GPU as a Service and AI Business Models

It was virtually inevitable that the rise of artificial intelligence would create new roles for data centers and computing “as a service.” Consider firms that now offer access to graphics processing units “as a service,” including virtually all leading cloud computing as a service suppliers, as well as newer entrants in the “GPU as a service” or “cloud GPU” space:

• 
  

Amazon Web Services (AWS)

Microsoft Azure

Google Cloud Platform (GCP)

NVIDIA DGX Cloud

IBM Cloud

Oracle Cloud Infrastructure (OCI)

CoreWeave

Jarvis Labs

Lambda Labs

Paperspace CORE

Perhaps notable on that list are NVIDIA DGX Cloud and firms such as CoreWeave. The former is notable because it represents a foray by a major GPU infra supplier into the “services” space, while the latter represents a new type of data center, namely focused on GPU as a service. 

Those efforts are part of a broader creation of revenue and business models for AI in general. 

As you might guess, a key question for generative AI and AI overall is “what is the revenue model?” The answers depend on what part of the internet ecosystem a firm operates in, and whether the “customer” is a business or a consumer. 

Chipmakers and server suppliers will sell infrastructure. Connectivity suppliers will sell bandwidth. Data centers will sell compute cycles, storage, interconnection and security “as a service.”

Virtually every firm will embed AI into its existing core business processes, creating revenue models for software and platform suppliers. The point is that direct revenue models (selling chips, servers, software, bandwidth, compute cycles, storage, security, payment processing, analytics) are likely to be common in business-to-business settings. 

The cost of those tools, in turn, will be monetized indirectly in the form of higher profit margins, higher sales, lower operating costs, lower churn, higher add-on sales, greater awareness and longer customer life cycles, for example. 

Indirect models are likely to dominate in consumer markets. Subscriptions, transactions and advertising are the basic consumer revenue models. So, in most cases, AI is unlikely to be a “product” sold separately to consumers. Rather, it is going to be embedded in some other revenue-producing process. 

That is probably what ChapGPT founder Sam Altman means when he says the costs of intelligence are on a path to near-zero costs. Revenue from applying AI will be embedded in the consumer’s cost to buy things, watch things, listen to things, read things, communicate with others, use social media and find things. 

Some pay-to-play consumer models could develop, much as some firms rely on subscriptions for access to content, services and features. But consumers are price conscious, so advertising and transactions  is likely to remain a key alternative to pay-to-play and subscriptions. 

But GPU as a service is among the new direct revenue and business models developing around AI.

Why TCP/IP Also Was a Business Choice

There are many reasons why the connectivity industry that might well have preferred asynchronous transfer mode instead of TCP/IP as the next-generation network protocol, there are lots of obvious reasons why TCP/IP was chosen. 

Dimension	TCP/IP	ATM
Cost	Relatively low	Relatively high
Scalability	Very high	Not as high
Openness	Open standard	Proprietary
Complexity	Relatively simple	More complex
Suitability for a variety of network technologies	Yes	No

On the other hand, the choice of TCP/IP also had serious implications for innovation and distribution of profit within the app ecosystem supported by multi-purpose, multimedia networks. By creating a layered, loosely-coupled architecture, permissionless innovation was possible.

In other words, so long as an app or service is compliant with TCP/IP network protocols, no business relationship needs to exist between any internet service provider and any app creator. So app creation is “open” and permissionless, not closed. App and service innovators do not need permission, or direct business relationships, to be used on any public IP network. 

In other words, nearly all apps and services become a matter of “direct to consumer,”

Factor	Advantage for Telcos	Advantage for App Creators	Disadvantage for Telcos	Disadvantage for App Creators
Loose coupling	Increased flexibility and scalability	Increased flexibility and agility	Increased complexity	Increased complexity
TCP/IP protocols	Ubiquitous and interoperable	Easy to use and develop for	Less control over network performance	Less control over network performance
Software layers	Easier to upgrade and maintain	Easier to innovate and develop new features	Less control over network performance and security	Less control over network performance and security

If you want to know why connectivity providers such as telcos worry so much about being "dumb pipes," the loosely-coupled, layered architecture of TCP/IP and modern software is a chief reason. 

No developer, app or content provider needs an ISP's permission to sell products directly to en users. Everything is "direct to consumer." 

source: TechTarget

5G is to Edge Computing as WANS are for Cloud Computing

One way to look at 5G is to examine its role in supporting computing operations, rather than the role of enabling communications, much as global data transport networks can be looked at as essential parts of the cloud computing infrastructure, or Wi-Fi can be viewed as a key part of the internet access function. 

Simply stated, in the internet era most “computing” is inseparable from “internet access.” In other words, data processing, storage and app consumption depend on communications. 

And some use cases demand low latency, which drives demand for  edge computing. In other cases, edge computing adds value by reducing the amount of wide area network investment that has to be supported. 

Use Case	Value from Low Latency	Value from Reduced Bandwidth
Self-driving cars	Critical	Significant
Augmented reality/virtual reality	Critical	Significant
Industrial process control	Critical	Significant
Smart cities	Moderate	Moderate
Content delivery networks (CDNs)	Significant	Critical
Internet of Things (IoT) devices	Moderate	Significant

Edge computing value also varies among use cases. It is hard to imagine successful widespread use of self-driving vehicles without very-low-latency data processing “on the device,” accessible without wires. 

On-device is the place to put real-time language translation activities, such as translation during a voice call or during video or audio playback of content. Untethered access is essential when the devices are smartphones. 

5G also enables new cloud-based gaming services. These services allow gamers to play high-end games on their mobile devices without the need to download or install any software. These use cases typically require both low latency. 

5G is being used to automate industrial processes by connecting untethered devices to local servers for real-time process control. 

Proposed virtual reality use cases typically require both on-device and remote computing support, but with low latency crucial for realism. 

But many edge computing scenarios actually benefit from a mix of low latency and bandwidth-reduction value. 

5G is used to stream high-definition and ultra-high-definition video to mobile devices from edge content servers. Live streaming of sporting events and concerts to mobile devices also requires untethered access. But there the value mostly is bandwidth reduction, not so much latency. 

Likewise, 5G is being developed to support smart city applications, such as traffic management, public safety, and utility monitoring, though most of these apps will initially use remote computing rather than on-device computing. 

In some cases, as when delivering a self-contained on-device app, perhaps the greatest need is simply downloading the app and updating the app, ad delivery and uploading of usage and behavior profiles. 

In many other use cases, virtually every keystroke for a document, every frame of a video, every note of a song, every pixel of an image has to be transported to a remote server location. 

Our computing architecture includes processing on-device, on the premises, metro or regional data center and remote data center venues. 

Traditionally we’d describe the various key parts of the connectivity network as involving the inside home network (local area network using Ethernet or Wi-Fi); access network (home broadband); middle mile (connection between local network and nearest internet point of presence) and wide area network (long distances between points of presence). 

In all these cases, connectivity to nearby or remote resources is required. 

Edge computing, in general, is driven by the need for processing with low latency, and sometimes by the added advantage of reducing network bandwidth demand. 

Language translation “on the fly” is an example of the former; video content delivery an example of the latter. 

Edge Use Case	Description
Object detection and recognition	AI can be used to detect and recognize objects in images and videos. This can be used for a variety of applications, such as security, surveillance, and quality control. For example, AI-powered cameras can be used to detect intruders in a secure facility, or to identify defects in products on a production line.
Natural language processing (NLP)	NLP can be used to understand and respond to human language. This can be used for a variety of applications, such as customer service, chatbots, and voice assistants. For example, NLP-powered chatbots can be used to answer customer questions about products and services, or to help customers book appointments.
Machine learning (ML)	ML can be used to train AI models to learn from data and make predictions. This can be used for a variety of applications, such as fraud detection, predictive maintenance, and medical diagnosis. For example, ML-powered models can be used to detect fraudulent transactions, or to predict when a machine is likely to fail.
Computer vision	Computer vision is a field of AI that deals with the extraction of meaningful information from digital images and videos. This can be used for a variety of applications, such as self-driving cars, facial recognition, and medical imaging. For example, computer vision-powered systems can be used to identify pedestrians and other objects on the road, or to detect cancer cells in medical images.
Self-driving autos	AI at the edge is used to power the self-driving features in cars, such as lane keeping assist and adaptive cruise control, which, by definition, must use untethered mobile access.
Smart homes	AI is used to power smart home devices, such as thermostats and security systems, which use untethered access and mobile networks for connectivity.
Smart cities	AI is used to collect and analyze data from IoT devices, such as sensors and actuators, which use mobile networks for network connectivity.
VR/AR	Uses a mix of edge and remote computing

The point is that 5G can be viewed through the lens of “compute platform” rather than “communications,” just as cloud computing, data centers, edge computing and devices can be assessed as computing venues.

When "Growth" Capex Becomes "Maintenance" Capex

In a fundamental sense, much capital investment previously described as "growth" capex (infra investments supporting capacity, speed, new products such as SD-WAN or

One almost-perverse reality in the connectivity business is that legacy products often are more profitable than the newer products intended to replace the legacy offerings, even as demand for the legacy products dwindles.

Almost perversely, the strategic rationale for many investments, such as next-generation mobile networks or fiber-to-home, is driven less by expectations of highly-profitable new services but by the necessity of doing so to remain in business.

To be blunt, "you get to keep your business" is the investment rationale, more so than "you will boost revenues significantly." In fact, even when new products or services are created, profit margins will be lower than for legacy products.

Cable TV companies and telcos face precisely that problem with video streaming services, compared to linear video, for example. Telcos faced the same problem with VoIP and messaging. Home broadband and mobile internet access seem to be faring the best, though each of those services is precisely a "dumb pipe" offer: access but not apps; bandwidth but not "services;" flat fees for usage but not participation in app and service revenue models dependent on internet access.

There are several reasons why products with declining demand are more profitable than newer services, starting with the strategy of “harvesting” the products. Since the investments to create legacy products often have already been amortized, there is relatively little additional capital investment or operating cost required to run the business supporting those products, compared to building infrastructure and demand for new products.

Also, revenue upside from more-advanced products can actually be lower, per unit or per user, than the legacy products. Cloud computing, for example, lowers the cost of creating new software products. Open source and multimedia, general purpose networks do too. 

Legacy public switched network voice was generally more expensive to create as it used proprietary technology to create a single-purpose network. Voice over IP is generally less costly as it uses a multi-purpose network, more open platforms and more-generic hardware, with resources that can be more centralized (so less investment is required). 

Factor	PSTN Voice Services	VoIP Services
Upfront costs	High	Low
Ongoing costs	Medium	Low
Scalability	Medium	High
Flexibility	Medium	High
Complexity	High	Medium

In a classic example, service provider profits and gross revenue from VoIP can be lower than what was the case for legacy voice products. 

Product	Type	Revenue Expectation	Profit Expectation
Voice over IP (VoIP)	Newer	Low	Low
Mobile data	Newer	High	Medium
Cloud computing	Newer	High	Medium
Traditional landline service	Legacy	Medium	High
Cable TV	Legacy	Medium	High
DSL internet	Legacy	Medium	High

Almost perversely, profit margins from selling digital subscriber line home internet access, though an “inferior” product, might actually be higher than selling home broadband using fiber-to-home or other platforms. 

Service providers used to make high profits from text messaging, but they have almost no ability to monetize the multimedia messaging alternatives such as WhatsApp. In principle, one might ask why development of new products makes sense, if the financial returns are low. 

The issue, strategically, is that harvesting does not solve the “what is my business of tomorrow” problem. As with any industry, connectivity services have product lifecycles. Each product eventually faces declining demand. So a company that only harvests will eventually go out of business.

So new products, even when they feature lower-profit margins, must be created. 

One is left with the almost-inescapable conclusion that often, when new products are envisioned as substitutes for legacy products, they become “features.” 

Voice communication once was the primary value provided by a mobile network. These days, though, voice is more often a feature of a mobile service. 

A mobile service unable to handle voice calls or text messaging features would not be a competitive product, but service providers earn less revenue from voice over time. 

Voice and text/multimedia messaging might be an essential feature of the service. But spending lots of effort and money to create a new voice and messaging experience might not produce satisfying financial returns. 

In other cases, such as upgrading networks from DSL to FTTH, the strategic rationale is quite clear: a telco has no future without that upgrade. But the actual revenue and profit impact might vary quite a lot in the near term. 

In other cases, such as the transition from linear to streaming versions of entertainment video, the outcome for service providers remains unclear. In the transition from dial-up internet to DSL and cable modems, a whole class of internet service providers was forced out of business, as success shifted to ownership of access facilities. 

It’s an open question right now whether connectivity service providers will retain a role, and what sort of role, in a future where most video entertainment has shifted to streaming delivery. Can network service distributors be disintermediated, and to what extent? 

Will distribution shift to the streaming video providers who go direct to consumer? And will that shift the connectivity provider role to that of mere sales agent? It remains unclear. 

The broad point to be made is that new substitutes for legacy products are not uniformly as revenue producing or profitable as the legacy products, often because demand has shifted away, because new forms of competition limit pricing power or because the new products require high levels of investment. 

“Doing nothing” is a strategic death sentence for any copper-access-based fixed network provider. But upgrading to FTTH is not a panacea, either. FTTH creates a sustainable platform for competing, but is not an automatic net generator of higher revenue and profits. 

In a nutshell, that is the problem with much technology innovation. Some strategic imperatives that allow a firm to remain in business are not automatically going to solve the business problem of replacing legacy products. 

VoIP is not a strategic answer if the product itself faces declining demand (people shift to mobile phones for calling, for example). And there are cases where changing demand actually destroys a business opportunity. Dial-up ISPs went away when home broadband emerged, for example. 

Nor is it clear what roles might be left for connectivity providers when a fuller shift to video streaming has happened.

The point is that investments in new platforms, networks and services, often without huge expectations of higher revenues and profit margins, is a necessity in the connectivity business. In a business sense, almost all infra capital investment is for "maintenance" rather than "growth," in a fundamental sense.

5G has to replace 4G, often less because revenue will be higher but because additional capacity must be added to remain competitive and meet customer desires for data consumption and app experience. The same holds for FTTH.