Mirendil signs a Google Cloud deal worth more than $100 million
A weeks-old lab just parked more than half its seed on one cloud bill. Per Mirendil’s announcement and CEO Behnam Neyshabur in TechCrunch, the company signed a multi-year Google Cloud partnership worth upward of $100 million for training, inference, and research on Google’s AI Hypercomputer. That is roughly half of a $200 million seed raised at a $1 billion valuation in late June. Compute is not a line item here. It is the product roadmap.
What the multi-year Cloud deal covers
Mirendil frames the pact as capacity for self-accelerating AI: systems meant to automate and continuously improve the research loop itself, not just train another chat model. Workloads span pre-training through post-training, large-scale reinforcement learning, and parallel experiment batteries that would otherwise sit behind human queueing.
Google gets a named frontier customer and a software partner that claims to wring more from mixed accelerators. Mirendil gets managed training clusters co-designed across compute, storage, networking, and control planes. Google Cloud’s post says Mirendil is already live on a TPU v5P cluster, with NVIDIA systems coming online soon, and that provisioning rides managed clusters in Gemini Enterprise Agent Platform. That is Google tying a young lab into its agent ops story as much as selling chips.
TPU and GPU mix on Hypercomputer
The hardware mix is the strategic tell. Mirendil wants both Google Cloud TPUs and full-stack NVIDIA AI infrastructure under one Hypercomputer roof. Co-founder Harsh Mehta told TechCrunch the point is matching workloads to chips instead of forcing every stage onto one vendor’s silicon. Pre-train here, RL there, inference wherever the cost curve bends.
Amin Vahdat, Google’s SVP and chief technologist for AI and infrastructure, leaned on orchestration over single-chip bragging rights: systems of intelligence, not just FLOPS. For Google, a dual-stack customer is a wedge against clouds that still sell one religion. For Mirendil, dual-stack is insurance against TPU or GPU scarcity locking the research loop.
Self-improving training claims to test
Mirendil’s long-term line is blunt: eventually take on the work of an entire frontier lab, then hand that loop to medicine, biology, and materials teams that lack hyperscaler budgets. Neyshabur’s Alzheimer’s example is the ambition slide. Point a self-improving system at a hard scientific goal and let it keep researching and raising its own performance over time.
That is still a claim, not a shipped eval. Recursive self-improvement is the category several labs are naming. What Mirendil can prove next is narrower: do agent-run experiment loops on Hypercomputer cut wall-clock research cycles, or burn $100M learning that human scientists were the cheap part? Demand published research-loop metrics, not mission prose.
Why a young lab lands this scale
Seed at a billion-dollar mark, founders from Anthropic and Google DeepMind, and a cloud giant hungry for logos that are not OpenAI or Anthropic. Neyshabur previously helped invent Sharpness-Aware Minimization and worked research at Anthropic and Google; Mehta helped stand up Anthropic’s internal automated-research push. Investors already priced the team. Google is pricing the compute commitment.
Analysis: burning roughly half of seed on a single multi-year cloud deal is either disciplined lock-in or a burn-rate confession dressed as strategy. If Mirendil’s software layer truly lowers cost for Google’s other customers, this is distribution for both sides. If the self-improving loop stays slideware, Google still books commitment revenue and Mirendil still needs another raise before the Hypercomputer invoice peaks. Watch whether NVIDIA capacity actually lights up beside those TPU v5Ps, and whether Mirendil publishes anything that looks like an automated research win outside its own blog.
