🏢 AI's data center problem | Axios AI+ Summary

1. Key Themes

Theme 1: Data Center Backlash as a Political and Existential Risk to AI

Public opposition to data centers has moved from environmental concern to active political force, threatening the bipartisan coalition that has historically supported AI infrastructure buildout.

"America's sharply rising public opposition to data centers — which fuel our everyday use of the internet — could pose an existential threat to AI's growth."

"70% of Americans opposing data center construction in their area."

"Republicans are warning that toxic views of U.S. data centers are directly impacting political outcomes, potentially upending years of support from the GOP and investors for their construction."

The backlash has already surfaced in competitive political races across Michigan, Wisconsin, Texas, and Ohio — signaling this is no longer a coastal/niche issue.

Theme 2: AI's Energy and Water Footprint Is Materially Larger Than Legacy Data Centers

The AI buildout has materially reversed progress toward greener computing, pushing the industry back toward fossil fuels and dramatically increasing both power and water consumption.

"The AI buildout has pushed some developers toward gas turbines and other fossil-fuel sources."

"A Berkeley Lab report updated in June estimates that U.S. data centers could consume 9.5% to 15.3% of U.S. electricity by 2030."

"Data centers can need 5 million gallons of water a day for cooling alone. In 2023, U.S. data centers consumed about 17 billion gallons of water, mostly for cooling."

"The Berkeley Lab says data centers also have an indirect water footprint of about 211 billion gallons to help produce the electricity that's consumed by data centers."

Theme 3: Agentic AI Is Creating a New Attack Surface — and a New Investment Category

Industry insiders believe agentic AI represents a platform-level reset for cybersecurity, creating white space for category-defining companies that don't yet exist.

"Cognition's thesis is that agentic AI is a platform-reset moment for cyber."

"We have this new attack surface that's being brought on by AI and agents." — Elia Zaitsev, former CrowdStrike Global CTO

"AI security didn't exist five years ago. ... But as the adoption of AI dramatically ramps up in the enterprise, it requires a new set of tools." — Gur Talpaz, Cognition co-founder

Theme 4: Data Privacy Architecture Is Becoming a Competitive Differentiator in AI

OpenAI and Anthropic are diverging sharply on how to balance safety monitoring with customer data privacy — and this architectural difference could become a major enterprise procurement decision point.

"We're seeing with more capable frontier models that often risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions." — Aleah Houze, Head of Product Policy, OpenAI

"We have recently announced our plan to require 30-day data retention on our most capable models — a decision we believe will be unpopular with customers who have come to expect zero retention, and pose real risks to our business success (especially if competitors do not follow), but which we believe is essential to detect and prevent sophisticated attacks that span multiple requests." — Anthropic Risk Report

2. Contrarian Perspectives

Perspective 1: Data Centers Are Significantly Less Resource-Intensive Than the Backlash Suggests

While 70% of Americans oppose local data center construction, the article notes important mitigating context that is almost entirely absent from public debate.

"Compared to other industries, data centers use far less water."

"Newer closed-loop designs can reuse water, cutting consumption."

Additionally, globally, data centers only consumed 1.5% of global electricity in 2024 — a number that the IEA projects to double (to 3%) by 2030, which is still a small share of total global consumption. The domestic numbers are more alarming (up to 15.3% of U.S. electricity by 2030), but the public discourse treats all data centers as equivalent, when AI-specific facilities are a subset of the total.

Perspective 2: Anthropic's Mandatory Data Retention Policy Is a Deliberate Self-Harm Play — and May Be the Right Call

Anthropic is openly acknowledging that its 30-day retention policy for its most advanced models is bad for business in the short term, but argues it is necessary for long-term safety. This is a rare instance of a frontier lab explicitly choosing security over revenue.

"A decision we believe will be unpopular with customers who have come to expect zero retention, and pose real risks to our business success (especially if competitors do not follow), but which we believe is essential to detect and prevent sophisticated attacks that span multiple requests." — Anthropic Risk Report

The contrarian take: If Anthropic is right that sophisticated attacks require longitudinal pattern detection, OpenAI's "Private Safety Processing" — however elegantly engineered — may be trading real security for commercial positioning.

Perspective 3: The Smartest Cyber Money Is Betting Against Incumbent Platforms, Not With Them

Cognition's thesis is implicitly bearish on existing cybersecurity platforms (including CrowdStrike, where its founders just spent 13 years). They believe agentic AI is disruptive enough to enable a new entrant to build a broad security platform from scratch — a bet that rarely pays off in cyber.

"The three former CrowdStrikers behind the fund are specifically hunting for the rare technical team capable of building the next broad security platform."

"They oversaw the largest Series A exit in cyber history: SGNL, which they entered at $105 million and exited at $740 million."

The evidence: The SGNL outcome suggests this team has at least one confirmed large outcome from a platform-reset thesis, lending some credibility to the bet.

3. Companies Identified

OpenAI

Description: Leading AI lab and ChatGPT developer
Why mentioned: Previewing "Private Safety Processing," a new zero-data-retention safety architecture for enterprise and API customers; also paused some model training work due to safety concerns
Quote: "We're seeing with more capable frontier models that often risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions." — Aleah Houze, Head of Product Policy

Anthropic

Description: AI safety-focused frontier lab, maker of Claude models (referenced as Fable 5 and Mythos 5)
Why mentioned: Taking the opposing position on data retention — mandating 30-day retention for its most capable models, explicitly acknowledging the business risk
Quote: "We believe [30-day data retention] is essential to detect and prevent sophisticated attacks that span multiple requests."

CrowdStrike

Description: Major publicly traded cybersecurity platform company
Why mentioned: As the prior employer of all three Cognition founders; implicitly represents the incumbent generation of cyber platforms that agentic AI may disrupt
Quote: Elia Zaitsev is "leaving CrowdStrike after 13 years to launch Cognition."

Cognition (new venture fund, not to be confused with the AI coding startup of the same name)

Description: New $170M venture fund targeting AI-native cybersecurity platforms at seed and Series A
Why mentioned: Launched by CrowdStrike's former Global CTO; thesis that agentic AI is a platform-reset moment for cyber
Quote: "Cognition plans to lead or co-lead seed and Series A rounds, and expects to make three or four concentrated investments a year... average checks of $7 million at seed and $12 million at Series A."

SGNL

Description: Cybersecurity startup (identity/access management)
Why mentioned: Cited as the largest Series A exit in cyber history, achieved by Cognition co-founders Talpaz and Sipperly during their time at Brightmind
Quote: "They oversaw the largest Series A exit in cyber history: SGNL, which they entered at $105 million and exited at $740 million."

Brightmind (predecessor fund)

Description: Early-stage venture fund run by Talpaz and Sipperly before Cognition
Why mentioned: Track record vehicle that validates Cognition co-founders' investment acumen
Quote: "Talpaz and Sipperly worked in corporate development at CrowdStrike before launching their own early-stage fund, Brightmind."

Meta

Description: Social media and AI conglomerate
Why mentioned: Launched a new Mac app targeting influencers and small businesses for AI-powered marketing and operations
Quote: "Meta launched a new Mac app to help influencers and small businesses use AI to improve the way they promote products and run their businesses."

Digital Realty

Description: Major data center REIT and operator
Why mentioned: Cited as a visual example (photo) of data center infrastructure next to a power substation in Ashburn, Virginia
Quote: (Descriptive caption only — "Cooling fans and generators at a Digital Realty data center next to a power substation in Ashburn, Va.")

Liquid Death / Garage Beer

Description: Canned water brand and craft beer brand, respectively
Why mentioned: Cultural signal of how broad the anti-data-center backlash has become — a celebrity-backed consumer brand campaign encouraging Americans to mail urine to data centers in protest of water use
Quote: "Former NFL star Jason Kelce announced a partnership between his Garage Beer company and Liquid Death to encourage Americans to send actual urine to data centers in protest of their water use."

Teleport (sponsor)

Description: Identity and access security company focused on AI agents
Why mentioned: Newsletter sponsor; promoting "Agent Trust" framework for securing AI agents at runtime, not just at instantiation
Quote: "Runtime is the new perimeter."

4. People Identified

Elia Zaitsev

Description: Former Global CTO of CrowdStrike (13 years); co-founder of Cognition
Why mentioned: Departing a major platform company to lead a focused AI-cyber venture fund; described as the fund's "superpower" by co-founders
Quote: "We have this new attack surface that's being brought on by AI and agents."

Gur Talpaz

Description: Former CrowdStrike corporate development executive; Cognition co-founder; previously co-founded Brightmind
Why mentioned: Architect of the SGNL investment (largest Series A exit in cyber history); articulating Cognition's market thesis
Quote: "AI security didn't exist five years ago. ... But as the adoption of AI dramatically ramps up in the enterprise, it requires a new set of tools."

Tayler Sipperly

Description: Former CrowdStrike corporate development executive; Cognition co-founder; previously co-founded Brightmind
Why mentioned: Describes Zaitsev's value to portfolio founders and the fund's differentiation
Quote: "Being able to offer [Zaitsev's expertise] to founders is 'deeply unique in the market.'"

Aleah Houze

Description: Head of Product Policy at OpenAI
Why mentioned: Articulating OpenAI's rationale for Private Safety Processing and the emerging nature of cross-session AI risks
Quote: "We're seeing with more capable frontier models that often risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions."

Jason Kelce

Description: Former NFL star; entrepreneur (Garage Beer)
Why mentioned: Emblematic of the celebrity-driven, populist character of the anti-data-center backlash
Quote: "Former NFL star Jason Kelce announced a partnership between his Garage Beer company and Liquid Death to encourage Americans to send actual urine to data centers in protest of their water use."

5. Operating Insights

Insight 1: Zero-Data-Retention is Becoming a Required Enterprise Feature, Not a Premium Enterprise AI buyers are conditioning procurement on ZDR protections — and OpenAI is engineering around this constraint rather than asking customers to give it up. Operators building AI products for regulated industries (finance, legal, healthcare) should design for ZDR from day one, with customer-controlled encryption keys, rather than treating it as a future upgrade.

"Customer data can stay on customer-controlled infrastructure, or be stored by OpenAI with encryption keys controlled by the customer."

Insight 2: Concentrated, Operator-Led VC Is the New Template for Domain-Specific Cyber Funds Cognition's model — 3-4 deals per year, $7M seed / $12M Series A checks, led by a former practitioner CTO — is a replicable structure for any domain where AI is creating new attack surfaces (robotics, autonomous vehicles, agentic workflows). Founders in these spaces should seek investors with hands-on operator credibility, not just pattern-matching generalists.

"They see Zaitsev as their new fund's 'superpower,' Sipperly said, adding that 'being able to offer' his expertise to founders is 'deeply unique in the market.'"

6. Overlooked Insights

Insight 1: The Secrecy Surrounding Data Center Development Is Accelerating the Backlash The article briefly notes that many data centers are "developed in secrecy" — a tactical choice by developers that is almost certainly backfiring politically. The absence of community engagement or disclosure is amplifying distrust, particularly in states like Alabama where non-disclosure agreements have reportedly been used. For any company siting or investing in AI infrastructure, proactive community engagement and transparency may now be a material risk-management imperative, not a PR nice-to-have.

"Many [data centers] are developed in secrecy."

Insight 2: OpenAI Has Paused Model Training Work Due to Safety Concerns — A Significant but Under-Emphasized Disclosure Buried in the data-retention story is a significant admission: OpenAI has actively paused some training work on new models to address safety concerns. This is a material operational fact that suggests frontier capability timelines may be slipping for safety-related, not just technical, reasons — with implications for competitive positioning against Anthropic, Google DeepMind, and others who are not making similar pauses.

"OpenAI said Tuesday that it has paused some work in the training of new models to deal with safety concerns, while Anthropic says it doesn't see a current need to do the same."