So, I made this video uh or I did this

research kind of out of my own curiosity

and I realized that it would make a good

video. And basically, the question is

why isn't AI going faster and how can we

accelerate it faster? But really, this

is just a deep dive into what is the

holdup with artificial intelligence as

of the beginning of 2026. So, let's get

into it. The TLDDR upfront is that

everything uh the the biggest kind of

barriers is the thermodynamic wall. of

power, grid interconnection, physics,

that sort of thing. This comes down to

the grid capacity. This comes down to

transformers on the grid. This comes

down to actual generation capacity. So

basically everything to do with energy

is there. Then there's the supply which

is also structural. Um mostly the the

co-as so the chip on wafer on on

substrate. So if you see the COS, this

is basically the whole package getting

delivered. that is now the the actual

bottleneck and so this is why like

memory is sold out and and that sort of

thing. So then you have operational

friction which is further down the

stack. So this is enterprise execution,

data quality and ROI and then finally uh

things like AI safety and security and

public opinion that's actually just

noise. So when you have people saying oh

well you have to prove that AI is safe.

It's like actually and I mentioned this

in in other videos when and I said that

like serious researchers, policy people,

nobody takes extra seriously anymore. Um

it's basically just been

operationalized. The actual conversation

sphere has not caught up with reality.

Um which is really interesting and it's

kind of frustrating and so it's it's

basically kind of become a dog and pony

show of like you know but Ellie Azer

Yukowski said this and it's like nobody

who actually has decision power actually

cares. Um the and one thing I will say

is that in in layer three here is most

interestingly is insurance and we'll

talk about that but insurance doesn't

know how to price AI risk and that's

actually become more of a source of

friction than people being like AI is

just slop nobody cares that AI is slop

from a from an acceleration standpoint

so next infinite incentives abundant

capital what I have always said is that

acceleration is the default policy and

there's a lot of reasons that

acceleration is the default policy first

and foremost most being geopolitical

competition between America and China.

Uh all the great powers know that

artificial intelligence is the defining

technology of the next century or at

least the next few decades. And so

looking at the past uh as as guidance

for the future,

America and China and to a lesser extent

Russia, although they're not really a

player in this, but they all know that

artificial intelligence is the way to

go. Then you have market competition

which is basically the gold rush of all

the big tech companies wanting to get in

on artificial intelligence. So you say

okay well how much is being spent? So

$350 billion annual uh spend on on

hyperscalers. So that's um basically

data centers. $22 billion projected um

2025 VC funding that'll probably be

higher in 2026 which amounts to about

1.9% of GDP rivaling the Manhattan

project. So we are basically living

through a decentralized Manhattan

project. Um so the interstate highway

system, the Apollo program, the

Manhattan project, the 2025 and 20 and

beyond AI buildout is similar in scale.

We have achieved that. So the idea

however has been if we need it we can

buy it. The assumption that money

instantly dissolves physical

bottlenecks. It's just the bottlenecks

are not where we thought that they would

be. So that's the that's the level of

spend that we're seeing and the economy

is already restructuring around these

new problems. So how does that look?

Energy is the hardest bottleneck. This

is the hardest stop. So the average US

interconnection weight is 7 years. So if

you want to build a new data center, you

have to wait on average 7 years to

connect that new data center to the

grid, which is insane because that would

put us in 2032 or 2033 by that by which

point we should be at like super

intelligence and beyond. So the demand

is is expected to grow from 4 gawatt

total of data centers. Um, and when you

when you say 4 gawatts total in 2024,

consider that Microsoft and Meta and XAI

and OpenAI are all trying to spin up one

to two gawatt data centers, multiple

each of them. So just in the next few

years, they're they're wanting to more

than double what the current demand was.

Uh, and so the aggregate demand is

expected to hit 134 gawatts of data

center over the next 5 or 6 years. uh or

I guess the six years between 2024 and

2030. It's 2026 now. So we are on track

to have 134 gawatts of demand. For

comparison, your average run-of-the-mill

nuclear reactor produces about 1 gawatt.

Uh so that's a lot of juice. And I was

skeptical about these numbers. I'm like,

it can't possibly be that much. But when

you look at the rate of buildout and the

amount of backlog, like yes, that this

is not hypothetical future demand. This

is basically what is on deck and what

people are trying to get permitted and

approved right now. Uh we're going to

hit that and we might exceed it. That's

of course speculation on my part. We

might not hit that especially if we find

new efficiencies. Uh but energy is one

of the biggest bottlenecks. Now there

are solutions that didn't make it into

this presentation. So like um micro

grids where you have a combination of

solar batteries um and natural gas

turbines on site. Uh that's one answer.

uh directly connecting nuclear power

plants to data centers so that they're

not attached to the grid. Um basically

the grid is the bottleneck. So a lot of

well a lot of what these companies are

doing is becoming their own utility

companies as well. And we've seen this

like um there's a there's a site in the

far east of Russia where there's a

natural there's a natural dam or not a

natural dam there they use hydroelectric

to power aluminum smel sme sme sme sme

sme sme sme sme sme sme smelting and the

energy is just used primarily for the

aluminum because it's really expensive

so that's the only reason it exists now

obviously we don't have Siberia in

America where we just have thousands of

miles of unused space to spread out like

that but we do have lots and lots of

desert so we can build solar in that

anyways riffing on a little side tangent

on where things are at in terms of

energy. So, next is physics. So, this is

this is the highest latency part. So,

the the idea of getting the grid, you

know, because we've got we've got the

whole United States to power. So,

getting the grid up to speed is going to

take a lot of time and a lot of work. Um

because that means new transmission

lines, new transformers, and that sort

of thing. the transformers themselves,

the delay, the lead time is up to 210

weeks, which I thought was nuts. 210

weeks. Um, so for com for for

comparison, uh, a year is what, 56

weeks, 54 weeks? I don't remember off

the top of my head, but that's basically

four years almost. Um, and these are not

the little transformers that you see at

like the power substation. These are the

gigantic transformers that you don't

see. These these are the house-sized

transformers. Um, and so we're we've got

a backlog of those in demand. And then

the nuclear delusion. So the idea is

that nuclear the earliest that nuclear

could get spun back up is like 2030s. So

what really is emerging from the data

and the research is that the crisis

window is going to be over the next

couple years. So 2026 through 2028 is

really when it's going to be the hardest

because uh as the infrastructure

projects ramp up and as the orders for

transformers ramp up, those will

eventually get solved over the next 2 to

3 years. But nuclear is not going to get

solved that quickly. Uh so small modular

reactors, they're going to be too late.

So what we really need is stuff that you

can spin up very quickly. So that's

natural gas turbines, that's solar,

that's uh grid scale batteries. And we

do have cheap grid scale batteries now,

which is the um uh iron air batteries.

So if you haven't heard of those,

they're not very efficient, but they're

dirt cheap and they last forever. Um and

when you don't when they don't need to

be portable, they can be really heavy.

So iron air batteries are probably the

way to go if I had to guess um as the as

the best grid scale uh battery

technology. Cuz a lot of people like,

"Yeah, but you need to cover base load."

It's like you don't really need to cover

base load if you have enough batteries.

Um, and that's, you know, solar, right?

A lot of people say, "Well, solar can't

do everything." Solar can do everything

when it's combined with grids scale

batteries. So, moving on, the supply

chain. So, this is this is this is

interesting. GPUs are no longer the

bottleneck. Um, so the logic dies. So,

that's that's the actual wafer. That's

no longer the bottleneck. High bandwidth

memory is now the bottleneck. So, this

is sold out through the end of the year,

which is why people are making memes

about how expensive memory is. Um, and

part of the reason for this is that um

all the all the companies that can make

memory, instead of making memory for

your laptop and your desktop, they're

saying, "Well, we can make a lot more

money on memory for uh for the for the

uh accelerator. So, we're going to make

memory for that instead." Um, so this is

basically a price signal saying, "Hey,

we need a lot more memory, so the people

who can make it are abandoning it." So

the legacy DDR3 and DDR4 they're

abandoned. Um and that means that

robots, autos, laptops and that sort of

thing are all facing memory crunches. AI

nodes uh so that the hyperscalers are

consuming all of the available capacity

to build memory and then Nvidia books

greater than 50% of all cos capacity. So

that's the chip on wafer on substrate.

So what we're seeing here is the entire

package. So if you hear packaging, this

is the the the GPU, the onboard memory,

the high bandwidth memory, and then the

rest of the substrate that you put it

on. Uh and Nvidia is booking most of it.

So Nvidia is dominating this space as

well. It will take time. You know,

remember a year or two ago, we said,

"Oh, we're going to have a GPU

shortage." Well, it took a year and a

half to two years to figure out the GPU

shortage shortage, sorry. So now it'll

take another year and a half to two

years to sort out the bandwidth and

substrate shortage. There we go. So

those are again the the market will fix

these problems. That's kind of what I

want to drive home here is uh energy.

This is this requires state

intervention. This requires lots of uh

cutting through red tape. This is not

something that the market on its own

will sort out uh post haste. This needs

regulatory help. This is purely market.

So the the the the processors and the

memory the market will sort that out.

It'll just take time. Now the other

bottleneck is data. So people keep

talking about data. Uh and you know the

the thing is I'm not convinced by this

argument because we are seeing a very

very sharp rise of synthetic data. Um

however it is still a concern because AI

trained on its own data does tend to

result in model collapse. But the latest

generation of AI models are trained on

more and more synthetic data. So I'm not

convinced that this is actually going to

be a bottleneck. Now at the same time,

it is true that the amount of raw human

data we're going to run out of it. But

as many of you have pointed out over the

past, humans do a lot more with a lot

less data. Like a human brain is trained

on like less than 1% of the of the

amount of data that a uh that an AI

model is trained on. And we're still

better. We're still smarter. So, we are

clearly doing something different. We're

not getting the most out of the data

that we have. So, and I always say like

necessity is the mother of invention,

but constraints are the father of

creativity. So, if we run out of data,

I'm convinced that we will find better

algorithms to make better use of the

data that we do have because we clearly

have enough data to generate, you know,

super intelligent human level

performance. Again, if a human can

spend, you know, just 20 years or 30

years learning less than 1% of the data

that we do have available and advance

physics, then AI should be able to do

the same. So AI is just not that

efficient at learning. So we will get

better at making AI learn. So again, I

don't I don't really see this as a

barrier, but this is one of the things

that came up in the research that people

are concerned about. So I wanted to at

least include it. Now, the $600 billion

question. So this is the infrastructure

spend and the ROI gap the market

rotation. So investors are fleeing

infrastructure for software

productivity. This is the most

interesting thing where um investors are

getting more and more wary of you know

hey write me a check for hundred billion

to build data centers. When are the data

centers going to turn a profit? Oh maybe

5 to 8 years. Investors don't like that.

They want to see returns on investment

very quickly. Uh so the hyperscalers

that raised hundred billion in debt in

2025 the investors are starting to look

a scance at them which is why you see

Sam Alman going to places like Saudi

Arabia saying hey can you can you

finance our data centers [snorts]

the question can spending 20% of GDP

sustain on belief alone? This is this

goes back to the bubble fears. Um, so

you know, yes, some of the most

profitable companies on the planet

already are using AI. Although you might

say, well, where's the evidence? We're

in Solo's paradox or the J curve of

productivity where AI is is uh the

profusion is happening. AI is getting in

into the market and it's saturating.

But, you know, while most companies are

using AI, it's not saturated to the

degree that you'd want it to be

saturated because it's simply still too

expensive. In the same way that in the

80s the personal computer was not as

mature as it is today. Um you know it's

like okay you have a PC on your desk and

it you know has a a monochrome screen

that can connect to a central database.

Great. That's definitely more productive

but it's not nearly as productive as

even just a cell phone is today. We are

at that era where yes it's it's there.

It's useful but it's not saturated. So

it'll take another few years to really

find all the uses to saturate the AI,

but in the meantime scaling up the

production of the AI. So we've got like,

you know, this log jam between it's not

as mature as it could be. It's still

expensive to run. And this is also very

similar to early internet. Um where, you

know, when you still had like dialup and

ISDN and that sort of thing. Um then you

had to or I no ISB in his book. um the

ISDN I'm remembering incorrectly DSL

basically digital subscriber lines um so

when the internet was new it was useful

but it was still too too expensive

because the the cost per packet was too

high now of course the internet is

basically free you just have a flat

monthly subscription and you get as much

internet as you need and it's at gigabit

plus speeds um so it'll take a it'll

take a little bit of time for us to

build up that much infrastructure until

the point that AI is just kind of a

background expense and you're not

thinking about the cost per token. So,

and in the meantime, the the enterprise

ROI is not necessarily manifesting as

quickly as they would like, although

most business leaders understand that AI

is very competitive and compelling and

and that they need to not be asleep at

the wheel. So, then of course the ROI is

about, you know, why 88% of pilots fail.

And of course, this is different from

the 95% of pilots that fail um from the

MIT study. I've talked about this plenty

of times, but I'll I'll go over it

briefly, which is basically uh most

pilots fail. Uh AI is no different.

That's the point of a pilot is to see

can we realize value. Um so you start

with 100% of AI pilots that you initiate

integration complexity and data mess and

then 12% reach production. So that's

still better than the 95% from the MIT

study. Uh but it's still the vast

majority fail. So what are the barriers?

One is number one is data quality. Um

and getting good data is always the

problem. I mean getting good data

integration and data governance is

always step one. Uh number two is

integration with legacy systems. This is

another thing is big companies are

running stuff that is 20 30 plus years

old. Um and it it many of them don't

even have modern APIs. Some of them

don't even have modern operating

systems. Uh some of them are running on

old operating systems like SCO Unix. Um

and then they also have a lack of

talent. So the lack of talent, the lack

of integration, the lack of being AI

ready, it has nothing to do with safety

or ethics. The barriers are very mundane

which is cost and ROI and then also

insurance oddly enough. Um so then when

we say what what actually gets talked

about. So the the vast majority of

public opinion about AI is safety and

you know uh AI art is crap and but what

about copyright and what about the what

about the actors and general regulation

about consumer protections? That's where

most people are talking. That's where

the conversation is. But the reality is

that it's like less than 5% like of what

of of what actually gets discussed. So,

Frontier Labs accelerate despite the

resignations that happen. Why? Because

there's a lot of money to be made if

you're if you're a talented AI engineer,

and there will be for a while. Um, one

of the numbers that I saw is that

there's globally there's about 22,000

high-end AI engineers out there, which

is just not enough. So, if you want to

make bank in the next few years, become

a crack AI engineer. Um, and and help

help with the acceleration. That's

that's one of the biggest bottlenecks

that any individual can help overcome.

B2B adoption ignores public sentiment.

So this is basically like, yeah, you

guys can whine about AI art and all this

other stuff and copyright infringement.

The businesses don't care. It doesn't

impact them whatsoever. And then US

federal regulation is largely noise. So

the EU is slowing things down, which is

like, okay, Europe, you're you're

hurting yourself to, I don't know, spite

America. I don't know. But um but the

regulation is not really slowing down

artificial intelligence. Neither is

safety discussions.

So however where regulation is the real

friction. So number one is the USChina

um uh rivalry. So there's export

controls and import controls and this is

if you take a step back and you say

globally what is the barrier to friction

and it is the compute gap. Already

America has like two to three to four

times or five times the amount of

compute that China has. But our compute

is pulling ahead of Chinese compute not

only in terms of quality like per unit

quality but overall volume which means

by 2027 America is expected to have 17

times the compute of China and that is

due to uh you know export controls which

from a geopolitical perspective that

means America is winning that's our

moat. Um now you might say globally

that's not a good thing because that

slows down Chinese research although as

I said constraints are the father of

creativity and the Chinese are very

creative. they're able to do a lot of

research with a lot less resources. Now,

when you look at the center for AI

security and innovation, which is uh is

that the one that is the subset of NIST?

Um I think I think that's the department

inside of NIST. Their studies basically

show that the Chinese models are

inferior across the board. Some of them

are cheaper, some of them are more

efficient, but in terms of uh in terms

of cyber security risk, they're much

higher. In terms of intelligence,

they're not as good. Um, so they don't

really represent a threat uh on a

military level. Then you've got the EU

AI Act. So this is the break. This is

the compliance wall. So the licensing

cost for any high-risk systems is 52,000

a year, which means that you're

basically going to stifle any startup.

No startups dealing with higher risk AI

use cases will be in Europe. And this is

already true. Um, and so where do they

go? They go elsewhere. They go to

America. Some of them go to China. Some

of them go to Saudi Arabia. Basically,

Europe is very good at ensuring that

frontier business does not happen in

Europe, which is kind of silly in my

opinion. It's like, you know, they're

they're very very proactive and there's

a lot of regulatory capture and and and

what I call a vetocracy. So, basically

the department of no, that's what Europe

is. And then finally, liability. This

was the most interesting finding from my

my research is that uh many insurance

policies have absolute AI exclusions.

Meaning that if you use AI and there is

um if there's an OSHA violation, if a

patient dies, then the insurance company

completely washes their hands. They say

if AI touched that incident, we are not

taking any responsibility whatsoever.

And the reason is because it's new. It's

high risk. It's high variance. They

don't know how to price it. I used to

work at a at a workman's comp insurance

company. When you have a and this was at

the very beginning of my career, when

you have a domain that is well

understood and the risks can be

controlled um and and and understood and

measured, then you then they know how to

price it. And so like we had a we had a

division of our of our insurance company

that would handle like the one-off cases

where it's like oh hey you're a

physician who operates out of a yacht or

something like that like you know cuz

you have onboard onboard physicians and

that is not as well understood as a

physician operating in say a hospital or

a clinic. So then it's like, okay, well,

how do you ensure a physician if they're

working under suboptimal conditions on,

you know, a really expensive yacht and

with really expensive clients, how do

you price that insurance policy? And so

you can price those things, but it takes

a lot of work and a lot of effort to do

a one-off, and that's not how you

operationalize and how you scale. So

likewise, when you have you, you know,

your your average corner store or your

mom and pop shop or your average

enterprise that really just wants a a a

templated insurance policy to say, "Hey,

you know, we're using AI for this. Make

sure it covers our customers. Make sure

it covers our own butts." The insurance

company says, "We don't know how to

price that." And so we're not going to

cover it, and we're going to ensure that

our policy has an exclusion saying, "We

will not be on the hook." which and if

if there's no insurance then the

companies then the enterprises simply

won't do it. Um and so this is this is I

think the most ironic uh kind of barrier

uh uh legal friction for adopting AI. So

if I have any insurance people any

insurance pros in the audience who want

to like help explain how we could make

this better, I think that would be great

because the this this to me is the

dumbest reason to slow down [laughter]

AI. Um, I get it, but it's it's like

really like we're just gonna we're

because because we don't know how to

price it. Like moving on. So what this

is being called this this phase that

we're in 2026 through 2028 is the

digestion phase. So the hype cycle was

2023 to 2025. Now we're catching up with

reality. So reality says, you know, we

need grid interconnects, we need

transformers, we need high bandwidth

memory, we need the chips and wafers, we

need verifiable synthetic data, and we

need to figure out the insurance

policies. So it's now from bigger models

and scale is all you need to efficiency

and distillation and make do with what

you've got. Just do the best with what

you've got. And that's the that's the

paradigm that we're going to be in over

the next couple years. Now, that's not

to say that no new capacity is coming

online. And obviously there's new

capacity coming online all the time. Um

but we're not the the the the distance

between what we would like to install

and deploy and where we can actually

deploy is still growing. So then uh

after 2028 that's when acceleration and

the the economic pivot will uh continue.

So the friction map so this is basically

a recap is the critical and binding

constraints are power availability grid

interconnection and HBM supply moderate

and addressable so that's data quality

again I'm not concerned about that the

packaging um that seems like it's going

to be solved by the market deployment

friction um deployment friction again

that it just takes time for enterprises

to learn how to deploy these things and

to operationalize it and then liability

the insurance liability this to me is

the silliest thing um the things that

are not uh uh a friction at least for

America is number one safety and x-risk,

number two federal regulation, and

number three capital availability. So

that's it's a really interesting place

to be. Um you know there is there is no

bubble to speak of. Um there are

frictions, but as I've talked about in

previous videos, the difference between

a typical bubble and what we're seeing

is everything is sold out, meaning we

still have unmet demand. With a bubble,

that's pure speculation. uh with with

this it's like everyone wants more AI

and AI is not even like at its prime

time yet. It's not even fully matured

yet and it and people want more than

than we can give them. So this is very

very different from a bubble and that I

believe is it. So Adams eat arguments to

accelerate AI stop arguing about

philosophy and start pouring concrete.

The future belongs to those who master

the physical world. grid permits, fab

capacity, energy generation, focus on

atoms, and then we have the sources and

references. So, with all that being

said, thanks for watching to the end,

and I'll check you all later. You are

now aware of why AI is going to go

slower than we would prefer over the

next few years, and it has nothing to do

with the actual research being done. The

research is continuing a pace. Now, we

are entering where the the phase where

the rubber meets the road. So, the the

friction with reality is the barrier

now. Cheers.