From Scott’s Brain: Watermarks, Water Claims, and What “AI” Is Actually Worth

I won’t spend a lot of time this week dissecting the overall corporate moves by Google as they shift the leadership of their AI division yet again. While reorganizing DeepMind and moving executives like Demis Hassabis and Koray Kavukcuoglu is interesting from a management perspective, the reality remains that Google continues to slip as a foundation model player. Right now, their tools are sitting well behind the capabilities delivered by OpenAI and Anthropic.

The development that did catch my eye this week is the recent statement by Anthropic regarding watermarking. They have announced that content created by Claude will soon contain a machine-readable watermark. As someone who writes this newsletter independently and uses Claude to format the final layout, this is highly relevant. Under this new protocol, even though every word of this section (aside from the market data) is written entirely by me, the formatting pass will suddenly cause the document to be flagged as AI-generated.

There are, of course, many ways to remove watermarks from documents, but it is important to note that most of them are not legal. Watermarks often act as a digital representation of a copyrighted work, and actively stripping a watermark generally violates copyright fair use protections. It introduces a frustrating friction point for creators who use human-designed machine intelligence purely for editorial formatting rather than generation.

The Data Center Debate: Separating Fact from Fiction

To wrap up this week’s expert column, we need to talk about data centers. Anti-data center sentiment is growing rapidly in local municipalities, while pro-data center sentiment remains highly stable at the enterprise and government levels. What is deeply concerning to anyone architecting these solutions is the sheer volume of false claims being propagated by both sides.

When infrastructure planning is driven by misinformation, both the community and the technology suffer. Below is a breakdown contrasting the false claims made by both sides of the argument:

Expanded AI & AI-Infrastructure IPO and Capital-Markets Table

The capital markets continue to aggressively price AI infrastructure. The table below separates companies that are already trading from those with confidential filings, proposed SPAC combinations, and private financing discussions. Note: Read the valuation column as a description of the transaction basis, not as a universal measure of comparable market value.

Conclusions for Investors and Strategic Planners

The AI capital-markets cycle has moved firmly from anticipation into uneven price discovery. An “AI” label is no longer a uniform investment thesis. Public outcomes differ sharply—companies with massive fundamental capital requirements are being scrutinized heavily, while scarcity in tradable float can trigger massive, disconnected surges.

As we look toward the potential public debuts of OpenAI and Anthropic, dropping inference prices introduce a vital valuation question. Falling API costs stimulate demand and improve customer economics, but they compress revenue per token. Growth in raw usage and higher-value agent services must outpace this price decline; revenue growth without corresponding gross-margin discipline will struggle to support trillion-dollar public valuations. Furthermore, public float and capital structure matter just as much as headline numbers. Staggered lockups and limited floating shares fundamentally alter how a stock trades long after the IPO celebration ends.

Finally, we must recognize that AI infrastructure is now a distinct capital-markets category. From memory and networking to launch systems and power conversion, the shared risks are uniform: extraordinary capital requirements, energy availability, and the looming possibility that compute supply eventually expands faster than economically sustainable demand.