Can Anthropic's $50 Billion DC Plan Work? The Reality of Power, ROI, and the 'CapEx Treadmill'

AI infrastructure investment has entered a phase of 'unprecedented' scale. Reports indicate that Anthropic plans to invest $50 billion to build its own data centers within the U.S. We discussed the deployment across multiple locations like Texas and New York, relationships with hyperscalers, power constraints, and the outlook for return on investment from a multidimensional perspective. This article explains the key points from the broadcast from the viewpoints of investors and business managers.

1. What happened (Summary of facts)

• Anthropic: $50 billion investment to build its own U.S. data centers. Multi-year, multi-state project. Details such as total power (gigawatts) have not been disclosed.

• While reliance on AWS and Google Cloud continues, there is an implication of wanting to 'increase internal control.' Host Ed pointed out that they remain 'heavily reliant' on hyperscale partners.

• Host Caroline touched on the stance of Anthropic, which has a valuation of around $83 billion, noting that while they understand the scaling requirements, they are becoming more conscious of financial discipline.

2. Putting investment scale into perspective: Comparisons with Meta and others

• Meta has a new DC plan in Wisconsin worth approximately $1 billion. In addition, they have a policy of pursuing large-scale infrastructure investments in a chain over several years. Kurt Wagner explained that the motivation is to regain control—'owning it themselves rather than renting other companies' DCs.'

• The broadcast introduced the move to deploy DC clusters in parallel across the U.S., including 'Ohio, Texas, Louisiana...' It was also pointed out that there is pressure to deliver 'first-class models' (top-tier LLMs) within the next six months.

3. Keys to supply and demand: Power, sites, and construction capacity

• The biggest bottleneck is power. The guest commented, 'Nuclear power takes time. In the near future, gas-fired power will likely increase.' Pressure to restart closed nuclear plants is also rising, but safety reviews and construction periods are long.

• Data centers are homogeneous 'specialized buildings,' and progress cannot be made unless construction companies, equipment, cooling, and transmission/distribution infrastructure are all available simultaneously.

4. Focus on monetization: Will the CapEx 'pay off'?

• The market's question is consistently about the recovery curve. The guest voiced the investor perspective, asking, 'Please tell us about the pace of investment and how it reflects on revenue,' and 'Is it necessary to spend everything upfront now?'

• In Meta's case, early signs of ROI, such as improvements in advertising effectiveness, are visible, but a time lag between data center investment and revenue contribution is inevitable.

• The interest rate environment is also important. The consensus that 'rates will fall' is wavering, and in the short term, offsetting tailwinds and headwinds are mixed.

5. Market update: AMD, NVIDIA, Cisco

• AMD announced at its Investor Day that it expects 25% annual revenue growth, with data centers growing at 80% annually, and its stock price rose by about 10%. The host summarized this as a positive reaction to the 'five-year accelerated growth outlook.'

• NVIDIA has a market capitalization of approximately $4.7 trillion. The view is that the 'future tone' in the next earnings report will be the biggest point of focus.

• Cisco is focused on recovering its presence in AI networking. The question is whether it can 'consistently win deals' in competition with Broadcom and others.

6. Counterarguments and Dissent: Is There Really Such a Need for DCs?

• Guest Kimberly Forrest is skeptical, stating, 'Current LLMs are brute force. Humans will inevitably innovate and find more efficient methods. We won't need as many DCs as planned.'

• However, she is bullish on the actual demand for AI, holding positions in NVIDIA, AMD, Micron, and Intel. She points out, 'Even if methods evolve, the role of high-performance chips will remain.'

7. 2026 Discussion Points: A One-Off or a 'CapEx Treadmill'?

• The guest identified the biggest issue as whether this massive CapEx is a one-time event or a 1990s telecom-style 'continuous cycle of capital investment.'

• If the former, the room for shareholder returns expands; if the latter, upside valuation will remain suppressed. The outcome depends on progress in Power Purchase Agreements (PPAs), cooling technology (liquid/immersion), and inference efficiency (distillation/quantization/caching).

8. Implications for Investors and Business Leaders (Checklist)

8-1. Builders (Business/Infrastructure Side)

1. Power Reality: Grid connections, substations, fuel, and regulatory progress in local municipalities.

2. Quality of Scale: The optimal mix of hyperscaler dependence versus in-house construction.

3. Future Lightweighting: Avoiding 'over-engineering' through inference efficiency, modular design, and reusability.

8-2. Investors (Capital Side)

1. Recovery Curve: Alignment of contracts (utilization rates) and operational start dates.

2. Technology Inflection Point: The risk/opportunity that model efficiency changes the nature of demand and could shrink DC requirements.

3. Interest rates and cash: The combination of rates and depreciation determines free cash flow.

Summary

Anthropic's $50 billion is not just a 'stretch,' but a challenge to secure self-sovereignty and optimize long-term costs. At the same time, the realities of power, construction, and interest rates are clashing with a future focused on AI efficiency. The keys are 'visibility of utilization rates,' 'tracking technical efficiency,' and 'capital allocation discipline.' Building data centers does not guarantee a win. 'Only what is needed, when it is needed'—how thoroughly this basic principle can be implemented will determine the next winners and losers.

Recommended Articles

Next Big Wave (Growth Stocks, Seeds of Ideas, Deep Dives into Trends)