OCP Global Summit 2026 | Alibaba Group | Zhiping Yao | Rethinking AI Scale-up with Optics: From Pluggable to (n)PO

Oct 22

7 min read

1. Why Alibaba Is Rethinking Scale-up

Yao opened with a distance map: scale-in inside the server is under 1 m, scale-up inside a super node is under 5 m, scale-out from POD to cluster runs from 50 m to 2 km, and campus and beyond is scale-across. The slide labels the bandwidth multipliers as 1,000x, 100x, 10x and 1x: the closer to the chip, the more bandwidth and the shorter the reach.

NVIDIA's evolution is the textbook case: NVL8 in the 2023 DGX relied on PCB, NVL72 in 2025 used a massive copper cable cartridge, then NVL144 in 2026, and NVL576 in 2027 moves to a PCB backplane. All copper. The slide's verdict is blunt: with 100G-class SerDes, practical copper reach is only a few meters, so scale-up stays tied to one rack or a few tightly coupled racks.

Once scale-up has to grow from hundreds to thousands of accelerators, it hits copper's reach bottleneck. That matches the consensus we saw in OCP APAC Summit 2026 | CPO / NPO / XPO Panel: the question is no longer whether to use optics, but which form of optics.

2. Breaking Down the Scale-up Domain: Port Count Sets the XPU Count

Alibaba calls its architecture the Ultra Performance Network: switch nodes on top, each switch ASIC surrounded by NPOs; XPU nodes below, four XPUs per node, each with its own NPO; a fiber shuffle box in between spreads the links. The whole network is a single switch tier, so the size of a scale-up domain is essentially the port count of one switch chip.

▲ Alibaba's Ultra Performance Network: switch nodes and XPU nodes are decoupled via NPO and a fiber shuffle; the table below lists bandwidth, SerDes and supported XPU count for each switch-chip generation (Image source: Alibaba Group | 2026 OCP Global Summit presentation (7:26))

The switch table on the slide spells it out:

Switch chip	Timing	Bandwidth	SerDes	XPUs supported
CN-25.6T	2024	25.6T	112G	256
TH6P	2025	102.4T	112G	512
TH6C	2026	102.4T	224G	512
TU2	2027	115T	224G	576
TH7	2027	204.8T	224G	1,024
NG-SW1	2028	230T	224G	1,152
NG-SW2	2028	409.6T	448G	1,024

Yao added that 512 GPUs can share one domain today, and the next 1,024-port switch takes that past a thousand. Going from 72 to 512 is about 7x, and to 1,024 about 14x (STT calculation).

▲ In a single-tier switch architecture, scale-up domain size is set by switch-chip port count: from 72 in NVL72 to 1,024 in the TH7 generation (Image source: Simple Tech Trend; data from Alibaba Group's 2026 OCP Global Summit presentation)

The second benefit of optics is often underrated: compute nodes and switch nodes are decoupled. A copper cartridge welds both into one rack design; with optics, compute boxes and switch boxes can be designed and delivered separately, improving availability and resilience. For a hyperscaler that defines its own racks, that matters more than saving a few watts.

3. (n)PO Is Not One Option but a Spectrum

Yao deliberately titled the talk (n)PO rather than NPO. His taxonomy has four optical-engine locations:

Option	Engine location	Electrical channel	Serviceability	Key challenge
XPO	Pluggable module	Longest, highest loss	Field-replaceable	Linear drive difficult, complex thermals
NPO	Main board	Medium, lower loss	Replaceable in production	Hard to support 448G
IPO	Interposer board	Relatively short, lower loss	Replaceable by manufacturer	Elastomer reliability, complex mechanics
CPO	Co-packaged with ASIC	Shortest, lowest loss	Not replaceable	Deep OE-ASIC coupling, limited ecosystem, vendor lock-in

The slide rates engineering complexity in stars: XPO three, NPO two, IPO four, CPO five. In other words, NPO is the easiest to build, CPO the hardest, and IPO buys near-CPO performance for near-CPO difficulty.

▲ Alibaba's high-density optical interconnect roadmap: LPO/XPO in 2025, NPO/UPO in 2026, IPO in 2027 and CPO in 2028, collectively called (n)PO (Image source: Alibaba Group | 2026 OCP Global Summit presentation (10:26))

Yao was blunt about pluggables: XPO and the high-density pluggable NVIDIA and Cisco just announced will extend the pluggable success story, but linear drive at 200G will be difficult and at 400G nearly impossible. That points the same way as Huawei at ECOC, see ECOC 2026 | Huawei Scores CPO vs NPO, NPO Leads 4:2: both Chinese players rank replaceability and multi-sourcing ahead of peak power efficiency.

4. Bandwidth Roadmap: 3.2T in Production, 6.4T in Validation, 12.8T in Early R&D

3.2T NPO (32x100G) in production: OIF form factor, silicon photonics PIC, 2D package, linear, external laser source (ELS), co-developed with the switch.
6.4T NPO (32x200G) in validation: UPO standard Type-1/2/3 form factors, 2.5D package, linear, ELS; the 6.4T SiPh PIC is being validated as UPO Type-2. Yao described UPO (Ultra Performance Optics) as an MSA among Chinese hyperscalers, covering 6.4T and 12.8T.
12.8T NPO/IPO (32x400G) in early R&D: grating couplers, Cu BEOL and TSV, hybrid bonding, with a SiPh MZM plus TFLN (thin-film lithium niobate) optical engine and the IPO engine sitting directly on the ICA board.

The table reveals two things. First, 400G per lane is treated as the dividing line between NPO and IPO: Alibaba already assumes NPO will not reach 448G, which is why IPO enters at 12.8T. Second, there is no single material bet: SiPh MZM and TFLN sit side by side, meaning the 400G modulator is still open.

5. The Real Lead: IPO and a 102.4T Switch That Is Already Running

IPO brings the switch chip's high-speed signals out of the top of the package, through an elastomer socket to an ICA board above, where the optical engines sit. The engines are closer to the chip than NPO, yet not soldered into the package like CPO, so the manufacturer can replace them.

Alibaba's hardware is a 102.4T TH6C-ICA IPO switch; Yao said this ICA version of the Broadcom switch chip was powered on about a month ago:

224G signals exit the top side of the TH6C-ICA package; verbally, all 512 200G-class lanes passed traffic
Each ICA card supports 25.6T of NPO optics, four ICA cards total 102.4T, shown in the rendering as sixteen 6.4T UPO engines
Light comes from sixteen 8-channel ELSFPs, 128 laser channels in total, or about 4 lanes per laser channel (STT calculation)
Claimed low-risk migration path to 204.8T and 409.6T ICA switches

▲ The 102.4T TH6C-ICA IPO switch hardware: four ICA optical cards surround the switch chip, with light supplied by 16 eight-channel ELSFPs (Image source: Alibaba Group | 2026 OCP Global Summit presentation (17:22))

On signal integrity, Yao's claim is clearly better than NPO, slightly behind CPO but comparable, a verbal conclusion from simulation and testing; the slides give no eye diagrams or loss numbers.

▲ Electrical channel length and serviceability trade-offs across four (n)PO options, and the makeup of the 102.4T IPO switch: 4 ICA cards, 16 6.4T engines and 16 eight-channel ELSFPs (Image source: Simple Tech Trend; data from Alibaba Group's 2026 OCP Global Summit presentation)

6. Risks and Counterarguments

Elastomer is a new reliability question. The slide itself lists elastomer reliability and complex mechanical design as IPO's challenges. CPO is criticized for being soldered in; IPO must prove the contact holds up through thousands of thermal cycles, and no public data exists yet.
The IPO ecosystem is still narrow. The demo uses Broadcom's ICA switch chip; if other switch vendors do not adopt top-side signal exits, IPO's multi-source advantage stays on the optical-engine side only.
No signal-integrity numbers. Better than NPO, close to CPO is a verbal statement, with no per-lane loss or pJ/bit comparison.
The cost of split standards. OIF, Open CPX and UPO coexist, and Yao himself acknowledged there is no single form factor. Suppliers serving both sides face roughly double the qualification cost.
The roadmap years are Alibaba's own cadence. IPO in 2027 and CPO in 2028 do not mean the whole industry moves in lockstep.

7. What It Means for Taiwan's Supply Chain

Dual-track standards are the most practical challenge. UPO is the Chinese hyperscalers' own MSA, running in parallel with OIF and Open CPX. Taiwanese suppliers serving customers on both sides need separate form factors, pinouts and test specs, which is both an engineering-resource question and a way to spread regional shipment risk.
IPO shifts value toward board-level and package interconnect. Interposer boards, ICA boards, elastomer sockets and top-side-exit switch packages are new work for PCB, substrate, connector and system-assembly makers, areas where Taiwan already has a base, but the slides name no Taiwanese company, so this remains a technology direction for now.
External lasers are clearly mainstream. Both 3.2T and 6.4T are marked ELS, and the demo uses sixteen 8-channel ELSFPs. That contrasts with the integrated-laser route we covered in CPO Lasers Don't Have to Chase 400 mW: NPO with Integrated Lasers; at least in Alibaba's design, ELSFP demand is explicit.
The window for linear drive is narrowing. Alibaba expects linear pluggables at 400G to be nearly impossible, so for LPO and high-density pluggable vendors, the 200G generation is the last sweet spot.

8. Conclusion

The value of this talk is not a new acronym but a hyperscaler stating its trade-offs clearly: replaceability and multi-sourcing first, peak power efficiency second. NPO is enough below 200G, IPO takes over at 400G, and CPO sits in 2028 as the last option rather than the starting point. That is consistent with our view in NPO Is Not a Stepping Stone to CPO.

Keep the timelines apart. 3.2T NPO is in production, 6.4T UPO is in validation, and the 102.4T IPO switch has been powered on for just a month, with real deployment still some way off. Over the next 12 months, watch three things: elastomer socket reliability data, whether a second switch vendor ships an ICA-style top-side-exit package, and whether the UPO spec is made public.

Once copper no longer sets the size of scale-up, switch radix does; and whether optics can be replaced and multi-sourced decides where they go.

This article is for technology and industry trend analysis only and does not constitute investment advice.