Is AI Trending Up or Down in 2026? | AI Reality Check

What It Covers

Cal Newport and tech commentator Ed Zitron review three major AI stories from early 2026: the OpenClaw agent framework hype and OpenAI's acquisition of it, Anthropic's military contract dispute with the Department of Defense, and the growing evidence that announced AI data center construction is vastly overstated, with only 15.2 of 115 planned gigawatts actually under construction.

Key Questions Answered

• •Media accountability gap: AI news cycles move fast enough that no publication follows up on previous predictions. Stories like OpenClaw's "singularity moment" coverage vanish without correction when outcomes fail to materialize. Readers can calibrate their reaction to new AI headlines by actively searching what happened to the last three major AI stories — almost none resolved as dramatically as initially reported.
• •LLM prompting bias toward sci-fi: Research shows that when any prompt signals the responder is an AI, the output reliably shifts toward dystopian, self-aware narratives. This means OpenClaw agents posting on social network Multbook were simply generating what LLMs predict AI social posts look like — not demonstrating emergent behavior. Recognizing this pattern helps filter genuinely novel AI developments from prompt-induced theater.
• •OpenClaw's real cost exposure: Anthropic's Claude Max subscription ($200/month) was briefly connectable to OpenClaw agents, allowing users to burn an estimated $2,700 worth of API compute per $200 paid. Anthropic cut off this access shortly after closing a $30 billion funding round in February. This cost structure reveals that frontier model API usage remains economically unsustainable even at subscription pricing, not just pay-per-token rates.
• •Anthropic's revenue discrepancy: Under sworn court affidavit during its Department of Defense lawsuit, Anthropic's CFO stated total lifetime revenue of $5 billion. This conflicts sharply with separately reported figures including $4.5 billion in 2025 annual revenue alone. No major financial publication has reconciled these numbers. Investors evaluating AI company valuations should treat annualized revenue projections with skepticism until companies file public S-1 disclosures.
• •Data center construction reality check: Of 115 gigawatts of AI data centers announced for completion by 2028, only 15.2 gigawatts are actually under construction according to Sightline Climate research. At a 1.35 PUE efficiency ratio, that 15.2 gigawatts represents roughly $285 billion in GPU capacity — far below Nvidia's stated forward sales visibility of $500 billion by end of 2026, suggesting significant GPU inventory is warehoused with no installation destination.
• •AI startup exit problem: AI startups are structurally difficult to acquire because their core product is a wrapper around a model owned by a third party, with no proprietary IP. Recent acquisitions like Windsurf and Inflection AI transferred founders and talent, not products. With an estimated $200–300 billion in venture capital locked in AI startups, and LLM compute costs rising with user volume rather than falling, a forced fire-sale exit cycle becomes increasingly probable.