Written by Chris Walsh, Partner @ 7GC

S-Curve Shift from Infrastructure to Applications

The question has shifted from with generative AI from not if this is the next compute cycle but at what scale will this compute cycle be. Validating that it is defined as the next clear compute cycle rinses away the taste of previous bubbles like crypto and drives the conversation to a question of ROI and what companies are willing to invest to achieve that upside.

While it is still early to answer that question, we are in the camp that demand will likely continue for the foreseeable future based on clear signs of a) increased ramp, b) increase in utilization, and c) clear ROI from the leaders in the gen AI race.

Positive Generative AI Demand Tailwinds

Given these tailwinds, we share Coatue's sentiment that the cold water thrown on the opportunity set at various points over the past twelve months underestimates the potential upside of the total addressable market.

Previous Compute Cycles Throughout History

Supercycles from personal computers, the Internet, and the cloud have been consistently underestimated on average by 38%. The initial forecast in 1996 for internet users called for 152M users by 2000 – fast forward to the future, and that figure stood at 361M users in 2000. Given the large unpredictability of applicability for these technologies at low utilization levels, the potential upside is impossible to predict. This unpredictability is ratified by the outperformance from leaders in each compute cycle and massive underperformance by incumbents incapable of visualizing the upside until it's too late.

Across these most recent compute cycles, we looked to understand the s-curve cycle better and assess generative AI opportunities in today's market.

Desktop Internet

While the Internet officially launched in 1993, the late 1990s to 2000s saw widespread adoption of the broadband Internet, which provided faster and more reliable access to online services. The Internet unlocked two-way telecommunications networks, with 1% of information flowing in 1993. By 2000, that figure was 51%; by 2007, more than 97% of households were online.

Infrastructure Drives Utilization That Makes the Application Layer Possible

The interplay between key infrastructure providers and hardware manufacturers in the 1990s was critical to growing this rapid two-way telecommunications accessibility over the decade. We analyzed the ecosystem's essential components to understand better the innovation cycle and the time until platform winners materialized and then application leaders.

Tim Berners-Lee's creation of the first web server in 1990 set the foundation for the Internet - growing from 1 million to 19 million households in seven years. The Netscape and Yahoo IPOs in 1995 and 1996 are widely considered the official start of the dot-com boom, but to get 10 million households internet access in six short years required infrastructure build-out across ISP, router, processor, broadband, memory companies to lead to this application layer Cambrian explosion.

Growth Rates Across Critical Infrastructure Sub-Verticals During Desktop Internet Era

*Source: CapitalIQ, SEC Filings, 7GC research*

The growth of many of these critical infrastructure pieces subsequently led to outsized demand from investors and subsequently drove multiple premia for many of these businesses through 1997. As exhibited above, the capex cycle finally hit a pause in 1997, and growth reverted downward, driving a collapse in multiples across each critical infrastructure piece.

Cumulative Trailing Multiple of Major Infrastructure Companies During Cycle

This infra-reliability created a flywheel of increased utilization, which led to the emergence of web technologies like HTML, CSS, and Javascript, which laid the foundation for software development and more dynamic websites. This infrastructure began to translate into radical platform shifts that led to the proliferation of e-commerce and online advertising for the first time. These platforms, like Google, Amazon, Yahoo, and eBay, quickly grew into billion-dollar revenue businesses before laying the foundation for a wave of web-based application tools that were created later. Like the infrastructure build-out mobilization from 1990 to 1996, the adoption rate of Internet platforms and core applications grew even faster.

Netscape, Amazon, eBay, and Yahoo were all founded and went live between 1994 and 1995. From then on, there was a vicious ramp in revenue, with both Google and Amazon breaking over $1B in revenue in under 5 years and eBay surpassing the hurdle in 7 years.

Revenue Ramp for First Wave of Platform / Application Startups

The adoption curve for desktop internet was swift in the 1990s and created a substantial sub-set of winners (and also losers) across the infra and application layer stack. What you see is a capex cycle that at some point did come to a halting end, and that was ultimately the inflection point to the application layer euphoria that promulgated in the back half of the decade before cresting and collapsing in 2000-2001.

Desktop Internet Compute Cycle Recap / Timeline

This cycle largely depended on the overall adoption rate, and the de-acceleration in the growth of new # of households going online began to shrink due to saturation and the law of big numbers. Infrastructure grew rapidly and then hit a point of overbuild as adoption plateaued at around 49% of households before ramping back up further in 2004 following the digestion period of 2000.

Cloud Computing

Unlike the desktop internet, the actual founding of the technology dates back over a decade before adoption hit meaningful levels that eventually validated pure-play cloud companies to ramp in demand.

Amazon Web Services (AWS) was introduced in 2006, offering very basic cloud services. The official unlock to begin the adoption curve occurred in 2007 with the release of the Elastic Compute Cloud (EC2), delivering the world infrastructure-as-a-service (IaaS) for the first time. The response to this release was a cascading effect of new entrants like Microsoft Azure (2008), Google Cloud (2008), VMware (2008), IBM (2011), and Oracle (2012).

Adoption Rate and Key Milestones During Cloud Compute Cyle

The adoption rate catapulted to 10% within 4 years compared to 6 years for desktop internet, driven by supply chain innovation and technological efficiency between periods. Within nine years, we had surpassed ~50% adoption, which followed a similar trajectory to the desktop era.

What stands out immediately when comparing the two cycles is that the actual stock-run-up is delayed despite a similar adoption ramp for the infrastructure/platform technology. This phenomenon can be attributed to the creation of the late-stage venture market in the 2010s. While the average time from founding to IPO was ~3 years across our core sample set in the 1900s, the average time for software companies that IPOed between 2016 and 2021 was ~10 years. The party was not delayed in this cycle; a different audience just enjoyed it.

When evaluating the infrastructure build-out for this compute cycle, we evaluated server/storage infrastructure, networking, virtualization, DC build-out, chips, and MSPs. In this period, many new infra companies succeeded and took advantage of new market needs like Palo Alto Networks (firewall networking), Arista Networks (networking), Fortinet (networking), and VMware (virtualization). However, unlike the desktop compute cycle, there was already a strong infrastructure foundation to leverage. With this reality and the fact that we had challenging comps during the GFC in 2008 and 2009, the broader infrastructure top-line revenue ramp cycle lacks visibility relative to other compute cycle builds. What exemplifies the infrastructure build the most is the creation and ramp of cloud service providers' top-line.

Cloud Providers Annual Revenue (2007-2022)

*Source: CapitalIQ, SEC Filings, Press Releases, 7GC research Note: Cohort set included AWS, Microsoft Azure, Oracle Cloud, and Google Cloud*

While many infrastructure constituents (application companies) enjoyed the birth of cloud computing and the subsequent breakout growth, the outsized winners on the infrastructure side were more concentrated for the cloud computing era - Amazon, Google, and Microsoft.

While the infrastructure winners were more concentrated, the application layer beneficiaries were widespread. The SaaS model was born in this cycle, as Marc Andreessen famously coined in 2011, "Software is eating the world." The broad takeaways of his essay touch on lower upfront costs and faster time-to-market, which will enable startups to enter new markets faster and disrupt established players more quickly. These tailwinds created a powerful feedback loop, which inevitability sparked more startups to launch and grow the technology ecosystem.

Completed Series A Deals (US) (2009-2021)

We saw the number of completed Series A startups double in five years from 2009 to 2014, then continue to grow incrementally until the peak of the market in 2021. While cloud computing lowered barriers for fixed costs and allowed many startups to touch a global customer base more quickly, it also increased competition. For any single problem, ten companies looked to compete for customers and talent, driving up variable costs incrementally. This competition shifted the normalized cost structures of startups while at the same time spurring dozens of startups to reach $100M in ARR at a record pace, many surpassing $1B in ARR in under 10 years of product launch.

ARR Ramp of Cloud Application Layer Companies

*Source: CapitalIQ, S-1 Filings, 7GC research*

While these may not compare to Google in terms of scale, the frequency across many different sub-verticals within SaaS across cybersecurity, DevOps, vSaaS, storage, and databases provided a substantial set of winners with 43 businesses with a greater than $5B market capitalization today, based on the Bessemer Cloud Index.

Cloud Compute Cycle Recap / Timeline

To summarize, the cycle we are about to round out in cloud computing lasted ~16 years, relatively in line with the desktop internet cycle. The ramp to utilization was faster in cloud computing due to pre-existing compute infrastructure that was leveraged, which allowed many legacy application layer companies quickly shift to this new cloud infrastructure (Salesforce, Workday, ServiceNow). At the same time, a whole new set of entrants rise quickly post the release of AWS EC2 in 2006. Given the removal of the barriers to creating a startup and the capability to scale globally without the need for physical geographic expansion, we saw a much more significant sub-set of businesses rise and get funded in the 2010s.

This expansion became bloated with the need to correct in 2019-2022, which continues to be digested in the private markets today. Not unlike the desktop compute cycle, there are generational companies that will be coming out of this glut. These cloud infra leaders will be well positioned to adopt generative AI analogous to Facebook and their versatility to traverse nimbly through the mobile compute cycle.

Where Does This Leave Us?

The earliest beginnings of the Generative AI cycle is the famous 2017 paper "Attention is All You Need," which introduced the notion that Nvidia's transformer technology could be used for generative AI. OpenAI's Chat GPT-1 was released in 2018, which should be considered the official kick-off to the cycle. This delayed fuse is directly comparable to Tim Berners-Lee's release of the web browser in 1991; it wasn't until the launch of Mosaic's web browser in 1993 that popularized the World Wide Web. Not dissimilar to the past; we did not see popularization for generative AI until the launch of ChatGPT in 2022, four years after GPT-1.

Reviewing the infrastructure layer, there are no signs of slowing across the stack from the foundry level to the application layer:

Foundry:

TSMC remains a monopolistic power in the ecosystem and exhibits bottlenecks based on demand from Nvidia and other leading chipmakers. Simultaneously, all three hyperscalers - Google, Amazon, and Amazon - are building their own chips, and OpenAI is also evaluating a similar strategy. The result has been a cyclical growth curve with revenue growth in tandem with chip ramp from new generational chips from Nvidia. The previous two quarters in 2024 showcased 39-40% YoY growth, which aligns with growth levels seen in 2H 2022 as H100 chips ramped in production.

Cloud Providers:

Growth has re-accelerated top-line for all three major hyperscalers based on increased demand from existing cloud companies for generative AI GPUs and support.

Utilities:

Utilities remain among the most prominent winners, bannered as genuine "picks and shovels" companies. Data center energy demand will continue to grow in the foreseeable future, but today, we remain supply-constrained.

Semis / Nvidia:

Data center revenue for Nvidia has now surpassed $31B in quarterly revenue, showing no signs of decline. Growth continues to accelerate across the supply chain as Nvidia preps for releasing Blackwell GB200 - the company's newest chip iteration.

Foundation Models:

Despite ChatGPT and Anthropic's Claude classification as consumer products, the foundation models that power those simple products are critical to the Gen AI infrastructure stack. Foundation Models saw a 6.5x increase in spending in the last year alone, increasing to $6.5B. This growth rate is accelerating, not decelerating.

Generative AI Compute Cycle Infrastructure Growth Snapshot

*Source: CapitalIQ, Nvidia Investor Relations, 7GC researcH*

On the consumer side, despite the endless noise on generative AI, the data points to the fact that we are also in the early innings. Analyzing Reuters data, only ~11% of the population is using OpenAI on a weekly basis. Some argue that would showcase a lack of real applicability, but that ramp is far quicker than we saw in the cloud and desktop internet eras. The data also shows that 84% have used ChatGPT once or have never heard of it. That would indicate, if we were looking at this on an apples-to-apples basis, that we are currently in 1995 or 2010, looking at previous compute cycles.

The consumer and enterprise-level momentum doesn't seem to be slowing either, evaluating Morgan Stanley's sentiment checks over the past 12 months. Web visit growth continues to accelerate and hit new all-time highs each month for ChatGPT, with Google trends pointing to robust demand. ChatGPT reached 100 on their 1-100 sentiment index, which is an all-time high. On the enterprise side, CIOs surveyed stating they had no plans to use fell to 3%, down from 31% a year ago.

Generative AI Compute Cycle Recap / Timeline

We are searching for decay! However, the overall trend remains intact outside of excess capital, potentially investing in lower-quality businesses and squeamishness on continued infrastructure ramp. We saw a similar picture in 2023, where the SOXX Index sat flat from June through November 2023, only to gain over 50% in the first half of 2024.

We will keep searching for signs of weakness. Still, today, we struggle to see what signs are pointing to a slowing in any shape, way, or form across all data points, especially when benchmarking trends relative to previous compute cycle timelines.

Compute Cycle Reflections

Convergence of “Non-AI” vs. “AI” Investing