Gavin Baker: Why TSMC Capacity Is the Only AI Metric That Matters

Gavin Baker, a prominent voice in the artificial intelligence investment landscape, has issued a stark warning and a clear directive for investors watching the current tech boom. He identifies TSMC’s capacity decisions as the single most critical indicator for determining whether the AI sector is facing an imminent bubble burst.

Baker’s analysis suggests that while hype is high, fundamental supply constraints may prevent a traditional market crash. His insights offer a nuanced view of the economic mechanics driving today’s generative AI revolution.

Key Takeaways from Baker’s Analysis

• Anthropic’s Efficiency: Anthropic likely burns 80% less capital than OpenAI, indicating superior operational efficiency in model training.
• Revenue Shift: The industry is moving from unlimited plans to pay-per-use models, potentially pushing ARR beyond $200 billion this year.
• Power Constraints: Electricity shortages will persist until 2027-2028, when orbital computing may offer a structural solution.
• Bubble Defense: Current infrastructure build-outs are funded by operating cash flow, unlike the debt-fueled dot-com bubble of 2000.
• Hardware Longevity: Older GPUs remain valuable for inference tasks as prefill and inference workloads are decoupled.
• The TSMC Indicator: Investors should monitor TSMC’s expansion plans closely; aggressive capacity growth signals confidence, while restraint suggests caution.

The Economics of Token Pricing and Efficiency

The financial dynamics of major AI labs are shifting dramatically. Baker highlights that Anthropic operates with significantly lower burn rates compared to its primary competitor, OpenAI. This disparity suggests that not all players are spending at the same unsustainable pace. Some companies are optimizing their training processes to stretch every dollar further.

This efficiency is crucial as the market transitions from experimental adoption to enterprise-scale deployment. Baker predicts that OpenAI and Anthropic could see annual recurring revenue (ARR) exceed $200 billion. This projection relies on a fundamental change in how services are sold.

From Unlimited to Pay-Per-Cup

The era of flat-rate, unlimited access to AI models is ending. Instead, the industry is adopting a pay-per-use structure, akin to buying coffee by the cup rather than an all-you-can-drink subscription. This model allows providers to charge premium prices for frontier tokens used in complex reasoning tasks.

Enterprise clients are willing to pay more for higher accuracy and reliability. By segmenting users based on usage intensity, AI companies can maximize margins. This pricing power helps offset the massive costs associated with GPU clusters and energy consumption. It also creates a more sustainable revenue stream that is tied directly to value delivery.

Infrastructure Bottlenecks and Energy Realities

Physical constraints are becoming the primary limiter of AI growth. Baker notes that electricity shortages will likely continue through 2027 and 2028. Data centers require immense power, and grid infrastructure in many Western regions cannot keep up with demand.

This bottleneck forces a strategic pause on unlimited expansion. Companies must prioritize efficiency over raw scale. The solution may lie in unconventional places, such as orbital computing. While still theoretical for mass adoption, space-based data centers could bypass terrestrial energy limits.

The Role of TSMC in Market Stability

Unlike the dot-com bubble, which was fueled by easy credit and speculative debt, the current AI boom is largely financed by operating cash flow. Tech giants like Microsoft, Google, and Meta have deep pockets and strong balance sheets. They are building infrastructure based on actual demand, not just future promises.

Baker argues that the fundamental wafer shortage, controlled primarily by TSMC, acts as a natural circuit breaker. TSMC manages the supply of advanced chips. If they expand capacity cautiously, it prevents an oversupply of AI hardware that could lead to a price collapse. Conversely, rapid expansion would signal strong demand validation.

Investors should watch TSMC’s capital expenditure reports. A conservative approach indicates a measured market, while aggressive scaling confirms the sustainability of the boom. This dynamic makes TSMC the central bank of the AI economy.

Hardware Lifecycle and Technical Evolution

Technological progress does not necessarily render old hardware obsolete. Baker explains that the separation of prefill and inference stages in large language models changes how GPUs are utilized. Prefill involves processing the input prompt, while inference generates the output tokens.

Older GPUs can handle inference tasks efficiently, even if they struggle with the heavy compute requirements of prefill or training. This division of labor extends the useful life of existing hardware fleets. It reduces the pressure to constantly upgrade to the latest NVIDIA chips for every task.

Algorithmic Breakthroughs and Future Scenarios

If a significant algorithmic breakthrough occurs, the need for brute-force computing could diminish. More efficient models might achieve similar results with fewer parameters. However, Baker remains cautious about predicting exactly when or if this will happen.

For now, the trend is toward larger, more capable models. The decoupling of workload types ensures that the entire ecosystem, from legacy chips to cutting-edge accelerators, remains relevant. This diversity stabilizes the hardware market and prevents sudden crashes in chip valuations.

Industry Context and Strategic Implications

The broader AI landscape is maturing rapidly. Early-stage experimentation is giving way to integrated enterprise solutions. Companies are no longer just testing chatbots; they are embedding AI into core business workflows. This shift drives consistent, predictable revenue streams.

Western tech firms dominate this phase due to their access to capital and talent. However, global competition remains fierce. Regulatory frameworks in the US and Europe are also shaping how these technologies are deployed, adding another layer of complexity to the market dynamics.

What This Means for Stakeholders

Developers and businesses must adapt to the new pricing models. Budgeting for AI usage requires careful monitoring of token consumption. Relying on flat-rate assumptions is risky. Understanding the difference between prefill and inference costs can help optimize application architecture.

Investors should look beyond headline revenue numbers. Focus on unit economics and cash flow sustainability. The companies that manage their burn rate effectively, like Anthropic according to Baker, are better positioned for long-term success. Avoid firms that rely solely on speculative funding rounds.

Looking Ahead: The Path to 2028

The next few years will define the stability of the AI market. Power infrastructure development will be critical. Until 2027, expect continued constraints on expansion. After that, new energy solutions, including orbital options, may alleviate some pressures.

TSMC’s role will remain pivotal. Their decisions will ripple through the entire supply chain. Monitoring their产能 (capacity) strategies provides the clearest signal of market health. As the industry evolves, the focus will shift from pure scale to sustainable efficiency.

Gogo's Take

• 🔥 Why This Matters: The shift to pay-per-use models validates AI as a serious enterprise tool, not just a novelty. It means AI is generating real economic value, supporting the high valuations of leading firms.
• ⚠️ Limitations & Risks: Dependence on TSMC creates a single point of failure for global AI infrastructure. Geopolitical tensions or supply chain disruptions could stall progress abruptly. Energy constraints also limit geographic flexibility for data centers.
• 💡 Actionable Advice: Monitor TSMC’s quarterly earnings calls for capex guidance. Optimize your AI applications to separate prefill and inference workloads to reduce costs on older hardware. Prioritize vendors with transparent, usage-based pricing structures.