W H I T E PA P E R

Rethinking Cloud Strategies

for Advanced AI

Sponsored by
Rethinking Cloud Strategies
for Advanced AI

Artificial intelligence (AI) is providing companies with HIGHLIGHTS

capabilities that would have been unimaginable only
a short time ago, including advances in self-driving
Advanced artificial intelligence (AI)
vehicles, rapid drug development in health care, natural technologies such as generative AI—
language processing, and better risk management in which are self-learning algorithms
that can create code, text, video,
finance. Advanced AI technologies such as generative and audio content—are requiring
AI—which are self-learning algorithms that can create companies to take a more strategic
view of their cloud adoption so they
code, text, video, and audio content—are requiring have the IT foundation required to
companies to take a more strategic view of their cloud make full use of state-of-the-art AI.
adoption so they have the IT foundation required to make
full use of state-of-the-art AI. Forward-looking companies are
increasingly using a purpose-built,
The AI InfrastructureView 2021 benchmark survey in April 2021 by Needham,
full-stack cloud infrastructure for
Mass., technology research firm IDC identified inadequate purpose-built the intense requirements of these
infrastructure capabilities as the most frequent reasons that AI projects new AI-powered applications.
fail. Thus, forward-looking companies are increasingly using a purpose-
built, full-stack cloud infrastructure for the intense requirements of these
new AI-powered applications, which require massive amounts of data and Today, companies often have
computing power for training models. Witness Wayve, a London startup, is inelastic processing power,
creating autonomous driving technology based on cutting-edge technology, insufficient storage capacity,
such as the latest advances in computer vision. “Advanced AI, the latest ineffective resource allocation
and greatest, is absolutely pivotal to what we’re doing,” says Jamie Shotton, management, and inadequate
chief scientist at Wayve. “We have to train the algorithm on petabytes and networking capabilities for
potentially greater amounts of data that we’ve captured from our fleet of generative AI.
cars, which is a radically different approach to autonomous self-driving than
anyone has done before.”
Indeed, the computing requirements for large-scale AI models doubled
every 10.7 months from 2016 to 2022, according to a study presented at the 2022

White Paper | Rethinking Cloud Strategies for Advanced AI

1
 Harvard Business Review Analytic Services

“What’s unique about generative AI is copyright ambiguities, lack of

truthfulness and accuracy issues, and increase in misuse and biases,”
says Ritu Jyoti, group vice president, worldwide AI and automation
research, with IDC’s software market research and advisory practice.

International Joint Conference on Neural Networks in Padova, scientific technical high-performance computing, which
Italy. Using the cloud for these computing requirements can uses supercomputers and other advanced technology to
enable organizations to rapidly scale their infrastructure solve complex computational problems. “Many industries are
as necessary. looking at how classical deterministic or scientific technical
Generative AI raises key issues around not just scale and high-performance computing can be used in conjunction
performance but responsibility, too, according to Ritu Jyoti, with AI or machine learning to come up with a blended model
group vice president, worldwide AI and automation research, that’s more efficient,” says Addison Snell, CEO of Intersect360
with IDC’s software market research and advisory practice. Research, a market intelligence, research, and consulting
“What’s unique about generative AI is copyright ambiguities, advisory practice in Sunnyvale, Calif. “The analytics part may
lack of truthfulness and accuracy issues, and increase in discard outcomes that are really unlikely, allowing companies
misuse and biases,” she says. “You need to make sure on the to save huge amounts of computation by focusing on the best
back end that you get the appropriate logs and alerts, so if solutions. This is leading to the new types of applications that
something goes wrong, you can correct the action right away. are doing everything from reducing shoplifting in stores to
Essentially, along with being proactive, you need to embrace identifying which credit card applicants deserve a lower or
generative AI with all the right guardrails.” higher rate because they present risks.”
This report will examine the opportunities and demands The enhanced ability to use analytics to ultimately home in
state-of-the-art AI presents and how both are causing on the best solutions and develop new types of applications
companies to reexamine their cloud strategy. It will explore is enabling organizations to make breakthroughs that are
how companies are effecting change within their industries by reshaping how industries are addressing problems and finding
implementing AI infrastructure that allows them to operate opportunities.
in an efficient and cost-effective manner. It will also address
the many strategic, technological, and human issues that
accompany the redefining nature of advanced technology
like generative AI.

The New World of Artificial Intelligence

Savvy customers are increasingly happy with AI-infused
offerings like intelligent virtual assistants that can automate
support, thus quickly meeting and even anticipating their
needs. Companies then are using AI to create new offerings,
predict trends, and, for other strategic purposes, reap greater
benefits from their overall digital transformation, according
to a December 2021 study conducted by Technology Horizon
Research that was sponsored by the London-based consultancy
EY. Nearly two-thirds (63%) of these advanced AI users report
their digital transformation is exceeding expectations,
compared to only 37% of companies that are using AI for
operational improvements such as cutting costs. FIGURE 1
AI-enabled applications are a refinement of traditional
programming techniques. These approaches include classical
deterministic computing, which is an algorithm that always
produces the same output when given the same inputs, and

White Paper | Rethinking Cloud Strategies for Advanced AI

2
 Harvard Business Review Analytic Services

For example, today, fashion designers can spend six to 12

months of intense work coming up with new styles, unsure
if their creations will sell or not. But Fashable, based in Viana
do Castelo, Portugal, built a generative AI application that
can create dozens of original AI-generated clothing designs
in minutes, without the need for actual material. “Many industries are looking at how
The algorithm ingests data from multiple sources to learn classical deterministic or scientific
about trends, styles, and clothing types. Using social media
to do A/B tests directly with customers lets designers gauge technical high-performance
interest and forecast demand for their particular creations computing can be used in conjunction
before going into production. “We can share the collection with with AI or machine learning to come
customers before they are produced, avoiding the problem of
overstock,” says Orlando Ribas Fernandes, CEO and cofounder
up with a blended model that’s more
of Fashable. efficient,” says Addison Snell, CEO of
Wildlife Protection Services (WPS), a nonprofit organization Intersect360 Research.
in Golden, Colo., is using AI to deal with issues related to
habitat loss, forest fragmentation, and poaching and other
illegal uses of wildlife. “Conservation is a huge challenge
globally, and we’re not necessarily winning the war,” says Eric
Schmidt, executive director of the organization. but democratization of a standard of care, allowing more
To improve its odds in the fight, WPS is arming itself with AI advanced protocols to be deployed in regions of the world
models that search images from thousands of camera feeds, that lack the human capital to do so now.”
looking for humans and vehicles that may be engaged in
suspicious activities or animals that may be encroaching on
human populations. It sends alerts to wardens and rangers Cutting-Edge AI Infrastructure
at national parks, private reserves, and other protected areas AI projects require a level of computer power and scalability
in sub-Saharan Africa, Southeast Asia, and Latin America. that is difficult and costly to implement on premises.
Health care is another field being transformed by advanced Companies often lack the right systems to handle the demands
AI. “Around six years ago, we identified that AI was having a of complex model training and keep pace with the rapid
disruptive influence in several adjacent verticals, most notably technology changes in the AI space.
in radiology,” says Rui Lopes, director of new technology Advanced applications like generative AI must be supported
assessment at Elekta, a Stockholm-based Swedish maker by cutting-edge infrastructure that provides the performance,
of precision radiation therapy solutions. “We realized the flexibility, and scalability that these applications demand.
tsunami of AI innovations that were happening in the According to the IDC report, though, AI infrastructure
computer vision and text recognition fields were eventually remains one of the most consequential but the least mature
going to find their way into the medical field, as well.” of infrastructure decisions that organizations make as part
While radiology is used to diagnose cancer and has long of their future enterprise. As a result, today, companies often
embraced AI, Elekta focused on a related but more involved have inelastic processing power, insufficient storage capacity,
area: radiotherapy, which is used to treat cancer. Patients with ineffective resource allocation management, and inadequate
cancer might need as few as five or as many as 45 radiotherapy networking capabilities for generative AI.
treatments, which must be constantly adjusted to ensure Organizations have three basic approaches to deploying
cancer cells are attacked, while healthy cells are spared. cloud services—infrastructure as a service (IaaS), platform
Currently, many people around the world do not have access to as a service (PaaS), and software as a service (SaaS)—each
this therapy—not because of a lack of technology but because of which divides which technology capabilities the cloud
of insufficient medical personnel from the diverse disciplines provider and company are responsible for managing in
that must collaborate to ensure the correct adjustments are different ways. FIGURE 2
made to treatment plans. These approaches are not mutually exclusive, and all three
“AI is well suited to address this problem because we are often used by enterprises and medium-sized companies.
can embed some of that intelligence, and the protocolized There is SaaS, which is on-demand access to cloud-hosted
ways of working, into the devices themselves to increase the application software that is ready to use. Another approach is
access to treatment for a larger swath of patients worldwide,” PaaS, which is on-demand access to a cloud-based platform on
Lopes says. “This provides not just personalization of care which companies can develop, run, and manage applications

White Paper | Rethinking Cloud Strategies for Advanced AI

3
 Harvard Business Review Analytic Services

On Premises Infrastructure as a Service Platform as a Service Software as a Service

(IaaS) (PaaS) (SaaS)

Applications Applications Applications Applications

Data Data Data Data

Runtime Runtime Runtime Runtime

Middleware Middleware Middleware Middleware

O/S O/S O/S O/S

Virtualization Virtualization Virtualization Virtualization

Services Services Services Services

Storage Storage Storage Storage

Networking Networking Networking Networking

Source: Guru99

using tools and programming languages that the cloud vendor says. “These were things we didn’t really think about when
supports. The customer controls the applications but not the we initially launched, but now it’s become an issue. We are
underlying cloud infrastructure. Then there is IaaS, involving trying to be more flexible in how we launch AI projects now.”
on-demand access to cloud-hosted physical and virtual Large enterprises might start with SaaS but move to an
storage, networking, and servers. The customer controls the IaaS or PaaS model. The choice will be driven by the size of
applications, storage, networking, and operating systems but the company, the agility requirements, the available talent,
not the underlying cloud infrastructure. and the use case for the technology. For some use cases,
Small and medium-sized businesses often favor the SaaS the decision is dictated by the lack of off-the-shelf SaaS
model and buy off-the-shelf components for their AI projects, solutions. “The one big lightbulb for me was to understand
even if they don’t meet 100% of their requirements. that there was no off-the-shelf, one-size-fits-all solution that’s
As a small nonprofit, WPS started with a SaaS model. While available out there today unless you’re in a fairly commodity
this move brings technology efficiencies and cost savings, the industry,” Lopes says.
SaaS model can also raise issues. “If we choose a provider that Most of the top data-centric organizations are using the
has a great tool we want to work with, we have to be concerned cloud to leverage more IaaS and PaaS capabilities, taking
if the tool will be available in the future” because the provider advantage of the speed and agility of these models to
might discontinue it or go out of business altogether, says innovate where they could not have done so with traditional
Matt Morrisette, a software engineer for the organization. technologies, while reducing costs.
WPS is also running into non-SaaS issues where it needs “We use a different mix of tools and technologies, some we
regional support to ensure it remains in compliance with build in-house, some we have acquired from partners,” Wayve’s
the General Data Protection Regulation, better known as Shotton says. “Using a managed platform gives us the ability to
GDPR, the law on data protection and privacy in the EU, as scale quickly and reliably. It also allows us to focus our efforts
its operations expand into new areas. “If you don’t host your doing the research and solving problems around autonomous
data in Europe, you can’t launch projects there,” Morrisette self-driving rather than building additional tools ourselves.”

White Paper | Rethinking Cloud Strategies for Advanced AI

4
 Harvard Business Review Analytic Services

Preparing for Scale of AI Development Proper Data Stewardship

Generative AI requires input from a continuous, larger data Data is at the core of every AI project. “Collecting, curating,
set to refine the algorithms. The more data, the more the and cleaning data sets” was ranked as the top AI infrastructure
algorithm learns. To train a model to make predictions, element that should receive the most resources, according
generate text, and handle the other chores of advanced AI, to the AI Infrastructure Ecosystem, a 2022 study by the
engineers use a process called “inference.” This process can be AI Infrastructure Alliance, an industry group in San
expensive since it needs to run millions of times, consuming Francisco. FIGURE 3
large amounts of computing time. Companies need to be specific about what they are trying
The different stages of AI development—from data analysis to accomplish with their AI initiatives and determine if they
and experimentation to model training and inference—can have enough of the correct kind of data. They then must choose
require different capabilities, tools, and AI infrastructure. how to prepare the data for use by AI applications, move it
The workload cost and scale can change quickly from one across the network, and store it. “The biggest challenges are
stage to the next. not technical but strategic,” Intersect360 Research’s Snell says.
“Since advanced AI applications can be costly to train, you “People have magical thinking that they can throw infinite
need to have rock-solid engineering under the hood to avoid computing power at a limited amount of imperfect data and
wasting money,” Shotton says. “This is resulting in a blurring learn something they didn’t already know.”
of the lines between research and production.”
There are ways to reduce these costs and speed innovations.
For example, organizations can build their own models or use
pretrained models for some applications and then fine-tune
these models using their own customer data.
An “AI-first” approach to infrastructure can accelerate
model training and accelerate AI innovation, as well. Cloud
infrastructures leverage graphical processing units, an
approach designed originally for 3D gaming and which since
has incorporated specialized hardware to execute calculations
in parallel and significantly reduce training time.
A common infrastructure can increase the velocity of
AI innovation. “If you have the data in one place and you
offer services around that data, it becomes more compelling
for people to use those standardized services,” Lopes says.
“Instead of spending two weeks having your local workstation
churn through creating a model, a developer can do that in a
couple of days with the more-powerful computing resources
in the cloud. We wanted to unleash our developers to be able
to use the data, while still maintaining the right security
constraints.”
As the Elekta example shows, the training workload
will execute much faster in the cloud than on a local
workstation, allowing the solution to be found sooner. In
addition, some models are too large to even run on a local
workstation, making the cloud the only viable option for
some advanced AI.
On a small scale, models can be trained in small clusters
on workstations on premises. However, according to Adam
Moore, director of global cloud solutions for Elekta, this
approach can disguise problems that will arise when the
models are scaled up to production. “At that point, you
will need to radically scale up the amount of data you use,”
he explains. “By training the models in the cloud, you can
identify those problems earlier and build resilience into your
compute infrastructure so you avoid hardware failures.”

White Paper | Rethinking Cloud Strategies for Advanced AI

5
 Harvard Business Review Analytic Services

a journey, and we partner closely with our finance team to

build up from small to larger investments in compute spend.”
AI can not only draw on data and resources across the
organization but also can impact every aspect of it. Elekta
began its AI efforts with a dedicated cloud team to generate
“People have magical thinking that the initial spark of enthusiasm, and then brought aboard
they can throw infinite computing other departments, including engineering, technology, and
security, as well as sales, customer, and product teams, to
power at a limited amount of get their feedback.
imperfect data and learn something “We brought everyone around the table, addressing their
they didn’t already know,” says concerns and ensuring we had an effective feedback cycle
across the entire organization,” says Elekta’s Moore. “We
Intersect360’s Snell.
maintained a clear focus on understanding the value to the
customer in everything we did.”
The custom-built cloud infrastructure needed to support
generative AI can also have tremendous implications for an
Many of the issues revolve around data movement. As organization’s talent strategy. Just barely behind the always
Snell explains, “Let’s say I’ve decided to use the cloud. Now present concerns about cost (56% of respondents) on the
I’ve got to move all my data up there, presumably for the issue of AI implementation challenges identified by the IDC
analytics to happen there. Is that a secondary data copy? Do I AI StrategiesView study was a lack of skilled personnel (55%)
also have the data somewhere else on premises? If I have it in and machine learning operators (55%). FIGURE 4
both places, then how do I keep it in sync? If I have it only in “Sometimes the perception is your on-premises skill set
one place, then where is it? Where is it backed up? And what is analogous to what you will need for the cloud,” Moore
happens if I want to move it back on premises or to a different says. “People think it’s just taking your computing and data
cloud at a different time?” Prospective cloud customers must resources to the cloud. From an architectural perspective,
also understand the ingress/egress fees that come with data there’s a number of design patterns and thinking that you
movement. This dizzying array of questions has important need, such as operating at scale and resilience.”
implications for infrastructure choices. In shifting its AI development to the cloud, Elekta employed
Snell says that AI initiatives must be based on a long- consultants in DevOps, tools, and other technical areas. “There
term data stewardship strategy. This time frame transcends
a three-to-five-year service-level agreement (SLA) with a
cloud vendor. “You need to maintain long-term control and
possession of your data and think of ‘long-term’ as at least
100 years,” he says. “Let’s say you are a pediatric hospital,
and you take a scan of a two-year-old. You want to keep that
data for the life of a patient, so that could be 100 years. If
that information could matter to the person’s children and
grandchildren, you can easily see instances where some data
you want to keep forever.”

Getting Everyone around the Table

The latest approaches to AI can fundamentally change
an organization’s processes and business model. For this
reason, companies must think beyond just the technology
implications to such matters as how advanced AI will impact
budgeting and talent strategies.
Tensions can arise between the finance and engineering
departments, like whether to spend scarce capital to develop
applications quicker. “We must be agile and move quickly if
we are to succeed in our mission, but as a startup, we need
to work within our financial constraints,” Shotton says. “It’s

White Paper | Rethinking Cloud Strategies for Advanced AI

6
 Harvard Business Review Analytic Services

“There will be a lot of hybrid environments—not just the cloud

and on-premises, but the cloud and the edge. Edge locations, the
intersection of 5G and real-time streaming, are shaping up and will
empower newer AI capabilities for businesses,” says IDC’s Jyoti.

are generalists who will help you do ‘a lift and shift’ to the on-premises but the cloud and the edge,” Jyoti asserts. “Edge
cloud, but they don’t necessarily understand what you’re locations, the intersection of 5G and real-time streaming,
trying to achieve,” Moore says. “We leverage an external are shaping up and will empower newer AI capabilities for
system integrator to help us speed up the journey, but we were businesses.”
also focused on making sure that we had the right balance The computing power needed for generative AI could be
of in-house and external skills so we didn’t lose all our skills changed by experimentation going on now about the right
when we finished the project.” number of training data sets (something called training
AI applications, which devour huge amounts of data from tokens) and model size or capacity (known as parameters) that
every corner of an organization, can require heightened levels are used. “A greater number of parameters is more compute-
of collaboration. In the past, Elekta’s distributed R&D centers intensive, which is more costly,” she says. As researchers
would use their own processes, tools, and data repositories. identify the right balance of data, model architecture, hyper
Consequently, the company’s data was cached and guarded parameters, and inference runtime, the cost equation of
by different departments. advanced AI could shift again, changing the infrastructure
For Elekta, Lopes, the director of new technology assessment, requirements and providing more opportunities to discover
says this “fractured approach” along business lines provided breakthrough solutions.
“the opportunity to reevaluate our infrastructure for how we Companies should see generative AI as a journey and
could consolidate not just our way of working in our processes develop relationships with vendors that will grow as their
but also the fuel that drives AI, the data.” AI use does. “Everyone won’t shut off their on-premises
technology today and go in the cloud tomorrow,” Jyoti says.
“You need to make sure you’ve matured your SLAs and
Choosing a Cloud for Growth governance practices, while ensuring there is automation
The diversity of cloud offerings allows organizations to choose at each step in the process.”
the best approach for their unique AI needs. In the past, there
were many legacy reasons for why people were hesitant about
moving to the cloud. With generative AI technologies, IDC’s Conclusion
Jyoti says the question has shifted from whether to use the Today, most companies rely on an infrastructure foundation
cloud for AI applications to which cloud provider is best designed for general-purpose components rather than
aligned with a company’s strategic vision for AI. developing an infrastructure stack honed for the complex
“The fight now is over which platform will become and demanding needs of advanced technologies like
the platform of choice for generative AI,” Jyoti says. “As generative AI.
companies make their strategic investments, they are going to “People are going gaga over the ChatGPT aspect of AI, but
bet on which cloud they use. We suggest companies work with that’s only a small part of generative AI,” Jyoti says about the
trusted hyperscalers or suppliers and do their due diligence. AI-driven chatbot based on language model Generative Pre-
But don’t sit back because the train has left.” trained Transformer (GPT) 3.5. “AI is also about creating new
The selection depends on both the capabilities of the cloud content like video, audio, images, and code. The technology
vendor and the ecosystem of partners and vendors that is built can have a transformational impact on almost every industry.
around the vendor’s offerings. The cloud will always be the For example, you can use it to not just create marketing blogs,
center of gravity for generative AI because of the scalability, sales proposals, social media posts, summarize customer
agility, and cost advantages it brings. feedback and/or code generation, but also to design toy
In the rapidly changing AI space, organizations need an images or create digital twins for design and simulation of
architecture that provides flexibility to allow them to take products. The benefits of generative AI can be remarkable,
advantage of modern technologies and approaches. “There from improving knowledge worker productivity to increasing
will be a lot of hybrid environments—not just the cloud and innovation velocity and/or accelerating decision velocity.”

White Paper | Rethinking Cloud Strategies for Advanced AI

7
 Harvard Business Review Analytic Services

“Just as word processing programs became commonplace tools to

increase general productivity, AI will become a common tool for
organizations to increase innovation. AI will be a superpower that
will democratize content creation. AI will not replace jobs, it will
transform work,” says Orlando Ribas Fernandes, CEO and cofounder
of Fashable.

The AI journey, as Fashable’s Fernandes puts it, is driven Those on the cutting edge of AI say companies should be
by constantly needing new inputs, new training, and new exploring technologies like generative AI now. “AI will be a tool
learning that continually increase not just in quantity but for every company,” Fernandes says. “Just as word processing
also in quality. He points out that the quality-over-quantity programs became commonplace tools to increase general
aspects could be demonstrated by having an AI algorithm productivity, AI will become a common tool for organizations
generate personalized clothing lines for individual shoppers. to increase innovation. AI will be a superpower that will
“In the near future, you will have a digital closet of clothing democratize content creation. AI will not replace jobs—it
designs that you can ask a manufacturer to produce just for will transform work.”
you,” Fernandes says. “We will use the metaverse to create
physical goods that are exclusive to each person.”

White Paper | Rethinking Cloud Strategies for Advanced AI

8
ABOUT US

Harvard Business Review Analytic Services is an independent commercial research

unit within Harvard Business Review Group, conducting research and comparative
analysis on important management challenges and emerging business opportunities.
Seeking to provide business intelligence and peer-group insight, each report is
published based on the findings of original quantitative and/or qualitative research
and analysis. Quantitative surveys are conducted with the HBR Advisory Council,
HBR’s global research panel, and qualitative research is conducted with senior
business executives and subject matter experts from within and beyond the Harvard
Business Review author community. Email us at [email protected].

hbr.org/hbr-analytic-services

Copyright © 2023 Harvard Business School Publishing. MCCRE37600423