SpaceX and CoreWeave are planning to change how they buy AI servers, according to a new supply-chain survey from Jefferies. The investment bank said the two buyers may stop sourcing all of those systems through Dell and instead purchase directly from Taiwan-based ODMs starting in 2027. The companies named in the report were Hon Hai, Quanta, Wiwynn, and Wistron.
Talks are centered on four Taiwan manufacturers
Jefferies said SpaceX and AI cloud computing company CoreWeave are in talks with Hon Hai, Quanta, Wiwynn, and Wistron as part of a possible shift in server procurement. The bank said the new model could begin as early as 2027, with initial deployment focused on VR200 racks based on NVIDIA’s Vera Rubin architecture.
Direct procurement from Taiwan ODMs is not unusual in the server supply chain. Jefferies noted that hyperscale cloud operators such as Google, Meta, and Microsoft have long used this approach instead of relying only on branded OEM vendors. What stood out in this survey was that SpaceX and CoreWeave, both closely watched AI compute buyers, appear to be considering the same path.
Jefferies sees Taiwan ODMs as the main beneficiaries
Jefferies said Hon Hai, Quanta, Wiwynn, and Wistron will still compete with one another over order allocation. Even so, it described Taiwan’s ODM sector as the broader winner if this supply-chain restructuring moves ahead.
The report tied the procurement shift to the production cycle for NVIDIA’s Vera Rubin generation. Jefferies estimated the average selling price of a Vera Rubin rack at $8 million to $9 million, compared with roughly $3 million for the current Blackwell-generation GB200 NVL72 rack. That implies an increase of about three times.
For the Taiwan ODMs, VR200 would mean bidding for orders at a much higher ASP level than in the prior generation. Jefferies also cautioned that while higher ASPs can lift net profit, gross margin may still face pressure because component costs are rising as well.
Dell risk is tied to share loss, not weaker demand
Jefferies said the report carries a clear negative read for Dell over the short to medium term. It noted that Dell’s AI server business has historically run on thin gross margins, but large shipment volumes have still made the segment a meaningful revenue contributor.
If orders from SpaceX and CoreWeave shift away, Dell’s AI server revenue line would shrink. Jefferies added that the higher ASPs expected in the Vera Rubin cycle would make every lost rack order more significant in revenue terms.
The bank stressed that this is a supply-chain share shift rather than a sign of declining overall AI server demand. It said large buyers rarely switch 100% of suppliers at the initial stage, and dual-source procurement is usually the norm. Jefferies said investors will be watching Dell’s next earnings call for details on the customer mix within its backlog, and whether peers deliver similar customer signals in upcoming results.
GB300 shipment forecast points to stronger visibility for Hon Hai
Beyond its comments on the Vera Rubin generation, Jefferies also gave a near-term forecast for Blackwell-based systems. The report said GB300 shipments have recently started to accelerate and projected that Hon Hai will ship 15,000 to 20,000 GB300 racks in the second half of 2026, followed by further growth in 2027.
Using the unit price of a GB300 NVL72 rack, the report said that shipment range would translate into AI server revenue support for Hon Hai in the second half of 2026 on the scale of tens of billions of US dollars. Jefferies said the projection provides a more concrete quantitative marker for near-term business visibility.