Lumentum chief executive Michael Hurlston told Bloomberg Television on 9 October 2026 that the company is “almost sold out for 2029,” and 3M said in a Form 8-K filed on 19 October 2026 that most of a warrant it issued to an Amazon subsidiary vests as Amazon pays for “optical connectivity and other Data Center products.”
Between 14 and 19 October, Marvell, Intel and Meta each described new places where light is carrying AI data: memory, chips and racks, and whole campuses. Put side by side, the five disclosures each place optics at a different layer of the AI stack, and Lumentum, one supplier, says it is under-shipping demand on some products.
Business Pill · BACKLOG
A short explainer on backlog: orders taken but not yet delivered. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the five disclosures map optics onto five different layers of the AI stack, from memory sitting up to 50 meters from compute (Marvell) to buildings linked over 100 kilometers apart (Meta), while one supplier, Lumentum, says it is almost sold out for 2029.
Marvell said on X on 19 October 2026 that “HBM stacks today sit 1.5 millimeters from a GPU,” and described its Photonic Fabric Memory Module as “capable of delivering up to 32 terabytes of disaggregated memory capacity up to 50 meters away from compute.” Marvell calls it the first optically interconnected high-capacity, high-bandwidth memory solution and says memory capacity can now scale independently from compute.
Intel said on X on 14 October 2026 that what comes next is “silicon photonics connecting compute across chips, nodes, and racks with more bandwidth and lower latency.” At its OCP 2026 keynote, Intel said it is the only high-volume fab that can deliver 300-millimeter indium phosphide silicon photonics at scale, according to Wccftech’s 18 October report.
Meta wrote on 19 October 2026 that its Prometheus aggregation network connects building fabrics “spanning 2 km to over 100 km,” and that the Prometheus network as a whole involves “millions of miles of fiber strand, tens of thousands of switches, and millions of optics.”
Where light is carrying AI data, one sourced figure per layer: memory up to 50 meters from compute (Marvell, 19 October 2026; HBM today sits 1.5 millimeters from a GPU); 300-millimeter indium phosphide photonics (Intel at OCP 2026, as reported by Wccftech); building fabrics spanning 2 km to over 100 km (Meta Engineering, 19 October); Lumentum almost sold out for 2029 (CEO on Bloomberg Television, 9 October); 85% of 3M’s warrant to Amazon vesting with optical connectivity and other data center purchases, full at $6 billion (3M Form 8-K, 19 October). 33,000 times is our arithmetic.
What Was Posted on X
HBM stacks today sit 1.5 millimeters from a GPU. What if they didnโt have to
In this clip from a recent conversation with @theCUBE, Marvell Executive Vice President and General Manager of Data Center Networking Dave Lazovsky walks through the Photonic Fabric Memory Module, the first optically interconnected high-capacity, high-bandwidth memory solution, capable of delivering up to 32 terabytes of disaggregated memory capacity up to 50 meters away from compute.
For the first time, memory capacity can scale independently from compute. Watch the full interview: https://mrvl.co/4s5SC86
Marvell posted this clip on X on 19 October 2026: Dave Lazovsky, its executive vice president and general manager of data center networking, walks through the Photonic Fabric Memory Module, which Marvell says is capable of delivering up to 32 terabytes of disaggregated memory up to 50 meters away from compute. It is the source of the memory layer in this piece’s table; the post gives no shipping date, customer or price.
Marvellโs Dave Lazovsky is holding a tiny module that uses photonics to move high-bandwidth memory (HBM + DDR) as far as 50 meters away from the GPU.
this can sit up to 50 meters from a GPU instead of the usual 1.5 mm.
For the first time, memory no longer has to live right next to compute, letting AI clusters share massive capacity without being locked to the accelerator package.
AI commentator Rohan Paul shared Marvell’s clip on 20 October, writing that memory and compute โnow scale on separate axes instead of being locked togetherโ. He also says latency stays in the hundreds of nanoseconds; that figure is his and is not in Marvell’s post, so this piece does not use it.
Where Supply Stands
Hurlston told Bloomberg Television that Lumentum is under-shipping demand on some of its products by 30% into 2028 and, for other products, by about 70% next year. “We’re almost sold out for 2029,” he said.
3M’s 8-K says the warrant lets the Amazon subsidiary buy up to 3,162,015 3M shares over seven years. It says 85% of the warrant shares vest as Amazon makes payments for optical connectivity and other data center products, with full vesting once cumulative payments reach $6 billion.
Our Earlier Coverage
9 October: Lumentum’s chief executive said the company is almost sold out for 2029, as we reported. 19 October: Meta published how Prometheus is wired, covered here. 19 October: 3M disclosed the Amazon warrant tied mostly to optical connectivity and data center purchases, covered here. 19 October: Marvell described its photonic memory module. 14 October: Intel posted on X about silicon photonics at OCP 2026, and on 18 October Wccftech reported Intel’s keynote claim on 300-millimeter indium phosphide.
What Is New Here
By our arithmetic, 50 meters is about 33,000 times the 1.5 millimeters that Marvell says HBM sits from a GPU today. The table we built maps one sourced figure to each layer: 50 meters at the memory layer (Marvell), 300-millimeter indium phosphide at the chip-to-rack layer (Intel, as reported), over 100 kilometers at the campus layer (Meta), 2029 on supply (Lumentum) and $6 billion on the buyer side (3M’s filing).
As we read it, the common thread is distance: the sourced figures run from 50 meters at the memory layer (Marvell) to over 100 kilometers between buildings (Meta).
The Structural Read
As we read it, Marvell’s module addresses the distance between memory and compute: its post says HBM today sits 1.5 millimeters from a GPU and that the module delivers memory up to 50 meters away.
As we read it, the buyer side shows up in 3M’s filing, where 85% of Amazon’s warrant vests with payments for optical connectivity and other data center products.
As we read it, the supply side shows up in Lumentum’s comments: under-shipping some products by 30% into 2028 and others by about 70% next year, by its chief executive’s account.
Marvell, on X, 19 October 2026
“For the first time, memory capacity can scale independently from compute”
Three Implications
MEMORY MOVES OUT Marvell says its Photonic Fabric Memory Module is capable of delivering up to 32 terabytes of memory up to 50 meters from compute.
CAMPUSES LINK UPMeta says its Prometheus aggregation network spans 2 km to over 100 km, and that the Prometheus network as a whole uses millions of optics.
ONE SUPPLIER’S ORDER BOOK Lumentum’s chief executive said the company is almost sold out for 2029 and is under-shipping demand on some products by 30% into 2028.
The framework puts it this way: “This is the critical interface layer where compute meets data, and its physics-driven limitations now determine the ceiling on AI capability scaling”.
As we read it, Marvell’s photonic memory module is aimed at that interface, since its post says memory capacity can now scale independently from compute. The framework was published in December 2025 and does not discuss Marvell’s module.
What Is Not Established
Marvell’s post gives no shipping date, customer or price for the memory module. We have not read Intel’s keynote slides directly; the 300-millimeter claim is as Wccftech reported it. The 3M filing does not say how much of Amazon’s spending goes to optical connectivity rather than other data center products.
Lumentum’s comments describe its own order book, not the optics market as a whole. Nothing here links any one of these disclosures to another company’s purchasing, and nothing here is a forecast.
Business Engineer Framework
The AI Memory Chokepoint
A Business Engineer analysis of why the interface where compute meets data sets the ceiling on AI scaling.
Lumentum said on 9 October 2026 it is almost sold out for 2029, and 3M’s 19 October 8-K ties most of Amazon’s warrant to purchases of optical connectivity and other data center products, while Marvell, Intel and Meta each described light carrying AI data at a different layer of the stack.
A note on sourcing. This piece is built on Marvell’s 19 October 2026 post on X, Intel’s 14 October posts on X, Wccftech’s 18 October report of Intel’s OCP 2026 keynote (we have not read the keynote slides directly), Meta Engineering’s 19 October post (read in full), the transcript of Lumentum chief executive Michael Hurlston’s 9 October Bloomberg Television interview, and 3M’s Form 8-K filed on 19 October.
Rohan Paul’s 20 October post on X is embedded as commentary. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
