Google backs financing for Anthropic’s $19B AI data-center leases

Google backstops financing for Anthropic’s $19B data-center leases

TeraWulf signed Google credit-support agreements tied to roughly $19 billion in contracted Anthropic leases at its Kentucky AI campus. A separate construction agreement with Fluor covers three data-center buildings. The 401 MW project is scheduled for phased delivery from late 2027 into 2028.

Why it matters. Google’s balance sheet is being used to reduce financing risk for infrastructure serving a rival AI lab. The $19 billion is contracted lease value over time, not money raised or immediately spent. The credit backstop has contractual triggers and does not amount to an unconditional guarantee of every project cost.

US bank regulator tells examiners to judge AI agents by consequences, not model sophistication

The Office of the Comptroller of the Currency published its first examiner framework for what it calls “super intelligence,” encompassing AI and agentic systems. Examiners are directed to weigh autonomy, human oversight and the consequences of failure rather than treating AI as a wholly separate risk class.

Why it matters. A basic agent authorized to move money may warrant stricter controls than a more advanced model used for low-stakes research. The framework guides examiners; it creates no new enforceable requirements for banks.

Tencent unveils Hy4 with defensive-cybersecurity training restrictions

Tencent introduced Hy4 and says its cybersecurity training data were restricted to defensive work, including detecting and fixing vulnerabilities. It excluded instructions for offensive hacking and says internal evaluations did not find increased cyberattack capability.

Why it matters. Model developers are experimenting with limiting dangerous capabilities at the training-data level. Tencent’s safety results are company-reported, not independent proof that the model cannot be misused; planned publication of weights and a safety report will allow closer scrutiny.

India warns banks of AI-agent-assisted attacks at machine speed

India’s cybercrime coordination centre warned banks and non-bank lenders about financially motivated intrusions assisted by AI agents. Its advisory describes the potential for automated compromise and urges the financial sector to strengthen defenses.

Why it matters. Security teams need controls that can respond at automation speed, particularly around credentials and privileged workflows. The advisory is an official warning, not independent measurement of how many attacks were carried out autonomously.

Figma says in-house AI models are helping lower inference costs

Figma says it has more than doubled the proportion of AI work handled by its proprietary design model, while more than 80% of large paid customers used AI credits weekly through October. The company raised its 2026 revenue forecast alongside its quarterly results.

Why it matters. Vertical AI products may improve margins by routing common tasks to specialized models instead of relying entirely on frontier APIs. Claims of better quality and latency come from Figma; the revenue forecast is guidance, not realized results.

Watch whether Google’s credit guarantees become a standard way to finance AI campuses, and whether Tencent releases Hy4’s promised safety materials for independent inspection.