"Is there an AI bubble?” Gavin Baker and David George

Is This an AI Bubble?

The current AI boom is fundamentally different from the 2000 telecom bubble. The defining feature of the 2000 era was “dark fiber”—at its peak, 97% of the laid fiber optic cable was unused. In stark contrast, today there are “no dark GPUs.” The intense demand for compute means that GPUs are utilized to their maximum capacity, often to the point of overheating during training runs.

While the planned $3-4 trillion investment in data centers over the next five years sounds immense, the usage metrics support it. Google, for example, has seen a 150x increase in tokens processed in the last 17 months. Crucially, the return on invested capital (ROIC) for the biggest spenders on GPUs has been strong, with companies seeing approximately a 10-point increase in ROIC since ramping up their capital expenditures. This demonstrates a clear and positive return on AI spending so far, a key differentiator from the speculative, zero-return investments of the dot-com era.

The Infrastructure Race: A Battle of Titans

The primary investors in AI infrastructure are the world’s most profitable companies, collectively generating around $300 billion in free cash flow annually with $500 billion in cash reserves. This financial strength provides a massive buffer against market volatility. The mentality, as reportedly articulated by Larry Page, is that this is an existential race they cannot afford to lose.

Competitive Dynamics and “Round-Tripping”

The main competitive battle in AI infrastructure is not between Nvidia and AMD, but between Nvidia and Google. Google’s TPU is currently the only viable alternative to Nvidia for training large models and is a leading inference solution. Google’s ownership of both the chip (TPU) and a leading lab (DeepMind with Gemini) makes it a formidable, integrated competitor.

Concerns about “round-tripping” deals—where a company like Nvidia invests in a model lab that then uses the funds to buy its chips—are overblown. These investments are not driven by a need to finance GPU purchases but by strategic necessity. With labs like Anthropic aligning with Google (TPUs) and Amazon, Nvidia must support other key players like OpenAI and xAI to ensure a competitive ecosystem for its hardware. It’s a rational strategic move to maintain market balance.

The Broader Chip Ecosystem

The market is shaping up as a contest between Nvidia’s full-stack solution (chips, software, networking) and an alliance forming around Google’s TPU, Broadcom, and AMD. Broadcom is positioning itself as an open alternative, offering custom Ethernet-based networking fabrics and ASIC development to compete with Nvidia’s proprietary NVLink and InfiniBand. In this arrangement, AMD serves as a fallback option if a company’s custom ASIC fails. However, many of these custom ASIC programs are likely to be canceled, especially if Google begins selling TPUs directly to external customers.

The Application Layer: New Models, New Margins

It is still very early to predict winners at the application layer, analogous to the pre-Google era of the internet. While big tech companies have a “right to win” due to their vast resources in data, compute, distribution, and talent, execution is not guaranteed.

A fundamental shift is occurring in the business model for software. Unlike traditional SaaS companies with 80-90% gross margins, AI-native applications are inherently more compute-intensive due to scaling laws. This means their gross margins will be structurally lower.

This shift should not be viewed as a negative. For existing SaaS companies, declining gross margins should be seen as a “mark of success,” indicating that customers are genuinely using their compute-heavy AI features. The transition from on-premise software to the cloud, where Microsoft and Adobe successfully navigated lower margins to achieve massive growth, serves as a powerful precedent. Application SaaS companies can leverage their profitable existing businesses to run new AI products at break-even, but few are aggressively pursuing this strategy yet.

The Future: Outcome-Based Models and Robotics

Consumer and Business Models

On the consumer side, the battle for distribution is fierce. While new AI-native browsers have emerged, they face an uphill battle against Google Chrome’s 5 billion users. The advent of sophisticated reasoning capabilities in frontier models is a game-changer, creating the classic internet flywheel effect: more users lead to better data for reinforcement learning, which improves the model, attracting more users.

The dominant business model will likely shift from advertising to being paid for outcomes. Instead of paying for clicks, businesses will pay for resolved customer support tickets, closed sales, or completed tasks. This will squeeze out the inefficiencies of the advertising model, where advertisers systematically overpay for customer acquisition.

Robotics

The field of robotics is advancing rapidly, with a future likely dominated by a competition between Tesla and Chinese companies. The debate between humanoid and non-humanoid robots is largely settled; humanoids have a distinct advantage because they can learn from the vast corpus of human-centric video data available online (e.g., YouTube) and can be more easily trained by humans. The progress shown in videos of Tesla’s Optimus performing diverse tasks points to a very real and tangible future for humanoid robotics.

• Tags:
• Robotics
• Feature
• SaaS
• AI Infrastructure
• Data Centers
• Nvidia
• Business Models
• AI Bubble
• Google TPU
• Gross Margins