Table of Contents
Microsoft: Maia’s 2027 deployment target and the infrastructure around custom silicon
Reynold D’Sa (Corporate VP, Silicon, Cloud Hardware & Infrastructure Engineering, Microsoft); opening remarks from George Tchaparian (CEO, Open Compute Project).Meta: NPO pilots start in scale-out, not wholesale replacement of copper scale-up
Ge Nan (Technical Sourcing Manager, Meta).NVIDIA Networking: CPO, denser pluggables and faster recovery from network failures
Gilad Shainer (SVP, Networking, NVIDIA), including the follow-on Q&A.LightCounting: Strong optical growth, but fewer transceivers per GPU complicate the extrapolation
Vlad Kozlov (Founder & CEO, LightCounting).Intel: Model routing, open-model substitution and a heterogeneous inference stack
Nic Dubé (Corporate VP, Data Center Systems & Solutions, Intel).Oracle Optics: Field reliability, laser placement and why DSP functions do not disappear
Kannan Raj (AI Infrastructure Architect, Oracle).Lumentum: The laser opportunity across CPO, NPO and optical I/O
Matt Sysak (CTO, Cloud & Networking Platform, Lumentum).Alibaba: Interposer-packaged optics as an alternative to tightly integrated CPO
Zhiping Yao (Principal Engineer, Alibaba).Astera Labs: Where copper, NPO and CPO fit as scale-up domains expand
Chris Blackburn (Director, Solutions Architecture, Astera Labs); Jignesh Shah (Senior Director, Product Marketing, Signal Connectivity Group, Astera Labs).Ciena: OpenCPX, longer-reach copper and the economics of denser switch racks
Peter Winzer (VP, Ciena).Lattice / AMI: More control silicon, security retrofits and autonomous infrastructure
Ford Tamer (CEO, Lattice Semiconductor); Sanjoy Maity (SVP & GM, AMI Business Unit, Lattice); Naga Gurumurthy (Microsoft; title not stated in the recording).Arista: A proposal to make AI campuses smaller, cheaper and faster to build
Andy Bechtolsheim (Co-founder & Chief Architect, Arista Networks).
1. Microsoft: Maia’s 2027 deployment target and the infrastructure around custom silicon
Speaker: Reynold D’Sa, Corporate VP, Silicon, Cloud Hardware & Infrastructure Engineering, Microsoft. Opening remarks: George Tchaparian, CEO, Open Compute Project.
Summary: Microsoft’s most consequential disclosure was a multi-gigawatt deployment target for Maia systems beginning in 2027. Maia 300 is expected to deliver up to 7x Maia 200’s rack-scale token throughput, rather than a 7x per-chip improvement. The broader presentation showed how Cobalt CPUs, DPUs, networking, security, power distribution and cooling support the custom-silicon deployment. The Official Microsoft Blog
1. Maia’s deployment scale matters more than the headline benchmark
D’Sa described Maia as moving beyond a standalone accelerator program toward a deployable infrastructure platform. Microsoft’s accompanying announcement says Maia is intended to become the primary inference system for its products and services as multi-gigawatt deployment begins in 2027.
For investors, the questions are how much capacity actually enters service, which workloads migrate and how consistently Microsoft can use that capacity. That is more consequential for GPU-versus-custom-silicon spending than an isolated benchmark. The deployment commitment remains forward-looking, and the up-to-7x comparison is against the preceding Maia rack configuration, not every competing GPU system. Source
2. Cobalt and Azure Boost address the work surrounding inference
Microsoft positioned Cobalt 200 around the CPU demands of agent execution and Azure Boost around infrastructure offload. These functions include moving data and handling storage, networking and security rather than generating every model token directly.
The architecture reinforces the CPU thesis from the first OCP batch: an agent can consume accelerator time while also generating substantial conventional compute and infrastructure work. Microsoft’s custom-silicon strategy therefore extends beyond replacing an external GPU purchase.
3. Networking and security are part of the same system design
Microsoft highlighted multipath networking as a way to connect more than 100,000 GPUs using a two-tier architecture while maintaining resilience. The objective is to avoid adding layers of switches, latency and failure exposure each time a cluster grows.
It also discussed integrated hardware security modules and the Tessera collaboration. The distinction is between Microsoft’s own hardware implementation and the broader open specification intended to make similar capabilities available across vendors.
4. Power and cooling determine how quickly the silicon becomes usable
Mount Diablo and higher-voltage power distribution separate power conversion from valuable compute space. Microsoft also described a third-generation heat-exchanger design delivering nearly twice the cooling capacity in a similar footprint.
The broader goal is a fleet that increasingly diagnoses and recovers from problems automatically. Tchaparian’s opening remarks framed OCP’s expansion around energy, open clusters and security, matching Microsoft’s argument that scaling AI requires repeatable infrastructure standards, not just more chips.
2. Meta: NPO pilots start in scale-out, not wholesale replacement of copper scale-up
Speaker: Ge Nan, Technical Sourcing Manager, Meta.
Summary: Meta’s most useful clarification was that its near-packaged optics pilot is aimed first at scale-out. Optical scale-up remains an additional opportunity, but the presentation did not establish a decision to replace copper throughout the rack. The sourcing priorities were power, system reliability and preservation of a competitive supplier ecosystem.
1. The pilot’s location changes the read-through
Meta described scale-out as the initial deployment environment for NPO, with existing pluggables and other optical approaches providing alternatives. That lets the company evaluate a new architecture without making the entire accelerator interconnect depend on it immediately.
For optical investors, a scale-out pilot is evidence of customer engagement, but not equivalent to an optical connection replacing every copper scale-up link. The size, timing and supplier content differ materially.
2. Power savings matter, but require a system-level comparison
One example compared an approximately 16W conventional 800G module with an approximately 10W NPO/CPO implementation. The important mechanism is shortening the electrical path and reducing the processing required between the switch and optical engine.
Those figures are an illustrative link-level comparison, not a reduction in total data-center power. Some electrical functions remain inside the ASIC or elsewhere in the system, so removing a module DSP does not eliminate every related power cost.
3. A module that passes its own specification is not enough
Meta emphasized the combined behavior of the switch, electrical channel, optical engine, connector and thermal environment. A component can meet its individual specification while the assembled system remains difficult to qualify or unreliable in operation.
That makes co-design, common testing and comparable telemetry important purchasing requirements. It also means qualification can take longer than a supplier’s successful demonstration suggests.
4. Supplier choice is part of the architecture
Meta wants to retain the competitive sourcing and replaceability that made pluggables successful. That is a meaningful argument for standardized NPO interfaces.
The counterweight is that greater integration can improve power and density. Meta’s presentation was therefore a practical evaluation of options, not an unconditional endorsement of one optical architecture across every network.
3. NVIDIA Networking: CPO, denser pluggables and faster recovery from network failures
Speaker: Gilad Shainer, SVP, Networking, NVIDIA. Includes the follow-on Q&A.
Summary: NVIDIA described CPO as shipping, with scale-out adopting it before broader optical scale-up. The important nuance came in the Q&A: NVIDIA continues investing in pluggables and leaves room for NPO where appropriate. Its differentiation argument combines optical integration with hardware-level load balancing, fault recovery and software across the network.
1. CPO adoption does not mean the end of pluggables
Shainer described CPO as particularly attractive inside the AI factory, where NVIDIA can optimize an optical technology for a defined set of distances and requirements. Longer-distance connections and connections to other infrastructure still require different optical options.
The new USFP form factor is designed for 400Gb/s electrical lanes, with a path to 6.4T and 12.8T module capacity and support for up to 100W. NVIDIA and Cisco are developing the ecosystem with TE Connectivity and Amphenol. The Q&A pointed toward demonstrations followed by production in 2027, not immediate broad availability. NVIDIA Investor Relations
2. The reliability claim is about the complete network
NVIDIA claimed up to a 5x improvement in optical power efficiency and roughly 10x better mean time between interruptions in the discussed architecture. Shainer paired CPO with multiplane networking rather than attributing the entire improvement to the optical package.
Component failure rates, link interruptions and job-level availability are different measurements. Better optics helps, but the system must also avoid letting one failed link disrupt a much larger accelerator pool.
3. Hardware recovery is intended to preserve usable capacity
NVIDIA described moving failure detection and load balancing into hardware, reducing response times from close to a second toward less than three milliseconds in its examples. A failed plane can then be bypassed while most available bandwidth remains usable.
The commercial objective is not merely a faster network under ideal conditions. It is maintaining productive GPU time when failures inevitably occur.
4. Networking content broadens beyond the accelerator fabric
Shainer described separate functions for scale-up, scale-out, scale-across, context storage and secure access. BlueField plays a role in the context-storage and security layers, including moving KV-cache processing away from host CPUs.
For scale-across, NVIDIA discussed connecting sites over tens of kilometers today and expanding to longer distances. Its approach emphasizes measured path characteristics and coordinated traffic distribution. That is one architectural answer, not proof that every competing buffered network design is obsolete.
4. LightCounting: Strong optical growth, but fewer transceivers per GPU complicate the extrapolation
Speaker: Vlad Kozlov, Founder & CEO, LightCounting.
Summary: LightCounting remained strongly constructive on optical demand, discussing approximately 120% market growth in 2026 and roughly 70–80% in 2027. The more differentiated warning was that network redesign can reduce transceiver intensity, so investors should not extrapolate today’s optical share of AI capex indefinitely. These are LightCounting forecasts, not customer purchasing commitments.
You are reading TMTB Pro. For more info or a trial, please reach out to sales@tmtbreakout.com