Daily industry note
The 1.6T race is a 2027 story. 800G cash, scarce optics, and MRC already in production will decide the next two quarters
Arista just printed its first $3 billion quarter on Etherlink that is still almost all 800G. Cisco booked $9.3 billion of hyperscaler AI-infrastructure orders and expects $7.5 billion of that book as FY27 revenue. LightCounting, via IEEE ComSoc, says datacom optics demand is about 30 percent above supply, with indium-phosphide EML lasers as the bottleneck. NVIDIA restated Spectrum-X Ethernet Photonics as in production. Against that, HPE’s June claim of shipping a liquid-cooled 1.6T QFX5250 is a real first-to-market window — but there is no public air-cooled 64×1.6T SKU, no independent volume proof, and Multipath Reliable Connection on Junos is still marked FRS+.
Competitor moves
What changed
These are the items that are not already baked into the June Discover / OFC / GTC Taipei cycle.
Arista posted its first $3 billion quarter. The 4 August Q2 2026 print was $3.036 billion, up 37.7 percent year on year and 12.1 percent quarter on quarter. The press release pointed at the June 7060XE7 1.6T Etherlink launch and at open MRC plus SRv6. Call color, via Converge Digest — treat as management commentary, not audited fact — said Etherlink now has more than 100 cumulative customers, versus four or five discussed in 2024. 1.6T systems enter trials with a single-digit number of large customers in the second half of 2026. Volume production is expected in 2027. Optics constraints are expected through 2028. Arista is converting 800G mindshare into cash now. 1.6T is still a 2027 volume story.
Cisco printed the AI-infra order book. On 12 August, for the period ended 25 July, Cisco reported Q4 revenue of $17.3 billion and FY26 revenue of $63.3 billion. Hyperscaler AI-infrastructure orders were $4 billion in Q4 and $9.3 billion for the year. About $4 billion of that book became FY26 revenue. Cisco expects $7.5 billion in FY27. Secondary call color (Fierce Network / Converge Digest, attributed to Robbins) said about 60 percent of those AI orders were Silicon One systems and the rest optics; Acacia AI-infra bookings were more than $1 billion in Q4; three Q4 hyperscale design wins spanned Silicon One P200 (scale-across), G200 (scale-out), and optical line systems.
Optics, not switch silicon, is the binding constraint. IEEE ComSoc summarized LightCounting’s August 2026 report on 14 August: datacom optics demand is outstripping supply by about 30 percent; transceiver plus hardware demand is projected up 53 percent to $39 billion; the bottleneck is InP epitaxial / EML laser capacity, with MOCVD tools back-ordered. Hyperscalers are expanding qualified vendor pools from two or three to five to seven. Every TH6 1.6T box — QFX5250, Arista 7060XE7, Nokia H6, DriveNets 2600, Celestica DS6000, Edgecore AIS1600 — collides on the same OSFP / ORHS / EML supply. The winner is whoever has qualified module SKUs and allocation, not whoever announced the switch first.
NVIDIA restated Spectrum-X Ethernet Photonics as full production. The original production claim is 31 May 2026 at GTC Taipei. Mid-August trade coverage (StorageReview, 15 August) restated it. NVIDIA’s own claims — label them as vendor claims — include CPO on 200G SerDes, 5× better network power efficiency versus pluggable transceivers, Spectrum-X Ethernet Photonics switches available in the second half of 2026, and chassis up to 409.6 Tb/s (SN6800 class). Spectrum-6 is a 102.4 Tbps chip pairing with ConnectX-9, BlueField-4, and Vera/Rubin. NVIDIA claims up to 1.6× higher AI networking performance than off-the-shelf Ethernet and up to 95 percent network efficiency above 100k GPUs. This is the closed Ethernet stack HPE already resells inside HPE AI Factory. NVIDIA is the only vendor with a named, in-production CPO Ethernet switch.
DriveNets TH6 platforms are due to ship this quarter. Announced 1 July; ship window is Q3 2026, which is now. DriveNets 2600SL is 100 percent liquid-cooled; 2601S is air. Broadcom TH6, 102.4 Tbps, 64 × 1.6T. The software story is a full stack — NIC driver, kernel, CCL, plus Cluster Orchestrator — not just a NOS. Another 100 percent DLC TH6 box is competing for the same neo-cloud RFPs as QFX5250-64OE-L, and DriveNets already has an air SKU.
UEC ticked to v1.0.3. The specification is dated 16 July 2026 and sits in the consortium’s August folder. Prior: 1.0.2 on 28 January 2026; 1.0 launched 11 June 2025. “UET-ready” on a datasheet is a moving target. Interop and compliance matter more than a logo.
Inference and frontend are becoming half the switch growth story. Dell’Oro, via the same 14 August ComSoc note (Sameh Boujelbene): more than half of the growth in front-end data-center switch sales over the next few years will be driven by AI-related opportunities. Named beneficiaries include Accton, Arista, Celestica, Cisco, HPE/Juniper, H3C, Huawei, and NVIDIA. Discover’s QFX5140 plus MX301 plus CX-into-Mist is the right shape. The risk is that Arista and Cisco take the easy frontend and campus money while HPE over-indexes on backend 1.6T.
The Next Platform, on 17 August, put a finer point on Cisco: it is selling AI systems plus scarce 400G and 800G optics, not just Nexus. Switch share without module allocation is incomplete.
Optics and speeds
LPO, CPO, and the next lane
Industry consensus from earnings and OFC — directional, not a standard.
| Speed | Where it is |
|---|---|
| 400G | Mature |
| 800G | Volume cycle 2026–2027. This is the production AI port. |
| 1.6T | Systems sampling or early-orderable in 2026. Production scale 2027. Arista: trials H2 2026, volume 2027. |
| 3.2T | Launch talk 2028. Volume 2029–30. |
SerDes. 112G PAM4 is production everywhere (400G / 800G). 224G / 212.5G PAM4 — “200G per lane,” so 1.6T equals 8 × 224G — is in production in TH6 Condor (BCM78914), Cisco G300, Spectrum-6, and NVIDIA CPO. OIF CEI-224G, including CEI-224G-Linear for LPO / CPO / NPO, is active work. 448G is research and IA start: OIF CEI-448G-VSR, a live electrical demo at ECOC 2025, a Semtech 3.2T ACC demo at OFC 2026. Demo, not a product.
Pluggable versus integrated, 2026 reality.
| Architecture | What it is | 2026 reality |
|---|---|---|
| FRO | Fully retimed; DSP in the module | Default 800G. Volume, power-hungry, InP/EML constrained. |
| LRO | Linear-retimed; partial DSP | Shown in multi-vendor 1.6T demos. |
| LPO | No module DSP; host SerDes equalizes | The 2026 power story. Arista claims ~60% interconnect power cut. Cisco claims 50% module / 30% switch. Needs TH6 / G300 / Spectrum-6 analog front end. No public LPO 1.6T claim found for QFX. |
| XPO / NPO | Near-package / external-laser variants | Transition architectures. |
| CPO | Optical engine on the switch package | NVIDIA Spectrum-X / Quantum-X Photonics in production (vendor claim). Broadcom TH6 is CPO-capable. Arista: open/socketed CPO; copper + pluggable dominant for several years; broader CPO ~2028–29. |
Intra-DC stays IM-DD PAM4 (DR / FR / VR) at 800G and 1.6T. Coherent ZR / ZR+ is DCI, scale-across, and campus-long — not the GPU rail. Cisco leans on Acacia here. Juniper PTX and QFX5230 already do ZR / ZR-M.
OFC 2026 (15–19 March) takeaway that still holds: 800G is what is being installed. 1.6T is what is being qualified. CPO is what NVIDIA is shipping to its own stack.
Silicon
Who ships the switch brain
Vendor claims labeled as such. No invented node or ship dates.
| ASIC | Vendor | Bandwidth | Status as of 18 Aug 2026 |
|---|---|---|---|
| Tomahawk 5 (BCM78900) | Broadcom | 51.2T | Volume. Current AI 800G workhorse. QFX5240, Arista Etherlink, Nokia H5. TSMC 5 nm in merchant summaries. |
| Tomahawk 6 (BCM78910 / 78914) | Broadcom | 102.4T | Production volume 12 March 2026. 3 nm. 1024 × 100G or 512 × 200G. ~267 MB shared on HPE / Edgecore system docs. |
| Trident 5 (BCM78800) | Broadcom | 16T (QFX5140) | Shipping in the 1RU inference SKU. |
| Silicon One G200 | Cisco | 51.2T | Shipping. Cisco says programmed for UEC after tape-out. |
| Silicon One G300 | Cisco | 102.4T | Announced 10 February 2026. 512 × 224G LR. 252 MB fully shared (Cisco / reseller). First systems targeted 2H26. Customer ship as of today: not verified. |
| Silicon One P200 | Cisco | 51.2T | Deep-buffer scale-across / DCI. Shipping and expanding. |
| Spectrum-6 | NVIDIA | 102.4T | Arriving with Vera Rubin. CPO option. Vendor: H2 2026 availability. Spectrum-4 remains the 800G Spectrum-X chip in the field. |
| Express 5 | Juniper | 28.8T (BXX) | Shipping on PTX10002-36QDD and PTX10k LC1301. 7 nm at 2022/23 launch. No 2026 node update found. HBM. 36 × 800G, inline MACsec. |
| Trio 6 | Juniper | 1.6T device / 9.6T LC | Shipping on MX301 and MX10K. 7 nm at launch. |
| FP5 | Nokia | 6.0 Tb/s NPU | Shipping (2022+). No FP6 found. |
| Teralynx 10 | Marvell | 51.2T | Volume since July 2024. A 102.4T “T100 sampling June 2026” line appeared only in secondary synthesis — not verified on Marvell.com. |
Scheduled Ethernet and packet spraying are software plus ASIC tables — DLB / WPS / GLB (HPE), Cognitive Routing (Broadcom / Arista), path-based LB (Cisco G300), Spectrum-X adaptive routing, MRC entropy spraying. Not a separate chip category.
Juniper “Triton” was not found as a current shipping ASIC name. Custom routing chips are Express 5 and Trio 6. QFX AI boxes are merchant Broadcom.
AI Ethernet
RoCEv2, MRC, UEC
MRC means Multipath Reliable Connection. It is an OCP-published transport that extends RoCEv2 Reliable Connection. It is not a marketing synonym for “multipath RoCE.”
OpenAI’s May 2026 note, the OCP MRC 1.0 spec, Microsoft Azure HPC, and AMD all describe the same thing: per-packet entropy spraying across hundreds of paths and multi-plane Clos; out-of-order delivery; SACK / NACK plus packet trimming; path retirement and probes on loss; NSCC (UEC sender-based congestion control; AMD says it authored NSCC / SMaRTT); optional SRv6 so the endpoint can bypass a failed path without a dynamic control plane.
OpenAI’s production claim: already running on its largest NVIDIA GB200 supercomputers, including OCI Abilene and Microsoft Fairwater, and used to train multiple OpenAI models. Partners named by OpenAI: AMD, Broadcom, Intel, Microsoft, NVIDIA. Deployment partners: Microsoft Azure, OCI, NVIDIA, Arista.
HPE status: QFX5250 software lists “MRC with SRv6 uSID” as FRS+, not Day-1. That is the most important software gap versus Arista, which is already talking MRC on the earnings stage.
RoCEv2 plus DCQCN is still the default production Ethernet AI transport in 2026 for anyone not on Spectrum-X or MRC. Every vendor in this note ships PFC, ECN, and DCQCN. Nokia explicitly tested RoCEv2 alongside UET.
UET-ready is not UET deployed. UEC 1.0 published 11 June 2025; 1.0.3 is dated 16 July 2026. Hardware is being marked UET-ready. End-to-end production UET clusters were not documented as the default in the sources reviewed. Nokia and Keysight ran an 800GE UET plus RoCEv2 coexistence lab test. Cisco says G200 was programmed for UEC 1.0 after tape-out. HPE QFX5250 lists trimming, UET forwarding, and CSIG; LLR is FRS+.
Spectrum-X versus InfiniBand. Spectrum-X is a closed NVIDIA Ethernet loop: Spectrum switch, ConnectX SuperNIC, NVIDIA congestion control, adaptive routing, packet spray, NIC reordering. Marketed as 1.6× versus “off-the-shelf Ethernet” and about 95 percent efficiency — NVIDIA marketing; no third-party cluster paper was found in this pass. InfiniBand (Quantum-X / Quantum-X800) is still the reference fabric for tightly coupled, SHARP-offload, SuperPOD-class training. Hybrid IB backend plus Ethernet frontend remains common. Open Ethernet (UEC / MRC) is the multi-vendor alternative. It is not a 2026 default replacement for IB on the largest NVIDIA-designed pods.
| Stack | Evidence of production use |
|---|---|
| InfiniBand NDR / XDR | NVIDIA SuperPOD references; widespread OEM |
| Spectrum-X Ethernet | NVIDIA and OEM AI factories; xAI Colossus cited third-party |
| RoCEv2 + DCQCN on merchant Ethernet | Default enterprise / neo-cloud Ethernet AI |
| MRC over Ethernet + SRv6 multi-plane | OpenAI + Microsoft Fairwater + OCI Abilene (primary) |
| UET 1.0 | Lab and multi-vendor tests; “ready” SKUs. Not shown as the majority training fabric |
Competitive desk
HPE Networking versus the field
QFX5240 is the shippable 800G SKU. 64 × 800GbE, Tomahawk 5, 51.2 / 102.4 Tbps, 700–750 ns store-and-forward on the datasheet, 2U air (ORv3 DC SKUs exist). Hardware parity with Arista’s 800G Etherlink. The gap is mindshare and a named titan install base.
QFX5250-64OE-L is a vendor-claimed shipping liquid 1.6T. 64 × OSFP-RHS 1.6TbE, Tomahawk 6, 102.4 / 204.8 Tbps, 100 percent liquid, 2OU ORv3, 48 V busbar, fanless. Typical / max power 3375 / 3435 W loaded with DR optics. Junos Evolved 25.2X100-D20. RoCEv2, DCQCN, DLB / GLB / RLB, WPS, packet trimming, UET forwarding, CSIG. MRC + SRv6 uSID is FRS+. The 17 June TechPost says HPE Juniper “has started shipping” and calls it industry first. That is a first-party claim. Independent volume-shipment evidence was not found. There is no public air-cooled 64 × 1.6T QFX.
Arista’s 7060XE7: air 64 × 1.6T in Q4 2026, liquid 2OU ORv3 in Q1 2027, 128 × 800G air in Q1 2027. Nokia 7220 IXR-H6-64 is a 3RU air plus DLC 64 × 1.6T (available 1Q26 per the November 2025 launch; silicon is highly likely TH6 but Nokia pages fetched do not print “Tomahawk 6”). DriveNets 2601S air ships Q3. Celestica DS6000 has been orderable since 29 April. A liquid-only ORv3 SKU concedes every air-cooled hall and every non-OCP rack.
No named titan win was found for HPE in this cycle. Arista’s 9 June 1.6T PR carried Meta, Microsoft (Maia / Fairwater), and Oracle quotes. Cisco has a $9.3 billion FY26 AI-infra order number and three Q4 design wins. HPE’s June Discover PR did not put an equivalent named AI-fabric win on the page.
PTX Express 5 and MX301 are under-sold as scale-across and inference edge. PTX10002-36QDD is 36 × 800GbE with inline MACsec — stronger WAN / DCI silicon than Arista is using to sell the 7800 as scale-across, and the counterpart to Cisco’s P200 motion. MX301 is 1.6 Tbps of Trio 6 in 1RU for inference-edge and metro. Dell’Oro says more than half of frontend switch growth is AI. Those boxes exist. They are not packaged the way Cisco sells Silicon One plus Acacia as one AI-infra PO.
| HPE Juniper | Arista | Cisco | NVIDIA | |
|---|---|---|---|---|
| Shippable AI port today | QFX5240 64×800G | Etherlink 800G, 100+ customers (call) | Nexus G200 + N9100 Spectrum-X | Spectrum-X 800G widely deployed |
| 1.6T system | QFX5250-64OE-L liquid ORv3 — vendor: shipping. No public air 64×1.6T | 7060XE7 air Q4’26, liquid Q1’27 | N9364F-SG3 (G300), 2H26 target | Spectrum-X Photonics CPO, H2’26 availability (vendor) |
| Transport proof | MRC + SRv6 FRS+; UET features on the list | MRC on the earnings stage; OpenAI lists Arista as deploy partner | UEC-ready G200 / G300 (Cisco claim) | Closed Spectrum-X stack; IB still SuperPOD default |
| Optics | ORHS liquid list; no captive DSP; no public LPO 1.6T claim | LPO ~60% interconnect claim on XE7 | Acacia; >$1B Q4 AI optics bookings (call) | CPO in production (vendor) |
| Public AI proof | Discover + TechPost. No named 1.6T titan win found | Meta, Microsoft, Oracle quotes; 100+ Etherlink | $9.3B FY26 orders; three Q4 design wins | Spectrum-X in production AI factories |
| Routing / DCI | Express 5 36×800G MACsec; Trio 6 MX301 | 7800 as scale-across spine | P200 51.2T deep buffer + 800G ZR/ZR+ | Not a WAN vendor |
HPE SKUs that matter this cycle:
| Platform | Silicon | Ports | Role | Caution |
|---|---|---|---|---|
| QFX5240 / 5241 | TH5 | 64 × 800GbE | Production AI leaf / spine | This is the shippable SKU |
| QFX5250-64OE-L | TH6 | 64 × 1.6T OSFP-RHS | Scale-out leaf / spine, 100% DLC | Vendor shipping claim; no independent volume proof; MRC is FRS+ |
| QFX5252 | Dual TH6 per ServeTheHome | Helios scale-up tray | UALoE + SONiC (STH) | Not confirmed on an HPE datasheet |
| QFX5140-24CD8O | Trident 5 | 24 × 400G + 8 × 800G | Inference / edge / border-leaf | Discover 16 June |
| PTX10002-36QDD | Express 5 | 36 × 800GbE, MACsec | Core / peering / DCI / scale-across | Under-sold as the AI scale-across box |
| MX301 | Trio 6 | 1.6 Tbps, 400G | Inference edge / metro | Under-sold as the inference-edge box |
QFX5250 FRS optics (TechPost, 17 June) are OSFP-RHS riding-heat-sink parts, not standard air-cooled OSFP: ORHS-1600G-DR8-2-P, ORHS-1600G-DR8-2, ORHS-2x800G-FR4-P, ORHS-800G-DR8-2-P, ORHS-800G-VR8-P. Differentiator only if those SKUs are actually orderable.
For product managers
Where to be first
1. Ship an air-cooled 64 × 1.6T QFX in calendar 2026
Arista’s 7060XE7-64PS is Q4 2026 air. Nokia H6-64 is already a 3RU air box. DriveNets 2601S ships this quarter. A liquid-only ORv3 SKU concedes most enterprise and neo-cloud halls. First on DLC is real. First on the port customers can actually buy is air plus 800G breakout.
2. Make MRC + SRv6 a Day-1 feature, and put it on the box in Apstra
OpenAI, Microsoft, AMD, and Arista have already made MRC the public language of Ethernet training fabrics. Junos already lists it on QFX5250 and QFX5240. If it stays FRS+, HPE will be explaining “UET-ready” while competitors demo the protocol that is already training models in Abilene and Fairwater.
3. Publish one named, measured AI-fabric win before Arista’s Q4 1.6T trials
Not GPU-utilization copy. A titan, neo-cloud, or sovereign with GPU count, rail design, QFX5240 versus 5250 mix, JCT or NCCL allreduce versus the incumbent (Spectrum-X or Arista 800G), and the optics used. Arista’s June PR had Meta, Microsoft, and Oracle on the page. HPE’s June Discover PR did not.
4. Treat optics allocation as a product, not a commodity
Expand qualified 800G / 1.6T module vendors from the usual two or three to five to seven, including ORHS and standard OSFP224. Publicly qualify LPO on TH6 QFX — Arista and Cisco already claim 50–60 percent module-power cuts. Do not wait for CPO; NVIDIA owns that narrative through 2027. PTX / QFX ZR inventory is the scale-across analog of Cisco’s Acacia print.
5. Catalog a dual-stack AI Factory item: QFX Ethernet or Spectrum-X
HPE already resells Spectrum-X (December 2025 PR). Fighting NVIDIA on GB200 / Vera Rubin scale-out is a losing motion. Winning motions: AMD Helios plus QFX5252 / UALoE; multi-vendor Ethernet plus MRC; Spectrum-X when the customer is buying NVIDIA end-to-end. One Data Center Director / GreenLake motion. Make the choice a catalog item, not a services project.
6. Package inference and scale-across, not just backend 1.6T
Ship QFX5140 + MX301 + Apstra + Mist as a named inference-edge bundle — already the Discover shape. Sell PTX Express 5 800G MACsec as the AI scale-across / multi-factory DCI box with the same MRC / SRv6 story Arista is using on the 7800. That is how HPE uses routing silicon Cisco and Arista do not have in the same way.
Open questions
Could not verify
Independent evidence that QFX5250-64OE-L is in customer production. Official QFX5252 datasheet. A Juniper “Triton” ASIC (not found; current custom chips are Express 5 and Trio 6). A Nokia product named “SRE.” Nokia H6 quoted as Tomahawk 6. Marvell Teralynx T100 on Marvell.com. Any Cisco G300 N9364F-SG3 customer ship as of 18 August. NVIDIA CPO metric wording (5× power vs 10× MTBI vs 5× runtime). UEC 1.0.3 changelog. An HPE AI-infra order number comparable to Cisco’s $9.3 billion. Spectrum-X 1.6× / 95 percent as a third-party measurement. LightCounting’s ~30 percent shortage in the underlying paid report.