Karman Industries launches HPU to cut data center cooling energy by 25–100% using supercritical CO₂

Summary

Karman Industries launched its Heat Processing Unit (HPU) on January 15, 2026, targeting the rapidly expanding market for gigawatt-scale AI data center cooling. The Long Beach-based startup's core claim is a minimum 25% reduction in cooling energy consumption versus current best-in-class systems, rising to 60–100% improvement in moderate climates such as Virginia. Cooling currently consumes 20–30% of total data center energy, making it a direct drag on compute capacity rather than revenue-generating infrastructure.

The technology and unit economics

The HPU uses supercritical CO₂ as its working fluid, running through an in-house-designed high-speed turbo compressor operating at 30,000 RPM. Each unit is a 10-megawatt modular system — five to ten times larger than the 1–2 MW incumbent solutions from traditional HVAC vendors — and ships in a standard shipping container footprint. Pricing is approximately $750,000 per megawatt, meaning a single 10 MW unit runs roughly $7.5 million. Karman claims comparable upfront capex to legacy systems, with a net cost advantage once reduced mechanical yard requirements (80% smaller footprint), faster deployment, and lower piping complexity are factored in.

The fundamental value proposition is not primarily opex savings but compute headroom. Freeing up electricity previously consumed by cooling directly expands billable GPU capacity, which is where hyperscaler and neocloud economics are actually made.

Manufacturing and scale targets

Karman is standing up Gigaworks 1, its first factory, this fall, targeting an annual run rate of 100 units — equivalent to 1 gigawatt of cooling capacity per year. The company is targeting approximately 1 gigawatt of actual production in 2027 and is already allocating that capacity to customers. The team of 23 is scaling toward a generalist-then-specialist technician model, drawing manufacturing philosophy explicitly from SpaceX's incremental vertical integration playbook. The COO previously led production at Millennium Space Systems.

Funding and backers

Karman closed a Series A in September 2025, announced publicly alongside the HPU launch. Riot Ventures led the round, joined by Space VC, Wonder Ventures, and Sunflower Capital. Pat Gelsinger, former CEO of Intel, joined as an angel investor. The company was co-founded by CEO David Turus, an entrepreneur-in-residence alumni of Riot Ventures, and CTO Dr. C.J. Kalra, whose career began at GE Research working on supercritical CO₂ systems and who most recently served as head of technology at Antora Energy. Roughly a third of the engineering team comes from SpaceX and Rocket Lab.

Why the market is receptive now

Rack power density has jumped from 5–20 kW historically to 120 kW today, with Nvidia's Vera Rubin architecture expected to push that to 600 kW per rack and roadmaps trending toward 1 MW per rack. A gigawatt-scale data center built on legacy 1–2 MW cooling units requires 500 to 1,000 separate systems and a mechanical yard that, by aerial view, exceeds the footprint of the data center itself. Existing water-cooled systems also face regulatory and community pushback: a gigawatt facility can consume over 500,000 gallons of water per hour. Karman's closed-loop CO₂ system consumes no water, produces no PFAS compounds (unlike Honeywell's HFO refrigerants), and is nonflammable, a combination that hyperscalers have flagged as directly aligned with their forward build specifications.

The company is also positioning heat reuse as a longer-term revenue layer — waste-heat-to-electricity conversion is technically viable in colder climates during winter months using temperature differentials, which would feed additional power back into the facility's compute budget. Residential district heating applications, already more advanced in Europe due to data center proximity to urban areas, represent a further optionality.

Deployments are scheduled to begin later in 2026. The competitive moat rests on a proprietary compressor design, aerospace-grade engineering talent, and a form factor that directly addresses the deployment bottleneck constraining hyperscaler and neocloud build timelines.