OCP Global Summit 2026 | Molex, Broadcom | Clay Miller, Rajesh Annapillai | 224G AECs for AI Data Centers

1. Why Molex and Broadcom Shared the Stage

At the same OCP Global Summit, Credo, the company that built its business on AECs, argued that µEmitter optics can beat copper on reliability, power and cost at once. Molex and Broadcom took the other side and answered a different question: how does copper keep going at 224G?

The division of labor is clear. Molex's Clay Miller covered cables and the market: Molex has already shipped millions of active cables at 112G, with a common deployment range of 1 to 7 m, and 224G is Molex's first known good die (bare die) design after only doing packaged DSP solutions before. Broadcom's Rajesh Annapillai covered the retimer silicon and specs: Broadcom has 25 years of SerDes experience, and its AEC retimers use the same SerDes as its Tomahawk switches.

This differs from Credo's vertically integrated model of selling DSP and cable together. It is a horizontal split where the chip vendor sets the spec and the cable maker differentiates. We broke down Credo's AEC revenue mix in Earnings Highlights: Credo (CRDO) | Q1 FY2027, and this talk is effectively the Broadcom camp's direct answer.

The copper camp is split into two roads: Credo builds the whole cable, Broadcom hands specs and reference designs to cable makers.

2. Reach Breakdown: 224G Pushes Passive Copper Back to 1 m

As each lane doubles from 112G to 224G, loss grows and the reach of passive copper keeps shrinking. The Molex slide draws the lines plainly:

Option	Within 1 m	1–2 m	2–6 m
DAC (passive copper)	Good	Not viable	Not viable
ACC (linear amplified copper)	Good	Limited	Not viable
AEC (retimed active cable)	Good	Good	Good

▲ At 224G per lane, DAC reaches only 1 m, ACC is limited at 1 to 2 m, and AEC maintains good BER from 1 to 6 m (Image source: Molex, Broadcom | 2026 OCP Global Summit presentation (p. 5))

The implication: at 224G, most in-rack and adjacent-rack links become active. Molex's power scenario assumes only 25% of ports can stay passive and 75% need active modules. Compared with the 1 to 7 m deployment range typical of 112G AECs, AEC moves from an option to the default after 224G.

3. Power Breakdown: The 28% Is Versus AOC, Not Versus DAC

Molex's scenario: 4 GPU trays per rack, 18 ports per tray, 72 ports in total, of which 18 are passive and 54 active. Scaled to 5,000 racks:

• 270,000 active modules

• AOC at 25 W each, 6.75 MW in total

• AEC at 18 W each, 4.86 MW in total

• A gap of 1.89 MW, or 28% lower power (module power only, excluding other IT and cooling)

▲ In a 5,000-rack scenario with 270,000 active modules, AEC at 18 W versus AOC at 25 W gives total power of 4.86 MW versus 6.75 MW, 28% lower (Image source: Molex, Broadcom | 2026 OCP Global Summit presentation (p. 3))

Two things to read carefully. First, the baseline is AOC, not passive copper; the 18 W an AEC draws is itself the price of passive copper failing at 224G. Second, the saving is 7 W per module, about 378 W per rack across 54 ports, which is small inside an AI rack drawing more than 100 kW.

Applying the roughly $100 per watt opportunity cost that Credo presented at the same summit, 1.89 MW maps to about $190 million. This is STT's arithmetic combining the two companies' numbers, not a Molex claim, but it shows why Molex stresses that the gap only becomes visible at scale.

▲ The 224G copper landscape: DAC retreats to within 1 m, AEC holds 1 to 6 m, saving 7 W per module versus AOC and 1.89 MW across 5,000 racks (Image source: Simple Tech Trend; data from Molex and Broadcom's 2026 OCP Global Summit presentation)

4. The 6 m Breakdown: The Bottleneck Shifts from Insertion Loss to Skew

Broadcom splits the challenges beyond 6 m into five items: insertion loss, crosstalk, intra-pair skew, reflections and power supply noise. Rajesh's point: insertion loss still matters, but at 224G crosstalk and skew matter just as much.

• Legacy metrics fail: time-domain (TDR/TDT) correlates poorly with system performance, and raw frequency-domain (VNA) skew varies widely across bands.

• New metric, EIPS: frequency-domain skew integrated against a weighting function, combining time- and frequency-domain views. Broadcom says it took 12 to 18 months to define a spec suited to cables.

• Two CMC masks: long, tightly coupled twinax of 3 m and above uses the IEEE CR mask; short, loosely coupled traces such as packages and paddle cards use a skew below 0.5 ps mask.

Following this spec, Broadcom first optimized the retimer package (P/N skew, impedance, ground shielding), then optimized the paddle card with Molex, and finally co-simulated retimer, package, paddle card and twinax together. The slide's simulation of a 1.6T AEC at 6.5 m shows per-lane pre-FEC BER between roughly 3E-10 and 7E-10; Broadcom said verbally that simulation-to-silicon correlation exceeds 99%.

Put simply: 6 m is not reached with thicker wire, but with a skew budget for every segment. That is why Broadcom says the work starts at the retimer.

5. The Real Moat: Telemetry and Channel Modeling

In the second half of the talk, Broadcom moved from signal integrity to operational efficiency.

Real-time diagnostics: the retimer itself measures pre-FEC BER, SNR, EMAX, FEC tail (frame-error statistics), power and temperature at both ends of the cable, near end and far end. For cable makers, this shortens bring-up, production test and failure analysis; for hyperscalers, it catches a bad cable before it hurts the fabric.

Real-time channel modeling: measuring the frequency, step and pulse response of the whole AEC channel at the hyperscaler site, per lane on both host and media sides. The interface integrates into CMIS, so diagnostics and modeling share one software stack.

▲ Broadcom retimers measure pre-FEC BER, SNR, FEC tail, power and temperature in real time at both ends of the cable (Image source: Molex, Broadcom | 2026 OCP Global Summit presentation (p. 12))

The product line is already laid out:

Generation	Chip	Process	Spec	Status
400G / 800G	AGERA2LP 8:8	5nm	Same SerDes as Tomahawk5, >45 dB, 8 m+	In production
1.6T	AGERA3 / 3U 8:8	3nm	Same SerDes as Tomahawk6, >45 dB, 6 m+	In production
3.2T	AGERA4 8:8	2nm	400G SerDes, >40 dB	Coming soon

▲ Broadcom's AGERA AEC retimer family: the 1.6T AGERA3/3U is in volume production, while the 3.2T AGERA4 uses 2nm and 400G SerDes (Image source: Molex, Broadcom | 2026 OCP Global Summit presentation (p. 14))

This matches Credo's move to make telemetry a selling point with its PILOT platform: AECs no longer compete on wire, but on how much the retimer can see and how far the software can plug in.

6. Market: LightCounting's 2030 Crossover

Molex cited LightCounting's AEC shipment forecast (July 30, 2026):

• 800G AECs shipped 1.74 million units in 2025 and, read from the chart, peak at roughly 11 million in 2027 before settling around 8 million.

• 1.6T AECs grow from 1.37 million units in 2027 to 8.85 million in 2030, 6.5x in four years.

• 2030 is the first year 1.6T outships 800G.

▲ LightCounting forecasts 1.6T AECs growing from 1.37 million to 8.85 million units between 2027 and 2030, overtaking 800G for the first time in 2030 (Image source: Molex, Broadcom | 2026 OCP Global Summit presentation (p. 6))

Get the timing right: 2027 is only the ramp year for 224G AECs, and 800G still dominates that year. Clay said 224G AECs will see substantial growth in the next two years, but by the numbers 1.6T becomes the main revenue source only in 2028 to 2029.

7. Risks and Counterarguments

• The power baseline favors AEC: the 28% is versus AOC; against low-power optics such as LPO or NPO the gap would shrink, and the slides give no such figure.

• Per-rack savings are small: about 378 W per rack matters in a power-limited hall, but is not enough on its own to decide the architecture.

• 6.5 m is still simulation: the BER chart is co-simulation; production yield and measurements at temperature extremes are not public, and the chart lists only 7 TX lanes.

• Optics is catching up: at the same summit Credo claimed its 1.6T ALC runs 12 W per end over 50 m with volume production in 2027; if its cost really approaches copper, the 2 to 6 m AEC zone gets squeezed from above. For the copper-versus-optics debate in scale-up, see 2026 OCP APAC Summit | SemiAnalysis | Scale Up Sophistry.

• Reference designs compress cable makers' differentiation: once the paddle card baseline is shared, cable makers can only compete on twinax, connectors and manufacturing, and price competition intensifies.

8. What It Means for Taiwan's Supply Chain

• Going active is a value upgrade for cable assemblers. With 75% of ports needing active modules at 224G, each cable adds a retimer, a paddle card and a test flow, raising both unit price and the technical bar above DAC. Among Taiwanese cable assemblers, BizLink is already on Credo's SRMO workstream list, meaning it is betting on both active copper and active optics.

• Broadcom's reference design lowers the entry barrier. Cable makers without their own DSP can take Broadcom's paddle card baseline and CMIS telemetry tools straight to hyperscalers, a shortcut for Taiwanese and Chinese cable makers entering 224G AECs, but it also means differentiation is limited to twinax, connectors and production yield.

• Connectors and twinax are home turf for Amphenol and Molex. Both are highly vertically integrated in high-speed cabling; we covered Amphenol's AI interconnect revenue in Earnings Highlights: Amphenol (APH) | Q2 FY2026. Taiwanese firms are competing for assembly and test capacity, not for control of the spec.

9. Conclusion

The point of this talk is not that copper is still alive, but that 224G redefines copper: passive copper retreats within 1 m, everything from 1 to 6 m goes active, and once it goes active, value shifts from the wire to the retimer, the skew spec and the telemetry software.

Broadcom uses EIPS, CMC masks, co-simulation and a reference paddle card to turn 6 m into repeatable engineering; Molex turns it into cable and takes it to hyperscalers. On timing, 2027 is the 1.6T AEC ramp year and 2030 is when it overtakes 800G. Three things to watch over the next 12 months: measured BER and yield of 6.5 m AECs, whether CMIS telemetry becomes a hyperscaler procurement requirement, and how far 3.2T AGERA4 can reach within a >40 dB budget.

After 224G, the copper cable race is not decided by copper, but by how much the retimer at each end can see.

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