The math is stark, almost absurd: a quant researcher can dream up trading signals in minutes, but proving they work—running them against years of tick data, testing hundreds of parameter variations, stress-testing across asset classes—can eat days. Machine learning may generate hypotheses faster than ever, but the computational grunt work of validation hasn't kept pace.
Ibrahim Rabbani thinks that gap is costing firms money. Real money.
His answer is Zibra Labs, a fledgling startup that emerged from Y Combinator's Spring 2026 batch with infrastructure designed to collapse backtesting cycles from days into minutes. The pitch, unveiled publicly on May 19, 2026, centers on orchestrating millions of parallel tasks across tens of thousands of CPU and GPU nodes—a scale that would let quant shops run the exhaustive parameter sweeps they currently have to trim for time.
It's an audacious claim from a two-person team with no disclosed customers. But Rabbani isn't coming out of nowhere.
The Distributed Systems Pedigree
Before Zibra, Rabbani was the tech lead on Ray, the open-source distributed computing framework now maintained by the PyTorch Foundation. Earlier, he worked on database infrastructure at LinkedIn—Venice, Espresso, systems built to handle data at LinkedIn-scale. His co-founder, Zac Policzer, brings a similar résumé in large-scale databases and open source tooling.
Now they're aiming that experience at a corner of finance where compute has become the chronic constraint. As one former hedge fund researcher put it (speaking on condition of anonymity because they still work in the industry): "You end up making compromises you hate. Smaller grids, shorter windows, coarser data. You know there's signal you're missing, but the backtest won't finish before Monday."
Zibra's proposition is straightforward, if technically demanding: let quant firms run millions of backtests in parallel without wrestling with orchestration overhead, spot-instance interruptions, or the chaos of multi-cloud deployments. The platform spans traditional hyperscalers and specialized GPU clouds like CoreWeave, abstracting away the messiness while promising dispatch latency under 50 milliseconds.
The Numbers—and the Caveats
The technical specs read like catnip for anyone who's ever babysat an overnight parameter sweep. Clusters from 100 to 50,000 nodes. Native support for spot instances across availability zones and cloud providers. Orchestration of up to 6.4 million parallel tasks.
That last number comes with a footnote, though. An earlier version of Zibra's launch materials cited 5 million tasks. The discrepancy is minor, perhaps just marketing copy in flux, but it underscores how new this operation is. Either way, the target is clear: enable the kind of exhaustive Monte Carlo simulation and parameter exploration that currently gets sacrificed on the altar of compute budgets.
For firms anxious about leaking proprietary strategies—and in quantitative trading, where edges are measured in basis points, that anxiety runs deep—Zibra deploys inside the customer's own cloud account or on-premises hardware. No data crosses borders. It's a security posture that acknowledges the paranoia baked into the industry.
Rethinking the Stack
Zibra positions itself as a modern successor to tools like SLURM and MPI, the workload managers and distributed-memory standards that have ruled scientific computing since the era of purpose-built supercomputers. Those systems work, reliably even, but they predate the cloud's chaotic realities: instances that vanish without warning, pricing that shifts by the minute, workloads that need to span regions and providers on the fly.
Zibra's architecture assumes those headaches as defaults rather than exceptions.
The competitive landscape is cluttered. Managed platforms like Anyscale and Coiled already tackle distributed compute for Ray and Dask users. Cloud providers have their own HPC offerings. Quant-specific services like QuantConnect and QuantRocket offer backtesting infrastructure. Zibra's differentiation, at least on paper, hinges on three things: raw scale, sub-50ms latency, and the ability to keep everything behind the customer's firewall.
Whether that's enough remains an open question.
Beyond Backtesting
Quantitative finance is the wedge market, but Zibra's launch materials gesture toward broader ambitions. Reinforcement learning post-training. Large-scale robotics simulation. Molecular dynamics and drug discovery. The connective tissue is embarrassingly parallel workloads—problems that scale horizontally but choke on traditional HPC infrastructure.
Whether those adjacent markets materialize is anyone's guess. Infrastructure startups often sketch expansive visions; execution is where the story gets written.
For now, Zibra is two people in San Francisco with no publicly named customers, no disclosed pricing (the website offers a "Schedule a Call" button), and no announced funding beyond the Y Combinator batch. That's standard for early-stage enterprise infrastructure, where the first customers are design partners and revenue models get negotiated in confidence.
The Real Bottleneck
Still, the problem Zibra is chasing is genuine. As AI accelerates the generation of trading signals, the computational burden of validating them grows exponentially. Researchers generate hypotheses faster than they can disprove them, and the result is either inflated budgets or compromised rigor.
If Rabbani and Policzer can deliver on their latency and scale promises—admittedly a big if—they'll be offering something quant researchers didn't realize was within reach. And in an industry where milliseconds matter and every basis point counts, that's the kind of product people don't give up easily once they've tasted it.
