Brief · ENT-BDB-2026-02 · April 2026 Edition
Copper, optics, and hybrid paths for AI infrastructure 2026–2029.
A buyer decision brief on where capital should move in the AI data center interconnect layer. Written for corporate strategy, corporate development, product strategy, and infrastructure investment decision-makers evaluating how interconnect architecture shifts should translate into capital, roadmap, and partnership commitments through 2029.
How should a corporate strategy team, corporate development team, or infrastructure investor evaluate the emerging interconnect architecture shift, and how should that evaluation translate into capital, roadmap, or partnership decisions between now and 2029?
The question is no longer theoretical. Between late 2025 and April 2026, the interconnect layer moved from a background engineering issue to a foreground capital allocation category, with more than $9 billion in strategic and financial capital committed across copper extension, co-packaged optics, photonic fabric, and the optical supply ecosystem.
This brief provides a structured way to distinguish the architectural paths, decompose scale-up from scale-out exposure, and decide where capital should move before roadmap milestones and partnership commitments lock in.
This brief is written for buyers making capital, roadmap, and partnership decisions in the interconnect layer. It is not designed as chip-level engineering analysis, nor as a substitute for specialist physics, packaging, or device-level research.
The following is the unedited executive summary from the full brief. Additional preview chapters are available on request.
The AI infrastructure stack has reached a new binding constraint in 2025 and 2026: interconnect. At 224 Gbit/s signaling, copper interconnects consume roughly 30 percent of total cluster power, with effective reach limited to less than one meter without amplification. Nvidia itself has stated that the GB200 NVL72 rack, had it used optical transceivers and retimers instead of direct-attach copper cables, would have drawn an additional 20 kilowatts per rack — 120 kW instead of 100 kW. That single data point reframes the interconnect debate: it is no longer an engineering preference, it is a capital and power allocation decision with measurable economic weight per rack.
Between December 2025 and April 2026, more than $9 billion in strategic and financial capital has moved into the interconnect layer. Kandou AI raised $225 million at a $400 million valuation for copper extension technology backed by SoftBank, Synopsys, Cadence, and Alchip. Ayar Labs raised $500 million at a $3.75 billion valuation for co-packaged optics, with Nvidia, AMD, MediaTek, and Qatar Investment Authority participating. Marvell completed its acquisition of Celestial AI for $3.25 billion upfront, with up to $5.5 billion in total consideration, for photonic fabric technology positioned at the memory disaggregation layer. Nvidia committed $4 billion across Lumentum and Coherent to secure laser and optical component capacity for future scale-up architectures. Credo agreed to acquire DustPhotonics for approximately $750 million in cash plus stock.
These are not four companies competing for the same dollar. They represent three architectural paths — copper extension, co-packaged optics, and photonic fabric — plus one enabling ecosystem layer that benefits from any of the optical paths. The commercial question for enterprise and investor buyers is not which company wins. It is: which path should my capital, my roadmap, and my partnership commitments align with, and at which point in the 2026–2029 window?
The dominant strategic mistake in 2026 is treating this as a single binary decision. Copper is not uniformly obsolete, and optics is not uniformly ready. Scale-up networking (tight, intra-rack, latency-sensitive) and scale-out networking (looser, inter-rack, bandwidth-dense) are distinct problems with distinct economics. They will tip toward optics on different timelines. Treating them as one decision is how enterprises and investors misallocate capital.
Bottom line: Scale-up and scale-out are not the same decision. The real market question is not copper versus optics in absolute terms — it is: on which layer does the economics tip first, and what does that mean for capital, roadmap, and partnership commitments through 2029?
The full brief includes the Interconnect Exposure Matrix and the CROSS path logic. The matrix below shows the six factors used to determine whether interconnect is a 2026 commitment or a 2027–2028 watch item.
Most interconnect commentary collapses the category into a single question: copper or optics. This brief addresses a different question first: where is the buyer actually exposed, on what timeline, and with what capital consequences if the wrong layer is prioritized first.
The purpose of the matrix is not engineering precision. It is capital-allocation clarity. It gives corporate strategy, corp dev, product strategy, and infrastructure investment teams a shared language for deciding whether the interconnect transition is an immediate archetype commitment or a monitored future move.
Is the buyer or customer base building or financing pods of 64+ GPUs where scale-up beyond a single rack is a near-term requirement?
Is the buyer operating under explicit per-rack, per-hall, or per-site power constraints that materially affect capacity planning?
Has the buyer or a key customer committed to a specific fabric — NVLink, UALink, Ethernet, or proprietary — that constrains the interconnect choice?
Is the scale-up layer a meaningful share of the buyer's strategic surface area, versus scale-out only?
Is the buyer aligned with the Rubin Ultra or Feynman window, where interconnect architecture decisions have committed deadlines?
Is the buyer exposed to supply concentration in lasers, EMLs, or wafer-level optics packaging?
In the full edition, these factors are combined with the CROSS path logic — Copper continuation, Retimer bridging, Optical pluggable, Scale-up CPO, Scaled photonic fabric — to produce an archetype-level capital allocation stance.
Why interconnect became the binding constraint in 2025–2026, with power, roadmap, and capital evidence.
The Interconnect Exposure Matrix and the CROSS path logic for translating architectural shifts into capital and roadmap decisions.
Copper extension, co-packaged optics, photonic fabric, and the incumbent optical ecosystem — with distinct positions, timelines, and risk structures.
Who owns this decision internally, and how it should flow across strategy, corp dev, product, and investment teams.
Decision matrix, 90-day interconnect sprint, and capital-allocation framing that survives committee review.
The strategic mistake to avoid, the Northfold buyer stance, and how to use the brief.
Plus appendices: Glossary, First 90-Day Position KPIs, Methodology and Sources.
Frames interconnect as a capital allocation and roadmap question rather than a generic technology debate. Distinguishes scale-up from scale-out, separates archetypes that solve different problems on different timelines, and gives buyers a structured way to align capital, roadmap, and partnership commitments through 2029.
It does not replace chip-level engineering research, photonic device analysis, or detailed implementation work on specific fabrics. It is not written for readers who only need vendor news aggregation without archetype-level decision logic.
That distinction is deliberate: most interconnect commentary is either technical or promotional. Fewer sources explain where capital should move and why scale-up and scale-out should not be collapsed into one decision.
If the objective is to gain optical-transition exposure without committing to one architectural winner too early.
If the buyer's exposure is mainly intra-rack scale-up and the economic floor of copper remains attractive through one more generation.
If the buyer is aligned with the Rubin Ultra or Feynman timeline and cannot avoid the move beyond rack-scale copper economics.
If the strategic horizon is 2028–2029+, and the buyer has explicit exposure to memory disaggregation or multi-rack coherent scale-up.
The single most expensive mistake is to treat them as one decision. Their economics tip on different timelines and through different vendor ecosystems.
This brief is available under Northfold's licensed Single User, Team, and Enterprise tiers, with optional Standard and Extended Calibration. Current market-specific pricing (EUR / GBP / CHF) is on the Pricing page.
Not sure whether the full brief or calibration is the better fit? Email us referencing ENT-BDB-2026-02 and we will indicate which format fits your situation.
B2B only; requests require confirmation that the requester acts in a commercial or professional capacity. Current market-specific pricing is on the Pricing page. Licensing terms are detailed in the Terms of Sale and Licence. Northfold Research publications do not constitute legal, tax, investment, or implementation advice.