報告_廣發香港_VR200展望_20261026

PDF 原檔：報告_廣發香港_VR200展望_20261026_original.pdf

原始內容

Robust VR200 Outlook

What's New?

We provide our forecasts for GB/VR rack shipment TAM, supply chain analysis, and key stock recommendations.

Comments

We forecast 73k/96k NVL72 racks TAM in 2026E/2027E: Coming off early yield improvement across cooling and power components (inner busbar/manifold & QD), we believe Vera Rubin racks production is ramping with 85%+ first-pass rate and set to further accelerate in 1Q27. For 2026E, we forecast total NVL72 rack shipments of 73k (67k GB200/GB300 + 6k VR200). In 2027E, we expect the total TAM to grow 31.5% YoY to 96k racks, led primarily by VR200 and initial VR300 (Ultra) volumes, totaling 86k units. Of note, we expect SpaceX to be the largest VR200 buyer in 2027E with 20k racks (on 30k+ demand, under a 10GW net addition).

CoreWeave's near-term funding risk manageable: Concerns over CoreWeave's CDS spread and >15% yields on its unsecured notes have weighed on the stock (-30% since Oct 13). However, Blackstone COO Jon Gray was constructive on the company, saying CoreWeave will 'find a way through' its financing challenge given the strong underlying business, citing compute pricing up ~50% YoY and legacy H100s holding value better than expected, factors that suggest capacity pricing may rise much faster. Structurally, no unsecured notes (~$3.8bn, ~10% of total debt) mature before June 2030, and secured funding has continued to clear ($2.6bn DDTL 5.5 in August at SOFR+5.50%, oversubscribed). We thus view near-term funding risk as manageable, though rising capital costs are worth monitoring.

Major 6 neoclouds of 19k/34k NVL72 racks in 2026E/2027E: To fund its $500bn buildout (10GW at $50bn/GW), SpaceX is expected to tap further debt financing, alongside an estimated ~$200bn in operating cash inflow next year. Near-term, SpaceX is well-covered to support the immediate 4Q26 CapEx step-up ($35-$40bn) and early 2027 spend, with: 1) Apollo-led $40bn GPU debt ($10bn loan + $30bn IG bonds, per Financial Times), 2) $89.9bn cash, and 3) large upfront prepayments from CSPs. Meanwhile, tier-2 neoclouds such as Nebius/Lambda/IREN/Nscale are also stepping up their buildouts. Together with SpaceX/CoreWeave, we expect these 6 operators' net capacity addition of 3.5GW/12GW in 2026E/2027E, translating into 19k/34k rack shipments, respectively (27%/35% of industry total).

ODM/OEM Vendor Dynamics: Per our earlier report in June, SpaceX is partially turning to ODM-direct model, a trend we now see gaining stronger traction. Hon Hai is winning more allocation on the Rubin platform at SpaceX (up to a 50% share, but we model 40-45% given capacity constraints), with Pegatron also in the supply chain, albeit with limited share (<5%). By contrast, Wistron would be hurt by lower SpaceX share (via Dell), even with new customers like HPE, Lenovo and Oracle. On the CSP front, we expect Hon Hai to gain share from Quanta at Google (100%) and Meta (>50%).

Recommendations

• ODM: Hon Hai (2317 TT Buy) with higher market share on expanding SpaceX allocation and CSP gains.
• OEM: Lenovo (00992 HK Buy) with multiple AI server wins across IREN, Nscale, and other tier-2/3 Neoclouds.
• Maintain neutral view for Dell (Hold) on more prominent share loss at SpaceX.

Risks

• 1) AI demand slowdown, 2) supply constraint, 3) competition.

Figure 1: Major 6 Neoclouds NVL72 racks shipment (k units) forecast

Sources: Company data, GF Securities (Hong Kong) Brokerage

Figure 2: Major 6 Neoclouds year-end capacity (GW)

Sources: Company data, GF Securities (Hong Kong) Brokerage

Figure 3: NVL72 racks TAM (k units) forecast

Sources: Company data, GF Securities (Hong Kong) Brokerage

Risks

1) AI demand slowdown, 2) supply constraint, 3) competition.

Rating definitions Benchmark: Hang Seng Index (Hong Kong)