2026 Showcase

OCP Global Summit Showcase

The Open Compute Project held its OCP Global Summit at the San Jose McEnery Convention Center from October 12th to 15th, 2026, under the theme “Scaling Innovation for the AI Era.” The event reached its 15,000-person capacity, with a 3,000-person waiting list; to handle future capacity, 2027 moves to Moscone Center under contracts through 2030. This year's summit had 22 tracks, nearly 600 sessions, and about 950 speakers.

OCP's mission is unchanged: taking hyperscale innovation to all, guided by openness, efficiency, sustainability, impact, and scale. What's growing is the scope — from the data hall to the full AI infrastructure stack, spanning silicon diversity, chiplets, facilities, energy, and the AI computing continuum from hyperscale data centers to regional and smaller sites. Looking to 2027 and beyond, OCP is launching three initiatives: system technology co-optimization for silicon design, modular interfaces that reduce grid-to-chip redesign, and solutions for smaller racks under 100 kilowatts.

Jim Carroll of Converge Digest and Roy Chua of AvidThink were onsite gathering video interviews with speakers, vendors, and community leaders. We appreciate our multi-year partnership with OCP and thank them for letting us capture video.

While you're here, feel free to grab our latest research brief right from this page.

Latest Research Brief

Download Our Latest Report

Data Center Networking in 2026 — Scaling to New Heights

More about this report →

No personal info needed to download

Download the Latest Report

By downloading this report, you agree to our Terms of Use and Privacy Policy.

Subscribe to our newsletter and receive new report updates

The monthly AvidThink infrastructure brief. No spam — unsubscribe anytime.

Analysis by AvidThink

What we learned at the 2026 OCP Global Summit

Across 28 interviews at a sold-out San Jose show, the open hardware community is rebuilding the AI data center around light, 800-volt power, and open standards for scale-up.

  1. An optics show first — and CPO vs. NPO is no longer theoretical

    650 Group's Alan Weckel called OCP 2026 fundamentally an optics show, with laser demand heading from hundreds of millions to billions. Meta has logged 225 million hours on Broadcom's Tomahawk 5 CPO, NVIDIA's co-packaged optics is shipping with nearly 5x better power efficiency and 10x resiliency, and Ayar Labs ran TSMC COUPE-based optical engines for 200+ Tbps of scale-up bandwidth in a liquid-cooled rack. Near-packaged optics keeps the pluggable model alive: Molex's connectors scale toward 448G, and Credo is building NPO PICs with customers.

  2. Scale-up gets open standards at every layer

    Broadcom's Ethernet for Scale-Up Networking group grew from 12 founders to 175 members with a 1.0 spec, five hyperscalers, and 10+ ODMs on board. The OCI MSA — Microsoft, Meta, OpenAI, AMD, NVIDIA, and Broadcom — standardizes scale-up optics, and Xscape Photonics showed Falcon X as the first OCI-compliant pluggable laser. Upscale's Token Fabric unites scale-up and scale-out under one operating model, while Lumentum, Astera Labs, and Corning stretch optical PCIe across racks and rows.

  3. Scale-across: the AI cluster spans buildings

    Meta's keynote confirmed clusters now span buildings 2–10 km apart, linked by trenches carrying around a thousand fiber strands per conduit — validating 650 Group's raised coherent-lite forecast. Ciena's RLS Hyper Rail packs up to 128 fiber pairs per rack at 32x the density, Broadcom offers a Jericho 4-based scale-across portfolio, and Keysight now validates ZR, ZR+, and coherent-lite links alongside scale-up and scale-out.

  4. 800 volts, from the grid to the tray

    Advanced Energy laid out both paths: rack power shelves that convert 800 V down to 50 V for today's trays, or 800 V straight into the tray with onboard conversion. Flex's EPC Power acquisition converts 34.5 kV directly to 800 V DC for roughly five points of efficiency, with solid-state transformers adding one to two more, and its pre-manufactured AI reference platform cuts deployment time by up to 50%. NVIDIA ties it together with tokens per watt as the metric.

  5. Inference hits the memory wall — and the industry pushes back

    Credo's OmniConnect Weaver chiplet delivers HBM4-class bandwidth with terabyte-class capacity, and Marvell's second-gen Structera X CXL controller cuts time to first token by 50%. d-Matrix disaggregated attention onto NVIDIA H100s and experts onto its Corsair nodes for 800+ tokens per second per user, and Rebellions reports most customers buying inference at rack scale rather than server by server.

  6. AI is a system now — test, secure, and open it as one

    Keysight argues qualifying any component means emulating the whole system around it, MultiLane notes one faulty connection among hundreds can stop a rack from shipping, and VIAVI tests from fiber to AI job performance. Marvell's Project Tessera rethinks cloud security for a world where AI agents outnumber employees 100 to one. Solutional calls this an "IBM moment" dominated by a few players — which is exactly why OCP's open alternatives matter.

More about this report →

Highlights from industry thought leaders

Video interviews

650 Group

Why Coherent-Lite Demand Is Taking Off

  • Meta's day-one keynote confirmed AI clusters no longer fit in one building and now span sites 2–10 km apart.
  • Those sites are linked by trenches of conduits, each carrying on the order of a thousand fiber strands.
  • 650 Group raised its coherent-lite forecast for scale-across about two months before the show, and Meta's talk signals the market is taking off as projected.
Abstract
Chris DePuy, Co-founder & Analyst at 650 Group, highlights the key takeaway from the OCP show: Meta's day one keynote revealed that AI clusters no longer reside within a single building but now span many buildings 2 to 10 kilometers apart, linked by massive trenches of conduits each carrying around a thousand fiber strands. He explains that 650 Group raised its coherent-lite forecast for scale-across applications about two months before the show after seeing earlier trench pictures, and views Meta's presentation as broad industry acknowledgement that this market is taking off in line with those projections.
650 Group

OCP 2026 Takeaways: CPO vs. NPO

  • OCP 2026 was fundamentally an optics show, with heated NPO-versus-CPO debate settled by manufacturing scale and fit with XPU roadmaps.
  • Laser demand is moving from hundreds of millions to billions, opening room for high-power lasers, VCSELs for scale-up and microLEDs for memory.
  • Weckel expects an "all of the above" outcome, matching each technology to the workload for cost- and energy-optimized deployments.
Abstract
Alan Weckel, Founder and Technology Analyst at 650 Group, shares his takeaways from a sold-out OCP 2026, describing it as fundamentally an optics show marked by heated debate over near-packaged versus co-packaged optics, with manufacturing scale and alignment with XPU roadmaps shaping adoption. He expects an "all of the above" outcome as laser demand grows from hundreds of millions to billions, with high-power lasers, VCSELs for scale-up networks, and microLEDs for memory each matched to the right workload for cost- and energy-optimized deployments.
Advanced Energy

Powering AI Data Centers: Two Paths from 50V to 800V

  • Advanced Energy supports both today's 50 V AI infrastructure and the industry's move to 800 V architectures.
  • One path keeps 50 V trays and adds a rack power shelf that converts 800 V down to 50 V.
  • The other brings 800 V into the tray, where a hot-swap power distribution board and onboard converters step down to 50, 12 or 6 V.
Abstract
Conor Quinn, Vice President Marketing, High Volume Products at Advanced Energy, presents the company's power conversion technology at the Open Compute Global Summit 2026 in San Jose, explaining how it supports both 50-volt AI infrastructure and the industry's transition to 800-volt architectures. He outlines two approaches: rack power shelves that convert 800 volts down to 50 volts for existing trays, and power distribution boards that bring 800 volts directly into server and networking trays with hot swap control and onboard conversion, all backed by global technical support and a robust manufacturing footprint.
Ayar Labs

Co-Packaged Optics Demo in a Liquid-Cooled Rack

  • Ayar Labs, Wiwynn and Corning show an upgraded optical scale-up rack aimed at proving co-packaged optics can be built and delivered reliably at scale.
  • It is the first live rack demo of Ayar Labs' TSMC COUPE-based optical engines, integrated with GPU and HBM in GUC multi-chip packages for more than 200 Tbps of scale-up bandwidth.
  • Hours of fully liquid-cooled operation show stable micro-rings, lasers held near 50°C, and all components within their temperature ranges.
Abstract
At the OCP Global Summit, Vishal Chandrasekar, Director of Product Management at Ayar Labs, presents an upgraded optical scale-up rack built with Wiwynn and Corning, featuring the company's first live demonstration of TSMC COUPE-based optical engines in multi-chip packages developed with GUC that deliver more than 200 Tbps of scale-up bandwidth. The fully liquid-cooled demo shows stable micro-ring resonators and well-controlled thermal profiles over many hours, demonstrating that co-packaged optics can be manufactured and deployed reliably at data center rack scale.
BE Networks

How AI Powers Outcome-Driven Networking for Enterprises

  • BE Networks positions itself as a control plane for large fabrics, edge deployments, private clouds and neoclouds.
  • As hyperscaler technology flows down to enterprises, the company helps them build better networks.
  • Its new outcome-driven networking moves beyond intent, using AI to deliver the result customers actually want rather than just the starting point.
Abstract
Amir Elbaz, CEO and founder of BE Networks, speaks at a sold-out OCP event with 14,000 attendees and describes his company as a control plane for large fabrics, edge deployments, private clouds, and neoclouds that helps enterprises build better networks as technology flows down from the hyperscalers. He introduces outcome-driven networking, which moves beyond intent-based networking by retooling BE Networks products and applying AI to deliver the outcomes customers seek, and he notes that open networking is gaining significant momentum.
Broadcom

Why Ethernet Is Winning AI Networking

  • The Ethernet for Scale-Up Networking work group has grown from 12 founding companies to 175 members and published its 1.0 specification.
  • Five hyperscalers now use Ethernet for scale-up, with at least 10 ODMs building scale-up Ethernet racks.
  • Tomahawk 6 is in volume production, joined by the Thor Ultra NIC (UEC- and MRC-compliant) and the Jericho 4-based scale-across portfolio.
Abstract
At the OCP Global Summit, Hasan Siraj, VP, Product Management, Core Switching Group at Broadcom, highlights growing momentum for open, Ethernet-based AI networking, including the Ethernet for Scale-Up Networking work group, which has grown from 12 founding companies to 175 members with a published 1.0 specification and adoption by five hyperscalers and at least 10 ODMs. He also showcases Tomahawk 6, now in volume production as the only 100-terabit platform supporting more than 100,000 XPUs in a two-tier topology, alongside the production Thor Ultra NIC, the Jericho 4-based scale-across portfolio, and Broadcom's leadership of OCP's AI cluster design track.
Broadcom

Making the Case for OCI

  • Meta reports 225 million hours of operation on Broadcom's Tomahawk 5 CPO platform, and the Tomahawk 6 Davidson CPO platform is expected in production next year.
  • The OCI MSA — founded by Microsoft, Meta, OpenAI, AMD, NVIDIA, and Broadcom — standardizes optics for scale-up by decoupling electronics from optics.
  • Its DWDM-on-a-CWDM-grid approach uses standard lasers rather than comb lasers, and the MSA plans to open to more members soon.
Abstract
At OCP, Rajiv Pancholy, Director, Hyperscale Strategy & Products, Optical Systems Division at Broadcom, highlights the company's co-packaged optics momentum, citing Meta's 225 million hours of operation on the Tomahawk 5 CPO platform and partner showcases of the Tomahawk 6 Davidson CPO platform, which Broadcom expects to enter production next year. He emphasizes the Optical Compute Interconnect MSA, founded by Microsoft, Meta, OpenAI, AMD, NVIDIA, and Broadcom, which standardizes optics for scale-up by decoupling electronics from optics and using DWDM on a CWDM grid with standard lasers, and which the founders plan to open to additional members soon.
Ciena

Scaling AI Across Buildings: 32x Denser Networks for Hyperscalers

  • Hyperscalers spreading AI clusters across buildings need massive scale-across capacity, which Ciena serves with coherent transceivers and new photonic line systems.
  • The RLS terminal shelf, deployed in volume for single-span links, pairs 800G coherent pluggables across C and L bands with EDFA/Raman amplification and optical protection.
  • RLS Hyper Rail supports 16 fiber pairs per system – up to 32x the density and 128 fiber pairs per rack – so hundreds of fibers share one amplifier hut.
Abstract
At the OCP Global Summit, Helen Xenos, Senior Director, Portfolio Marketing at Ciena, showcases coherent transceivers and photonic line systems built for hyperscalers distributing AI clusters across multiple buildings in "scale-across" networks, including the RLS terminal shelf, which delivers high capacity, low latency, and low power with 800G coherent pluggables across the C and L bands. She also highlights the RLS Hyper Rail system, which supports 16 fiber pairs in a compact form factor for up to 32 times more density and up to 128 fiber pairs per rack, letting customers run hundreds of fibers within a single amplifier hut.
Cisco Systems

New USFP Open Optical Form Factor Scales to 12.8T

  • USFP is a new open-standard optical module with 16 or 32 lanes at up to 400 Gbps per lane, scaling from 3.2T to 12.8T.
  • It extends the OSFP model in manufacturing and field deployment, growing through faster lanes and wider widths.
  • The module supports air cooling and system-level liquid cooling via a top-mounted cold plate, with no fluid fittings in the module.
Abstract
Tom Williams, VP, Marketing at Cisco Systems, discusses the USFP at the close of the 2026 OCP Global Summit in San Jose, describing a new open-standard optical form factor with 16 or 32 lanes at up to 400 Gbps per lane, scaling from 3.2T to 12.8T while extending the OSFP model in manufacturing and field deployment. He highlights support for both air cooling and system-level liquid cooling through a top-mounted cold plate, noting strong interest from end users and vendors as full specifications roll out over the coming months.
Credo

Breaking the AI Memory Wall: HBM4-Class Bandwidth

  • AI inference is now memory bound rather than compute bound — processors sit idle waiting on memory bandwidth, the so-called memory wall.
  • Credo's OmniConnect Weaver memory gearbox chiplet links an XPU to LPDDR memory, delivering HBM4-class bandwidth and terabyte-class capacity, and is available now.
  • A two-board AMD Versal FPGA emulation lets customers run their own workloads and inject link errors on real hardware before final silicon.
Abstract
Vishal Shah, VP of Product at Credo, explains that AI inference has become memory bound rather than compute bound and demonstrates a hardware emulation of Credo's OmniConnect Weaver memory gearbox chiplet, built on two AMD Xilinx Versal-based FPGA boards that emulate an XPU linked serially to the Weaver chiplet and LPDDR memory. Using a GUI to issue memory transactions, monitor data volumes, and inject errors, Shah shows how customers can test their own workloads on real hardware before final silicon, highlighting the Weaver chiplet's HBM4-class bandwidth, terabyte-class capacity, and current availability.
Credo

Scaling Silicon Photonics: 1.6T Optics, NPO, and 400G per Lane

  • With DustPhotonics' PICs now part of Credo, the company offers a full 800G and 1.6T silicon photonics lineup ramping to very high volumes.
  • Credo has spent six months scaling wafer supply and test capacity to serve external customers as well as its own internal product needs.
  • A booth demo paired the Credo DSP with Credo silicon photonics at 1.6T, and next-generation PICs for near-packaged optics and 400G per lane are on the way.
Abstract
At OCP, Ronnen Lovinger, VP of Silicon Photonics at Credo and former CEO of DustPhotonics, discusses how the recent integration of DustPhotonics' photonic integrated circuits gives Credo a full lineup of 800G and 1.6T silicon photonics products ramping to high volumes, backed by expanded wafer supply and test capacity for both external and internal customers. He also highlights a 1.6T demo pairing the Credo DSP with Credo silicon photonics for strong optical performance and interoperability, along with upcoming PICs for near-packaged optics and a 400G-per-lane product.
d-Matrix

d-Matrix Attention Feed-Forward Disaggregation at 800+ Tokens/Sec

  • d-Matrix announced Squadron, its cluster-scale solution, at the 2026 summit.
  • A live attention feed-forward disaggregation (AFD) demo ran attention on NVIDIA H100 GPUs and experts on Corsair nodes, topping 800 tokens per second per user on a 20B-parameter model.
  • Memory-bound expert workloads suit Corsair's memory-bandwidth-optimized design, enabling larger batches and keeping nodes efficiently utilized.
Abstract
Sree Ganesan, VP of Product at d-Matrix, announces the company's Squadron cluster-scale solution and demonstrates the industry's first live attention feed-forward disaggregation on heterogeneous compute, running a 20 billion parameter model with attention on NVIDIA H100 GPUs while experts execute on d-Matrix's Corsair nodes at over 800 tokens per second per user. The AFD approach optimizes performance by leveraging Corsair's memory bandwidth architecture for memory-bound expert workloads while keeping attention operations on GPUs, enabling higher batch processing and efficient node utilization.
Flex

Grid to Rack as One Power System

  • Flex spans grid-level power through rack and board power to integrated rack and server manufacturing, so customers can treat the data center as one system.
  • A new pre-manufactured, payload-agnostic AI reference platform cuts deployment times by up to 50%.
  • The EPC Power acquisition converts 34.5 kV straight to 800 V DC for roughly five points of efficiency gain, with solid-state transformers adding one to two more.
Abstract
Chris Butler, President, Embedded and Critical Power at Flex, explains how Flex brings a system-level approach to the OCP community through a portfolio spanning utility-level power, power distribution, rack and board-level power, and integrated rack and server manufacturing, letting customers treat the data center as a single system for better efficiency, faster deployment, and modular construction. He highlights a new pre-manufactured, payload-agnostic AI reference platform that shortens deployment times by up to 50% and the acquisition of EPC Power, whose direct 34.5 kV-to-800-volt DC conversion yields roughly five points of efficiency gain, with solid-state transformer technology on the roadmap adding at least another one to two points.
Keysight Technologies

Validating Scale-Up, Scale-Out, and Scale-Across

  • Scale-up brings a mix of high-speed copper and optical links plus new transports like Ethernet-based standards and UALink, all needing validation at speed.
  • In scale-out Clos fabrics, congestion and incast cut GPU utilization – issues Keysight helps customers resolve in pre-deployment labs.
  • AI is now about systems, not devices, so qualifying any component means emulating the entire system around it at scale.
Abstract
At OCP 2026, Ram Periakaruppan, VP and GM, Network Applications and Security at Keysight Technologies, explains how customers struggle to validate AI infrastructure at scale across scale-up, scale-out, and scale-across domains, from high-speed copper and optical links and protocols like UALink within racks, to congestion and incast in Clos fabrics, to coherent optics such as ZR, ZR+, and coherent lite linking data centers. Emphasizing that AI is now about systems rather than individual devices, Periakaruppan notes that Keysight helps organizations emulate the entire surrounding system of any component so they can qualify it at scale in pre-deployment lab environments.
Lightmatter, Flex

200-Terabit Optical Switch: Collaboration Advances AI Networking

  • Lightmatter and Flex built a 1RU, 200-terabit optical switch through an OCP workstream, pairing a dense laser system and all-optical connectivity with Flex's modular fiber attach units and fiber trays.
  • The work has moved from a contributed white paper to prescriptive specifications covering fiber connectors, optical loss, and other critical parameters.
  • The prototype is a test vehicle for building supply-chain know-how with partners such as Senko (connectors) and JetCool (liquid cooling), and both companies invite wider participation in the OCP work group.
Abstract
Bijan Nowroozi, Head of Ecosystem Development at Lightmatter, and Janos Veres, Senior Director, Assembly & Test Technologies at Flex, present a collaborative 1RU, 200-terabit optical switch developed through their OCP workstream that advances AI networking infrastructure with high-density photonics and modular architecture. The project has evolved from a white paper to developing prescriptive specifications for fiber connectors and critical parameters, with the current prototype serving as a test vehicle to build supply chain understanding among partners including Lightmatter, Flex, Senko, and JetCool.
Lumentum

Slow and Wide Optical PCIe

  • Lumentum's 1060-nanometer VCSEL builds on its backside-emitting platform, already shipping in high volume for 3D sensing.
  • A demo with Astera Labs and Corning extends optical PCIe Gen 5 and Gen 6 across multiple racks or rows.
  • Lumentum argues a "slow and wide" approach to scale-up connectivity gives the best balance of performance and reliability.
Abstract
Christian Urricariet, Senior Director, Product and Strategic Marketing at Lumentum, discusses the industry's growing focus on scale-up connectivity for AI accelerators at the OCP Global Summit 2026, highlighting the company's 1060-nanometer VCSEL technology built on its high-volume backside-emitting platform. He describes a demonstration with Astera Labs and Corning of extended-reach optical PCIe connectivity for Gen 5 and Gen 6 that spans multiple racks or rows, using a "slow and wide" approach that balances performance and reliability.
Marvell

2nd-Gen CXL Cuts Time to First Token by 50%

  • Second-gen Structera X is the first CXL memory expansion controller with six DDR4 and six DDR5 channels, adding 50% more capacity and bandwidth.
  • Four-host connectivity, 5x better LZ4 compression, and built-in post-quantum cryptography raise utilization and security at rack level.
  • By easing the AI memory wall, the new generation cuts time to first token by 50% over its predecessor.
Abstract
Khurram Malik, AVP Custom Cloud Solution Group at Marvell, introduces the second-generation Structera X CXL memory expansion controller, the industry's first with six channels each of DDR4 and DDR5, delivering 50% more memory capacity and bandwidth, four-host connectivity, 5x enhanced LZ4 compression, and integrated post-quantum cryptography. He positions CXL as a solution to the AI memory wall, citing a 50% improvement in time to first token, and highlights how Marvell's portfolio of Structera X, Structera A, and a photonic fabric appliance enables rack-scale memory pooling and sharing with XPUs.
Marvell

Introducing OCP's Project Tessera

  • With AI agents outnumbering employees 100 to one, cloud security needs a fundamental rethink.
  • Tessera is a new OCP initiative from Microsoft, Marvell, and Oracle, with Marvell's LiquidSecurity HSMs exchanging keys with Tessera on compute nodes.
  • Both sides use NIST FIPS 140-3 Level 3 certified modules — the first step in a multi-year effort to secure AI and confidential computing.
Abstract
Bill Hagerstrand, Director, Security Business at Marvell, explains that AI agents now outnumber employees 100 to one, which requires a fundamental rethinking of cloud security, and introduces Tessera, a new Open Compute Project initiative to address this challenge. He describes how Marvell's LiquidSecurity HSMs, deployed across the top hyperscalers, securely exchange keys with Tessera on compute nodes using NIST FIPS 140-3 Level 3 certified modules, a first step in a multi-year effort to secure AI and confidential computing workloads.
Molex

Copper vs. Fiber in AI Scale-Up: Why Optics Are Moving Inside the Box

  • Larger AI scale-up clusters push optics into shorter parts of the channel that copper has traditionally owned, moving the fiber-versus-copper decision off the front panel.
  • Near-package optics keep a pluggable, modular approach while avoiding the challenges of co-packaged optics on the substrate.
  • Molex's Mirror Mezz Enhanced, Gemini and other mezzanine connectors scale from 224G toward 448G, alongside its second-source Si-Fly HD for the OpenCPX MSA.
Abstract
Bill Wilson, Group Product Manager at Molex, explains at the OCP Global Summit how AI scale-up is reshaping the fiber-versus-copper decision, as larger clusters push optics into shorter channel segments that copper has traditionally dominated, and highlights industry efforts such as OCP's Short Reach Optical Interconnect initiative, OIF work, and the OpenCPX MSA. He showcases Molex copper solutions that enable pluggable near-package optics modules, including Mirror Mezz Enhanced, Gemini, and other mezzanine connectors scaling from 224G toward 448G, along with the second-source Si-Fly HD product supporting the OpenCPX MSA.
MultiLane

One Faulty Connection Can Stall GPU Revenue

  • MultiLane, part of the Teradyne group, offers wafer-to-rack test aimed at reducing the risk of delayed GPU revenue.
  • Every AI system element – ASIC, cartridge, cable harness, GPU and switch trays – has hundreds of connections, and one fault can stop the whole system from shipping.
  • Its backplane cable cartridge tester uses an Amphenol connector to verify every GPU-to-GPU pin at a functional data rate.
Abstract
Hani Daou, VP Sales and Business Development at MultiLane Test Products, speaks from the Open Compute Project Global Summit about how the company, part of the Teradyne group, delivers a wafer-to-rack test solution that reduces the risk of delayed GPU revenue, since a single faulty connection among the hundreds in every AI system element can stop an entire system from shipping. Daou highlights the backplane cable cartridge tester, which uses an Amphenol connector to interface directly with backplane cartridges and cable harnesses and verifies every pin supporting GPU-to-GPU connectivity at a functional data rate.
NVIDIA

NVIDIA's Power Efficiency Leap: CPO & 800V Racks

  • NVIDIA optimizes for tokens per watt, using liquid cooling and 800-volt racks to raise compute density and steer more power to GPUs.
  • NVIDIA Photonics co-packaged optics is shipping, improving scale-out network power efficiency by nearly 5x and resiliency by 10x.
  • Future scale-up generations will use CPO beyond the rack, while the new USFP pluggable, built with Cisco and Amphenol, carries 12.8 Tb/s in one module.
Abstract
Gilad Shainer, SVP of Networking at NVIDIA, presents the company's power efficiency optimization strategy through tokens per power, liquid cooling, and 800-volt racks that enable higher compute density. He highlights NVIDIA's co-packaged optics technology, already in production, which delivers nearly 5x better power efficiency and 10x improved resiliency, along with the new USFP pluggable optical module developed with Cisco and Amphenol that achieves 12.8 terabit per second in a single module.
Rebellions

Efficient AI Inference at Rack Scale

  • Rebellions builds inference infrastructure end to end, pairing its in-house Rebel 100 accelerator with a software stack tuned for serving tokens economically.
  • An open-source, no-fork software strategy lets customers keep their existing tools and models without learning new skills.
  • Because inference is a rack-scale problem, most customers deploy Rebel Rack or Rebel Pod rather than individual servers.
Abstract
Jennifer Glore, EVP of Product Management at Rebellions, explains from the OCP event in San Jose how the inference infrastructure provider optimizes both hardware and software, pairing its power-efficient Rebel 100 accelerator with an open-source, no-fork software stack that lets customers run their existing tools and models. She describes how Rebellions packages these components into complete solutions, including Rebel Server, Rebel Rack, and Rebel Pod, with most customers choosing rack-scale deployments as global adoption grows.
Solutional

Open Infrastructure, Neoclouds & Supply Chain

  • AI data centers are in an "IBM moment" led by a few integrated players, and OCP gives momentum to those seeking open, interoperable alternatives.
  • Robohn's top short-term wish is better delivery timelines, backed by more US silicon production and manufacturing capacity.
  • Neoclouds are not monolithic – large providers like CoreWeave and Crusoe face different multi-tenancy networking needs than smaller players like FarmGPU.
Abstract
At the OCP Summit 2026 in San Jose, Scott Robohn, CTO of Solutional, describes the AI data center market as an "IBM moment" dominated by a few players, with OCP driving momentum toward open alternatives and interoperability across every layer of the stack. He calls for better delivery timelines and more US silicon and manufacturing capacity, and explains how Solutional helps neocloud providers of varying sizes address multi-tenancy networking challenges.
Upscale

Uniting Scale-Up and Scale-Out in One Fabric

  • Token Fabric is Upscale's open standards-based fabric that unites scale-up and scale-out into one operational domain with a single software and operating model.
  • Unifying the fabric gives customers more control, visibility and resiliency, tackling the network bottlenecks that keep AI accelerators below peak utilization.
  • No single company can build the future of AI alone — Upscale partners with NVIDIA on scale-out and works across ASIC and optics form factors to push heterogeneity from silicon to software.
Abstract
Aravind Srikumar, SVP Product, Marketing & Customer Experience at Upscale, introduces Token Fabric, an open standards-based fabric that unifies scale-up and scale-out domains under one software and operating model to improve control, visibility, resiliency, and AI accelerator utilization, and unveils SkyFabriX with 115 Tbps of throughput today. He calls for industry collaboration and standards-based interoperability from silicon to software, highlighting Upscale's partnership with NVIDIA on scale-out and its role as a strategic engineering partner across ASIC and optics form factors to advance heterogeneous AI.
VIAVI

Why Test & Measurement Drives AI Job Performance

  • The industry is entering the steep part of the AI "hockey stick," and test and measurement is now seen as critical to scaling it.
  • VIAVI covers every layer from fiber, optics and transceivers through signal integrity and protocol compliance to the AI switching fabric.
  • Its portfolio spans PCIe compliance, Ethernet transceiver testers, AI workload emulation and optics manufacturing test.
Abstract
At the Open Compute Global Summit 2026, Kevin Chang, Engineering Director, Hardware Platforms at VIAVI, explains that the industry is entering the steep part of the AI growth "hockey stick" and that test and measurement is now recognized as critical to scaling AI infrastructure, from the physical plant to AI job performance. He describes how VIAVI addresses layers ranging from fiber interconnects, transceivers, and signal integrity to the AI switching fabric, with product lines spanning PCIe protocol compliance, Ethernet transceiver testers, AI workload emulation, and optics manufacturing testing.
Voyant Photonics

From LiDAR to Optical Circuit Switching

  • Voyant is bringing seven years of dense, high-yield silicon photonic switch work from LiDAR to optical circuit switching.
  • AI scale-up architectures need OCS with a very small form factor, fast switching and low cost – where Nouvel says Voyant has a head start.
  • Voyant sells engines and modules to OCS vendors chasing hyperscalers, positioning it for rapid scale-up deployments expected around 2029.
Abstract
Clement Nouvel, CEO of Voyant Photonics, explains how the silicon photonics startup is applying seven years of experience building dense, high-yield silicon photonic switches for LiDAR to optical circuit switching (OCS) for AI scale-up architectures that demand small form factors, fast switching, and low cost. Voyant Photonics supplies engines and modules to OCS vendors competing for hyperscaler business, a model Nouvel expects will position the company for the rapid scale-up deployments he anticipates around 2029.
Xscape Photonics

Solving the Data Center Bandwidth Bottleneck

  • Xscape's multicolor laser platform generates up to 100 wavelengths from a single piece of silicon to ease the data center escape bandwidth bottleneck.
  • Eagle X, its first scale-up product, carries 16 wavelengths on one fiber and has more than 300,000 device hours of reliability data.
  • Falcon X, billed as the first OCI MSA-compliant pluggable laser module, is shown driving a 50G NRZ compliant link.
Abstract
Vivek Raghunathan, Co-Founder and CEO of Xscape Photonics, demonstrates the company's multicolor laser platform at partner Keysight's booth, showing how up to 100 wavelengths generated from a single piece of silicon address the escape bandwidth bottleneck in data centers, including a 64-wavelength comb device sampling for memory cluster applications. He also covers Eagle X, a 16-wavelength scale-up product with more than 300,000 device hours of reliability data, and Falcon X, described as the industry's first OCI MSA-compliant pluggable laser module, which is shown delivering 50G NRZ eye-diagram performance in a compliant link.
Xsight Labs

Xsight's 200G DPU for Edge AI & Space Computing

  • The new E1L is a 200G DPU for edge environments, built on Arm Neoverse N2 cores and compatible with the existing E-series line.
  • Its small power budget lets it run agentic AI workloads at the edge by offloading the CPU in power-constrained settings.
  • SpaceX deploys Xsight's X2 programmable switches on V3 satellites, where remote programmability is essential.
Abstract
Jillian Kaplan, Senior Director of Product and Marketing at Xsight Labs, announces the E1L, a 200 gig DPU designed for edge environments including space applications, featuring ARM Neoverse N2 cores and a small power budget that enables Agentic AI workloads in power-constrained settings. The late-stage silicon startup serves customers like SpaceX with its E-series DPU and X-series programmable switch lines, where remote programmability is essential for equipment deployed in physically inaccessible locations such as satellites.
Reader feedback on Data Center Networking in 2026 — Scaling to New Heights: "Appreciate the broad coverage and unique insights." "Up, Out, Across, and Away we go." Download our latest report and send us your feedback!
Download the Latest Report

By downloading this report, you agree to our Terms of Use and Privacy Policy.

Continue exploring

More executive showcases across next-gen infrastructure.

2025

2025 OCP Global Summit

Exclusive videos from thought leaders covering data center networking (scale-up, scale-out, scale-across), photonics, liquid cooling, SONiC

20 video interviews
2026

2026 Technology Exploration Forum

Highlights from Ethernet Alliance's 2026 TEF. Topics: 224/448 Gbps/lane, 1.6/3.2 Tbps, AI fabric, InfiniBand, UCIe, CXL, PCIe Gen 6/7

20 video interviews
2026

2026 AI Infra Summit

Executives at AI Infra Summit 2026 in Santa Clara discuss full-stack AI infrastructure — chips, interconnects, networking, memory, storage, power, and the data centers behind AI training and inference

30 video interviews
Browse all topics →