OCP Global Summit 2026 | Point2 Technology, AttoTude | Sean Park, Mike Staskus | Alternate Scalable AI Interconnects

1. Two Companies, One Slogan, Two Architectures

The session sat in OCP's "Special Focus: Photonics" track, yet neither speaker talked about light. Point2 founder and CEO Sean Park opened with the claim that there is a third way to connect AI systems: not copper, not optics, but "RF or ASICs over dielectric waveguides." The slide lists four goals plainly: low cost, reliability, simplicity, scalability, or "overcome the copper wall, avoid the penalties of photonics."

Both companies have serious backing. Point2 has raised $147M, with investors on the slide including Maverick Silicon, NVIDIA, Arm, UMC Capital, Bosch Ventures, Molex and Samsung Securities. AttoTude was co-founded by Infinera co-founder Dave Welch and Joy Laskar; Welch is CEO and Laskar, a Georgia Tech professor, is CTO, and Arista's Andy Bechtolsheim is a board advisor. Note: the title slide lists Dave Welch as AttoTude's speaker, but Mike Staskus from product presented (per his on-stage introduction); the title follows the actual presenter.

Lay the two block diagrams side by side, though, and the differences are larger than the slogan suggests:

Item	Point2 e-Tube	AttoTude DiAx
Signal handling	RF transceiver chips at both ends, upconverted to mm-wave	No upconversion; SerDes baseband sent directly
Medium	Plastic waveguide (metal-shielded)	Hybrid of twinax plus dielectric waveguide
Added electronics	Yes (RF SoC, antenna integrated on package)	None (DSP compensates group delay)
Claimed reach	10–20 m	10–30 m
First application	1.6T OSFP Active RF Cable (ARC)	1.6T AEC running 448G over 10 m

▲ Point2 adds RF chips at both ends to upconvert before a plastic waveguide; AttoTude does not upconvert, sending low frequencies over twinax and high frequencies over a dielectric waveguide (Image source: Simple Tech Trend original; data from Point2 Technology, AttoTude 2026 OCP Global Summit presentation)

The moderator made the same point in Q&A: the principles look similar, but Point2 has a transmitter and receiver, while AttoTude simply sends the SerDes signal straight out. AttoTude's answer was telling: "Really what we're making is a cable."

2. Breaking Down the Gap: Why 5 to 30 Meters Became a Problem

This AttoTude slide was the most useful of the session, because it puts three media on the same ruler:

Lane rate	Twinax	DiAx	Fiber
112G PAM4	7 m	Not listed	500 m
224G PAM4	4 m	10–30 m	500 m
448G PAM4	Under 1 m	10–30 m	500 m

▲ As lane rates rise from 112G to 448G, twinax reach shrinks from 7 m to under 1 m; DiAx claims to fill the 10–30 m gap (Image source: Simple Tech Trend original; data from AttoTude 2026 OCP Global Summit presentation)

From 112G to 448G, twinax reach shrinks more than sevenfold. As we wrote in ECOC2026 | Fast-and-Narrow vs. Slow-and-Wide: 448G Is the Watershed, 448G is where passive copper drops out. In our view, the problem is that fiber arrives with a 500 m spec, and using it for a dozen meters between racks means paying long-haul costs to solve a short-reach problem.

AttoTude puts the cost of optics at 5–7 pJ/bit, roughly 1,000x the failure rate, and about 30x the cost. That is a competitor's framing, but the direction is right: on a 1.6T link, 5–7 pJ/bit works out to 8 to 11 W (STT estimate: 1.6 Tb/s × 5–7 pJ/bit). What the two companies share is a push to treat "5 to 30 meters" as a distinct system challenge, rather than jumping straight to optics whenever copper runs short.

3. Breaking Down Point2: Moving Upconversion from 200 THz to a Band Silicon Can Reach

Point2's e-Tube works like "wired wireless": a pure-silicon RF transceiver with the antenna built onto the package launches the signal into a plastic waveguide much like a radio. Park used a power-versus-frequency chart to explain the band choice: silicon delivers ample output power below roughly 500 GHz; above that lies the "THz gap," with no compact, reliable room-temperature source; higher still is laser territory.

▲ Point2 places e-Tube in bands reachable by silicon processes, avoiding the THz gap; the slide claims 10x copper's reach, 10x lower material cost and 5x lighter weight (Image source: Point2 Technology | 2026 OCP Global Summit presentation)

The cable metrics on the slide: 10x copper's reach, 10x cheaper material, 5x lighter, 2x more flexible; 224G is available now, 448G is expected within two years. Manufacturing is handled by partners such as Molex and Foxconn using existing copper cable equipment.

The roadmap has three steps: pluggable (OSFP/QSFP-DD), near-packaged (NPE) and co-packaged (CPE).

▲ Point2 e-Tube three-stage roadmap: density, energy-efficiency and cost targets for pluggable, near-packaged and co-packaged (Image source: Point2 Technology | 2026 OCP Global Summit presentation)

Reading the bubble positions on the slide (it carries no numeric labels and no dates): pluggable at about 4+ pJ/bit and about $0.20 per Gbps; NPE at about 3 pJ/bit, 1 Tb/mm and $0.15; CPE at about 2 Tb/mm and $0.10, with its energy bubble sitting between 1 and 2, close to 2. Park said CPE will eventually reach 1 pJ/bit; we conservatively use the slide reading of "about 2 pJ/bit or below."

Be clear about one thing: e-Tube is still an upconversion architecture, just with the carrier moved from optics' 200 THz down to mm-wave. At about 4+ pJ/bit, the pluggable is below the 5–7 pJ/bit AttoTude cites for optics, but by less than half.

4. Breaking Down AttoTude: From Terahertz to Baseband

AttoTude's story is the more interesting one, because it changed course. An audience member asked: at OFC, weren't you using a carrier? Mike Staskus confirmed that AttoTude had been upconverting to 700 GHz, found a waveguide material with very low loss at 700 GHz, extended it downward, discovered it stays flat to very low frequencies and complements twinax, and so is "no longer doing terahertz work, focused on baseband transmission only."

DiAx is a frequency split: copper loss rises monotonically with frequency because of skin effect, while a waveguide is a cutoff structure that stays flat above its cutoff. So low frequencies stay on the twinax, high frequencies go into the dielectric waveguide. The cost is that the two paths have different group delays, so the DSP must correct differential group delay; beyond that, no additional electronics are needed. 448G uses roughly DC to 110 GHz, and the next generation will not need terahertz either.

▲ AttoTude lab results for 224G over DiAx: BER 2e-8 and SNR 21.8 dB after equalization (Image source: AttoTude | 2026 OCP Global Summit presentation)

The only measured data so far is at 224G: equalized BER 2e-8, SNR 21.8 dB. Read the conditions carefully: the transmit-side group-delay pre-equalization was done in an arbitrary waveform generator, which Staskus said partners will move into the DSP in future; the 448G link is "the next step," held back by test equipment. The first product is a 1.6T AEC running 448G over 10 m, with a roadmap to 896G.

5. Density Is Not Its Battlefield

Point2's slide says an e-Tube backplane can raise GPU count per rack by 2 to 4x, with the same backplane serving 224G, 448G and beyond. But the co-packaged density target is only 2 Tb/mm. As we covered in ECOC2026 | Microsoft Writes Down the Optical I/O Spec Sheet, Microsoft's optical I/O target is >10 Tbps/mm; e-Tube is about one fifth of that (the two use different definitions, so treat this as an order-of-magnitude comparison).

That is not a flaw so much as positioning: dielectric waveguides win on reach and cost, not on chip-edge bandwidth density. We think the market it is most likely to take first is the rack-to-rack segment between AECs and short-reach AOCs; the density fight at the scale-up chip edge remains CPO/NPO territory, as discussed in 2026 OCP APAC Summit | SemiAnalysis | Scale Up Sophistry on the copper/optics split.

6. Risks and Counterarguments

• 448G has not been measured: DiAx lab results exist only at 224G; 448G is waiting on test equipment. The first product is 448G over 10 m, which makes this the biggest unverified point.

• Someone has to own group-delay compensation: it currently lives in an arbitrary waveform generator. To become a product, SerDes/DSP vendors must build it into silicon. "No additional electronics" assumes other people's chips cooperate.

• Point2's power advantage is limited: the pluggable is about 4+ pJ/bit and only CPE reaches about 2 pJ/bit or below, with no CPE timeline. The slide also does not say whether pJ/bit is per end or per link.

• Reliability is an argument, not data: "wavelengths ~2,000x longer, immune to dust" is physically plausible, but neither company gave MTBF or field data.

• The ecosystem does not exist yet: the closing call to action asks for shared targets on reach, BER, power, reliability, cost and interoperability, which means the specs are not there yet. Even the moderator said it "sounds too good to be true."

7. What It Means for Taiwan's Supply Chain

• UMC is the only Taiwanese company on the slides. UMC Capital is on Point2's investor list, and Park made a point of saying "UMC is a Taiwan foundry company." e-Tube uses standard pure-silicon CMOS/SiGe processes, so foundry is the most direct Taiwan position, though Point2 did not say where it fabs.

• Foxconn is a named cable manufacturing partner. Park said e-Tube cables are made by partners such as Molex and Foxconn on existing copper cable equipment. If dielectric waveguides take hold, the process barrier for Taiwanese cable assemblers is lower than for fiber.

• No other Taiwanese companies appeared in the talk. AttoTude's co-packaged illustration uses an Amphenol product image but states no partnership, so we draw no further pairing.

8. Conclusion

The real message of this session is not "the third path has arrived" but that the third path has already forked. Point2 moves upconversion from 200 THz to mm-wave, dropping the laser but keeping transceiver chips; AttoTude stops upconverting altogether and just builds a better cable. Our position: the passive path is the likelier one to reach the shelf first, because it most resembles the AEC business model and asks the least of anyone else's chips.

Separate the timelines. 224G has lab data; 448G does not. Point2 says 448G within two years; for AttoTude, 448G is the next step. Over the next 12 months, watch three things: AttoTude's measured BER for 448G over 10 m, which DSP vendor builds group-delay compensation into silicon, and whether OCP actually opens a 5–30 m workstream.

Dielectric waveguides will be won not by how high the frequency goes, but by whether they can quietly behave like a cable.

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