JPMorgan: NVIDIA VRU Rack Cooling Value Could Rise Nearly 49%
nashnova research
JPMorgan estimates thermal component value per Nvidia VRU NVL72 rack will climb from ~$66,300 to ~$98,500 — a nearly 49% increase that runs counter to market fears of a thermal-spec downgrade, suggesting Street earnings estimates for thermal suppliers are systematically too low.
What is driving this thermal upgrade cycle?
JPMorgan identifies three drivers: cold plates shifting from single-sided to double-sided with diamond-copper material, quick disconnects upgrading from MQD/UQD to higher-spec ZQD, and chip lids moving to a removable three-piece design.
In plain terms = each chip used to need one cooling plate on the front only; now both sides need one, built with a faster-conducting diamond-copper alloy. Connector specs double; the lid splits from one piece into three.
Combined, the three changes support a ~50% increase in thermal component value per rack versus the prior VR200 generation.
How much do cold plates and quick disconnects each contribute?
Cold plates: VRU's vertical power-delivery architecture moves power components and memory to the back of the chip board, requiring additional rear cold plates. This means → cold plates per compute tray rise from 2 to 6, a threefold jump.
Quick disconnects: some connectors upgrade to ZQD, which carries an average selling price roughly double the prior product. Total quick-disconnect value per VRU NVL72 rack rises ~100% versus VR200 / GB300 racks.
Per-tray quick-disconnect value climbs from $400–500 on VR to ~$750 on VRU.
Is the three-piece lid on schedule?
Nvidia's Rubin Ultra GPU uses a removable lid + stiffener frame + multiple screws design. JPMorgan confirms the program has not been delayed.
This means → versus the current one-piece lid, thermal lid value is expected to reach 4–5× the original.
Supply-chain checks suggest Nvidia may begin pilot production on a small number of Rubin GPUs as early as Q4 2026, with Rubin Ultra volume production potentially starting in Q1 2027 — slightly ahead of market expectations, opening a window for near-term earnings upside at lid suppliers.
Are cloud vendors' custom chips seeing the same thermal upgrade?
AWS Trainium 3 liquid-cooling system (Teton 3 Max) carries thermal component value of $3,000–3,500 per compute tray, roughly 30% above Nvidia's VR tray; per-rack total reaches ~$86,000, above VR NVL72's sub-$70,000.
Google TPU v8 thermal value runs $2,000–2,500 per tray, with a more complex cold-plate design — 12 cold plates per tray, plus gold plating required by the metallic thermal-interface material.
In plain terms = Nvidia is not alone in ramping thermal specs. AWS and Google custom silicon is pushing in the same direction, expanding the entire industry's thermal value pool.
How big is the TPU v9 lid opportunity?
JPMorgan notes that if Google TPU v9 ultimately adopts a removable-lid-plus-stiffener design, related component value could reach ~10× the original.
However, this design has not been confirmed — it is a potential uplift, not a firm order.
This reflects a broader pattern: the three-piece lid concept is spreading from Nvidia to cloud-vendor custom chips. If adopted, it would raise the ceiling for thermal suppliers yet again.
Which stocks does JPMorgan favor — and what is the risk?
In the thermal sector, JPMorgan prefers AVC (奇鋐) and Fositek (富世達), both rated Overweight.
AVC's appeal lies in GPU and ASIC system cold-plate upgrade trends and its leading position across programs; Fositek's appeal centers on near-term gross-margin upside and long-term quick-disconnect upgrade value.
JPMorgan flags a key caveat: early component production does not equal simultaneous system-level demand. This means → actual design adoption and volume-production timing are the critical checkpoints for validating the thesis above.
市场有风险，内容仅供研究参考，不构成投资建议。