The AI Infrastructure Buildout: Why the Bubble Narrative Misses the Mark

The current multi-trillion-dollar investment in AI infrastructure is not a replay of the 2000 telecom bubble. It is a rational response to intense competitive pressure. While skeptics point to massive data center spending as proof of a bubble, the system differs from the dark fiber era in one key way: GPUs are being used immediately rather than sitting idle. This spending acts as a defensive moat for the largest tech companies. Those who view this buildout through a traditional, short-term ROI lens will miss the strategic necessity of these investments. For founders and investors, the advantage lies in recognizing that current gross margin pressure is not a sign of failure, but a required cost of entry for the next generation of software dominance.

The Dark Fiber Fallacy and the Reality of Utilization

The most common critique of the current AI boom is that we are over-building infrastructure, a sentiment rooted in the trauma of the 2000 telecom bubble. Gavin Baker, CIO of Atreides Management, notes that the telecom bubble was defined by dark fiber, which refers to infrastructure laid down but never lit up because the necessary networking technology did not exist.

The current environment is different. There is no such thing as a dark GPU. Every chip purchased is immediately put to work, often pushing thermal limits during training runs.

The year 2000 internet bubble or telecom bubble was defined by something called dark fiber. Contrast that with today. There are no dark GPUs.

-- Gavin Baker

The system's response to this demand indicates its health. Unlike the internet era, which required the difficult task of building a two-sided network of websites and users, AI tools benefit from immediate distribution via existing cloud infrastructure. The large numbers, representing trillions in data center spend, are being underwritten by companies with 500 billion dollars in cash and 300 billion dollars in annual free cash flow. This is not speculative capital; it is a war chest being deployed to secure dominance.

The Hidden Dynamics of Round-Tripping

Skeptics often cite round-tripping, where companies fund customers to buy their own products, as evidence of an artificial bubble. Baker acknowledges this is happening, but he frames it as a rational competitive response rather than a sign of systemic instability.

In this system, Nvidia's primary competitor is not another chipmaker, but Google. Because Google owns the TPU and the Gemini model, they have a strong incentive to fund labs like Anthropic to ensure their own hardware ecosystem thrives. When a giant like Google funds a competitor, the system forces a response. Nvidia's support for other labs is a defensive maneuver to maintain its position. The round-tripping is the system routing around potential monopolies, not a sign of a market collapse.

Why Margin Pressure is a Signal, Not a Failure

A recurring point of tension is the decline in gross margins for AI-native software companies. Many SaaS companies are afraid to move away from the 80 to 90 percent margins of the 2021 era, fearing that investors will punish them. Baker argues this is a misunderstanding of the current shift.

I worry that application-sask companies are trying to preserve their existing gross margin structures because they believe that if their gross margins go down, their stocks will go down. It is definitely impossible to succeed in AI without gross margin pressure.

-- Gavin Baker

The transition to AI is similar to the shift from on-premise software to the cloud. Cloud margins were lower than perpetual license margins, yet Microsoft successfully navigated this transition to become a dominant stock. Companies that lean into lower margins to gain usage and data, which are the ingredients for reinforcement learning flywheels, will create a lasting advantage that high-margin incumbents cannot touch. Those clinging to legacy margins are choosing to be disrupted.

The 18-Month Payoff: Reasoning and User Flywheels

The most non-obvious insight is that reasoning capabilities have changed the economics of model companies. Before reasoning, a model without unique data or distribution was a rapidly depreciating asset. Now, the ability to use post-training data to create a verified reward loop creates a competitive advantage that compounds over time.

This creates a feedback loop: better products lead to more users, which leads to better algorithms, which leads back to a better product. While this is not yet fully spinning for all players, it is the mechanism that will separate the winners from the losers over the next 18 months.

Key Action Items

• Reframe Margin Expectations: If you are building in the AI space, stop treating gross margin compression as a badge of shame. It is a badge of usage. Over the next quarter, shift your internal KPIs to prioritize outcome-based metrics over traditional SaaS margin targets.
• Identify Your Reasoning Moat: Assess whether your product architecture allows for a verified reward loop. If your model is not getting smarter through user interaction, you are building a commodity, not a platform. This investment pays off in 12 to 18 months.
• Monitor the ASIC Landscape: Watch for Google's external TPU strategy. If TPUs become widely available, it will trigger a massive shift in the GPU-dependent competitive landscape. Prepare for a pivot in your infrastructure strategy if your primary dependency relies on a single hardware vendor.
• Ignore the Bubble Hype: Focus on the ROIC of the major spenders. As long as the biggest tech companies continue to see positive returns on their CAPEX, the infrastructure buildout is sustainable.
• Embrace Outcome-Based Pricing: Move away from seat-based or flat-fee models. Start experimenting with pricing based on task resolution or successful outcomes, as this is where the market is trending. This requires immediate groundwork but creates a more durable business model in the long term.