Detailed Analysis
Google's release of Gemini 3 marks a significant milestone in the competitive large language model landscape, with the model claiming top positions across major benchmarks including a reported 91.9% score on GPQA Diamond, a PhD-level science evaluation, and strong performance on mathematical reasoning evaluations. The model also introduces native multimodal capabilities spanning text, images, video, and code from the architecture level rather than as post-hoc additions, representing a meaningful design departure from earlier generational approaches. A companion release, referred to in the newsletter as "Nano Banana Pro," is highlighted for its particularly strong performance on text-within-image generation tasks, reportedly producing accurate rendered text from complex visual inputs such as chalkboards — a longstanding weakness across the AI model ecosystem.
The article's most consequential claim — and the one most directly relevant to Anthropic — is that Google trained Gemini 3 entirely on its proprietary Tensor Processing Units (TPUs) and is now lending that same TPU infrastructure to Anthropic and Midjourney for their own model training. If accurate, this represents a structural shift in how leading AI labs access compute. Anthropic has historically relied heavily on cloud infrastructure, and a deepening dependency on Google's custom silicon would reflect the existing investment relationship between the two companies — Google has made substantial investments in Anthropic — while potentially deepening that strategic entanglement. The research context provided does not independently corroborate the specific TPU-lending claim, and it should be treated as reported but unverified at this stage.
The infrastructure argument the newsletter advances is analytically significant regardless of those specific details. For years, competitive advantage in frontier AI development has been understood primarily through the lens of Nvidia GPU access — training clusters, chip allocation, and supply chain positioning. Google's demonstrated ability to reach or surpass frontier performance using internally manufactured TPUs challenges that framing directly. If TPUs can now match or exceed GPU-based training pipelines at the highest levels of performance, the hardware moat shifts from Nvidia's supply chain to whoever controls large-scale proprietary silicon manufacturing — a category in which Google holds a structural advantage that took more than a decade to build.
The broader competitive picture sketched in the newsletter reflects an accelerating fragmentation at the top of the AI model hierarchy. The piece notes that Grok-4.1's benchmark victory over GPT-5.1 was itself quickly superseded by Gemini 3's release within the same week, suggesting a pace of frontier displacement that is compressing the competitive half-life of top model status to days rather than months. For Anthropic, whose Claude model family competes directly at this tier, the emergence of a newly dominant Gemini 3 while simultaneously potentially training on Google-supplied infrastructure creates a strategically complex position — one in which a primary infrastructure partner is also a direct product competitor.
Google's concurrent launch of Antigravity, an agentic IDE built from the $2.4 billion acquisition of Windsurf's leadership team, further illustrates the company's strategy of pursuing vertical integration across the AI stack — from silicon to model to developer tooling. While the newsletter expresses skepticism about Google's IDE track record, the involvement of Sergey Brin in the launch and the competitive pressure from Cursor's $29 billion valuation suggest the product is being treated internally as a strategic priority rather than an experimental release. Taken together, Google's moves this week — frontier model, proprietary training infrastructure, and developer tooling — represent a coordinated effort to control multiple layers of the AI value chain simultaneously, a posture that will intensify competitive pressure on every major AI lab, Anthropic included.
Read original article →