OCP Global Summit 2026 | 650 Group | Alan Weckel | Industry Perspective

1. Why This Talk Deserves Its Own Breakdown

650 Group is a market research firm founded about ten years ago that describes itself as counting everything: ports, servers, NICs, SerDes lanes. It publishes about 50 reports a quarter and tracks more than 500 billion USD of equipment spend per quarter. At the OIF workshop held during OCP, co-founder Alan Weckel used just over 20 minutes to lay out the scale of the entire data center network.

This talk is the mirror image of the LightCounting session. LightCounting offered three scenarios and admitted 2028–2029 could be bumpy; 650 Group gave a single line that only goes up. Both agree optics will take over scale-up, but they disagree completely on whether growth needs a breather first.

2. Breaking Down Demand: How Many Ports Does a One-Million-XPU Campus Need

Weckel splits AI spending into four waves: Wave 2 (2022–2025, foundation models) at about 300 billion USD of equipment; Wave 3 (2025–2028, AI agents) at about 1 trillion; Wave 4 (2027–2035, physical AI) at more than another 1 trillion. One trillion dollars means roughly 1 million USD of spend and 300 data center switch ports shipping every minute.

He said that since this summer, traffic generated by AI agents has exceeded human traffic, a milestone originally expected in 2027. Agents run around the clock with chattier traffic, so bandwidth into and out of the data center surges.

Here is how a one-million-XPU AI campus compares with a traditional single-building data center:

One detail he mentioned: there is so much fiber that ceilings must be reinforced to carry the weight of the overhead trays.

Another demand variable is the neocloud. Nebius spent about 5.6 billion USD of capex in 2Q26 and CoreWeave about 9.4 billion USD. 650 Group expects neocloud spending as a category to outgrow the hyperscalers; the hyperscaler share of AI networking revenue falls from about two-thirds in 2025 to under 60% in 2030.

3. Breaking Down Silicon: Data Center Semis Grow 5x in Five Years

650 Group expects data center semiconductors in 2030 to be 5x their 2025 size, about 2 trillion USD, far above roughly 800 billion USD for non-data-center chips. Two notes on the slide are blunt: data centers alone need 12+ fabs at 3nm and below; and Apple is no longer TSMC's number one customer.

On accelerators, 2026 GPU/XPU/AI ASIC revenue is about 385 billion USD, with NVIDIA first and Google TPU second, and the rest split among AMD, Amazon, Meta, Anthropic and others. 2030 revenue reaches about 1.3 trillion USD, still mostly merchant silicon; but of roughly 60 million units shipped, custom ASICs are already more than half.

Weckel's inference: each large customer will run dozens of accelerators, racks cannot be redesigned for each one, and Ethernet becomes the shared fabric. He also flagged what makes optical modules special: memory can borrow capacity from consumer markets, but there is no consumer market for transceivers, so capacity has to come from the industry itself. That matches what we saw in ECOC2026 | No Tech Revolution in the Next Decade, Just a Scaling War: manufacturing capability decides the winners.

4. Breaking Down Networking: Scale-Up Is the Biggest Slice, but 70% Is NVLink

650 Group segments the network by distance: scale-in under 5 cm (in package), scale-up within 3 m of the rack, scale-out from 50 m to 2 km, and scale-across/DCI beyond 20 km. Each step inward carries roughly 10x the bandwidth.

Within switch equipment spend, the data center share rises from 48% in 2020 and 58% in 2025 to 83% in 2030; in 2031–2032 data center switch ports will exceed campus ports for the first time, at about 450–500 million ports a year, not counting scale-up.

▲ 650 Group's preliminary forecast puts data center networking revenue at about $1.25 trillion in 2035, with a 2030–2035 CAGR of about 29% and scale-up the largest share (Image source: 650 Group | 2026 OCP Global Summit OIF workshop presentation (17:54))

▲ CAGR by segment: scale-across is highest at about 147% through 2030, scale-up 68% and scale-out 62%; overall growth slows to 29% in 2030–2035, with scale-up still at 38% (Image source: Simple Tech Trend original; data from 650 Group 2026 OCP Global Summit presentation)

Scale-up revenue is about 17 billion USD in 2026 and about 105 billion USD in 2030, with NVLink at roughly 70% and Ethernet at roughly 20%. The path has three stages: single rack before 2026, multi-rack copper in 2026–2028, and multi-rack fiber from 2028. Weckel said Ethernet "will eventually take scale-up too, it just takes a little longer."

One thing to be clear about: these numbers count standalone switch hardware only, excluding devices embedded in servers or accelerators; scale-across is also outside the AI networking report and is added on top.

5. Breaking Down Optics: No Copper Cliff, but All the Growth Is Optical

▲ 650 Group preliminary forecast: scale-out CPO catches up with traditional optical modules around 2034; scale-up CPO overtakes copper around 2031, after which copper shrinks quickly and optical interposers take over from 2031 (Image source: 650 Group | 2026 OCP Global Summit OIF workshop presentation (21:38))

The two charts move in completely different ways:

• Scale-out plus front-end: traditional transceivers grow from about 35 billion USD in 2025 to about 110 billion USD in 2035 and are not displaced; CPO starts climbing in 2027 and draws level around 2034.

• Scale-up: copper-only revenue peaks at about 48 billion USD around 2030 and drops sharply by 2032; CPO passes copper in 2031 and reaches about 160 billion USD in 2035, with optical interposers adding about 90 billion USD more.

This is what underpins the trillion-dollar forecast: optical content per switch rack goes up, and "just as NVIDIA racks went from 1 million to 5 million USD, networking racks will follow the same path."

In Q&A someone asked whether there is a "copper cliff." Weckel said no: copper stops growing but holds its run rate and still exists ten years from now, "but in ten years I would want to be an optical company." That is close to the copper view in OCP APAC Summit 2026 | SemiAnalysis | Scale Up Sophistry, except that 650 Group marks the optical transition earlier.

SerDes is another timing signal. 100G is ramping now, 200G becomes meaningful only in 2H27, and 400G follows right after, with each generation shipping more than 2x the volume of the previous one. Weckel warned of "orphan technologies": when hyperscalers jump to 1.6T and 3.2T, suppliers will stop supporting the older sub-400G speeds, and enterprises will be forced to upgrade too.

6. Breaking Down DCI: 5% of the Ports, 147% Growth

The scale-across logic is simple: no heroic assumptions, just about 5% of in-building ports needing to leave the data center, and that alone creates a whole new market.

▲ ZR/ZR+/Coherent Lite forecast: AI scale-across reaches about 21.5 million ports and about $18 billion in 2030, versus about 1.5 million ports for traditional non-AI DCI (Image source: 650 Group | 2026 OCP Global Summit OIF workshop presentation (27:42))

• In 2029 AI uses about 7 million ports and about 7.5 billion USD; in 2030 it jumps to about 21.5 million ports and about 18 billion USD.

• Traditional non-AI DCI stays nearly flat at about 1.5 million ports and about 2.5 billion USD a year.

• The unlock named on the slide is 2nm 1.6T Coherent Lite, paired with high-radix switches, to bring cost and power low enough.

Weckel said nobody ever expected DCI to exceed 10 million ports, let alone 20 million. Several hyperscalers have already added RFPs or adjusted plans, and new builds are laying conduit and fiber between buildings up front. That contrasts with Cignal AI's view in ECOC2026 | 95% of OCS Shipments Are One Company's Business: Google: when a new market is just starting, customer concentration is often more worth watching than the forecast curve.

▲ 650 Group timeline breakdown: 200G SerDes in 2H 2027, multi-rack fiber in 2028, scale-across ramping in 2029–2030, scale-up CPO overtaking copper in 2031, and scale-out CPO matching optical modules in 2034 (Image source: Simple Tech Trend original; data from 650 Group 2026 OCP Global Summit presentation)

7. Risks and Counterarguments

• Values are chart readings: there is no official PDF for this session, most dollar values were read from video captures with possible errors in the billions, and all post-2030 charts are marked "Preliminary."

• The CAGR label is questionable: the growth column on the 2030 chart is labeled "CAGR (30-35)", but the chart stops at 2030; compared with the spoken "40% plus," STT reads it as 2025–2030.

• Single scenario: LightCounting gave three scenarios with a 4x spread and sees a possible bumpy 2028–2029; 650 Group offers no downside case. The split between Goldman Sachs and Eoptolink on CPO timing (see Goldman Raises, Eoptolink Delays) also shows how easily timelines slip.

• Scale-up is still NVIDIA's market: about 70% of 2030 scale-up is NVLink, so how much Ethernet captures depends on whether NVIDIA opens up.

• Macro assumption: Weckel himself said the forecast needs an extra 0.5 to 1 point of global GDP growth per year; in a recession, the x-axis stretches out.

8. What It Means for Taiwan's Supply Chain

• Foundry is the most certain beneficiary. The slide states data centers need 12+ fabs at 3nm and below and that Apple is no longer TSMC's top customer; with data center semis growing 5x in five years, leading-edge capacity is the first bottleneck.

• "Manufacturing wins" favors Taiwan's assemblers, but power is a new challenge. The slide notes manufacturing is as power-constrained as operators, and system test and burn-in need significant power investment; as Taiwanese server and switch assemblers expand, site power matters as much as production lines.

• Switch rack value is rising, so white-box switch ODMs are worth tracking. With 450–500 million DC switch ports a year, the data center share of spend heading to 83%, and neoclouds bringing a more diverse customer base, white-box switch makers that build for cloud operators, such as Accton, see a signal of more customers and higher value per unit at the same time.

• DCI and coherent are a gap for Taiwan. The key to 2nm 1.6T Coherent Lite is the DSP and coherent optical components, where Taiwan has no clear volume role today; that is the thing to watch before the 2029–2030 ramp.

9. Conclusion

The value of this talk is not the trillion-dollar headline but that it marks when each segment turns optical: 200G SerDes in 2H27, multi-rack fiber in 2028, scale-across ramping in 2029–2030, scale-up CPO passing copper in 2031, and scale-out CPO only in 2034.

But this is a single-scenario line. Compared with LightCounting, both agree optics takes over after 2029; the difference is that 650 Group sees no pause in between. Over the next 12 months, three things to watch: actual 200G SerDes switch shipments in 2027, the Ethernet share of scale-up outside NVLink, and the timing of the first 1.6T Coherent Lite shipments.

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