Morgan Stanley Raises AI Scale-Up Network Market to $73 Billion
0xBroomberg
Morgan Stanley lifts its 2030 AI scale-up networking market estimate to roughly $73 billion — about 4× its prior $17 billion 2029 forecast; copper interconnects remain the near-term play while co-packaged optics won't face its real proving year until 2029.
Where does a $73 billion market come from?
Morgan Stanley now sizes the 2030 AI scale-up network — the layer that links large numbers of GPUs for joint computation — at roughly $73 billion, with a ~46% CAGR from 2026 to 2030.
Three drivers stack: rising AI capex + single domains expanding from 72 GPUs to 576 or even 1,152 + higher switching bandwidth and interconnect ports per accelerator.
This means → no single variable is growing alone. "GPU count × network content per GPU" is compounding simultaneously, producing exponential rather than linear expansion.
A scope note: this figure excludes NICs, pluggable optics, and cables. It captures the switching-and-interconnect base only — the full AI networking value chain is larger.
How long can copper hold?
Morgan Stanley sees copper retaining its lowest-latency, lowest-power, lowest-cost triple advantage at short distances for at least about two more years.
In plain terms = copper is the side street next to your house — fast and cheap for short runs. No need to build a highway (optical interconnect) for a few meters.
On the tech side, 448G SerDes, PAM6 modulation, and 200G retimers can stretch copper's usable reach to roughly 6 meters, covering most intra-rack links.
Copper comes in three tiers: passive direct-attach (shortest) → active copper with analog redrive (~3 m) → active cables with retiming and equalization (typically 7–9 m, compressed to ~5 m at 200G per lane). Power draw is still roughly 50% lower than pluggable optics.
Who wins on the copper side?
Morgan Stanley names Astera Labs, Credo, Semtech, and Marvell as more direct 2026–2027 beneficiaries; Nvidia and Broadcom also feature.
This means → near-term capital flows toward the chip and solution vendors that make copper reach farther, not toward optical-module makers.
Nvidia's next-gen Rubin Ultra will likely adopt a copper-inside-the-rack, optics-between-racks hybrid. Copper's share will shrink as domains grow, but it will not vanish overnight.
Why does co-packaged optics still have to wait?
Co-packaged optics — integrating optical engines directly onto the switch die, replacing electrical signals with light — won't reach meaningful scale-up penetration before 2029. Morgan Stanley puts 2028 penetration at low single digits.
The blockers: cost, yield, serviceability, and supply-chain lock-in. Only after Nvidia's Feynman architecture pushes a single domain to 1,152 GPUs will cross-rack bandwidth density and I/O power become painful enough to override those obstacles.
This means → optics isn't failing on the technology; it's failing on the pain threshold — GPU domains aren't yet large enough to force the switch.
Near-packaged optics — engines close to the switch die but not fully co-packaged — serves as a bridge, preserving pluggability and vendor openness for customers wary of single-source lock-in. Full 3D co-packaging at scale is more likely post-2030.
When does the window open for optical companies?
Demand for fiber arrays, lasers, optical engines, optical-circuit switches, and test equipment from Corning, Lumentum, Coherent, and Keysight starts ramping only after 2028.
Morgan Stanley flags four downside risks: copper reach keeps extending / scale-up domains stay single-rack / inference workloads don't need ultra-large shared domains / co-packaged optics failure rates remain too high. Any one of these pushes the revenue inflection later.
In plain terms = the optical story is "the day after tomorrow looks great," but "tomorrow's" money is still on the copper side.
From 72 GPUs to 1,152, the optical-engine count per GPU may rise from 2 to 34–70 — once the inflection arrives, the leverage is enormous.
Five protocols coexist — who calls the shots?
Five protocols run in parallel today: Nvidia NVLink / NVSwitch (closed system), NVLink Fusion (third-party access but Nvidia retains control), Broadcom scale-up Ethernet + custom accelerators (open ecosystem), PCIe (2026–2027 transitional bridge), and UALink (the biggest variable after 2027).
This means → protocol choice directly determines where the profit pool lands — Nvidia's closed ecosystem captures the fattest margin; open ecosystems spread value across more players.
UALink's commercial switch silicon and large-scale deployment aren't expected until after 2027; on-time member delivery and software-stack stability remain uncertain.
Morgan Stanley's bottom line: hyperscalers won't bet on a single path, so protocol conversion, retiming, switch silicon, and test-and-measurement all capture incremental value. Key milestones to track: Rubin Ultra's actual rack architecture / 200G-per-lane copper distance / first scale-up optical-engine orders / whether Feynman holds its 1,152-GPU design.
Content is for reference only, not financial advice.