January 2022

AI and Compute DECEMBER 2021

How Much Longer Can Computing Power

Drive Artificial Intelligence Progress?
CSET Issue Brief

AUTHORS
Andrew J. Lohn
Micah Musser
Executive Summary

For the last decade, breakthroughs in artificial intelligence (AI)

have come like clockwork, driven to a significant extent by an
exponentially growing demand for computing power (“compute”
for short). One of the largest models, released in 2020, used
600,000 times more computing power than the noteworthy 2012
model that first popularized deep learning. In 2018, researchers at
OpenAI highlighted this trend and attempted to quantify the rate
of increase, but it is now clear this rate of growth cannot be
sustained for long. In fact, the impending slowdown may have
already begun.

Deep learning will soon face a slowdown in its ability to consume

ever more compute for at least three reasons: (1) training is
expensive; (2) there is a limited supply of AI chips; and (3) training
extremely large models generates traffic jams across many
processors that are difficult to manage. Experts may not agree
about which of these is the most pressing, but it is almost certain
that they cannot all be managed enough to maintain the last
decade’s rate of growth in computing.

Progress towards increasingly powerful and generalizable AI is

still possible, but it will require a partial re-orientation away from
the dominant strategy of the past decade—more compute—
towards other approaches. We find that improvements in
hardware and algorithmic efficiency offer promise for continued
advancement, even if they are unlikely to fully offset a slowdown
in the growth of computing power usage. Additionally, researchers
are likely to turn to approaches that are more focused on specific
applications rather than the “brute-force” methods that
undergirded much of the last decade of AI research. The release of
AlphaFold, which made incredible progress on a long-standing
problem in the field of biology without the need for record-
breaking levels of computing power, may be an example of this
new shift in focus.

Center for Security and Emerging Technology | 1

These findings lead to a few recommendations for policymakers:

• If continued AI advancement relies increasingly on improved

algorithms and hardware designs, then policy should focus
on attracting, developing, and retaining more talented
researchers rather than simply outspending rivals on
computing power.
• As a specific example of the above, we suggest that
institutions such as the National AI Research Resource
should not view computing power alone as the primary way
to support AI researchers. These institutions should also
invest in providing researchers with the skills to innovate
with contemporary AI algorithms and to manage modern AI
infrastructure, or should actively promote interdisciplinary
work between the AI field and other subfields of computer
science.
• Finally, policymakers should take proactive steps to ensure
that researchers with small or moderate budgets can
effectively contribute to the AI research field. Concentrating
state-of-the-art technologies among the small number of
research centers possessing extremely large compute
budgets risks creating oligopolistic markets and shrinking
the talent pool and opportunities for researchers.

Center for Security and Emerging Technology | 2

Table of Contents
Executive Summary......................................................................................... 1
Introduction........................................................................................................ 4
Modern Compute Infrastructure ................................................................. 7
Projecting the Cost and Future of AI and Compute .............................. 8
The Cost of Compute .................................................................................. 11
The Availability of Compute ..................................................................... 14
Managing Massive Models ........................................................................ 16
Where Will Future Progress Come From? ........................................... 20
Conclusion and Policy Recommendations ............................................ 23
Authors ............................................................................................................ 26
Acknowledgments ....................................................................................... 26
Endnotes.......................................................................................................... 27

Center for Security and Emerging Technology | 3

Introduction

In the field of AI, not checking the news for a few months is
enough to become “out of touch.” Occasionally, this breakneck
speed of development is driven by revolutionary theories or
original ideas. More often, the newest state-of-the-art model
doesn’t rely on any new conceptual advances at all, rather just a
larger neural network and more powerful computing systems than
were used in previous attempts.

In 2018, researchers at OpenAI attempted to quantify the rate at

which the largest models in AI research were growing in terms of
their demands for computing power (often referred to as
“compute” for short).1 By examining the amount of compute
required to train some of the most influential AI models over the
history of AI research, they identified two trend lines for the rate of
compute growth.

They found that prior to 2012, the amount of compute used to

build a breakthrough model grew at roughly the same rate as
Moore’s law, the long-standing observation that the computational
power of an individual microchip has tended to double every two
years. In 2012, however, the release of the image recognition
system AlexNet sparked interest in the use of deep learning
methods—the computationally expensive methods that have been
behind most of the AI advances of the past decade. Following the
release of AlexNet, the compute demands of top models began to
climb far faster than the previous trend, doubling not every two
years but rather every 3.4 months between 2012 and 2018, as
visualized in Figure 1.

The largest models in the early years of deep learning were

devoted to image classification, where researchers quickly realized
that increasing computing power reliably led to better
performances.2 After image recognition systems began to surpass
human-level performance on some tasks, research shifted to new
priorities even as the same trend in rising compute needs
continued. Around the middle of the 2010s, larger AI models were
playing games like Atari or Go using reinforcement learning
algorithms.3 Then, the emergence of a new architecture known as

Center for Security and Emerging Technology | 4

the transformer shifted attention again—this time to language
tasks.4 Over the past few years, the largest AI models have been
text generators like OpenAI’s GPT-3.5 Even with improvements to
algorithms and architectures enabling them to do more learning
with fewer calculations, the computing demands continued to
expand. That same 3.4-month doubling time for compute needs
has continued more or less uninterrupted from AlexNet to GPT-3.

Figure 1: Growth in compute demands over the past decade far

outpaces the historical norm

Source: OpenAI and CSET.

This compute demand trend only considers the most compute-

intensive models from the history of AI research. The most
impactful models are not necessarily the largest or the most
compute-intensive. Most AI projects are much smaller than these
large efforts, and even some famous breakthroughs, such as
AlphaFold, used more modest computing power. However, several
of the most well-known breakthroughs of the last decade—from
the first AI that could beat a human champion at Go to the first AI
that could write news articles that humans mistook for human-
authored text—required record-breaking levels of compute to
train.

These massive models tend to be adaptable in addition to being

capable, which means they can form the foundation for a wider

Center for Security and Emerging Technology | 5

range of applications and studies because of their general purpose
nature. Some researchers have begun referring to models like
these as “foundation models” and have suggested that the next
wave of AI research will emphasize such approaches.6 Whether or
not this framework is accepted, however, it is clear that with
sufficient compute, models can be developed that acquire skills far
beyond what they were explicitly trained to do. GPT-3, for
instance, not only learned how to write realistic-looking text—it
also learned how to generate passable programming code and
even rudimentary music compositions, despite not having been
explicitly intended to do so.

While we believe that the computing used in future models will

continue to grow, the current growth rate for training the most
compute-intensive models is unsustainable. 7 We estimate that the
absolute upper limit of this trend’s viability is at most a few years
away, and that, in fact, the impending slowdown may have already
begun. The implications of this finding are significant, as it means
that the future of AI progress will likely rely more on algorithmic
innovation and applications than simply scaling-up compute
usage. If correct, this projection could affect the tools at
policymakers’ disposal for promoting AI development.

Center for Security and Emerging Technology | 6

Modern Compute Infrastructure

In order to understand the sustainability (or unsustainability) of the

compute growth trend, it is helpful to understand the current
compute landscape.

GPT-3 and similar models such as the Chinese PanGu-alpha,

Nvidia’s Megatron-Turing NLG, and DeepMind’s Gopher are the
current state of the art in terms of computing appetite.8 Training
GPT-3 in 2020 required a massive computing system that was
effectively one of the five largest supercomputers in the world. 9

For large models like these, compute consumption is measured in

petaFLOPS-days. One petaFLOPS-day is the number of
computations that could be performed in one day by a computer
capable of calculating a thousand trillion computations
(specifically, floating point operations) per second. For comparison,
a standard laptop would need about a year to reach one
petaFLOPS-day.10 That laptop would need several millennia to
reach the 3,640 petaFLOPS-days it took to train GPT-3. On the
world’s hundredth-fastest supercomputer, GPT-3 could be trained
in two and a half years, and even on the world’s fastest
supercomputer, training would still take over a week.*

High-end AI supercomputers require special purpose accelerators

such as Graphics Processing Units (GPUs) or Application-Specific
Integrated Circuits (ASICs) such as Google’s Tensor Processing
Units (TPUs) or Huawei’s Ascend 910. These accelerators are
specialized hardware chips that are optimized for performing the
mathematical operations of machine learning. They are managed
by many general purpose computer chips (primarily Central
Processing Units, or CPUs) and data is passed using high
bandwidth interconnections. 11 The accelerator chips are common
for training even relatively small models, though inference—using
the model after it is trained—is a much simpler task that typically

* At the time of writing, the world’s fastest supercomputer (Fugaku) can

compute 442 petaFLOPS. The tenth and hundredth fastest can compute 23.5
and 4.1 petaFLOPS respectively. “TOP500 List – June 2021,” TOP500, accessed
December 4, 2021, https://www.top500.org/lists/top500/list/2021/06/.

Center for Security and Emerging Technology | 7

 uses fewer special accelerators or uses accelerators that are
specialized for low power, including another class of chip called
Field Programmable Gate Arrays (FPGAs). 12 All of this hardware
can be purchased outright, or—as is common for AI research—
rented through cloud services such as Google Cloud, Amazon Web
Services, Microsoft Azure, or IBM Cloud.

Cloud services offer economies of scale by sharing maintenance

personnel, building space, cooling, and other operational
necessities across many projects. Because these expenses are all
included in the cloud computing costs that researchers pay, we use
cloud computing costs to estimate how expensive the compute
demand trendline is both today and in the future. Given the price of
purchasing compute through the cloud, how much longer can this
growth trend continue, and when will the exponential growth
trend in compute demand become non-viable?

Table 1: A basic comparison of AI processors

Processor Type Uses in the AI Pipeline Other Uses

Central Processing Small models can be directly trained or Central unit of every
Unit (CPU) fine-tuned on CPUs; necessary in larger computing device; at least
models as a means to coordinate training one CPU is necessary for
across GPUs or ASICs. Sometimes every computer, phone,
needed to generate or manipulate smart appliance, etc.
training data.

Graphics Processing Optimized to perform certain Used for video game

Unit (GPU) mathematical operations that are also systems to render 3D
common in machine learning; can train graphics; commonly used for
models far quicker than CPUs cryptocurrency mining

Application-Specific Designed specifically for AI, to perform If designed specifically for AI

Integrated Circuit the types of matrix operations that are algorithms, no major uses
(ASIC) the bedrock of machine learning; can beyond the AI pipeline
train models far quicker than CPUs

Field Programmable Primarily used for model inference using Used in a wide variety of
Gate Array (FPGA) AI models that have already been trained applications, particularly in
embedded systems

Source: CSET.

Center for Security and Emerging Technology | 8

Projecting the Cost and Future of AI and Compute

One possible constraint on the growth of compute is expense.

Throughout this paper, we do not consider the compute needed to
generate or prepare the data, instead restricting analysis to the
training process itself. We use Google’s TPUs as a baseline to
calculate the expected cost of compute. These TPUs are among
the most cost-efficient accelerators being advertised at the time of
writing, though we obtain similar results when using state-of-the-
art GPUs in our calculations. 13 For these calculations, we make two
simplifying assumptions here that we will relax in the next two
sections: for now, we assume that (1) the cost of compute remains
constant, and (2) the only constraint on GPU and ASIC access is
willingness to pay (and not the number of physical chips that
actually exist in the real world).

With these assumptions in place, we can begin to make some

rough estimates. At the advertised maximum performance of a
Google TPU v3, it would take approximately 57 hours and cost
approximately $450 to reach one petaFLOPS-day of training. GPT-
3 required approximately 3,600 petaFLOPS-days to train, which
works out to a cost of around $1.65 million if trained on TPUs
performing continuously at their maximum speeds.* Even that was
somewhat less than the 20,000 petaFLOPS-days that the
compute demand trendline anticipated for the largest model as of
the day GPT-3 was released, which would cost $9.4 million at
current prices. By the end of 2021, the trendline predicted several
more doublings, for an anticipated model of just over one million
petaFLOPS-days. Training such a model at Google Cloud’s current
prices would cost over $450 million.

*
OpenAI did not release the actual costs, but estimates are typically higher than
ours because we have made conservatively low pricing assumptions and
assumed 100 percent accelerator utilization. An estimate of about $4.6 million
is probably more accurate. We use conservatively low cost estimates to ensure
that we do not overstate the rising cost of training models and the impending
slowdown in compute growth. Chuan Li, “OpenAI’s GPT-3 Language Model: A
Technical Overview,” Lambda Labs, June 3, 2020,
https://lambdalabs.com/blog/demystifying-gpt-3/.

Center for Security and Emerging Technology | 9

That is a large sticker price, to be sure, but not unobtainable.
Governments have, in the past, paid much more to fund basic
scientific projects. The National Ignition Facility (NIF) is said to
have cost $3.5 billion, finding the Higgs Boson was estimated to
have cost $13.25 billion, and the Apollo program’s annual
expense of 2.2 percent of gross domestic product (GDP) would be
about $450 billion today.14 But the trendline that described the
growth of AI models over the past decade quickly blows past
these benchmarks too, costing as much as the NIF by October
2022, the search for the Higgs Boson by May 2023, and
surpassing the Apollo program in October of 2024. In fact, by
2026, the training cost of the largest AI model predicted by the
compute demand trendline would cost more than the total U.S.
GDP (see Figure 2, below).*

Actually spending a U.S.-GDP-worth of money to train a single

mega-powerful AI model is highly unlikely. Indeed, even spending
as much as the entire search for the Higgs Boson to train a single
model seems improbable in the near term. This suggests that the
compute demand trendline should be expected to break within
two to three years at the latest, and certainly well before 2026—if
it hasn’t done so already.

*
Our calculations for reaching this conclusion, along with our calculations for
other figures in this and the next two sections, can be found in our GitHub
repository: https://github.com/georgetown-cset/AI_and_compute_2022.

Center for Security and Emerging Technology | 10

The Cost of Compute

While this projection seems pessimistic, the reader might object

that the cost of compute is not fixed in the way that we have
assumed. To explore the extent to which falling computing prices
can extend the viability of AI’s compute demand trendline, we
consider the historical trends in cost of computing.

The price of computations in gigaFLOPS has not decreased since

2017.15 Similarly, cloud GPU prices have remained constant for
Amazon Web Services since at least 2017 and Google Cloud since
at least 2019.16 Although more advanced chips have been
introduced in that time—with the primary example being Nvidia’s
A100 GPU, released in 2020—they only offer five percent more
FLOPS per dollar than the V100 that was released in 2017. *

During this period, manufacturers have improved performance by

developing chips that can perform less precise computations
rather than by simply performing more of them. A full floating
point operation, or FLOP, uses 32 bits for each number, but deep
learning methods do not always need that much precision and can
run faster without it. For example, GPT-3 only used half-point
precision, which requires half as much memory and can be
computed faster. Nvidia has used techniques to further reduce
precision that have allowed their newest processors to train two to
three times faster than in 2017. Nonetheless, only so much
precision can be shaved off of these calculations before AI
performance degrades, and these techniques are quickly reaching
practical limitations.

As for price per computation, there is a surprising dearth of

quantitative research. Some sources, however, suggest that the
amount of compute that could be purchased for a dollar (as
measured in FLOPs) has doubled roughly every 2.3 years on
average since the 1940s, with a slower doubling rate of between
three and five years over the past decade.17 Figure 2 shows that if

*
For single precision, the A100 advertises 19.5 teraFLOPS and costs $2.939 per
hour on Google Cloud. The V100 costs $2.48 per hour and advertises single
precision at between 14 and 16.4 teraFLOPS.

Center for Security and Emerging Technology | 11

we assume that compute per dollar is likely to double roughly
every four years (solid line), or even every two years (lower bound
of shaded region), the compute trendline still quickly becomes
unsustainable before the end of the decade.

Figure 2: Extrapolated costs will soon become infeasible

Source: CSET. Note: The blue line represents growing costs assuming compute
per dollar doubles every four years, with error shading representing no change
in compute costs or a doubling time as fast as every two years. The red line
represents expected GDP at a growth of 3 percent per year from 2019 levels
with error shading representing growth between 2 and 5 percent.

Center for Security and Emerging Technology | 12

Without any changes in the price of compute, the cost of a cutting-
edge model is expected to cross the U.S. GDP threshold in June of
2026. If the amount of compute that can be performed for a dollar
doubles every four years, this point is only pushed back by five
months to November of 2026. Even if compute per dollar doubled
at the rapid pace of every two years, this point is only delayed until
May of 2027, less than a year after it would be reached with no
changes in the price of compute. Relaxing the assumption that
compute per dollar is a stable value, then, likely buys the original
trendline only a few additional months of sustainability.

Center for Security and Emerging Technology | 13

The Availability of Compute

Rather than fall, price per computation may actually rise as

demand outpaces supply. Excess demand is already driving GPU
prices to double or triple retail prices.18 Chip shortages are stalling
the automotive industry and delaying products like iPhones,
PlayStations, and Xboxes, while creating long wait lists for
customers across the board. 19 Whether budgets grow fast enough
to continue buying them does not matter if there are not enough
chips to continue the trend.

Estimates for the number of existing AI accelerators are imprecise.

Once manufactured, most GPUs are used for non-AI applications
such as personal computers, gaming, or cryptomining. The large
clusters of accelerators needed to set AI compute records are
mostly managed in datacenters, but many of those accelerators
are better suited for low-power inference than high-performance
training.20 In what follows, our estimates attempt to count the
accelerators managed across all cloud datacenters without
separating inference chips from training chips, an approach that
likely overstates the number of accelerators actually available for
AI training.

Overall, 123 million GPUs shipped in the second quarter of 2021,

with Nvidia accounting for 15.23 percent of the total, which
suggests Nvidia sells approximately 75 million GPU units per
year.21 Thirty-seven percent of Nvidia’s revenue came from the
datacenter market, and if we likewise assume that approximately
37 percent of its units went to datacenters, this translates to about
28 million Nvidia GPUs going to datacenters annually.22 Nvidia
GPUs are not the only AI accelerators going into datacenters, but
they reportedly make up 80 percent of the market.23 Based on all
these figures, we estimate the total number of accelerators
reaching datacenters annually to be somewhere in the ballpark of
35 million. This figure is likely a substantial, but it does not need to

Center for Security and Emerging Technology | 14

be precise.* As in the previous section, large errors in estimating
the total available supply only result in small changes in the dates
at which large-scale models on the compute demand trendline
become unattainable.

Following the conventional three-year lifespan for accelerators,

we find that by the end of 2025, the compute demand trendline
predicts that a single model would require the use of every GPU in
every datacenter for a continuous period of three years in order to
fully train.† Since such a model would need to begin training at the
end of 2022 with the full utilization of all accelerators already in
cloud datacenters at that time, it would need to use all datacenter
accelerators produced since 2019. Just over two years have
passed since then, so it is natural to wonder: is the compute
demand trend even still alive today, and how much more compute
growth is possible if it is not?

*
For this calculation we assumed that 37 percent of Nvidia’s revenue coming
from the datacenter market implies that 37 percent of its units are shipped to
datacenters, but high-end AI processors are more expensive than most
consumer GPUs, which means that fewer Nvidia accelerators likely end up in
cloud datacenters each year than what we have calculated.

†
Specifically, December 2025. Even if our estimate for the number of
accelerators available in the cloud to train on is off by an order of magnitude,
this breaking point would still be reached by December of 2026. The reality may
even be more pessimistic than we claim here, because for our calculations we
assume that every accelerator in the cloud is capable of operating continuously
with a throughput of 163 teraFLOPs per second, a figure that has been obtained
experimentally on Nvidia A100 GPUs but that likely overestimates the average
performance of all accelerators available in the cloud. See Deepak Narayanan et
al., “Efficient Large-Scale Language Model Training on GPU Clusters Using
Megatron-LM,” arXiv [cs.CL] (April 2021): arXiv:2104.04473.

Center for Security and Emerging Technology | 15

Managing Massive Models

The only major increases in model size since GPT-3’s release in

2020 have been a 530 billion parameter model called Megatron-
Turing NLG, which was announced in October 2021, and a 280
billion parameter model called Gopher, which was announced in
December 2021. The developers of Megatron-Turing NLG
reported the size of their compute infrastructure, but they did not
report how long the model was trained for, making it impossible to
infer a total compute requirement for the model’s training
process.24 A useful estimate for how much compute such a model
might require to train came five months earlier, when the same
developers outlined a similar approach for training models with up
to one trillion parameters and included estimates for total training
time.25 They concluded that training a trillion parameter model
would take 42,000 petaFLOPS-days, which we conservatively
estimate would cost $19.2 million dollars on Google’s TPUs
training continuously at maximum performance. Had such a model
been released in October 2021, it would have fallen a year behind
the projected compute demand trend line. This, combined with the
fact that GPT-3 likewise fell below the curve, suggests that the
compute demand trend may have already started to slow down.

In other research from 2020, OpenAI derived a series of

mathematical equations to predict the minimum amount of
compute needed to train a variety of models, based on factors like
their number of parameters and dataset size. 26 These equations
factor in how machine learning training requires the data to pass
through the network several times, how compute for each pass
grows as the number of parameters grows, and how the data
needs to grow as the number of parameters grows.

The blue line in Figure 3 shows OpenAI’s equation representing

the minimal amount of compute required to effectively train
language models of various sizes extrapolated to very large
models.27 The green squares show the amount of compute that
was used to train several smaller versions of GPT-3—each of
which used larger training datasets than the optimal minimum,
and which therefore used more compute than the theoretical
minimum. Nvidia’s projection for a one trillion parameter model is

Center for Security and Emerging Technology | 16

shown as a purple diamond along with projections for GPT-4 and
a 100 trillion parameter model. For now, assuming that developers
can achieve near optimal efficiency, the equation estimates that
building GPT-4—which we define as one hundred times bigger
than GPT-3 (17.5 trillion parameters)—would take at least
450,000 petaFLOPS-days. That would require 7,600 GPUs
running for a year and would cost about $200 million. Training a
100 trillion parameter model would need 83,000 GPUs running for
a year and would cost over $2 billion. 28

Figure 3: Anticipated compute needs for potential AI milestones

Source: OpenAI, Nvidia, and CSET.

83,000 GPUs represents only 0.2 percent of the 35 million

accelerators we estimate go into the cloud every year, and $2
billion is a very high sticker price, though well within the
budgetary capacity of a nation-state. But for models over roughly
one trillion parameters to be trained at all, researchers will have to
overcome an additional series of technical challenges driven by a
simple problem: models are already getting too large to manage.
The largest AI models no longer fit on a single processor, which
means that even inference requires clusters of processors to
function. This requires careful orchestration on a technical level to

Center for Security and Emerging Technology | 17

ensure that multiple processors can run in parallel with one
another.

Parallelization for AI is not new. In prior years, AI training often

used data parallelization methods, in which many processors
worked simultaneously on separate slices of the data, but each
processor still stored a full copy of the model. Despite increases in
processor memory, this is no longer possible. To train these
cutting-edge models, the layers of a deep neural network are held
on different processors and even individual layers may be split
across processors, as illustrated in Figure 4.

Figure 4: Representation of highly parallelized model training

Source: CSET.

As one example, the 530 billion parameter Megatron-Turing

model used 4,480 GPUs in total. Eight different copies of the
model ran simultaneously on different slices of the data, but each
copy of the model was so big that it was stored across 280 GPUs.
The layers of the neural network were split across 35 servers, with
each layer itself being spread across eight GPUs.29 This example
shows the complexity of the problem, which only gets more

Center for Security and Emerging Technology | 18

difficult as the size of the model increases. Moreover, coordinating
all of this activity places additional compute requirements on the
training process while also requiring significant technical expertise
to manage.

Splitting the training process across multiple processors means

that the results of computations performed on one processor must
be passed to others. At large enough scales, that communication
can take significant time, and traffic jams arise. Managing the flows
so that traffic does not grind to a halt is arguably the main
impediment for continuing to scale up the size of AI models. Some
experts question whether it is even possible to significantly
increase the parallelization for transformer models like the one
used in GPT-3 beyond what has already been accomplished.30

Center for Security and Emerging Technology | 19

Where Will Future Progress Come From?

If the rate of growth in compute demands is already slowing

down, then future progress in AI cannot rely on just continuing to
scale up model sizes, and will instead have to come from doing
more with more modest increases in compute. Unfortunately,
although algorithms have been exponentially improving their
efficiency, the rate of improvement is not fast enough to make up
for a loss in compute growth. The number of computations
required to reach AlexNet’s level of performance in 2018 was a
mere 1/25th the number of computations that were required to
reach the same level of performance in 2012.31 But over the same
period, the compute demand trend covered a 300,000 times
increase in compute usage. Although algorithms improved
dramatically over the last decade, the growth in compute usage
has in general been a larger factor in improving the performance of
cutting-edge models.32

Estimating the rate of improvement in algorithmic efficiency is

much harder than estimating the growth in compute usage
because it varies across applications, with many major
architectures or subfields having only become popular recently.33
Over short time periods, some domains have improved at nearly
the same rate as the compute growth trend. 34 Nonetheless, an end
or even partial slowdown to the historical rate of increase in
compute usage would require major and continual improvements
to algorithmic efficiency in order to compensate. Additionally,
efficiency improvements have already been happening throughout
the deep learning boom. Making up for a reduced ability to simply
scale up compute usage would require not only finding major
additional gains in efficiency, but doing so at a rate that is faster
than researchers have already been doing. These improvements
would need to increase substantially from an already impressively
high rate.

Although these results may seem bleak, AI progress will not grind
to a halt. The trend in growing compute consumption that drove
many of the headlines for the past decade cannot last for much
longer, but it will probably slow rather than end abruptly. We
should also not discount ingenuity and innovations that could lead

Center for Security and Emerging Technology | 20

to new breakthroughs in algorithms or techniques, particularly
when financial incentives are so large. Indeed, the focus on
parallelization that enabled the compute explosion in the first
place is largely a byproduct of the looming end of Moore’s law and
the resulting fears of stagnating compute growth. Some current
and future theoretical approaches offer promise for advancing AI
research.

Leading algorithms—like the transformer—may be losing training

efficiency at the largest sizes, but other architectures are starting
to sustain larger models. For instance, Mixture of Experts (MoE)
methods allow for more parameters by combining many smaller
models together (which may themselves be transformers), each of
which are individually less capable than a single large model. This
approach permits models that are larger in the aggregate to be
trained on less compute, with Google and the Beijing Academy of
Artificial Intelligence both releasing trillion-parameter models in
the past year trained using MoE methods. 35 MoE approaches offer
some advantages but are not as capable in any one area as the
largest single models. Both compute and parameter size are
critical ingredients for increasing the performance of a model
under the current deep learning paradigm, and there are
diminishing returns associated with scaling up one without the
other.

More importantly, not all progress requires record-breaking levels

of compute. AlphaFold is revolutionizing aspects of computational
biochemistry and only required a few weeks of training on 16
TPUs—likely costing tens of thousands of dollars rather than the
millions that were needed to train GPT-3.36 Similarly, the current
top performing image classifier only needed two days to train on
512 TPUs.37 In part, these relative efficiencies are due to using
algorithms and approaches that have become more efficient over
time.38 But in part, these efficiencies come from simply focusing
more on application-centric problems (like protein folding) and
tailoring the approach to the task rather than simply throwing
more compute at the problem.

Major overhauls of the computing paradigm like quantum

computing or neuromorphic chips might one day allow for vast

Center for Security and Emerging Technology | 21

amounts of plentiful new compute.39 But these radically different
approaches to designing computing chips are still largely
theoretical and are unlikely to make an impact before we project
that the compute demand trendline will hit fundamental budgetary
and supply availability limits. In the meantime, progress will likely
involve more incremental improvements to the algorithms and
architectures that already exist.

In the nearer term, where the extremes of compute power are

needed, that investment can be shared. It may take years,
centuries, or millennia of computing time to train a very
generalized model, but far less time is needed to fine-tune such a
model for newer, more specific applications. 40 This provides an
alternate explanation for why GPT-4 has been slow to arrive:
rather than simply training a newer, bigger model, OpenAI appears
to have shifted its attention to adapting GPT-3 for more carefully
scoped, financially viable products such as the code-generating
program, Codex.

This shift from a focus on training massive “foundation” models to

fine-tuning and deploying them for specific applications is likely to
continue.41 But this type of shift in focus mainly benefits a
privileged few if such foundation models are kept as the carefully
guarded secrets of a small handful of companies or governments.
There may be some security benefits to having these models
controlled by a trusted few organizations, which would make it
more difficult for malicious actors to misuse models or develop
methods of attacking them. 42 On the other hand, if continued AI
research requires access to the largest models and those are held
by only the wealthiest or most powerful organizations, then AI
research will become increasingly difficult for the larger part of the
AI community.

Center for Security and Emerging Technology | 22

Conclusion and Policy Recommendations

For nearly a decade, buying and using more compute each year
has been a primary factor driving AI research beyond what was
previously thought possible. This trend is likely to break soon.
Although experts may disagree about which limitation is most
critical, continued progress in AI will soon require addressing
major structural challenges such as exploding costs, chip
shortages, and parallelization bottlenecks. Future progress will
likely rest far more on a shift towards efficiency in both algorithms
and hardware rather than massive increases in compute usage. In
addition, we anticipate that the future of AI research will
increasingly rely on tailoring algorithms, hardware, and
approaches to sub-disciplines and applications.

This is not to say that progress towards increasingly powerful and

generalizable AI is dead; only that it will require a partial re-
orientation away from the dominant strategy of the past decade—
more compute—towards other approaches. If correct, this finding
has a number of implications for policymakers interested in
promoting AI progress. We discuss a few of these implications
below:

(1) Shift focus towards talent development, both by increasing

investment in AI education at home and by actively competing
to attract highly skilled immigrants from abroad. Improving
algorithmic efficiency and overcoming parallelization bottlenecks
in training are difficult problems that require significantly more
human expertise than simply purchasing more compute. This
suggests that the path towards continued progress in the future
rests far more on developing, attracting, and retaining talent than
merely outspending competitors. Correspondingly, policymakers
who want to encourage AI progress at home should invest
significant resources in (a) bolstering AI and computer science
education, (b) increasing the number of H1-B visas available for AI
researchers specifically, and (c) striving to make the United States
a more attractive destination for immigrants generally. CSET
already has publications addressing each of these topics.43

Center for Security and Emerging Technology | 23

(2) Support AI researchers with technical training, not just
compute resources. The National Artificial Intelligence Research
Resource (NAIRR) Task Force is currently exploring the types of
support that it can provide to bolster AI research in the United
States, especially in the broad categories of “computational
resources, high-quality data, educational tools, and user
support.”44 Compute remains an extremely important factor in AI
progress, and the NAIRR should take steps where possible to
expand the access of researchers to compute resources—
especially academics, students, and those without access to multi-
million-dollar budgets.

It is unlikely that the NAIRR can provide sufficient compute to

researchers to keep the compute demand trendline alive, or even
to compete with the quantities of compute already used by major
research centers. Nonetheless, impactful results and educational
experience can come from even moderately sized models.
Significant attention should be paid to developing educational
tools that can help researchers build the skills necessary to
innovate with more efficient algorithms and better-scaling
parallelization methods. Programs that promote interdisciplinary
work between machine learning and other areas of computer
science such as distributed systems and programming languages
may be especially fruitful for generating broad efficiency gains.

(3) Promote openness and access to large-scale models

throughout the research community, especially for researchers
who cannot train their own. The future of AI research may come
to focus heavily on the intermittent release of massive, compute-
intensive “foundation models” that then become the basis for
extensive follow-on research and development. If this general
depiction is right, then the United States has an interest in
ensuring that these foundation models are not monopolized by
only a small handful of actors. There are likely to be other
researchers or entrepreneurs who could contribute meaningfully to
our understanding or application of these models even though
they may lack the compute resources to build similarly sized
models themselves.

Center for Security and Emerging Technology | 24

Policymakers, where appropriate, should seek to encourage the
owners of large foundation models to permit appropriately vetted
researchers access to these models. In many cases, however, this
must be balanced against the need to promote the security of the
models themselves, especially those with potentially dangerous
uses.45 Regrettably, there are unlikely to be hard or fast rules that
can govern when models should be made as public as possible or
when they should be deliberately made difficult to access. At this
stage, we limit ourselves to noting that efforts should be made to
ensure that AI remains a field where researchers of many
backgrounds can usefully contribute and where access to a few
key models does not rest entirely in the hands of a coterie of
powerful institutions.

Center for Security and Emerging Technology | 25

Authors

Andrew Lohn is a senior fellow with the CyberAI Project at CSET,

where Micah Musser is a research analyst.

Acknowledgments

For feedback and assistance, we would like to thank John

Bansemer, Girish Sastry, Deepak Narayanan, Jared Kaplan, and
Neil Thompson, all of whose experience and analytical comments
were tremendously helpful in developing this project. Melissa
Deng and Alex Friedland provided editorial support.

© 2022 by the Center for Security and Emerging Technology. This

work is licensed under a Creative Commons Attribution-Non
Commercial 4.0 International License.

To view a copy of this license, visit

https://creativecommons.org/licenses/by-nc/4.0/.

Document Identifier: doi: 10.51593/2021CA009

Center for Security and Emerging Technology | 26

Endnotes

1
Dario Amodei et al., “AI and Compute,” OpenAI, May 16, 2018,
https://openai.com/blog/ai-and-compute/.

2
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, “ImageNet
Classification with Deep Convolutional Neural Networks,” Communications of
the ACM 60, no. 6 (June 2017): 84-90.

3
Volodymyr Mnih et al., “Playing Atari with Deep Reinforcement Learning,”
arXiv preprint arXiv:1312.5602 (2013); David Silver et al., “Mastering Chess and
Shogi by Self-Play with a General Reinforcement Learning Algorithm,” arXiv
preprint arXiv:1712.01815 (2017).

4
Jacob Devlin et al., “BERT: Pre-training of Deep Bidirectional Transformers for
Language Understanding,” arXiv preprint arXiv:1810.04805 (2018); Yinhan Liu
et al., “RoBERTa: A Robustly Optimized BERT Pretraining Approach,” arXiv
preprint arXiv:1907.11692 (2019); Colin Raffel et al., “Exploring the Limits of
Transfer Learning with a Unified Text-to-Text Transformer,” arXiv preprint
arXiv:1910.10683 (2019); Alec Radford et al., “Language Models are
Unsupervised Multitask Learners,” Papers With Code, 2019,
https://paperswithcode.com/paper/language-models-are-unsupervised-
multitask; Tom Brown et al., “Language Models are Few-Shot Learners,” arXiv
preprint arXiv:2005.14165 (2020).

5
Brown, “Language Models.”

6
Rishi Bommasani et al., “On the Opportunities and Risks of Foundation
Models,” arXiv preprint arXiv:2108.07258 (2021).

7
For this paper, we mean to say that this trend is unsustainable in the sense
that the trend itself cannot continue. But it is worth mentioning that spiraling
compute demands are also unsustainable in an environmental sense. Training
GPT-3 released an estimated 552 tons of CO2 equivalent into the atmosphere—
the equivalent of 460 round-trip flights between San Francisco and New York.
David Patterson et al., “Carbon Emissions and Large Neural Network Training,”
arXiv [cs.LG] (April 2021): arXiv:2104.10350. It is easy to overstate the
importance of this value: even in 2020, roughly six times this many people flew
between San Francisco and New York every day. “SFO Fact Sheet,” FlySFO,
accessed December 4, 2021, https://www.flysfo.com/sfo-fact-sheet. This energy
consumption is also miniscule compared to other emerging technologies that
require enormous amounts of computing, most notably cryptocurrencies. In
2018, Bitcoin alone was estimated to have generated 100,000 times that
volume of CO2 emissions, and by 2021 the energy requirements of Bitcoin had
nearly doubled relative to 2018. That is more electricity than the entire nation of

Center for Security and Emerging Technology | 27

Finland and more than seven times as much as Google’s worldwide operations.
Christian Stoll, Lena Klaassen, and Ulrich Gallersdörfer, “The Carbon Footprint
of Bitcoin,” MIT Center for Energy and Environmental Policy Research,
December 1, 2018, https://www.jstor.org/stable/resrep34616; Jon Huang, Claire
O’Neill, and Hiroko Tabuchi, “Bitcoin Uses More Electricity Than Many Countries.
How Is That Possible?” The New York Times, September 3, 2021,
https://www.nytimes.com/interactive/2021/09/03/climate/bitcoin-carbon-
footprint-electricity.html. These caveats aside, the exponential trend we discuss
here—were it to continue—would quickly make training a cutting-edge AI
model an enormous expenditure, not only in terms of cost but also in terms of
CO2 emissions; Neil C. Thompson et al., “The Computational Limits of Deep
Learning,” arXiv preprint arXiv:2007.05558 (2020); Neil C. Thompson et al.,
“Deep Learning’s Diminishing Returns,” IEEE Spectrum, September 24, 2021,
https://spectrum.ieee.org/deep-learning-computational-cost.

8
Brown, “Language Models”; Wei Zeng et al., “PanGu-: Large-scale
Autoregressive Pretrained Chinese Language Models with Auto-parallel
Computation,” arXiv preprint arXiv:2104.12369 (2021); Paresh Kharya and Ali
Alvi, “Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, the
World’s Largest and Most Powerful Generative Language Model,” Nvidia
Developer Blog, October 11, 2021, https://developer.nvidia.com/blog/using-
deepspeed-and-megatron-to-train-megatron-turing-nlg-530b-the-worlds-
largest-and-most-powerful-generative-language-model/; Jack W. Rae et al.,
“Scaling Language Models: Methods, Analysis & Insights from Training Gopher,”
arXiv preprint arXiv:2112.11446 (2021).

9
Jennifer Langston, “Microsoft announces new supercomputer, lays out vision
for future AI work,” Microsoft AI Blog, May 19, 2020,
https://blogs.microsoft.com/ai/openai-azure-supercomputer/.

10
“Apple unleashes M1,” Apple, November 10, 2020,
https://www.apple.com/newsroom/2020/11/apple-unleashes-m1/; Jon
Martindale, “What is a teraflop?,” Digital Trends, June 14, 2021,
https://www.digitaltrends.com/computing/what-is-a-teraflop/.

11
Saif M. Khan and Alexander Mann, “AI Chips: What They Are and Why They
Matter” (Center for Security and Emerging Technology, April 2020),
https://doi.org/10.51593/20190014; Albert Reuther et al., “Survey of Machine
Learning Accelerators,” arXiv preprint arXiv:2009.00993 (2020).

12
Colby Banbury et al., “Benchmarking TinyML Systems: Challenges and
Direction,” arXiv preprint arXiv:2003.04821 (2020).

13
No accelerator is ever as efficient as its advertised full-utilization speeds
would suggest. Although our calculations assume that models are trained
continuously at the theoretical peak performance of the accelerators, AI

Center for Security and Emerging Technology | 28

researchers typically expect that reaching 30 to 50 percent of the full-utilization
speed of an accelerator represents roughly the upper bound on practically
attainable speeds. For this reason, all of our estimates in this section
underestimate the cost that models of various sizes would require on current
accelerators. Some expense might be saved by purchasing the chips directly
rather than through the cloud, but cloud providers operate at fairly narrow
margins and most major AI research projects have historically chosen to use
cloud services, which bolsters the claim that cloud compute prices are an
effective means of evaluating how expensive current AI approaches are.

14
“FAQs,” National Ignition Facility & Photon Science, accessed December 4,
2021, https://lasers.llnl.gov/about/faqs; Alex Knapp, “How Much Does It Cost to
Find a Higgs Boson?” Forbes, July 5, 2012,
https://www.forbes.com/sites/alexknapp/2012/07/05/how-much-does-it-cost-
to-find-a-higgs-boson/?sh=38b3b9e13948; Deborah D. Stine, “The Manhattan
Project, the Apollo Program, and Federal Energy Technology R&D Programs: A
Comparative Analysis” (Congressional Research Service, June 2009),
https://sgp.fas.org/crs/misc/RL34645.pdf.

15
Wikipedia has a historical account of lowest cost to performance ratios over
many years showing decades of rapidly falling prices that have stalled since
2017. See “FLOPS,” Wikipedia, last modified December 4, 2021,
https://en.wikipedia.org/wiki/FLOPS#Hardware_costs.

16
Compare Web Archive captures of “Amazon EC2 P2 Instances,” Amazon Web
Services for February 10, 2017 (representing the earliest available capture) and
November 23, 2021 (representing the latest available capture at time of writing)
at
https://web.archive.org/web/20170210084643/https://aws.amazon.com/ec2/ins
tance-types/p2/ and
https://web.archive.org/web/20211123064633/https://aws.amazon.com/ec2/ins
tance-types/p2/, respectively. For Google Cloud pricing, compare Web Archive
captures of “GPUs pricing,” Google Cloud for August 26, 2019 (representing the
earliest available capture) and November 17, 2021 (representing the latest
available capture at time of writing) at
https://web.archive.org/web/20190826211015/https://cloud.google.com/compu
te/gpus-pricing and
https://web.archive.org/web/20211117225616/https://cloud.google.com/compu
te/gpus-pricing, respectively. Of all GPUs available at both times, only the
Nvidia T4 has declined in price since 2019. It is not among the most powerful
GPUs offered by Nvidia and is not commonly used for training large models.

17
See “Trends in the cost of computing,” AI Impacts, accessed December 4,
2021, https://aiimpacts.org/trends-in-the-cost-of-computing/. Much of the data
discussed in this research is either outdated or of dubious quality. A historical
doubling rate of roughly two years with a more recent slowdown is nonetheless

Center for Security and Emerging Technology | 29

consistent with conventional wisdom in the AI community. Because our analysis
is robust to even very significant changes in this variable, we use these figures
in our analysis. Even if compute per dollar were to reliably double every single
year, the compute demand trendline would predict a model that would cost
more to train than the U.S. GDP by the end of this decade.

18
Sean Hollister, “The street prices of Nvidia and AMD GPUs are utterly out of
control,” The Verge, March 23, 2021,
https://www.theverge.com/2021/3/23/22345891/nvidia-amd-rtx-gpus-price-
scalpers-ebay-graphics-cards.

19
Stephen Wilmot, “The Great Car-Chip Shortage Will Have Lasting
Consequences,” The Wall Street Journal, September 27, 2021,
https://www.wsj.com/articles/the-great-car-chip-shortage-will-have-lasting-
consequences-11632737422; Stephen Nellis, “Apple says chip shortage
reaches iPhone, growth forecast slows,” Reuters, July 27, 2021,
https://www.reuters.com/world/china/apple-beats-sales-expectations-iphone-
services-china-strength-2021-07-27/; Paul Tassi, “PS5 and Xbox Series X
Shortages Will Continue Through 2023, Most Likely,” Forbes, September 4,
2021, https://www.forbes.com/sites/paultassi/2021/09/04/ps5-and-xbox-
series-x-shortages-will-continue-through-2023-most-likely/; Abram Brown,
“The War Between Gamers and Cryptominers—and the Scarce Global Resource
that Sparked It,” Forbes, May 24, 2021,
https://www.forbes.com/sites/abrambrown/2021/05/24/gamers-cryptocurrency-
cryptominers-gpu-microchip/?sh=33f44052dbf8.

20
Although a single inference usually requires markedly less computation than
training a model, inference can be much more computationally demanding over
the long run than the initial computational cost of training a model. Nvidia has
claimed that as much as 80–90 percent of compute used in AI applications is
used for inference, not training. George Leopold, “AWS to Offer Nvidia’s T4
GPUs for AI Inferencing,” HPC Wire, March 29, 2019,
https://www.hpcwire.com/2019/03/19/aws-upgrades-its-gpu-backed-ai-
inference-platform/. This provides another reason to think that our predictions
here represent overestimates of the amount of time left in the compute demand
trendline—we are focused only on the growing cost of training, but if the
demands of inference are factored into these equations, the overall cost of these
AI models would become even higher, and the supply of AI accelerators
available specifically for training would become even more strained.

21
Aleksandar Kostovic, “GPU Shipments Soar in Q2 with 123 Million Units
Delivered,” Tom’s Hardware, August 27, 2021,
https://www.tomshardware.com/news/jpr-gpu-q2-vendor-share.

Center for Security and Emerging Technology | 30

22
Jim Chien, “Nvidia, AMD see rising sales from server sector,” DigiTimes Asia,
June 30, 2020, https://www.digitimes.com/news/a20200630PD213.html.

23
“NVIDIA maintains dominant position in 2020 market for AI processors for
cloud and data center,” Omdia, August 4, 2021,
https://omdia.tech.informa.com/pr/2021-aug/nvidia-maintains-dominant-
position-in-2020-market-for-ai-processors-for-cloud-and-data-center.

24
Paresh Kharya and Ali Alvi, “Using DeepSpeed and Megatron to Train
Megatron-Turing NLG 530B, the World’s Largest and Most Powerful
Generative Language Model,” NVIDIA Developer Blog, October 11, 2021,
https://developer.nvidia.com/blog/using-deepspeed-and-megatron-to-train-
megatron-turing-nlg-530b-the-worlds-largest-and-most-powerful-generative-
language-model/.

Deepak Narayanan et al., “Efficient Large-Scale Language Model Training on

25

GPU Clusters Using Megatron-LM,” arXiv preprint arXiv:2104.04473 (2021).

26
Jared Kaplan et al., “Scaling Law for Neural Language Models,” arXiv preprint
arXiv:2001.08361 (2020); Tom Henighan et al., “Scaling Laws for
Autoregressive Generative Modeling,” arXiv preprint arXiv:2010.14701 (2020).

27
It is important to note that this scaling law applies specifically to the single-
transformer architecture. There are other approaches—for instance, mixture of
expert models, discussed a bit later—for which more parameters can be trained
using less compute but sacrificing aspects of performance. We analyze the
transformer architecture, as it is currently the favored approach for language
models and is versatile enough to perform well across a number of other
domains. Eventually, the transformer will likely be supplanted by other models,
but this discussion helps to ground the amount of compute that would be
required under the current paradigm to reach models of various sizes.

28
Lower prices per GPU-hour are available for long term commitments, but long
training times are less useful than splitting the model or increasing the batch
size, so it is not clear that long term commitments are beneficial. See Jared
Kaplan et al., “Scaling Law for Neural Language Models,” arXiv preprint
arXiv:2001.08361 (2020).

29
Kharya and Alvi, “Using DeepSpeed.”

30
Bommasani et al., “Opportunities and Risks.”

31
Danny Hernandez and Tom B. Brown, “Measuring the Algorithmic Efficiency
of Neural Networks,” arXiv preprint arXiv:2005.04305 (2020).

32
Note that efficiency improvements can also be due to improvements other
than those in the algorithms used to train models, such as improvements in

Center for Security and Emerging Technology | 31

methods for data collection, curation, or usage. See Sebastian Borgeaud et al.,
“Improving language models by retrieving from trillions of tokens,” arXiv
preprint arXiv:2112.04426 (2021) for an example of research in this area.

33
There have also been some analyses of algorithmic efficiency improvements
in fields beyond AI. See Yash Sherry and Neil C. Thompson, “How Fast do
Algorithms Improve?,” Proceedings of the IEEE 109, no. 11 (November 2021):
1768-1777.

34
Hernandez and Brown, “Measuring Algorithmic Efficiency.”

35
Noam Shazeer et al., “Outrageously Large Neural Networks: The Sparsely-
Gated Mixture-of-Experts Layer,” arXiv preprint arXiv:1701.06538 (2017);
William Fedus, Barret Zoph, and Noam Shazeer, “Switch Transformers: Scaling
to Trillion Parameter Models with Simple and Efficient Sparsity,” arXiv preprint
arXiv:2101.03961 (2021); Alberto Romero, “GPT-3 Scared You? Meet Wu Dao
2.0: A Monster of 1.75 Trillion Parameters,” Towards Data Science, June 5,
2021, https://towardsdatascience.com/gpt-3-scared-you-meet-wu-dao-2-0-a-
monster-of-1-75-trillion-parameters-832cd83db484.

36
John Jumper et al., “Highly accurate protein structure prediction with
AlphaFold,” Nature 596 (August 2021): 583-589.

37
Hieu Pham et al., “Meta Pseudo Labels,” arXiv preprint arXiv:2003.10580
(2020).

38
Hernandez and Brown, “Measuring Algorithmic Efficiency.”

39
“The cost of training machines is becoming a problem,” The Economist, June
13, 2020, https://www.economist.com/technology-quarterly/2020/06/11/the-
cost-of-training-machines-is-becoming-a-problem.

40
Danny Hernandez et al., “Scaling Laws for Transfer,” arXiv preprint
arXiv:2102.01293 (2021).

41
Bommasani et al., “Opportunities and Risks.”

42
Andrew J. Lohn, “Poison in the Well” (Center for Security and Emerging
Technology, June 2021), https://doi.org/10.51593/2020CA013; Benjamin
Buchanan et al., “Truth, Lies, and Automation: How Language Models Could
Change Disinformation” (Center for Security and Emerging Technology, May
2021), https://doi.org/10.51593/2021CA003.

43
Diana Gehlhaus et al., “U.S. AI Workforce: Policy Recommendations” (Center
for Security and Emerging Technology, October 2021),
https://doi.org/10.51593/20200087; Dahlia Peterson, Kayla Goode, and Diana
Gehlhaus, “AI Education in China and the United States” (Center for Security

Center for Security and Emerging Technology | 32

and Emerging Technology, September 2021),
https://doi.org/10.51593/20210005; Zachary Arnold et al., “Immigration Policy
and the U.S. AI Sector: A Preliminary Assessment” (Center for Security and
Emerging Technology, September 2019), https://doi.org/10.51593/20190009;
Tina Huang and Zachary Arnold, “Immigration Policy and the Global
Competition for AI Talent” (Center for Security and Emerging Technology, June
2020), https://doi.org/10.51593/20190024; Tina Huang, Zachary Arnold, and
Remco Zwetsloot, “Most of America’s ‘Most Promising’ AI Startups Have
Immigrant Founders” (Center for Security and Emerging Technology, October
2020), https://doi.org/10.51593/20200065; Remco Zwetsloot et al., “Keeping
Top AI Talent in the United States” (Center for Security and Emerging
Technology, December 2019), https://doi.org/10.51593/20190007; Remco
Zwetsloot, Roxanne Heston, and Zachary Arnold, “Strengthening the U.S. AI
Workforce: A Policy and Research Agenda” (Center for Security and Emerging
Technology, September 2019), https://doi.org/10.51593/20190003.

44
“The Biden Administration Launches the National Artificial Intelligence
Research Resource Task Force,” The White House, June 10, 2021,
https://www.whitehouse.gov/ostp/news-updates/2021/06/10/the-biden-
administration-launches-the-national-artificial-intelligence-research-resource-
task-force/.

45
See, e.g., Buchanan et al., “Truth, Lies, and Automation.”

Center for Security and Emerging Technology | 33