Are AI Data Centers Overbuilt? Prominent Investor Counters: The Real Problem is Underbuilding! $500 Billion Compute Can Break Even in 10 Months

Prominent investor and founder & chief investment officer of Atreides Management, Gavin Baker, has recently offered a radically different perspective on the future of AI infrastructure. He argues that the market's primary concern—"overbuilding" of AI data centers—may not be the real risk. Instead, the more pressing issue is insufficient compute supply. Baker points out that while there are approximately 1.5 billion knowledge workers globally, only a tiny fraction currently use AI extensively. Once AI adoption broadens, compute demand could far exceed existing supply.

On the a16z Podcast aired August 31, Baker stated that the diffusion of AI demand has only just begun. Multiple supply chain constraints—including power, chips, copper, data center construction, and government approvals—are collectively suppressing the pace of new compute capacity. He even suggests that if supply fails to keep up with demand over the long term, AI service prices may not continue to decline as the market generally expects. Instead, "compute inequality" could emerge, allowing large enterprises and well-funded users to access cutting-edge AI capabilities first.

Baker cited Atreides' own usage as an example of rapidly rising demand. He revealed that the company's internal token consumption increased 100-fold from March to August. After recently gaining access to Grok Enterprise, just two employees using it could increase the company's monthly token spending by another 10 to 20 times. Baker emphasized this is not wasteful spending but reflects genuine increased usage as AI becomes embedded in investment research and other knowledge work.

a16z partner David George noted that AI industry revenue is still primarily supported by fewer than 10 million heavy users. Compared to the global population of about 1.5 billion knowledge workers, AI adoption remains in a very early stage. He added that within AI-native companies, some high-frequency AI users—such as engineers—may consume 10 to 100 times more tokens than average engineers, indicating highly concentrated AI usage even within enterprises.

This demand structure leads Baker to believe that AI infrastructure may not be "overbuilt" but rather "not being built fast enough." He points out that the scale of AI data center construction has grown so large that it now impacts the raw supply capacity of industries such as power, semiconductors, copper, and others. As AI servers continue to increase, everything from chips to power equipment, transformers, cooling systems, and data center land could become bottlenecks limiting compute expansion.

Baker is particularly optimistic about the capital return speed of AI data centers. Citing experiences from new cloud providers like Nebius and CoreWeave, he notes that a 1GW data center built around NVIDIA (NVDA-US) GPUs costs approximately $40 to $50 billion to fully deploy. Some customers prepay about 50% to 60% of the fees, significantly reducing the capital operators must initially invest. If compute power is leased at spot market prices after the data center goes live, the payback period could shorten to about 9 to 10 months.

In other words, a 1GW-class AI data center costing around $50 billion could recover its initial investment in under a year under specific financing and leasing conditions. Baker considers this an exceptionally rare investment opportunity in his career—such massive capital deployment paired with such a short payback cycle.

Financing conditions for AI data centers are also lowering the capital barrier for large-scale projects. Baker notes that NVIDIA GPUs themselves already have a relatively mature financing market, with major asset management firms like Blackstone (BX-US), KKR (KKR-US), and Apollo Global Management (APO-US) participating in related financing. According to a16z's summary of Baker's interview, in a roughly $50 billion NVIDIA GPU data center deployment, about $35 billion could be financed through low-cost debt, reducing shareholder equity investment to around $15 billion, supplemented by residual value protection mechanisms.

Baker believes that as AI model capabilities improve and GPU usage cycles extend, the monetization efficiency of per-unit compute still has room to increase. Therefore, if AI service demand continues to grow rapidly, the actual equity payback period could be even shorter than the current 9–10 month estimate.

Resource allocation strategies among AI model providers are also beginning to diverge. Baker argues that frontier models, open-weight models, and application companies do not necessarily form a zero-sum competition. Different types of AI companies can simultaneously benefit from overall token demand growth.

In the frontier model space, OpenAI and Anthropic remain in intense competition. Both companies hold unreleased new models, allowing them to adjust release timing based on competitors' product rhythms and market demand. They can also reconfigure limited compute resources between training and inference. This resource allocation could directly impact short-term revenue, as the same GPUs can be used for model training or shifted to inference services that generate direct income.

Large tech companies' strategies are also beginning to diverge. Meta Platforms (META-US) is accelerating frontier model competition again, while Alphabet (GOOGL-US), Google's parent company, is being more cautious in some frontier model races and increasing cash flow by offering TPU compute externally.

Baker also warns that the market often confuses "open weights" with true "open source." Even if model weights are open, generating tokens still requires massive compute. Under the same usage conditions, open models do not eliminate computational costs—the main difference lies in the model provider's gross margin and business model.

He believes that falling prices for open models could actually stimulate overall AI usage. Companies could deploy lower-cost open models internally to handle large volumes of repetitive or standardized tasks, while reserving complex reasoning and planning tasks for frontier models.

As a result, enterprises may gradually adopt a hybrid architecture of "frontier models + privately fine-tuned open models." Companies can train proprietary models using their own data and use routing systems to determine which model handles each task. Complex reasoning tasks go to frontier models, while routine work is handled by lower-cost in-house models, simultaneously reducing data leakage and privacy risks.

Baker believes the next battleground may center on the so-called "smart abstraction layer"—a unified entry point through which enterprises and general users access AI capabilities. Users only need to state their needs, and the backend system autonomously decides which model to call, how to allocate compute, and how to protect data.

Microsoft (MSFT-US), Fireworks, Databricks, and various vertical agent companies are all competing for this entry point. Baker believes the real competition isn't just model capability but also seamless orchestration across models, data security, cost control, and execution reliability.

In discussions about an AI bubble, Baker acknowledges that historical experience is still worth referencing. Whenever a disruptive technology emerges, asset prices typically rise first, attracting massive capital and potentially causing overbuilding. However, he argues that this round of AI infrastructure investment differs significantly from past tech bubbles because a substantial portion of current spending is supported by strong operating cash flows from large tech companies, rather than relying entirely on high-leverage debt financing.

Nonetheless, this doesn't mean the market is entirely risk-free. Baker warns that even by 2028, if all currently announced AI data center construction plans are completed, the market might still lack significant idle capacity. Power supply, land, permits, and political factors could all cause actual construction speeds to fall below paper plans.

If supply remains tight, token price declines may be slower than market expectations. In extreme cases, prices could even rise during surges in demand. This would create a clear "compute inequality" problem, as large enterprises and well-funded users can afford expensive frontier models, while SMEs and general users may be forced to use weaker models.

Baker recently also pointed out on social media that the rapid expansion of AI usage scale will make compute spending a key factor in corporate competitiveness. He expects that in the future, many knowledge-based industries may establish a minimum AI token spending threshold. Companies unable to afford sufficient compute may gradually lose competitiveness. This is the "compute inequality" he warns of.

His views have also been echoed by some market participants. The CEO of Blackstone