Developments
in
generative
AI
over
the
last
12
months
have
begun
to
transform
the
way
people
live
and
work.
Language
models
are
being
used
to
develop
legal
strategies
in
court
cases;
image
diffusion
models
are
being
used
to
augment
the
workflows
of
major
entertainment
studios;
and
computer
vision
advancements
have
brought
fleets
of
self-driving
cars
on
roads
enmasse.
It
is
well
understood
that
the
primary
bottleneck
to
scaling
these
systems
is
access
to
compute
resources.
Wait
times
and
hourly
rates
for
spot
instances
of
Nvidia’s
A100
and
H100
chips
have
consistently
trended
upward
throughout
2023,
and
chip
production
capacity
simply
cannot
keep
up
with
demand.
The
ongoing
shortage
of
graphics
cards
stems
from
a
perfect
storm
of
materials
constraints,
supply
chain
disruptions,
surging
demand,
geopolitical
tensions,
and
the
long
production
cycles
inherent
in
fabricating
complex
GPUs.
Additionally,
key
materials,
like
the
advanced
silicon
used
in
GPU
chips,
specialized
substrates
for
the
PCBs,
and
memory
chips,
are
all
also
facing
shortages
amid
supply
and
demand
imbalances.
Datacenter
revenues
associated
with
AI
workloads
were
approximately
$100
billion
in
2023.
Datacenters
require
tremendous
upfront
capex
in
the
form
of
land,
electricity,
and
enterprise-grade
hardware.
New
data
centers
rely
on
external
financing
for
setup
and
operation,
but
rates
are
high
and
capital
is
tight.
AI
models
are
increasing
in
size
and
complexity,
and
although
the
price
per
unit
of
computational
performance
halves
every
thirty
months,
AI-specific
compute
requirements
double
every
six
months.
Demand
is
on
track
to
increase
orders
of
magnitude
faster
than
supply.
This
is
something
that
investors
dream
about:
a
tectonic
shift
in
innovation
that
impacts
practically
every
business
overnight,
driven
by
a
finite
resource
and
skyrocketing
demand,
causing
commodity
prices
to
spike.
NVIDIA’s
YTD
returns
of
231.5%
over
the
last
12
months
are
a
perfect
proxy
for
this
—
but
even
that
fails
to
represent
the
opportunity
at
hand.
We’re
still
in
the
early
innings
of
the
AI
renaissance.
Every
Fortune
500
company
is
figuring
out
their
AI
strategy
right
now,
and
the
demand
we
see
today
is
nowhere
close
to
demand
we’ll
see
tomorrow.
AI
will
augment
and
displace
workforces,
drive
productivity,
and
fundamentally
reshape
how
businesses
operate.
Compute
is
the
new
oil.
GPUs
are
the
currency
of
AI,
and
DePINs
are
here
to
deliver
it
There
is
an
answer
to
the
increasingly
vast
compute
shortage
problem:
find
un-utilized
supply.
A
new
form
of
crypto
network,
called
“Decentralized
Physical
Infrastructure
Networks,”
or
“DePINs”
for
short,
is
coming
to
the
rescue.
It
is
estimated
that
there
are
1.5
billion
freely
available
consumer
GPUs
globally,
and
another
six
million
datacenter
GPUs
in
datacenters
deployed
worldwide
outside
of
the
hyperscalers
(AWS,
GCP,
Azure,
Oracle).
Consumer
hardware
cards
often
have
comparable
computational
throughput
to
enterprise
grade
cards.
For
example,
the
consumer-grade
RTX
3090
is
capable
of
83
FP32
TFLOPS
whereas
enterprise-grade
A100’s
only
have
19.5
FP32
TFLOPS.
Currently,
there
are
over
330
million
consumer-grade
GPUs
in
personal
computers
(gamers,
designers,
video
editors,
etc.)
and
datacenters
that
could
be
brought
online.
The
problem
is,
it
historically
hasn’t
been
possible
to
incentivize
or
coordinate
these
disparate
GPUs
into
usable
clusters.
Recently,
specialized,
AI-focused
DePINs,
such
as
Render
Network
and
IO.net,
have
solved
this
problem.First,
they
are
giving
latent
GPU
operators
incentives
to
contribute
their
resources
to
a
shared
network
in
exchange
for
rewards.
Second,
they
are
creating
a
decentralized
networking
layer
that
represents
disparate
GPUs
as
clusters
AI
developers
can
use.
These
decentralized
compute
marketplaces
now
offer
hundreds
of
thousands
of
compute
resources
of
varying
types,
creating
a
new
avenue
to
distribute
AI
workloads
across
a
previously
unavailable
cohort
of
qualified
hardware.
In
addition
to
creating
net
new
GPU
supply,
DePIN
networks
are
often
significantly
cheaper
–
up
to
90%
cheaper
–
than
traditional
cloud
providers.
They
achieve
these
costs
by
outsourcing
GPU
coordination
and
overhead
to
the
blockchain.
Cloud
providers
markup
infrastructure
costs
because
they
have
employee
expenses,
hardware
maintenance,
and
datacenter
overhead.
DePIN
networks
have
none
of
those
expenses,
thus
they
can
pass
along
compute
costs
practically
at
cost
(with
insignificant
network
coordination
fees
on
top)
to
end
customers.
As
we
look
to
the
year
ahead,
we
expect
these
decentralized
networks
to
emerge
as
one
of
the
key
players
in
the
AI
race.
There
are
simply
not
enough
GPUs
(much
less
affordable
GPUs)
right
now
to
service
the
demand
of
every
major
company
in
the
world.
GPUs
are
the
currency
of
AI,
and
DePINs
are
here
to
deliver
it.
