The news

Mirendil, a young AI lab founded by former Anthropic researchers, has signed a multiyear compute partnership with Google Cloud worth upward of $100 million, TechCrunch reported exclusively on August 6. CEO Behnam Neyshabur confirmed the figure, which represents roughly half of the startup's seed round — the company was valued at $1 billion in late June.

Under the deal, Mirendil gets access to Google TPUs, Nvidia GPUs, and managed training clusters. The lab's stated ambition is unusual even by frontier-lab standards: it wants to build self-improving AI systems that can eventually take on the work of an entire frontier AI lab, automating scientific and AI research across medicine, biology, and materials science.

The multiyear structure also locks in demand for Google's cloud at a moment when every hyperscaler is racing to sign up AI labs. For Mirendil, the appeal is simplicity: one vendor, multiple chip types, and managed clusters that remove the infrastructure work from research.

Why a $100M compute bet on self-improving AI matters

Self-improving AI, also called recursive self-improvement, is one of the most aggressive ideas in the field right now. Mirendil's co-founders come from Anthropic, and peers like Recursive Superintelligence and Ricursive Intelligence are chasing the same goal. Neyshabur's pitch is simple:

"You can have a self-improving AI where you can point a problem at it and it keeps getting better with time." — Behnam Neyshabur, Mirendil CEO

Not everyone agrees self-improving systems will work as advertised, but the money is real: Mirendil's seed round valued the company at $1 billion in late June, roughly a month before this deal was signed. Investors are betting that a lab which can improve its own models will compound faster than one that cannot.

Co-founder Harsh Mehta says training is really about matching the right workloads to the right hardware, and Google's range of chip types lets Mirendil shuffle jobs around to lower cost. Google's SVP of AI and infrastructure, Amin Vahdat, frames the moment as a systems problem, arguing that progress is not just chip-level performance but:

"How we orchestrate entire systems of intelligence and break through the physical constraints of scaling." — Amin Vahdat, Google SVP, AI and infrastructure

What it means for Pakistan

Compute deals like this are a signal for anyone who buys AI capacity. When frontier labs lock in multiyear cloud contracts, the cost of training better models eventually shows up in API prices — sometimes higher, sometimes lower as efficiency improves. For Pakistani developers and agencies building on top of models, the practical move is to treat API credits as a managed cost rather than an afterthought. Enterprise AI API credits and our guide to buying AI API credits in Pakistan cover exactly this.

If self-improving AI delivers on its promise, the models you rent today could get noticeably better without you changing a line of code. The flip side is that demand for compute will keep climbing, which is why watching infrastructure trends early pays off. To check current API credit pricing and bundles, message AI Tools Pak on WhatsApp (+92 371 454 9245 or wa.me/923714549245).

For freelancers, the signal is subtler: labs that lock in cheap compute tend to pass efficiency gains down the stack, which is one reason API prices keep falling even as models get smarter. A $100 million deal today often becomes a cheaper, better model on the market tomorrow.

Practical takeaway

A quick checklist for developers and agencies watching the infrastructure race:

  1. Track which labs land big compute deals — they tend to ship faster, cheaper models later.
  2. Treat API spend as a budget line: pick models by task, not by hype.
  3. Keep a small credit buffer so you can test new models the week they drop.
  4. Buy credits in PKR through local sellers to avoid card and FX friction.

DEEPER DIVE

Frequently asked questions

What exactly did Mirendil sign with Google Cloud?
A multiyear partnership worth upward of $100 million, giving Mirendil access to Google TPUs, Nvidia GPUs, and managed training clusters for its self-improving AI research.

What is self-improving AI?
Self-improving AI, or recursive self-improvement, refers to systems designed to get better at a task over time without human rewrites. Mirendil wants such systems to eventually handle the work of an entire frontier AI lab, including scientific research.

Why should Pakistani developers care about compute deals?
Big compute contracts shape how fast frontier models improve and how their API prices move. Developers who track these deals can time their credit purchases and model switches better.

Related pages