Anthropic's Mythos model to use 3.5GW of Google TPUs by 2027

Anthropic's launch of its most powerful AI model, Mythos, is set to drive a massive 3.5 gigawatts of demand for Google's Tensor Processing Units (TPUs) by 2027, according to a new report from CITIC Construction Investment. The projection highlights a significant, long-term revenue catalyst for Google's cloud division and validates its specialized AI hardware strategy.

"Mythos is Anthropic's strongest model to date, with a leap in capabilities for code, reasoning, and research," the CITIC report noted. "The Google TPU industry chain is expected to continuously benefit from the expansion of cutting-edge model training needs."

The model demonstrates a pronounced advantage in identifying and exploiting cybersecurity vulnerabilities, including zero-day threats, a capability that has led Anthropic to restrict its release to a small group of core customers. Its power is particularly effective in complex, legacy IT environments common in the financial sector. In a technical demonstration, Mythos identified a 16-year-old vulnerability in the widely used FFmpeg software library.

This surge in specialized hardware demand provides a major endorsement for Alphabet's (GOOGL) expensive, long-term investment in custom silicon. As AI competition intensifies with rivals like Microsoft-backed OpenAI and Amazon, securing large-scale, multi-year hardware commitments from a leading model developer like Anthropic solidifies a crucial and high-margin revenue stream for Google Cloud.

A Step Change in Cyber-Offense Capabilities

The restricted access to Mythos stems from its advanced offensive security capabilities. A Cloud Security Alliance briefing warned that the model represents "a step change" in AI that "lowers the cost and skill floor for discovering and exploiting vulnerabilities faster than organizations can patch them."

This poses a particular risk to the banking industry, which often relies on a mix of modern and decades-old legacy systems. In response to the threat, government officials in the US, UK, and Canada have already met with banking executives. Anthropic has initiated "Project Glasswing," a private evaluation of the model with partners including JPMorgan Chase, major tech firms, and cybersecurity vendors to proactively develop defenses.

Google's TPU Ecosystem Sees Major Endorsement

The CITIC projection of 3.5GW in TPU demand by 2027 underscores the immense computational power required to train and operate frontier AI models. This forecast provides a tangible measure of the commercial benefit to Google from its close partnership with Anthropic, in which it is both an investor and a key supplier of cloud computing resources.

While Amazon is also a major investor in Anthropic, this specific, large-scale demand for Google's proprietary TPUs is a strategic victory. It carves out a dedicated and growing demand for Google's hardware in a market dominated by Nvidia's GPUs, which remain the industry standard. For investors, this signals that Google's multi-billion dollar TPU project is not just a research effort but a core pillar of its future cloud revenue, directly competing for the massive AI training and inference budgets of leading labs.

This article is for informational purposes only and does not constitute investment advice.