AI Drives GPU Channel Conflict

To the extent that Nvidia is creating an ability to operate as a “computing or software as a service” supplier, using its own or partner data centers, a logical question is whether Nvidia might have at least a window of opportunity, given its assumed lead in creating hardware, software, tools, and services that power AI applications, including code libraries, a base of developers, pre-built models and integration with important software frameworks. 

On the other hand, as the leading “cloud computing as a service” providers step up their own in-house GPU and chip capabilities, develop code libraries, pre-built models and software framework integrations, one might project that any existing Nvidia advantage will close. 

Even granting an Nvidia lead, AWS, Google Cloud and Azure, for example, have their own hardware, library, models and frameworks in place, and developing rapidly in response to the perceived need for “AI as a service” capabilities. 

Capabilities	NVIDIA	AWS	Google Cloud	Azure
Hardware	High-performance GPUs	FPGAs, GPUs	GPUs, TPUs	FPGAs, GPUs
Code libraries	cuDNN, cuBLAS, TensorRT	MXNet, PyTorch	TensorFlow, PyTorch	MXNet, PyTorch
Pre-built models	NVIDIA NGC catalog	SageMaker models	Vertex AI models	Azure AI models
Integration with software frameworks	Windows, VMWare, Kubernetes	Windows, Linux, Docker	Linux, Kubernetes	Windows, Linux, Kubernetes

As we often have seen, suppliers can become competitors to their customers, and such competition generally spurs efforts by customers to reduce their reliance on key suppliers. The issue is how long any perceived Nvidia advantage can be sustained, and whether the self-reliance efforts by major customers will eventually outweigh the shorter-term revenue benefits Nvidia gets as the key supplier of GPUs and other infrastructure supporting AI operations. 

AWS, Google Cloud, and Azure are all investing heavily in AI hardware, including GPUs. For example, AWS has its own custom-designed GPUs called Graviton2.

AWS, Google Cloud, and Azure are all developing their own code libraries and pre-built models for AI, ML, and data science. For example, AWS has a library called SageMaker Neo that helps developers optimize their AI models for deployment.

Also, AWS, Google Cloud, and Azure are all improving their integration with software frameworks such as Windows and VMWare. For example, Azure has a service called Azure Batch AI that helps developers run AI workloads on Windows and Linux VMs.

So Nvidia is gambling that, overall, becoming a competitor to its best customers will create near-term revenue advantages, almost-certain long-term sales declines of its GPU products and services revenues to former customers who become competitors and also compel Nvidia to change its business model.

Nvidia might wind up as a supplier of cloud computing and SaaS services rather than a supplier of GPUs and related services, at least in substantial part. 

The technology business is no stranger to former suppliers that emerge as competitors to their former customers, leading to a “coopetition” model.

Former Supplier	Products	Former Customers
Android	Operating system	Samsung, LG, HTC, Motorola
ARM	Processor designs	Apple, Qualcomm, MediaTek
Amazon Web Services	Cloud computing services	Netflix
Google	Search engine, web browser, operating system	Microsoft, Yahoo, Apple
Microsoft	Software, operating system	IBM, Oracle, Apple
Intel	Processors	AMD, NVIDIA
Samsung	Displays, memory, processors	Apple, Sony, Google
Sony	Displays, semiconductors, gaming consoles	Microsoft, Nintendo
Microsoft	Software, operating system	IBM, Oracle, Apple
IBM	Mainframes, software, consulting services	Hewlett Packard, Dell, Oracle
Oracle	Software, databases, cloud computing services	SAP, Microsoft, Amazon Web Services
Hewlett Packard	Computers, printers, servers	IBM, Dell, Lenovo

AI for Recommendations is a No-Brainer

The ways artificial intelligence can expand on present uses is nearly infinite. Already widely used to make recommendations when people are shopping, searching, watching a streaming media service or listening to streaming music, for example, AI should rather quickly emerge as a foundational tool for any “search or comparison” operation, whether the number of passengers going to the same destination, from the same starting point, is one or some other number. 

And though much of that activity will be conducted on behalf of service suppliers, it’s easy to see how AI also can be used by services targeted at consumers, or by consumers to choose their travel modes.  

For example, I frequently travel places that are between 850 and 1,000 miles distant, solo. As I basically do not enjoy driving, I won’t drive more than eight hours a day, and often restrict hours to five or so. That makes those trips two to three days, each way. 

I virtually always fly, as my manual calculations suggest flying actually is the same price or cheaper than driving, when including hotel, meal costs and fuel, with the avoided need for a rental vehicle at the destination, when necessary, ignoring transit time. 

So, eventually, I can see comparison shopping sites that use AI to allow me to compare the all-in cost of flying versus driving to destinations of that distance, for shorter or longer trips, using my own assumptions about how far I’m willing to drive in a day, my preferences for lodging and so forth. 

If I’m right, the business model is built on steering more traffic to airlines and rental car suppliers, for example, as I am pretty sure the AI is going to suggest that driving is more expensive than flying, if not by a wide margin. In my stated case I also save 3.5 to six days of commuting time, roundtrip, for each of those journeys. 

In principle, AI could allow other combinations of transit methods as well (rail or other public long-distance transportation). 

Indeed, some studies of airline pricing support the thesis that airline pricing algorithms are used to price domestic flights at levels that make fares equal to or better than the assumed cost of driving. 

With AI, the level of analysis available to potential customers and travelers should be even more detailed and blindingly fast. 

University of California, Berkeley (2018)	The average cost of flying round-trip from Los Angeles to San Francisco was $190, while the average cost of driving the same route was $215. The cost of flying was lower than the cost of driving for about 60% of the trips analyzed.
University of Chicago (2019)	The average cost of flying round-trip from New York City to Chicago was $185, while the average cost of driving the same route was $200. The cost of flying was lower than the cost of driving for about 55% of the trips analyzed.
Journal of Air Transport Management (2019)	A study of “Airline Pricing and the Cost of Driving” found that airline fares are often equal to or better than the cost of driving, especially for longer distances. The study looked at the cost of driving and flying between different cities in the United States, and found that the cost of driving was typically higher than the cost of flying for distances over 500 miles..
Journal of Transport Economics and Policy, (2018)	“Airline Pricing and the Cost of Long-Haul Travel” found that airline pricing algorithms can be used to price long-haul flights in a way that takes into account the total cost of travel, including the cost of hotels, meals, and fuel.
Transportation Research Part A: Policy and Practice	The "Airline Pricing and the Cost of Driving" (2005) study found that airlines price domestic flights to be competitive with the cost of driving, especially for shorter routes.

And no doubt AI will demonstrate that flying, whether short, medium or long distance, will frequently be cheaper than driving, for one passenger. Obviously, the economics change as the number of passengers to be moved increases. 

The point is that lots of useful AI enhancements will be made to comparison and search engines of all types, whether for restaurants, home furnishings, clothing and personal items, books and content or anything else bought at retail. 

Similar use cases will be found in business-to-business transactions as well. And there are bound to be implications for all activities and jobs that are based on expert knowledge of products, where the “value add” includes recommendations.

How Much Service Provider Revenue from Network Slicing?

Rarely in the networking or computing businesses is there but one solution for any given problem. And that almost certainly applies to 5G network slicing as well. There are many other alternatives offering functionality that is comparable or better, with costs that might also be lower. 

The issue is whether there are use cases where a 5G slice is the best-performing or only solution for a particular application and use case. 

One of the unknowns about 5G network slicing is the number of slices that can be created at any single base station. The number might be far more restrictive than some might suspect. 

Depending on the amount of bandwidth used by each virtual private network, as few as 10 or as many as 100 network slices might be supported at any single base station. 

The 3GPP (3rd Generation Partnership Project) standard does not formally limit the  number of slices that can be supported at any base station. However, capacity is limited in practice by the hardware and software resources available, and the need to respect a few requirements including::

• Each slice must be isolated from other slices to prevent interference.

• Each slice must have its own dedicated resources, such as bandwidth and processing power.

• Each slice must be able to be scaled up or down independently of other slices.

In addition,  the number of slices that can be supported is affected by

• Some slices, such as those used for mission-critical applications, may require more resources than other slices.

• The more isolated the slices are, the more resources may be required.

• The heavier the traffic load on a slice, the more resources it may require.

For such reasons, it is possible--and perhaps likely--that many enterprise customers would choose some other method of assuring features touted for 5G network slicing. 

Feature	Network slicing	Dedicated circuits	VPNs	QoS	Traffic shaping	Edge computing
Low latency	Yes	Yes	Yes	Yes	Yes	Yes
Guaranteed bandwidth	Yes	Yes	Yes	Yes	Yes	No
Reliability	In development	Good	Good	Good	Good	Good
Cost	High	High	Medium	Low	Low	Medium
Availability	In development	Fair	Good	Good	Good	Fair, growing

Just how much revenue mobile and other service providers could earn supplying network slicing is somewhat hard to forecast at the moment, since most enterprises still do not have the option to buy it. But most forecasts seem reasonable, putting network slicing revenue somewhere between three percent and 5.5 percent of total mobile service provider revenue by 2026 to 2030 or so. 

At least in principle, those sums should not include infrastructure revenues earned by the likes of Nokia, Ericsson and others selling gear to enterprises to run their own private networks, or revenues earned by system integrators operating such networks on behalf of enterprises. 

Forecaster	5G network slicing revenue forecast (USD billions)	Global mobile service provider revenue expected (USD billions)	Network slicing revenue as a percentage of total mobile revenues
ABI Research	30 by 2026	1,200 by 2026	2.5%
Ericsson	100 by 2030	1,800 by 2030	5.5%
Gartner	20 by 2025	1,000 by 2025	2.0%
IDC	50 by 2027	1,400 by 2027	3.6%
McKinsey & Company	100 by 2025	900 by 2025	11.1%
Nokia	100 by 2030	1,800 by 2030	5.5%
Omdia	40 by 2026	1,200 by 2026	3.3%
Ovum	30 by 2025	900 by 2025	3.3%
STL Partners	50 by 2027	1,400 by 2027	3.6%
UBS	100 by 2030	1,800 by 2030	5.5%