Google splits TPU strategy in two

Google is optimizing for bottlenecks now

Google just admitted one AI chip no longer fits the job.

At Cloud Next, it introduced TPU 8t for model training and TPU 8i for inference and agent workloads. TPU 8t scales to 9,600 chips with 121 exaflops per pod, while TPU 8i focuses on memory speed, lower latency, and 80% better performance-per-dollar than the prior generation.

This is the practical move the market has been heading toward. Training giant models and serving millions of live requests stress hardware in different ways, and Google is treating that as an engineering fact instead of a branding story.

If these numbers hold, Google Cloud gets a stronger pitch against Nvidia-heavy stacks and AWS custom silicon. It also suggests future AI infrastructure will be sold as full systems: chips, networking, cooling, schedulers, all tuned together under bright data-center lights.

The open question is simple: will customers switch clouds for better chips, or stay where their workloads already live?

AI helps people work, then scares them

The people getting the biggest AI boost are often the most worried about losing their jobs.

Anthropic surveyed 81,000 Claude users and found workers in AI-exposed roles reported higher concern about displacement, especially early-career respondents. Average self-reported productivity scored 5.1 out of 7, and 48% of users who mentioned gains said AI helped them do new tasks, not just finish old ones faster.

That split matters more than the headline productivity gains. If AI lets one person handle broader work, companies may need fewer entry-level hires, which helps explain why younger workers feel the pressure first.

The strongest users are sending a mixed signal: they value the tool and fear the labor market around it. You can almost hear the office keyboard clicks here, faster than last year, while hiring ladders get thinner.

If AI keeps raising output, who gets the benefit: workers, employers, or the people who never get hired?

OpenAI is borrowing IBM’s playbook

OpenAI is not waiting for enterprises to buy direct.

It partnered with Infosys to embed tools like Codex inside the IT giant’s Topaz platform, targeting software engineering, legacy modernization, and DevOps work. Infosys says AI services already generate ₹25 billion ($267 million) in quarterly revenue, while OpenAI says Codex has more than 4 million weekly active users.

This is a distribution deal disguised as a product story. Big companies rarely adopt new systems through flashy demos; they buy through trusted integrators who can handle procurement, migration, compliance, and late-night calls when something breaks.

Consultancies and outsourcing firms may become the real gatekeepers of enterprise AI spend. If that happens, model companies compete on margins and support terms as much as raw capability, with slide decks open in fluorescent meeting rooms.

The question now is whether AI vendors own the customer relationship, or hand it to the services firms that already do.

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.