Linear Drive Optics Six Weeks Later — What Hyperscalers Actually Ordered

In April I wrote about linear drive optics and called it a power-budget story. The argument: LPO trades modest reach for significant power savings, right for AI fabrics inside a data hall but not for anything traversing more than a few hundred meters.

Six weeks later, public order data tells a different story. The hyperscalers didn't order the SKUs I predicted. They ordered the longer-reach variants in much higher proportion than the power-budget thesis would suggest. The reach question matters more than I gave it weight for.

This is the honest follow-up.

What I argued in April

Three arguments.

LPO saves real power. Linear drive at the receive side eliminates the DSP tax — 3-5W per module at current speeds. Across a 51.2T high-density switch, hundreds of watts.

LPO trades reach. Without DSP equalisation and FEC, the link has tighter margin. Reach drops from the DSP-based equivalent. I projected sub-500m intra-data-hall as the sweet spot.

Hyperscalers would adopt LPO for AI fabric specifically. I projected volume orders concentrated on 100m and 500m reach class SKUs.

First two arguments held up. The third was wrong in an interesting way.

What the order data shows

Q2 2026, aggregated from two major optical-module vendor quarterly reports and two distributor supply-chain commentaries:

| Reach class | Share of LPO volume |
|---|---|
| 100m sub-rack | ~25% |
| 500m intra-hall | ~15% |
| 2km LPO+ | ~50% |
| 5km variant | ~10% |

The 2km class taking half the volume is the surprise. The power-budget thesis would have predicted the shortest reach SKUs taking the majority. They didn't. The 2km variant — trading less power saving for more reach — became the dominant SKU.

Why I was wrong about reach

Three reasons, in retrospect.

Hyperscaler data halls are bigger than I assumed. Flagship AI training facilities moved from single-building footprints to multi-building campuses. "Intra-data-hall" connections are now 1-2km in many configurations, not the 100-300m my model assumed.

Fabric topology consumes reach. A high-radix fabric routing through three or four switching stages accumulates connection length. A 100m physical span needs optics that handle the cable routing — path length exceeds straight-line distance.

Price-performance favours 2km LPO+. The cost premium over the 500m LPO is smaller than I modelled. Hyperscalers absorb the small extra cost for the deployment flexibility of one SKU across more use cases.

The combination made the 2km variant the operational sweet spot in a way the simple power-budget analysis missed.

What I had right

Aggregate power savings is real. LPO is taking meaningful share because watt savings at scale are real. That part was directionally correct.

Cross-hyperscaler adoption is real. Multiple hyperscalers ordering at scale, not just one. Moved from one-customer-specific to broad adoption.

DSP-vendor positioning was correct. The three remaining coherent DSP houses are seeing measurable revenue impact. Two have accelerated their low-power DSP roadmap. One is still betting LPO doesn't scale beyond AI fabric.

The directional argument held. The reach-class prediction didn't.

What the order pattern implies for adjacent markets

Intra-campus is the new intra-DC. The boundary between inside-the-hall and inter-DC has shifted. 1-2km is now intra-DC at hyperscaler scale. Single-mode becomes even more dominant. Multimode loses more of its remaining niche. The 500m-to-2km reach class becomes the volume-winning category across multiple module types.

Power-budget arguments without reach context are incomplete. The April post made that mistake. Any analysis arguing for a module class on power has to also be specific about reach. The orders combine power savings with adequate reach, not power savings at the cost of reach.

LPO+ variants get more development attention. Vendors saw the order pattern. Next-generation LPO products will emphasise 2km-class reach. Sub-500m becomes a niche segment within the LPO category.

The forecast revision

April forecast: LPO volume by 2027 heavily concentrated on intra-DC short-reach. Revised: 2km LPO+ as the dominant SKU class, short-reach being a smaller share than originally projected. For the broader market: 2km-class pluggable optics become the volume sweet spot across LPO and DSP-based. Sub-500m remains a category but smaller. 10km-plus DSP-based optics keep their inter-DC position.

What I'll watch for in Q3-Q4 2026

Three questions whose answers determine whether the Q2 pattern holds.

Does the 2km LPO+ share grow or stabilise? Growth signals consolidation into a reach-flexible category competing with DSP optics broadly. Stable means Q2 was a one-quarter peak from specific deployments.

Do second-tier hyperscalers adopt the same SKU mix? Mirror means the pattern holds. Clustering on shorter-reach variants means the first-tier orders were specific to their campus shape.

Does any vendor announce a 5km or 10km LPO variant? Would signal LPO aiming at inter-DC, currently DSP-only. Changes DSP-vendor revenue trajectory.

The lesson on forecast hygiene

Second time in 18 months I've publicly revised an optics forecast because actual order data revealed a pattern my a-priori analysis missed. Both revisions ran "the technology is more flexible than the simple analysis suggested, buyers found a sweet spot I didn't model".

Pattern. Buyers buy what works. Clean analytical frameworks lose to empirical order data once volume is real.

I'll keep writing forecasts. I'll keep being wrong about specifics in ways the data reveals six weeks later. The follow-up posts are how the analysis gets better. The forecast that's never revised in light of data wasn't tracking reality in the first place.

LPO won the Q2 surprise. Revised. On to Q3.