- Nvidia's new Blackwell GPU may - or may not - be delayed
- The timing of the chip's availability is significant for data centers scrambling to prepare for its liquid cooling requirements
- EdgeCore Digital Infrastructure thinks liquid cooling will dominate among its customers by 2026
Rumors began swirling earlier this month that availability of Nvidia’s latest artificial intelligence chip Blackwell will be delayed by three months or more. Much noise was made about what this might mean for the artificial intelligence ambitions of major hyperscalers. But what – if anything – would a delay mean for data center operators?
Blackwell's launch date matters so much for data centers because the chip is expected to usher in broad adoption of direct-to-chip liquid cooling.
Tom Traugott, SVP of strategy at EdgeCore Digital Infrastructure, told Fierce that deployments of Nvidia’s previous-generation H100 (aka Hopper) chips “aren’t quite pushing direct-to-chip liquid cooling everywhere.” Instead, he said, customers are running 30-40 kilowatts per rack to stay within the limits of air cooling systems.
But since the announcement of the Blackwell chip, Traugott said customers (including hyperscale players) have “almost universally” revised their specifications to require data center vendors to be ready for liquid cooling systems in anticipation of Blackwell’s needs. That takes some preparation.
Traugott said data centers with existing water lines need to splice in valves to accommodate the liquid-cooled racks of the future. They also need to order coolant distribution units (CDUs) for their facilities well in advance.
“The CDUs are kind of living in more of a 6-12 month lead time universe,” he explained. “So, we need to know in advance, we can’t just snap our fingers and make the switch.”
All that is to say, release timelines matter.
Asked about the reported delay, a Nvidia spokesperson told Fierce, "Broad Blackwell sampling has started, and production is on track to ramp in 2H [of 2024]. Beyond that, we don't comment on rumors.”
The chip-maker will report Q2 earnings this week on Wednesday, Aug. 29.
Air-cooled with 'rumblings'
From Traugott’s vantage point, 2024 and 2025 will still be mostly air-cooled with “rumblings” of direct-to-chip cooling deployments. But from 2026 onward he expects the market to swing primarily toward direct-to-chip.
To illustrate just how dramatic the forthcoming shift will be, Traugott noted that except for one hyperscaler, which has led with liquid cooling deployments, 90% of data center fleets today are air-cooled.
Looking ahead though, Traugott said, "What we’re understanding is that we’re still going to need some air management but it’s going to be trending towards 80% of the room is direct-to-chip and still 20% standard air.”
Though he couldn’t name names, Traugott said EdgeCore’s customer base is primarily comprised of “the big seven” hyperscale cloud players and “the next 15.”
Since January, the company has secured over $2.3 billion in debt financing to continue construction geared toward meeting the needs of these clients.
Of course, not every customer is a hyperscale client needing Blackwell chips. So, while Traugott’s predictions of a sudden ramp in liquid cooling demand are consistent with analyst forecasts, the mix of cooling types is expected to be slightly different for the broader data center market.
Dell’Oro Group previously predicted liquid cooling will account for roughly a third of the overall data center thermal management market by 2028.