High AI Capex is Worrisome, But "Winner Take All" is the Prize

It is not hard to find estimates of investment in U.S. artificial intelligence infrastructure (computing capabilities) in the range of $300 billion or more between 2023 and 2030. IDC analysts have suggested $300 billion in investments between 2023 and 2026.

Nor is it hard to find critics who worry about uncontrolled spending without a clear revenue model. On the other hand, leaders of firms attempting to become leaders in the generative AI model business are likely to keep in mind the “winner take all” dynamic we have seen in the recent internet era, where just one or a few firms emerged as leaders in new markets. 

They might point to:

• Amazon's years of heavy investment to dominate e-commerce

• Google's massive spending to establish search leadership

• Cloud providers' huge datacenter investments

• Meta's acquisition strategy in social media.

In fact, many markets show scant ability to support three providers, as the market leader has twice the share--and up to an order of magnitude more-share compared to  the number-two provider.

Market	Dominant Player	Market Share	Runner-up	Market Share
Search Engines	Google	91.9%	Bing	3.0%
Desktop Browsers	Chrome	65.72%	Safari	18.22%
Mobile Browsers	Chrome	66.17%	Safari	23.28%
Social Media	Facebook	2.9B users	YouTube	2.5B users
E-commerce	Amazon	37.8% (US)	Walmart	6.3% (US)
Video Streaming	YouTube	2.5B users	Netflix	231M subscribers
Music Streaming	Spotify	31%	Apple Music	15%
Ride-hailing (US)	Uber	68%	Lyft	32%
Cloud Services	AWS	32%	Azure	22%
Mobile OS	Android	71.8%	iOS	27.6%

So even if McKinsey estimates AI infrastructure spending will exceed $500 billion between 2023 and 2030, and even if many of those investments do nor produce the expected results, model suppliers have incentives to risk quite a lot, knowing that there is a small  prize for being second best. 

Gartner forecasts global AI infrastructure investments will surpass $250 billion annually by 2030. 

The OECD estimates investments in AI infrastructure across industries, will reach $1 trillion by 2030, across the OECD countries. Bloomberg predicts that the global AI infrastructure market will $700 billion by 2030.

On the other hand, most of that investment will be by end users and others in the value chain, not the generative AI model providers. 

And some estimates made in 2023 might be considered conservative in 2024. Morgan Stanley’s  "The Economics of AI” study, published in October 2023 suggested more than $200 billion in AI infrastructure investments by 2030, including:

• Data centers: $125B

• Networking infrastructure: $50B

• Chip fabrication: $25B

• Cooling systems: $10B.

Boston Consulting Group in December 2023 suggested there would be $235 billion cumulative investments in 

• Data center buildout: 45%

• Compute infrastructure: 35%

• Power infrastructure: 20%. 

The Goldman Sachs "AI Infrastructure Report," published in September 2023 estimated $275 billion in  cumulative investment, including:

• Semiconductor investment: $100B

• Data centers: $115B

• Power systems: $35B

• Network upgrades: $25B. 

The caution, though, is that early estimates of the size of new technology markets often lead to overinvestment across the value chain. 

Study/Report	Date	Publisher	Key Conclusions
The Dot-Com Bubble Burst: Causes and Implications	2001	U.S. Securities and Exchange Commission (SEC)	Overinvestment in internet startups led to a speculative bubble that burst in 2000. Many companies were overvalued despite having no profitability.
Boom and Bust: The Telecommunications Investment Bubble	2002	Federal Reserve Bank of San Francisco	Overinvestment in telecom infrastructure during the late 1990s led to a major industry downturn, with unsustainable levels of capital spending.
The Case for Less Innovation	2017	Harvard Business Review	Many companies overinvest in unproven technologies without clear demand, resulting in failed projects and wasted resources.
Lessons from the Clean Tech Bubble	2016	MIT Energy Initiative	Overinvestment in cleantech (2005-2011) led to massive failures, with many companies being too early to market and receiving excessive venture capital.
Investing in Innovation: Creating a Research and Innovation Policy That Works	2010	The NESTA Foundation (UK)	Over-investment in R&D for new technologies can create inefficiencies and fail to produce proportional economic benefits if not managed strategically.
The Nanotechnology Investment Bubble	2005	Journal of Nanoparticle Research	Speculative investments in nanotechnology during the early 2000s led to unmet expectations, as many products were not commercially viable.
Unleashing Productivity: Overinvestment in Information Technology	2005	McKinsey Global Institute	Overinvestment in IT during the late 1990s and early 2000s did not yield expected productivity gains, with firms often adopting technology prematurely.
The Illusions of Overinvestment in AI	2021	Brookings Institution	Many companies overinvest in artificial intelligence without clear applications, leading to inflated expectations and unrealized returns.
The Biotechnology Bubble: When Science and Finance Collide	2004	Nature Biotechnology	Excessive capital flow into biotech during the 1990s led to overvaluation, with many firms failing to achieve meaningful breakthroughs.

In recent years we have also seen examples of overinvestment by many platform suppliers as well. 

Technology	Company/Industry	Year	Description of Over-Investment
Artificial Intelligence	IBM Watson	2011-2022	IBM invested billions in Watson AI for healthcare, but struggled to generate significant revenue and ultimately sold off the health assets
Virtual Reality	Meta (Facebook)	2014-present	Meta has invested over $36 billion in VR/AR technology with limited returns, facing skepticism about the metaverse vision
Blockchain	Various	2017-2018	Many companies rushed to invest in blockchain during the crypto boom, only to scale back or abandon projects when the hype died down
Autonomous Vehicles	Uber	2016-2020	Uber invested heavily in self-driving technology, spending over $1 billion before selling the unit after a fatal accident and regulatory challenges
3D Printing	3D Systems	2013-2015	The company aggressively acquired 3D printing startups, leading to over $1.3 billion in losses and a stock price crash when consumer adoption didn't materialize
Cloud Computing	HP	2011-2012	HP's $11 billion acquisition of Autonomy for cloud services led to an $8.8 billion write-down

So the rationale for investing heavily to secure the leading position in the generative AI model business is a reflection of the possible “winner take all” character of application and platform markets, where the number-one provider dominates. 

And since market share and profit margin generally are related, the rewards for market leadership also are significant. In many capital-intensive markets, the profit margin of the top provider is double that of number two. 

And provider number two can have margins double that of provider number three.

Mobile Generative AI Will Be a Huge Driver of Usage

Eventually, the leading generative artificial intelligence apps used by consumers will winnow down from hundreds to a few, though many platforms will likely find niche uses in some industries, market segments, job functions and use cases.

One important difference could arise in the mobile domain, compared to larger-screen use cases, and the reason is simply the huge amount of interactive app usage that now happens on mobile devices, compared to all other screens. 

User Type	Mobile Devices	PCs
Consumers	5-6 hours/day	2-3 hours/day
Business Professionals	3-4 hours/day	5-6 hours/day

Perhaps more important is the amount of data consumed on mobile platforms. While it might be difficult to directly correlate “value” with “data volume,” data consumption is connected with usage volume. Generally speaking, the volume of consumer data used on mobile devices is twice as much as on PCs, for example. 

User Type	Mobile Devices	PCs
Consumers	10-20 GB/month	5-10 GB/month
Business Professionals	20-30 GB/month	10-20 GB/month

That usage is then correlated with the volume of advertising spending on mobile and PC platforms. 

Among the big shifts in U.S. advertising spending in the internet era is not simply the growth of share taken by digital media, but also the share taken by mobile venues, which already are as much as 65 percent of all digital media advertising. 

Venue	Spending  (Billions USD)	Market Share (%)
Digital	233.4	56.7
Television	73.7	17.9
Radio	21.9	5.3
Out-of-Home	19.5	4.7
Print	15.6	3.8
Other	12.7	3.1

source: Statista, Seeking Alpha 

That suggests a huge opportunity for generative AI  use cases on mobiles, as well, assuming that GenAI winds up being a core functionality of most highly-used consumer-facing apps. 

source: Andreessen Horowitz 

source: Andreessen Horowitz 

And to the extent that the costs of GenAI usage for app and experience providers matters, open source models should, over time, increase their share of the market. Already, in mid-2024, enterprise user /.,mnbvcz+net promoter scores for open source GenAI models have rapidly approached those of proprietary models. 

source: Andreessen Horowitz 

As always for new markets, early market share often is not predictive of developed or mature market leadership. Eventual leaders often are not among the early leaders. Also, definitions of “active” users might vary. Some might include users who have accessed a model only once; others will have varying levels of persistent usage that are minimums for the purpose of defining active users. 

Model	Estimated Active Users
OpenAI (GPT-4, GPT-3.5)	100+ million
Google (LaMDA, PaLM, Gemini))	50+ million
Meta (LLaMA)	30+ million
Microsoft (Azure OpenAI Service)	20+ million
Stability AI (Stable Diffusion)	10+ million
Midjourney	10+ million

The term "active user" can include:

• Frequency of Use: How often a user interacts with the AI model. This could be measured by the number of prompts or requests submitted within a specific timeframe (e.g., daily, weekly, monthly).

• Duration of Use: The amount of time a user spends interacting with the AI model during a session or over a period.

• Type of Interaction: The nature of the user's interactions, such as text prompts, image generation requests, or code completion.

• Conversion Rate: The percentage of users who take a desired action, such as creating an account, subscribing to a premium service, or sharing content.

• Engagement Metrics: Measures of user engagement, such as click-through rates, time spent on the platform, and social sharing.

For OpenAI, “active users” are typically defined as those who have interacted with the model within a specific timeframe, such as the past month or year.

Google's definition of "active user" considers frequency of use and engagement metrics. Meta's definition of "active user" might be similar to its definitions for other products, focusing on user engagement and interaction.