"This Time is Different": Will AI end up like the telecom bust?

There are historical parallels, but maybe this time really is different

Introduction

I’m not the first person to make this observation, but there are a number of parallels between the hyperscalers’ AI build-out and the telecom companies’ fiber build-out in the late ‘90s. The latter ended in a lot of bankruptcies and consolidation. It’s at least worth understanding the parallels. There is, of course, danger in using historical antecedents to predict the future: this time may actually be different. TSMC and Nvidia say they see no slowdown in demand for their chips.

That said, let’s put aside the “this time is different” narrative, and consider the parallels between today’s AI build out and the fiber build-out of 25 years ago. Capital expenditures (capex) refer to investments made by a company to acquire, maintian, or upgrade physical assets like buildings, machinery, technology, and infrastructure. For hyperscalers—companies like Google, Microsoft, and Amazon that build and operate massive cloud computing infrastructures—capex spending often includes the purchase of cutting-edge servers, data centers, and now, AI-specific hardware such as GPUs and custom AI chips. Hyperscalers are pouring billions into these investments to expand their AI capabilities, anticipating exponential growth in demand for AI services. This is similar to what happened with fiber infrastructure in the late ‘90s.

The Parallels Between Today’s Hyperscalers and Telecom Companies of the ‘90s

1. Why Hyperscalers are Investing in AI Capex

   1. Anticipated Demand: Hyperscalers are betting that demand for AI capabilities, such as natural language processing, image recognition, and predictive analysis, will skyrocket. These services require vast computational power.

   2. Competitive Differentiation: As AI services become key to their offerings, hyperscalers aim to set themselves apart by building custom AI infrastructure (e.g., Google’s TPUs or Amazon’s Inferentia chips), which gives them control over performance and cost efficiency.

   3. Economies of Scale: More AI infrastructure allows hyperscalers to offer cheaper services. This, in turn, attracts large enterprise clients. Therefore, the hyperscalers are making these large-scale investments because they hope to lock in future revenue streams.

2. Potential Risks—Parallels to the Fiber Overbuild

   1. Overestimation of Demand: Hyperscalers may be misjudging the growth curve of AI adoption. Similar to the 90s, when telecom overestimated data needs, they could end up with more AI infrastructure than the market demands.

   2. High Fixed Costs: Massive AI investments have high fixed costs, so if the anticipated demand fails to materialize, hyperscalers would have to absorb these costs. Writing off these massive expenses will put pressure on balance sheets.

   3. Comeptitive Pressures and Price Wars: The hyperscalers are already competing fiercely on AI servcies, and overbuilding infrastructure could lead to aggressive pricing strategies. If their infrastructure outpaces demand, they could undercut each other’s prices, eroding profit margins. And, unfortunately, it does not appear that the software for which this vast infrastructure is being built—the foundational large language models like Google’s Gemini—increasingly look like commidities, with costs to end users rapidly trending toward zero.

3. The Telecom Bust as a Cautionary Tale

   1. In the late 90s, telecoms vastly overbuilt fiber optic networks, fueled by the assumption that internet traffic would continue to explode. However, actual growth didn’t match inflated expectations.

   2. Immediate Impact:

      1. Massive Write-Downs: Telecom companies had to write down billions in assets as network usage fell short of projections, devastating their finances.

      2. Bankruptcies and Consolidation: Many telecom companies couldn’t sustain the debt and filed for bankruptcy, leading to a consolidation phase where only a few players survived.

      3. Lost Jobs and Economic Fallout: The telecom bust led to massive layoffs, and the broader economic fallout was a key factor in the early 2000s recession.

   3. Long-Term Effects: Over time, the overbuilt infrastructure became useful as internet demand caught up in the 2000s. Yet, the companies that initially built the networks weren’t the ones reaping the benefits. They were taken over or outcompeted by more nimble players.

4. Possible Implications for the AI Industry

   1. Financial Strain and Potential Consolidation: If the hyperscalers miscalculate demand, they could face financial strain, resulting in cutbacks and write-downs. They might even have to sell their AI assets to newer companies.

   2. Slowdown in AI Advancement: If companies become more cautious with their investments, we might see a slowdown in the rapid pace of AI innovation and expansion as companies focus on optimizing their existing infrastructure.

   3. Shifts in Industry Leadership: Just as telecom overbuilding paved the way for new entrants and a reshuffling of industry leaders, an AI overbuild might create opportunities for startups and new players that can leverage excess capacity without bearing the initial costs.

5. What This Might Mean for AI’s Future

   1. Delayed but Inevitable Growth: Similar to telecom, AI demand could ultimately catch up, but the hyperscalers may not be the primary beneficiaries. Startups or similar companies that capitaize on overbuilt AI infrastructure could capture significant value.

   2. More Cautious Investment Approaches: If a bust occurs, future AI investments may focus more on incremental, flexible scaling and shared infrastructure, rather than single-company dominance, leading to more sustainable industry growth.

   3. Regulatory Scrutiny: A bust will focus regulator’s attetion on the technology industry’s investment decisions, especially if layoffs and economic fallout ensue. Never underestimate a regulator’s capacity for foisting pain on companies when it appears politically advantageous to do so.

While the hyperscalers’ AI capex investments might indeed pay off if demand aligns with their forecasts, there’s a susbtantial risk of history repeating itself. If these firms overextend, we could see an AI bust followed by a restructuring period that might fundamentally change the landscape of the tech industry, similar to how the telecom sector transformed in the 2000s.

---

Why This Time Might Be Different

In spite of all of the points above, this time might really be different, and it’s worth examining why that might be the case. Here are several factors which might support the “this time it’s different” claims.

1. Diverse and Established Demand

   1. Unlike the speculative, nascent demand for internet services in the ‘90s, AI already has a wide array of applications with clear, measurable value. AI is being adopted across numerous industries, from healthcare to logistics, where machine learning and data-driven automation can lead to substantial cost savings and operational improvements. This is not just a consumer demand cycle but an industrial and enterprise one as well. Industrial and enterprise demand tends to be more stable and reliable than consumer demand.

   2. AI has been integrated into hyperscalers’ core products and services. Additionally, the hyperscalers use AI extensively to optimize their own systems (e.g., data center efficiency, predictive maintenance, customer service automation). Therefore, even if external demand were to soften, hyperscalers would still benefit from AI infrastructure for internal optimization.

2. Better Financial Health and Long-Term Investment Strategies

   1. Today’s hyperscalers have much larger balance sheets, are much more profitable, and have more diverse revenue streams, than the telecom companies of the ‘90s. With strong balance sheets, these firms can better absorb potential dips in demand. They are also strategically deploying capex with long-term horizons, often seeing AI as a decade-long investment that will evolve as new capabilities emerge.

   2. Additionally, these companies have robust cash flow from other services (cloud hosting, consumer products) that can subsidize AI infrastructure investments. These extensive revenue streams allow them to weather market fluctuations more sustainably than the telecoms could.

3. Rapidly Expanding Use Cases

   1. AI applications are expanding faster than anticipated in fields such as generative AI, autonomous driving, and predictive analytics. Generative AI, in particular, is creating new market segments and use cases that didn’t even exist a few years ago, from content creation to drug discovery. These use cases broaden the hyperscalers’ potential customer base. A broad and diverse customer base also provides some insurance that if demand slips in one segment, other segments might see demand increase.

   2. Open-source AI models and accessible APIs create a broader development ecosystem, which could accelerate the rate at which new AI-driven applications are developed. The faster pace of innovation means that demand could catch up with hyperscaler infrastructure investments far more quickly than it did with fiber.

4. Improved Infrastructure Utilization and Scalability

   1. Modern cloud infrastructure is highly flexible and is designed to adapt to changing workloads. Hyperscalers can allocate resources dynamically across various applications (cloud storage, compute, AI processing) in ways that weren’t possible in the more rigid telecom networks of the ‘90s. This adaptability means that, even if AI demand fluctuates, hyperscalers can repurpose infrastructure for other high-margin services, reducing the risk of underutilized assets.

   2. Edge computing and other decentralized models offer hyperscalers flexibility in deploying AI infrastructure closer to end users, reducing latency and broadening AI use cases. This flexibility in scaling and infrastructure distribution mitigates the risk of overcapacity.

5. Global Reach and Adoption

   1. The hyperscalers are not just reliant on the US market. They have global reach, allowing them to scale AI services in regions with varying adoption curves. Emerging markets may lag in adopting certain AI technologies but will inevitably catch up, providing another demand source for deployed AI infrastructure.

   2. Countries with aging populations might adopt AI more readily to offset labor shortages, driving demand in the longer term.

6. AI as a Platform Technology with Exponential Growth Potential

   1. AI isn’t a single application like high-speed internet, but a foundational technology with compounding benefits. Once companies begin using AI, the insights and efficiencies generated will fuel further demand for even more advanced capabilities. In this sense, AI has a self-reinforcing growth potential as businesses come to rely on increasingly sophisticated models to maintain competitive advantages.

   2. Platform technologies create ecosystems, and AI is no exception. Hyperscalers can use their AI infrastructure to create ecosystems around their platforms, encouraging developers to build AI-dependent applications that, in turn, increase demand for AI infrastructure in a more sustainable way.

7. Resilience Through Vertical Integration and Proprietary Models

   1. Hyperscalers are not only investing in raw infrastructure but are also deeply involved in developing proprietary AI models and software, which allows them to capture the entire stack from hardware to application. This vertical integration reduces dependency on third-party technology and lowers the risk of external market fluctuations impacting their AI services.

   2. Proprietary models such as OpenAI’s partnership with Microsoft, or Google’s Gemini and DeepMind, give hyperscalers the ability to offer unique, highly differentiated services that competitiors cannot easily replicate. This creates a stronger moat around their AI investments, potentially locking in demand for their unique offerings.

8. Regulatory and Government Support

   1. Governments worldwide are increasingly recognizing AI as a strategic priority, providing grants, tax incentives, and other support mechanisms for AI infrastructure. Unlike the largely unregulated fiber boom, AI benefits from direct government interest and support, providing and additional layer of demand stability. Of course it goes without saying that government involvement also raises the potential for government overreach and stifling of innovation. It remains to be seen whether government involvement is beneficial over the long term.

   2. AI is also being integrated into national infrastructure and defense initiatives. If governments adopt AI for large-scale operations or defense purposes, this public sector demand will add significant long-term stability to the market.

9. Lessons Learned from the Past

   1. The hyperscalers are aware of the telecom bubble and subsequent bust, and have likely proceeded more cautiously in their build outs.

   2. Hyperscalers have also created demand ecosystems through partnerships, developer resources, and subscription-based services. These features lock in recurring revenue streams that mitigate the risk of fluctuating demand.

Final Analysis: Is it Different This Time?

While the parallels to the fiber overbuild offer valid cautionary points, hyperscalers appear to be pursuing a more adaptable, resilient, and globally diversified strategy. If demand growth in AI continues its current trajectory, these investments will be foundational, and they will allow for a much faster AI-powered transformation across multiple industries. However, they still face risks from potential regulatory crackdowns, rapid technological shifts, and possible public resistance to AI adoption. All of these factos will shape whether “this time” proves different or not.