The arithmetic behind the build-out.
Working notes on the parts of AI infrastructure that are hard to get a straight answer on. Every figure is sourced, and where the published numbers disagree we show the disagreement rather than picking a side.
- Industry watchOct 18, 2026
A link budget looking for a link.
Coherent Lite is dual-polarization 16QAM on a fixed O-band laser, for 2 to 20 km inside a campus. Its estimated power is within the range of OSFP-family 1.6T modules, at about 1.6 to 1.9 times the power of IM-DD. Forecasts span ten times, and nobody placed a volume order. Its market is the set of links where IM-DD runs out of budget, and at 1.6T that set was almost empty.
9 min read20 sources
- Industry watchOct 18, 2026
A second source is a new part.
The first clause of JESD264 says who the silicon photonics qualification standard binds, which events restart a qualification, and what it leaves out. Read together, they tie a qualification to a whole supply configuration, not to a design. A second foundry is therefore a new qualification, with a schedule measured in months.
7 min read5 sources
- Industry watchOct 11, 2026
The test menu is public. The results are not.
On 1 October 2026 JEDEC published JESD264, the first industry-wide qualification standard for silicon photonics. It tells a foundry what to test. It cannot show which platforms passed, and today almost no foundry publishes that. This essay reads the standard, does the statistics on its sample plans, and follows shipping products back to the fabs that could be asked.
10 min read26 sources
- FundamentalsSep 16, 2026
Four of these are distances. One is a direction.
Scale-inside, scale-up, scale-out, scale-across, scale-in, scale-beyond: these six terms appear on one slide as if they were one ladder. Four name a distance that a network must cross. The other two are not distances: one is a direction, and one has no published definition. Sorting them out explains most of what an AI fabric is.
10 min read11 sources
- Industry watchSep 11, 2026
The build-out runs on three clocks
Who is actually fast? Every hyperscaler build now advances on three separate schedules: power, silicon and network. No builder leads all three. This is a scoreboard of who energizes gigawatts, who gets Rubin first, and who still buys 400G while a rival deploys 1.6T.
12 min read33 sources
- Industry watchSep 2, 2026
The DSP picks the modulator
The first 400G-per-lane optical DSP arrived at OFC 2026 with two lines in its product bullets. One is an integrated laser driver. The other is proven interoperability with the vendor's own electro-absorption modulator. Together, those lines make a design decision for the module maker. The reason is the analog bandwidth of a silicon Mach-Zehnder.
14 min read45 sources
- Industry watchAug 19, 2026
The wafer under the wafer
Every silicon photonics chip starts on a specialty substrate, and only one company documents shipping it at 300 mm into a qualified production line. Soitec discloses the revenue but not the wafers. This essay counts them from the outside, removes the one public wafer count that is circular, and finds that capacity is not the constraint.
11 min read17 sources
- How to calculateAug 5, 2026
How many optical modules does AI actually need?
Published 2026 forecasts for 800G module shipments are between 33.5 million and 63 million, a factor of two for the same product and year. This essay does not select one. It counts the market in three independent ways, then follows the winning technology down to wafers and substrates. The 2025 results are now available, a market model is now in a regulated filing, and one widely cited wafer count is circular.
15 min read33 sources
- FundamentalsJul 22, 2026
Trust the FEC, not the thermometer
An optical module measures its laser temperature, bias current, output power and modulator bias point, and reports all of it to the network. Almost none of it is accurate enough to act on. The measuring chip stopped being a dedicated part years ago. The number that operators actually use comes from the FEC.
12 min read22 sources
- How to calculateJul 8, 2026
What a gigawatt of AI data center actually costs
Four respected research houses priced the same one-gigawatt build between $35 billion and $60 billion. An average is tempting. It would remove exactly the information that the spread contains.
12 min read9 sources
- How to calculateJun 17, 2026
CapEx, OpEx and the depreciation argument
An AI cluster with a tenant can show a 98 percent margin while its owner reports a multi-billion-dollar loss. Both statements can be true. The most important number in AI economics is not a measurement. Management declares it.
12 min read14 sources
- Industry watchMay 27, 2026
A thousand gigahertz is not a product
In one postdeadline session at OFC 2026, five different modulator materials reached 400 gigabits on a single lane. So bandwidth no longer decides the 3.2T generation. The rest of the datasheet does. A census of nine routes shows that the leaderboard everybody quotes is almost the reverse of the one that matters.
13 min read31 sources
- Industry watchMay 13, 2026
Who makes what, and what you can actually second-source
When an AI build is a year late, the review rarely finds a GPU problem. It finds a transformer. A supply chain can be mapped in dollars or in lead times, and almost everyone plans with the dollar map.
12 min read11 sources
- Industry watchApr 28, 2026
Reading the XPO MSA, the pluggable fights back
Many analysts expected co-packaged optics to replace the pluggable module. Then Arista and its partners announced a 12.8-terabit, liquid-cooled, 400-watt pluggable that puts a whole switch generation on one rack unit of faceplate. The arithmetic shows why the pluggable stays.
9 min read8 sources
- Industry watchApr 21, 2026
Reading the OpenCPX MSA, the socket that makes co-packaging serviceable
The least visible document of OFC 2026 may have the largest effect. OpenCPX standardizes a connector. That connector decides whether an operations team accepts co-packaged optics for production, or keeps it as a demonstration.
8 min read6 sources
- Industry watchApr 14, 2026
Reading the OCI MSA, where scale-up optics chose NRZ
Six of the most powerful companies in computing formed a consortium to define optical scale-up links. Their first technical decision was to replace PAM4 with NRZ, the simplest modulation there is. That choice explains the whole specification, and a short calculation predicts it.
9 min read7 sources
- FundamentalsApr 7, 2026
The placement ladder, or what LPO, LRO, NPO, CPO and OIO actually mean
Pluggable, LPO, LRO, NPO, CPO and OIO are six answers to one question: how close to the silicon should the optics sit? Each step closer saves energy per bit and makes a repair harder.
12 min read10 sources
- FundamentalsFeb 19, 2026
Silicon has no Pockels effect, and everything about modulator materials follows from that
Silicon, the most manufacturable material in history, has a crystal symmetry that gives it no linear electro-optic effect. So silicon modulates light by injecting or removing carriers, with side effects. The materials that avoid this show why the best physics rarely wins, and why it does not need to.
13 min read14 sources
- FundamentalsJan 29, 2026
Copper is out of bandwidth. Optics is out of SNR. That asymmetry is the roadmap.
The SerDes world seriously debates PAM6 and PAM8. The optics world will not use either, and adds lanes instead. Both are right, and the reason explains most of the interconnect roadmap.
13 min read11 sources
- Industry watchJan 15, 2026
Retrofits, mining sites, and whether a new build is obsolete on day one
Experienced buyers pay nearly new-build prices to convert bitcoin mines that cost a tenth as much to build. They know what they are doing. They buy the only part of a data center that a factory cannot make.
11 min read10 sources
- FundamentalsDec 19, 2025
Inside an optical module, where the watts and the dollars go
An 800G module that fits in your hand has a bill of materials over a thousand dollars and a power budget over ten watts. One chip inside it takes a third of the first and half of the second. That chip explains how the module economy works.
13 min read15 sources
- FundamentalsDec 8, 2025
Why AI capacity is counted in gigawatts
The industry measures computing buildings in units of electricity, and nobody finds this strange now. Three physical limits explain how that happened. The third one is why optics now sits in the critical path.
11 min read9 sources