry forecaster Peter
9:18
Wallford ranks XAI and Meta several months behind the
9:22
virtual tie for first place held by Entropic, Google and Open AI.
9:26
The analysis places them in a clear secondary here, behind the
9:29
models that currently dominate complex reasoning, advanced
9:32
coding and multimodal tasks. But the claim in response to
9:35
that ranking is that XAI will quickly catch up and then exceed
9:39
competitors by such a massive distance that you would need the
9:43
James Webb Space Telescope, which is stationed 930,000 miles
9:47
from Earth, to see who is in second place.
9:50
The James Webb Space Telescope. That's the quote.
9:52
The assertion is an astronomical leap that redefines the entire
9:56
hierarchy of artificial. I mean looking at the recent
10:00
models like Grok 4.2, they are visibly falling short in
10:05
reasoning and coding benchmarks compared to the front runners.
10:08
It draws a clear parallel to the heavily delivered promises of
10:11
full self driving. We've heard claims of complete
10:14
autonomy being just around the corner for a very long time.
10:18
Reaching parity and artificial intelligence is 1 hurdle, but
10:20
asserting a lead of that magnitude seems disconnected
10:23
from the current benchmark reality.
10:25
Software development requires iterative problem solving.
10:28
Throwing more resources at a model does not guarantee a
10:30
sudden exponential leap in logic capabilities.
10:33
See, I look at it a bit differently.
10:35
You have to factor in the massive compute advantages XAI
10:38
could leverage through those Dojo supercomputers.
10:41
When we talk about training in AI, we talk about parameters,
10:44
which are essentially the digital brain connections the AI
10:47
uses to make decisions. OK.
10:49
So just raw volume? Exactly.
10:52
They have access to immense raw processing power that could
10:55
accelerate their training models far beyond what standard
10:58
external cloud resources allow. If they successfully integrate
11:02
their proprietary hardware with their massive data pipelines,
11:05
they could train larger parameters much faster.
11:08
The sheer volume of hardware advantage could theoretically
11:11
force a massive gap, even if the current benchmarks show a
11:14
deficit. Well, this escalating cometition
11:17
heavily intensifies the talent war right here in EU driving U
11:21
the cost of acquiring top tier machine learning engineers.
11:24
It also changes the focus of the global AI race, threatening to
11:27
leave international competitors from China and France further
11:30
behind as the resources required to compete just consolidate at
11:33
the very top. And to leapfrog these software
11:36
competitors, X AI and Tesla are merging their capabilities into
11:40
a single product that targets office workers.
11:42
Right. They call it Digital Optimist,
11:44
jokingly nicknamed Macro Hard. Macro hard Yeah, This is a
11:47
software based AI agent designed to automate complex office
11:51
workflows like accounting in HR by observing and replicating
11:55
human computer interactions. It is designed to act as a
11:59
virtual counterpart to the physical humanoid robots,
12:02
handling repetitive digital operations across an entire
12:04
enterprise. The idea is to basically divide
12:07
the labor. Software agents manage screen
12:09
based tasks while humanoid robots tackle physical ones.
12:13
It operates on a dual process architecture inspired by human
12:16
cognition, the specialized AI Access system One processing the
12:20
recent seconds of real time screen, video and keyboard or
12:23
mouse actions for extremely fast instinctive execution.
12:27
Then XA is Grok serves as System 2 acting as the master
12:31
conductor, providing high level reasoning, world understanding
12:34
and directional oversight. How does that actually work in
12:36
practice? Imagine system one as pure
12:39
muscle memory. It watches your screen at 60
12:42
frames per second and learns that when a specific text box
12:45
turns blue, it needs to move the mouse and click exactly there.
12:49
System 1 doesn't actually know it is doing taxes or processing
12:52
payroll, it just executes the physical digital movement based
12:56
on visual cues. Wow.
12:57
OK. And System 2.
12:59
System 2 is the brain overseeing the entire operation.
13:02
System 2 understands why the tax form requires specific
13:05
documentation, reads the context of the page and tells System One
13:10
where to navigate next. Hold on, why is a car company
13:13
building an AI agent to do HR paperwork?
13:16
It all connects back to the silicon.
13:17
Digital Optimus runs on the low cost AI4 inference chip, which
13:22
minimizes the need for expensive external resources.
13:25
They plan to deploy this by using millions of parked cars
13:28
equipped with AI4 to handle office work during downtime.
13:31
The cars Think about the sheer volume of computing power
13:33
sitting in parking lots around the world.
13:35
A car sitting in a driveway for 20 hours a day is essentially
13:38
just a very expensive idle computer.
13:41
They want to network all those idle comuters together.
13:43
They're also lanning dedicated units at Supercharger stations,
13:46
tapping into 7 gigawatts of available power.
13:50
Instead of building massive new data centers that require
13:52
entirely new power grids and cooling systems, they're
13:56
utilizing the compute hardware already embedded in consumer
13:59
vehicles and their existing charging infrastructure.
14:02
This opens up an entirely new utility for consumer vehicles.
14:06
It transforms idle cars into a massive distributed enterprise
14:09
workforce and vastly reduces the reliance on expensive third
14:13
party server farms. A vehicle parked outside your
14:16
house suddenly becomes a revenue generating node in a global
14:19
computational network. And funding and executing this
14:22
combined hardware and software ecosystem requires is a stable
14:26
business environment, which recently saw a massive obstacle
14:29
removed. And just a quick reminder for
14:31
you listening, we are strictly discussing the legal filings and
14:34
reports provided to us here, not taking political sides on the
14:37
feud between Musk and the Brazilian courts.
14:39
Looking strictly at the legal facts, Brazil's Supreme Federal
14:42
Court ordered the closure of a lengthy investigation into Musk
14:45
and the social media platform X. Right.
14:48
The inquiry had been examining whether the platform was used to
14:51
coordinate attacks against members of the judiciary and if
14:54
there was deliberate obstruction of justice regarding the
14:57
suspension of certain accounts. But Prosecutor General Paulo
15:00
Gonet recommended the closure after an exhaustive review.
15:04
The investigation found no evidence of fraudulent intent,
15:07
deliberate obstruction or attempts to coordinate attacks
15:10
against the judiciary. The irregularities identified
15:13
did not indicate any deliberate attempt to circumvent court
15:15
orders, which completely changes the legal standing of the
15:18
platform in that country. Authorities had previously
15:21
ordered a nationwide block, imposed daily fines reaching up
15:24
to $920,000, frozen Starlink accounts and seized roughly $3.3
15:30
million. They had even frozen financial
15:33
assets linked to SpaceX through Starlap to collect unpaid
15:35
penalties and established penalties of $50,000 per day for
15:40
regular citizens attempting to bypass the restriction using
15:43
virtual private networks. Resolving this clears a major
15:46
operational bottleneck in a massive market.
15:49
Brazil has roughly 17,000,000 users on X and over 1,000,000
15:52
Starlink subscribers. Removing that legal and
15:55
financial pressure allows resources to be focused back
15:58
onto the hardware and software expansions we just discussed.
16:02
A company attempting to build a $20 billion chip facility cannot
16:06
afford to have its satellite Internet and social media Ryan
16:09
News frozen in one of his largest global markets.
16:13
Yeah. The interlocking nature of these
16:14
companies means a financial block on one heavily impacts the
16:18
operational capacity of the others.
16:20
Exactly. So to sum this all up, Tesla is
16:23
attempting to own the entire artificial intelligence
16:26
ecosystem, from building unproven 2 nanometer chip
16:29
factories to turning parked cars into an automated digital
16:33
workforce. And if they actually managed to
16:36
network millions of parked cars to handle corporate accounting
16:39
and human resources data, it forces you to wonder what
16:42
happens to the global economy when the car sitting in your
16:45
driveway is actively doing someone else's accounting.
16:47
If you're not subscribed yet, take a second and hit follow on
16:50
whatever app you're using. It helps us keep baking this.
16:52
We appreciate you being here.