Infrastructure Appetite: Google Prepares the Next Generation of LLMs
Google has announced the release of Gemini 3.6 Flash and confirmed that it is already actively training Gemini 4. The release of faster, cheaper Flash versions traditionally signals a software stack optimization for current hardware. However, the primary infrastructure challenge lies in the preparation for the fourth generation of models.
Training LLMs on the level of Gemini 4 requires exponential growth in compute resources. In my view, the announcement of Flash versions is an attempt to maximize the ROI from already deployed GPU clusters until new batches of accelerators arrive to power the resource-intensive training of future architectures.
Compute Geopolitics: Chinese Models Split the Silicon Market
The success of Chinese AI developments has triggered a serious rift among President Donald Trump's AI advisors. Experts are publicly accusing leading American AI companies of inefficiency, which directly impacts export control policies and investments in domestic chip production.
This conflict goes far beyond software. The internal struggle within the White House determines how subsidies for data center expansion will be allocated and what restrictions will be placed on accelerator shipments. If American AI giants cannot prove the efficiency of their infrastructure, the government may redirect funding from cloud providers toward national semiconductor projects.
Data as a Bottleneck: $1.5 Billion to Expand Compute Bases
A judge has approved Anthropic's $1.5 billion settlement to resolve copyright disputes. Out of thousands of authors, only 350 managed to opt out of the class-action lawsuit.
For the hardware market, this event has an indirect but critical significance. The evolution of AI accelerators is hitting a wall not just in lithography, but in access to legitimate training data. Massive financial payouts for copyright royalties mean that AI companies must now factor the cost of data corpora into their CapEx alongside GPU purchases. In my view, this precedent will force compute providers to rethink their business models: the data centers of the future will have to provide access not just to hardware, but to legalized datasets.
Bottom Line
AI infrastructure is ceasing to be a purely engineering task. Compute power now sits at the intersection of a technological race, geopolitical pressure, and legal constraints. In the near term, expect a revision of accelerator procurement strategies: demand will be shaped not simply by peak FLOPS, but by the overall efficiency of clusters in the face of rising data legalization costs and export quotas.