Jane Street reveals AI lab’s evolution from six Dell boxes to liquid-cooled GPU data center

The quant trading giant built a 4,032-GPU facility in Texas and invented an internal currency called 'hive bucks' to auction off compute time.

Share

Add us on Google by Editorial Team May. 16, 2026

Jane Street, one of the most secretive and profitable quantitative trading firms on the planet, has pulled back the curtain on its AI infrastructure journey. What started with six Dell servers has grown into a purpose-built data center in Texas housing 4,032 liquid-cooled GPUs.

The firm also revealed something arguably more interesting than the hardware itself: an internal auction system called “hive bucks” that forces teams to bid against each other for GPU compute time.

Advertisement " document.getElementById("alkimi-leaderboard").innerHTML = iFrame var iframeDoc = document.getElementById(idIFrame).contentWindow.document pbjs.renderAd(iframeDoc, highestCpmBids[0].adId); } } setTimeout(function () { renderAds(); }, FAILSAFE_TIMEOUT);

From six servers to 4,032 GPUs

Jane Street’s AI ambitions didn’t start with a grand vision and a blank check. They started with six Dell boxes. The firm has since transformed that modest beginning into a dedicated Texas facility packed with thousands of GPUs designed specifically for AI research and trading model development.

Liquid cooling systems can be up to 15% more energy-efficient than their air-cooled counterparts, and water transfers heat far more effectively than air. Modern rack-scale designs can support up to 256 GPUs per rack with liquid cooling, a density that would be impossible with fans alone.

The internal economy of compute

The company created “hive bucks,” a virtual currency distributed to internal teams as a budget for GPU time. Teams don’t just request compute through a ticket system or wait in a queue. They bid for it in a live auction, competing against other teams who also need the hardware.

The system forces teams to make genuine tradeoffs. If a research group burns through its hive bucks on a speculative training run, it has fewer resources available for the next project. This creates natural prioritization without requiring top-down management decisions about which AI initiative matters most.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.