LABX

Morgan Stanley: Scale-Up Market

Massive Expansion of the Scale-Up Market

> 4x Market Surge: The AI scale-up network market is projected to reach a $70 billion+ opportunity by 2030, which is over four times the size estimated just a year ago ($17 billion by 2029).

> The Scale-Up Moat: Scale-up networks are essential because they allow thousands of GPUs/accelerators to operate seamlessly as a single supercomputer system, eliminating communication bottlenecks during modern frontier AI model training.

> Bigger Clusters: Cluster sizes are expanding rapidly. For example, NVIDIA’s roadmap has progressed from 72 GPUs in a single rack (Blackwell) to 144 GPUs (Vera Rubin), 576 GPUs (Rubin Ultra), and potentially over 1,000 GPUs with Feynman.

The "Copper Wall" vs. Optical Transition

> Copper is Staying Longer: Copper remains the preferred short-reach technology due to lower latency, lower power consumption, and lower cost. Innovations like PAM4/PAM6, advanced SerDes, and retimers are extending copper's life further than the industry previously expected.

> Optics is Inevitable but Back-Weighted: As cluster domains push past a single rack, managing electrical signal degradation at high speeds becomes too inefficient. Morgan Stanley expects Co-Packaged Optics (CPO) to see meaningful adoption starting with the NVIDIA Feynman generation around 2028–2029. Recent market sell-offs surrounding delayed CPO adoption are viewed as overdone.

Fabric & Ecosystem Protocol Debates

> NVIDIA Dominance: NVLink remains the dominant proprietary fabric. NVIDIA is also deploying NVLink Fusion, which allows third-party CPUs and custom ASICs to tie directly into the NVLink ecosystem.

> Open Alternatives Growing: Broadcom is spearheading Scale-Up Ethernet (SUE), while an AMD-led consortium is backing UALink as an open alternative. Morgan Stanley sees PCIe (e.g., via Astera Labs) acting as an important temporary bridge before hyperscalers migrate to native AI fabrics.

> AWS & Google Outliers: Google relies on an Optical Circuit Switch (OCS) Torus topology. AWS has shifted to a flatter, highly efficient mesh network called Resilient Network Graphs (RNG) using passive optics to reduce router counts and power consumption.

$Amazon(AMZN.US) $Alphabet(GOOGL.US) $AMD(AMD.US) $NVIDIA(NVDA.US) $Credo Tech(CRDO.US) $Semtech(SMTC.US) $Broadcom(AVGO.US) $Marvell Tech(MRVL.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US) $Corning(GLW.US)