Tag: AI networking

  • Cisco’s $9B Hyperscaler Forecast Shows AI Networking Is Now Its Own Budget Line

    Cisco’s $9B Hyperscaler Forecast Shows AI Networking Is Now Its Own Budget Line

    TL;DR · 30-second read

    The Short Version

    Cisco, whose equipment moves data inside offices and the giant buildings that run the internet, just had its fastest growth in more than 15 years: $18.5 billion in sales in three months, up nearly a quarter from a year ago.

    The surprise is where the money came from. About nine of every ten extra dollars came from its networking equipment, the gear that links thousands of artificial intelligence chips so they can work together as one machine.

    Cisco now expects the largest tech companies to spend about $9 billion with it on that artificial intelligence equipment this business year, more than double last year.

    Cisco Systems reported revenue of $18.5 billion for its fiscal first quarter of 2027, which ended October 24, 2026, up 24% from $14.9 billion a year earlier, according to the earnings release it furnished with an 8-K filed November 12. GAAP earnings per share rose 58% to $1.14 and non-GAAP earnings per share rose 40% to $1.40, which Cisco called its strongest revenue and non-GAAP EPS growth in more than 15 years. Networking product revenue grew 43% to $11.1 billion.

    Cisco raised its full-year guidance to $74.8 billion to $75.8 billion in revenue and $5.47 to $5.53 in non-GAAP EPS, and lifted its expectation for AI infrastructure revenue from hyperscalers, the largest cloud and internet companies, to $9 billion in fiscal 2027, more than double the fiscal 2026 total. For the second quarter it guided to $18.0 billion to $18.2 billion in revenue.

    Executive Summary

    Cisco’s quarter was, in practical terms, a networking quarter. Of roughly $3.6 billion in added year-over-year revenue, about $3.3 billion came from the networking product line, now 60% of total company revenue. Services, long Cisco’s ballast, grew 1%. Security, collaboration and observability grew in the single digits.

    The figure that matters most for infrastructure planners is the one Cisco chose to put in its headline bullets: a $9 billion expectation for AI infrastructure revenue from hyperscalers in fiscal 2027, more than twice the prior year. When a company that built its franchise on office and campus networks forecasts AI data center networking as a separate, fast-growing revenue stream, it signals that the network connecting AI accelerators is being purchased as part of compute capacity, not as background IT.

    The results also carry tensions worth watching: product gross margin fell nearly three points, contracted backlog grew far more slowly than product revenue, and second-quarter guidance sits below the first quarter’s actual result, implying that Cisco is counting on a stronger second half to reach its raised full-year range.

    Networking Carried Nearly the Whole Quarter

    Cisco’s 8-K exhibit breaks product revenue into four groups, and the distribution is lopsided. Networking delivered $11.13 billion, up 43%. Security brought in $2.12 billion (up 7%), collaboration $1.14 billion (up 8%) and observability, the software that monitors application and infrastructure performance, $279 million (up 2%). Services revenue was essentially flat at $3.85 billion, up 1%.

    Working backward from the 43% growth rate, networking was roughly $7.8 billion a year earlier, so the line added about $3.3 billion of the company’s $3.6 billion revenue increase. That is the core fact of the quarter: a diversified vendor that has spent years building software and subscription businesses grew almost entirely on hardware that moves data.

    Cisco describes demand as broad-based, with double-digit growth across all geographies and customer markets. The regional numbers support that: the Americas grew 28% to $11.5 billion, EMEA (Europe, the Middle East and Africa) 18% and APJC (Asia Pacific, Japan and China) 21%. The growth is not exclusively hyperscaler AI, which matters for reading the next section correctly.

    Why a $9 Billion Forecast Makes the Network Its Own Budget Line

    Large AI models are trained and served by spreading work across thousands of accelerators, the specialized chips that do the math, which must exchange data constantly. The network fabric linking them, meaning the switches and connections inside and between racks, scales with the size of the cluster. If the fabric is undersized, expensive accelerators sit idle waiting for data. That is why AI networking is bought alongside compute rather than as overhead.

    Cisco’s own disclosures put a number on that dynamic. The company raised its expectation for AI infrastructure revenue from hyperscalers to $9 billion in fiscal 2027, more than double fiscal 2026, which by implication was below $4.5 billion. Against the midpoint of the new full-year revenue guide, $75.3 billion, that is roughly 12% of Cisco’s expected sales from a customer group and use case that the company now forecasts separately. Vendors break out a revenue stream when it is large enough and distinct enough to move the whole business; Cisco has done that here.

    The consequence lands on several groups. Data center developers and operators building for AI tenants have to plan network gear as a capacity input with its own lead times, rack space and power draw, not a rounding error in the fit-out. Enterprises refreshing campus and data center networks are now buying from a supplier whose fastest-growing demand comes from the largest cloud companies. And investors evaluating Cisco are, increasingly, evaluating exposure to hyperscaler capital spending cycles. The caveat is scale: $9 billion is a minority of what networking now generates, so the broader 43% surge reflects enterprise and service provider demand too, not AI alone.

    Earnings Are Outrunning Margin, and Cost Discipline Is the Bridge

    Earnings grew much faster than revenue, but not because each sale became more profitable. GAAP product gross margin fell to 61.8% from 64.5%, and non-GAAP total gross margin fell to 65.6% from 68.1%. Product cost of sales rose about 43%, to $5.61 billion from $3.93 billion, roughly in step with networking growth. Cisco did not attribute the margin decline to a specific cause, and its second-quarter guidance of 65% to 66% non-GAAP gross margin suggests the lower level is expected to persist near term.

    The offset came below the gross margin line. GAAP operating expenses fell 3% to $6.19 billion, so GAAP operating income rose 66% to $5.6 billion and non-GAAP operating margin reached a record 37.8%. A slightly lower diluted share count, 3.98 billion versus 3.99 billion, added a little more per-share lift. CFO Mark Patterson framed this as operating leverage. The numbers support that claim for this quarter; whether flat spending can hold while revenue scales further is a question for future quarters.

    The Guidance Leans on the Second Half

    Second-quarter revenue guidance of $18.0 billion to $18.2 billion sits below the first quarter’s $18.5 billion. Subtracting the first quarter and the guided second quarter from the full-year range leaves roughly $38.3 billion to $39.1 billion for the back half, or about $19.2 billion to $19.6 billion a quarter. The raised annual outlook therefore depends on a sequential step-up later in the fiscal year, consistent with, but not proven by, the hyperscaler forecast.

    Two balance sheet lines are worth reading alongside that. Remaining performance obligations, revenue already contracted but not yet recognized, rose 8% to $46.2 billion, and deferred revenue rose 7%, both far slower than 32% product growth. That suggests much of the surge is shipping close to when it is ordered rather than accumulating as long-dated contracted backlog. Inventories rose to $6.4 billion from $5.7 billion at fiscal year-end, which is consistent with preparing for higher shipments. CEO Chuck Robbins called this the very early stages of a multi-year networking supercycle; the $9 billion forecast is the most concrete support for that view, while the backlog figures show how much of it remains a forecast.

    Background

    Cisco Systems, based in San Jose, California, built its business on routers and switches, the equipment that directs traffic across corporate offices, campuses and service provider networks. Over the past decade it expanded into security, collaboration software such as Webex, and observability, a category it deepened with its 2024 acquisition of Splunk. Its fiscal year ends in late July, which is why results for August through October are its first quarter.

    The AI buildout has shifted attention inside data centers from individual servers to the clusters that link thousands of accelerators together. That has made high-speed networking a core part of AI capacity and drawn established networking vendors into competition for hyperscaler budgets, the large capital programs that the biggest cloud companies run to expand their data center fleets.

    Sources

    Source: Cisco Reports First Quarter Earnings, Cisco’s fiscal first-quarter 2027 results release distributed via PR Newswire on November 12, 2026.

    Primary sources: Cisco Systems, Inc. Form 8-K filed November 12, 2026 (SEC EDGAR); Exhibit 99.1 to the 8-K: Cisco Reports First Quarter Earnings (SEC EDGAR).

  • Even AI Optics Wait on the Grid: Power Slows AOI’s US Build Toward 1M Units a Month

    Even AI Optics Wait on the Grid: Power Slows AOI’s US Build Toward 1M Units a Month

    TL;DR · 30-second read

    The Short Version

    The giant computer centers that run artificial intelligence often sit waiting for electricity hookups. That same wait is now slowing a factory that makes their parts.

    Applied Optoelectronics, a Texas company, makes small devices that turn data into pulses of light so computers can talk to each other over glass fiber. It says its planned American factory expansion will take longer, partly because getting enough electricity is taking longer than expected.

    It still aims to make more than a million of these devices a month by the end of 2027. That is roughly three times what its equipment can make today.

    Applied Optoelectronics (NASDAQ: AAOI) reported third-quarter 2026 results on November 5, 2026. It posted its sixth consecutive quarter of record revenue and returned to GAAP profitability. GAAP refers to the standard US accounting rules. The company said it has manufacturing equipment capacity approaching 330,000 units per month. It expects to have equipment installed for more than 500,000 800G and 1.6T transceivers per month by the end of 2026, and more than 1 million per month by the end of 2027. Its 2027 US capacity expansion will take longer than planned because obtaining the necessary electrical power is taking longer than expected.

    Yahoo Finance reported that the company expects AI demand to outrun its capacity through 2027. In an 8-K filed November 4, AOI also disclosed a new $200 million credit facility led by East West Bank. Most of it is earmarked for equipment purchases.

    Executive Summary

    The headline numbers in AOI’s third-quarter release describe a supplier in rapid expansion. Its equipment base is approaching 330,000 units a month. Its stated path runs past 500,000 high-speed transceivers a month by year-end and past 1 million a month by the end of 2027. CEO Dr. Thompson Lin also said the company recently secured a multi-year strategic agreement with a Tier 1 hyperscale datacenter customer, meaning one of the largest cloud operators. He believes that deal improves long-term visibility.

    The more consequential line came from CFO Dr. Stefan Murry. He said the company’s 2027 US expansion will take longer due to the “longer than expected time to obtain necessary electrical power among other construction related issues.” Power constraints are usually discussed as a problem for data center sites. Here, a manufacturer of the optical components those sites depend on is reporting the same constraint on its own factory schedule.

    For anyone planning AI cluster deliveries, this matters because optical transceivers sit on the critical path alongside GPUs, switches and power. When a transceiver supplier’s ramp depends on a utility hookup, that hookup becomes an input to the cluster schedule as well.

    The Grid Wait Has Moved Upstream

    Every AI training cluster is held together by optical links. Each fiber connection between a server, a switch and the next tier of the network needs a transceiver at each end. A transceiver is a module that converts electrical signals into light and back. The current generation runs at 800 gigabits per second (800G) and 1.6 terabits per second (1.6T). Those are the products AOI is scaling. As clusters grow, transceiver counts grow with them, which is why a supplier’s monthly output is a meaningful constraint on how fast clusters can be assembled.

    AOI’s November 5 release names electrical power as a reason its 2027 US expansion will take longer. The mechanism is the same one that delays data centers: a new industrial facility cannot run production lines until the utility can deliver the required service. AOI has not said how long the delay is, how much power the site needs, or which utility is involved. Still, the disclosure places grid access on the critical path of an optics ramp, not only a data center ramp. Those most directly affected are the hyperscale customers sizing their 2027 networks around AOI’s output, particularly any that value US-made supply.

    The company frames the slip carefully, and it should be read that way. AOI still expects equipment for more than 500,000 800G and 1.6T units a month by the end of 2026 and more than 1 million a month by the end of 2027. The delay therefore applies to the US build specifically. It is not a withdrawal of the headline target. What has not been disclosed is how much of the 1 million figure depends on the delayed US site. That share determines how much the power issue could move the 2027 number.

    Installed Equipment Is Not Shipped Product

    AOI’s capacity figures describe equipment, not output. Its current base is approaching 330,000 units a month of “total manufacturing equipment capacity.” The year-end 2026 figure is equipment “installed to produce” more than 500,000 pieces. Turning installed machines into shipped modules also depends on yields, trained staff, qualification by each customer, and a steady supply of lasers and other components. The release does not quantify any of those. Rising from roughly 330,000 to over 1 million in about five quarters is roughly a tripling. That is achievable on paper, but it is demanding in practice.

    If demand does exceed supply through 2027, then capacity, rather than orders, sets AOI’s revenue. In that case, every month of delay at a new site becomes deferred revenue. For buyers, a constrained supplier means allocation: whoever holds multi-year agreements is served first. That makes the newly announced Tier 1 hyperscale agreement as much about securing supply as about price.

    Borrowing to Buy the Machines

    The 8-K filed November 4 shows how the equipment is being financed. On October 30, AOI entered a credit agreement with East West Bank as administrative agent. It includes $170 million in initial term loans, $30 million in revolving commitments, and an option to request up to $50 million more in term loans. Term loan proceeds may be used to purchase or refinance eligible equipment. The facility matures October 30, 2030. It is secured by substantially all of the company’s assets, and borrowings price at Term SOFR plus 0.10% plus a margin of 2.50% to 3.25%, depending on leverage.

    Two details matter for reading the ramp. First, the initial term loans amortize at about 3.57% of the outstanding amount each quarter, so repayment begins well before maturity. Delays that push revenue out while repayments proceed would tighten cash flow. Second, AOI terminated its July 2025 facility with BOK Financial, which still had $35 million of unused revolving commitments, and paid a $350,000 termination fee to do so. Trading that facility for a larger, equipment-oriented one signals that management expects the capital needs of the build-out to be substantial.

    Visibility and Concentration Come Together

    The multi-year hyperscale agreement gives AOI what Dr. Lin called improved long-term visibility. AOI has not named the customer or disclosed volumes, pricing or duration. The company’s own forward-looking risk factors list reliance on a small number of customers for a substantial portion of revenue, potential pricing pressure and delays in shipments. A single large agreement can reinforce that concentration even as it reduces demand uncertainty. The 10-Q filed November 5 also continues to account for a customer warrant issued to an Amazon subsidiary in March 2025, a reminder of how closely the company’s fortunes are tied to large cloud buyers.

    The balanced reading is that AOI has substantiated three things: record revenue, a return to GAAP profitability, and committed equipment financing. Its capacity trajectory and the extent of unmet demand remain forward-looking statements. The power-related delay is the one operational fact in the release that cuts against the expansion narrative, and the company disclosed it plainly.

    Background

    Applied Optoelectronics is headquartered in Sugar Land, Texas, and was founded by its current President and CEO, Dr. Thompson Lin. It makes optical networking products for three main markets. In datacenters, it supplies transceivers linking servers and switches. In cable television (CATV), it supplies hybrid fiber-coaxial (HFC) broadband equipment. It also serves telecom networks. The company has operations in the US and in New Taipei City, Taiwan.

    Generative AI has transformed the transceiver market. AI clusters move vast amounts of data between processors, which drives demand for faster optical links at 800G and now 1.6T. Large cloud operators are the dominant buyers. That gives them significant leverage over suppliers, and it also gives suppliers that win multi-year agreements a clearer view of future demand.

    Sources

    Source: Applied Optoelectronics Says AI Demand Will Outrun Its Capacity Through 2027 — Yahoo Finance coverage of AOI’s capacity outlook.

    Primary sources: Applied Optoelectronics Form 10-Q for the quarter ended September 30, 2026 (filed November 5, 2026); Applied Optoelectronics Form 8-K (filed November 5, 2026); Exhibit 99.1: Applied Optoelectronics Reports Third Quarter 2026 Results; Applied Optoelectronics Form 8-K on new East West Bank credit facility (filed November 4, 2026).

  • Marvell’s $115B Lumentum Deal Is a Bet on AI Outgrowing Copper’s 1.25-Meter Reach

    Marvell’s $115B Lumentum Deal Is a Bet on AI Outgrowing Copper’s 1.25-Meter Reach

    TL;DR · 30-second read

    The Short Version

    Marvell, a chipmaker, has agreed to buy Lumentum, a maker of lasers and other light-based parts, for $115 billion paid in its own shares. Nvidia, the chip giant, has invested in both. The deal is not expected to be final until 2028.

    The reason is physical. Inside the computers that run artificial intelligence, data travels over copper wire, and at today’s top speeds the signal fades after about four feet. Light in glass fiber travels for miles. Marvell is betting that as companies link thousands of chips together, light wins.

    The catch: light-based links have used more electricity, a problem still being worked on.

    Marvell Technology announced on November 2, 2026 that it will acquire Lumentum Holdings, a maker of lasers, optical transceivers and optical circuit switches, in an all-stock transaction valued at about $115 billion. The Register reported that the deal is expected to close in the first half of 2028, and TechCrunch reported that Nvidia is a strategic investor in both companies. Marvell and Lumentum executives presented the deal on a joint investor call the same day, a transcript of which Marvell filed with the SEC.

    24/7 Wall St. reported that Marvell shares fell 3% and Lumentum shares fell 6% on the announcement, while shares of rival optics supplier Coherent rose 2%.

    Executive Summary

    Marvell is paying roughly $115 billion in stock for Lumentum to fill the parts of its photonics portfolio it could not build itself: high-performance laser sources, optical transceivers (modules that convert electrical signals to light and back), optical circuit switches (which redirect beams of light without converting them to electricity), and in-house manufacturing on indium phosphide, the compound semiconductor used to make the lasers. On the joint investor call, Marvell CEO Matt Murphy called it an opportunity that “comes along once in a generation” and said the combined company would be the pure-play leader in optical connectivity.

    The strategic logic is about distance. At 400 gigabits per second per lane, copper wiring holds a signal for only about 1.25 meters, and that reach halves every time lane speed doubles. As Nvidia, AMD and others design systems that tie hundreds or thousands of GPUs into one machine, that physical limit starts dictating how tightly racks must be packed. Optics removes the distance constraint but has historically cost power. Marvell is betting, at a price equal to nearly two-thirds of its own market value, that the power problem gets solved and that optics moves into the heart of the AI cluster.

    The 1.25-Meter Problem Behind a $115 Billion Price

    The title claim rests on a simple rule of signal physics. A “lane” is one electrical path carrying data between chips. At 400 gigabits per second per lane, copper can carry a clean signal about 1.25 meters, and each doubling of lane speed halves that reach. By the same rule, the next step to 800 gigabits per lane leaves roughly 60 centimeters. Every GPU that must talk to its neighbors at full speed has to sit within that radius, which is why copper-based designs push toward denser, hotter racks with more power delivered into a smaller footprint.

    Optical links do not obey that scaling rule; they can carry the same signal over kilometers. That lets a “compute domain,” the group of accelerators that behave as a single machine for training or inference, stretch across an entire aisle rather than being crammed into one rack, while lowering the power and cooling load on each individual rack. Nvidia and AMD are both moving toward larger rack-scale systems that stitch together hundreds or thousands of GPUs. At those counts and speeds, the 1.25-meter ceiling stops being an engineering footnote and becomes a design constraint, and that is the market Marvell is paying for. On the investor call, Murphy framed connectivity as central to what he called the top three bottlenecks in AI data centers: performance, bandwidth and energy efficiency.

    The people most affected are those who build and run the facilities. If optics moves inside the scale-up domain, the planning problem shifts from concentrating extreme power and heat into single racks toward spreading compute across rows connected by fiber. That changes power-density assumptions, cooling design and cabling plans for data centers being specified now for systems that will ship after this deal closes.

    Power Is the Argument Optics Still Has to Win

    The case for optics has always run into one number. When Nvidia unveiled its first Blackwell rack systems in 2024, Jensen Huang said the company deliberately avoided optics because they would have added 20 kilowatts to a system already drawing 120 kilowatts under load, roughly a 17% increase. For operators already struggling to secure grid capacity, that penalty is decisive.

    Marvell’s answer is integration. In February 2026 it paid about $3.5 billion for Celestial AI, whose near-packaged and co-packaged optics place the optical engine beside or on the same package as the chip, using a fraction of the power of the pluggable modules common in today’s networks. Lumentum adds the lasers and manufacturing those designs depend on. The thesis is coherent, but the payoff depends on co-packaged optics reaching volume production at the power levels promised. Until Marvell shows production power figures for integrated optical systems, Huang’s 20-kilowatt figure remains the benchmark customers will measure it against.

    There is also an internal tension worth watching. Lumentum sells optical transceivers, the pluggable products that co-packaged optics is designed to displace in some uses. Marvell will own both the incumbent technology and part of its intended successor. That can be read as a sensible hedge across a multi-year transition, but it also means some of the acquired revenue sits in the product category Marvell’s own roadmap aims to replace.

    Nvidia Sits on Both Sides of the Table

    Nvidia invested $2 billion in Lumentum roughly eight months before the deal to expand its optics manufacturing capacity, and it is also a strategic investor in Marvell. At Computex 2026, Huang predicted that Marvell’s photonics investments would lift it to a $1 trillion valuation; Marvell’s market value currently sits around $180 billion. The deal consolidates manufacturing capacity that Nvidia helped fund into a company in which it also holds a stake.

    For other buyers, including AMD and the hyperscalers (the largest cloud operators) building their own accelerators, the practical question is supply allocation. Optical components have been a constrained input to AI buildouts, and a combined supplier that Murphy describes as significantly larger than its next competitor will have considerable influence over who gets capacity and when. The move in Coherent’s share price suggests investors see value in the remaining independent supplier, though a single day’s trading is a thin basis for conclusions.

    An All-Stock Price Worth Nearly Two-Thirds of the Buyer

    The $115 billion price equals about 64% of Marvell’s roughly $180 billion market value, which means Lumentum shareholders will hold a substantial share of the combined company. Paying in stock avoids raising large amounts of debt, but it also means the deal’s value floats with Marvell’s share price over a closing period of up to 18 months. Both stocks fell on the announcement.

    Murphy’s headline claims, that the combination will be the pure-play leader in optical connectivity and one of the top five AI infrastructure semiconductor companies, are forward-looking statements. They will become testable once Marvell files its S-4 registration statement, the document that sets out terms, combined financials and risks for shareholders voting on the deal. Lumentum CEO Michael Hurlston and two other Lumentum directors will join Marvell’s seven-member board at closing, expanding it to ten.

    Background

    Marvell Technology designs chips for data infrastructure, including networking, custom processors and electro-optical components. CEO Matt Murphy describes a decade-long effort to build positions in the core technologies behind AI, including what he calls one of the industry’s largest engineering teams for electrical and optical connectivity. In February 2026 Marvell paid about $3.5 billion for Celestial AI, a developer of low-power photonic interconnects. Marvell’s market value is around $180 billion.

    Lumentum, spun off from JDS Uniphase in 2015, makes lasers, optical transceivers and optical switching gear for telecom and data center networks, and expanded in data center transceivers with its purchase of Cloud Light in late 2023. Nvidia invested $2 billion in the company earlier in 2026 to bolster optics manufacturing capacity. Lumentum’s fiscal year ends in late June; it filed its quarterly report for the period ended September 26, 2026 the day after the deal was announced.

    Sources

    Source: Marvell wants to be a copper topper, bets $115B on photonics leader to push optical networking — The Register’s report on the deal and the copper-to-optics physics behind it.

    Primary sources: Marvell Form 425: transcript of joint Marvell–Lumentum investor call (Nov. 2, 2026); Marvell Form 425: presentation to Marvell employees (Nov. 2, 2026); Marvell Form 425: presentation to Lumentum employees (Nov. 2, 2026); Lumentum Form 425: corporate LinkedIn and X posts (Nov. 2, 2026); Lumentum Form 425: corporate LinkedIn post (filed Nov. 3, 2026); Lumentum Form 10-Q for the quarter ended Sept. 26, 2026; TechCrunch: In one of the largest deals in tech history, Marvell buys Lumentum for $115B; 24/7 Wall St.: Marvell Slides 3%, Lumentum Drops 6% on $115 Billion All-Stock Merger Deal; Coherent Climbs 2%.

  • Marvell’s Reported $83B Lumentum Bid Would Fold AI Optics Into a Chip Vendor’s Price List

    Marvell’s Reported $83B Lumentum Bid Would Fold AI Optics Into a Chip Vendor’s Price List

    TL;DR · 30-second read

    The Short Version

    Marvell, an American chip company, is reportedly close to buying Lumentum for about $83 billion. That could make it one of the ten biggest technology takeovers on record.

    Lumentum makes parts that turn electrical signals into light, so data can travel over glass fiber between the thousands of computers that power artificial intelligence. Nvidia, the leading AI chipmaker, has backed it.

    Why it matters: companies building AI would buy both the chips and the light-based connections from the same seller, giving that seller more say over price and supply. Nothing is signed yet.

    Bloomberg reported that Marvell Technology is nearing a deal to acquire Lumentum Holdings, the Nvidia-backed maker of optical communications components, with an announcement possible as early as next week. Dealroom put the deal’s value at roughly $83 billion, a size that would place it among the largest technology acquisitions on record.

    Lumentum shares rose about 15% in after-hours trading on the reports, GuruFocus reported, while Marvell shares slid after hours. Neither company has announced a transaction. Separately, in an 8-K filed October 23, Marvell disclosed a new $3 billion revolving credit facility, expandable to $5 billion, replacing a $1.5 billion line.

    Executive Summary

    Marvell already sells much of the electronics inside data center networks: the signal-processing chips that sit inside optical modules and custom AI processors it designs for large cloud companies. Lumentum supplies the photonic side of the same links, meaning the lasers, detectors and modules that turn electrical signals into light so data can cross fiber between racks and buildings. A combination would put both halves of an AI cluster’s optical connection under one corporate owner.

    That changes the commercial relationship for the buyers who build AI “scale-out” networks, the fabrics that connect thousands of accelerators so they can work as one machine. Optics that are bought today as a component from a specialist supplier would instead be negotiated with a chip vendor that also sells the silicon at each end of the link.

    The deal is unconfirmed, the price and structure are unannounced, and Marvell has not said how it would pay for a transaction roughly sixteen times the maximum size of the credit line it just put in place. Those open questions matter as much as the headline number.

    Optics Would Land on the Silicon Vendor’s Price List

    Every optical link in an AI data center has two parts: electronics that clean up and encode the signal, and photonics that convert it to light and back. Marvell is a major supplier of the first part, including the digital signal processors (DSPs) that sit inside high-speed optical modules. Lumentum supplies the second. Today a cloud operator, or hyperscaler, typically qualifies and buys these from different companies, and module makers assemble them. Owning Lumentum would let Marvell offer the full link, and price it as a package.

    For the buyer, that has real benefits: fewer vendors to qualify, one point of accountability when a link fails, and potentially tighter engineering between the chip and the laser. The trade-off is leverage. A hyperscaler that also relies on Marvell for custom AI processors would be negotiating optics, networking silicon and accelerators with the same counterparty. When a single supplier controls several linked parts of the same system, the buyer’s ability to play suppliers off one another narrows.

    The Nvidia connection adds a complication. Reports describe Lumentum as Nvidia-backed, and Nvidia’s GPUs compete for the same data center budgets as the custom accelerators Marvell designs for cloud customers. How Nvidia’s relationship with Lumentum would survive a change of ownership is one of the most consequential unknowns in the reported deal.

    An $83 Billion Price Tag Against a $5 Billion Credit Line

    Marvell’s 8-K filed October 23 describes a Third Amended and Restated Revolving Credit Agreement led by Bank of America: up to $3 billion in revolving loans, expandable by another $2 billion, replacing a $1.5 billion facility from June 2025. A revolver is a corporate credit card for working capital, and the filing says the proceeds are for working capital and general corporate purposes. It does not mention Lumentum or any specific acquisition.

    One clause is worth noting. Marvell must keep a leverage ratio, broadly debt measured against earnings, of no more than 4.00 to 1. That limit can rise to 4.50 to 1 for the quarter in which an acquisition of at least $500 million closes and for the three quarters after. Such “acquisition holiday” provisions are standard in large-company credit agreements, so the clause signals flexibility rather than intent. Still, even fully drawn, the facility would cover only about 6% of a reported $83 billion price. The bulk of any consideration would have to come from Marvell stock, new bonds or term loans.

    The market’s first reaction reflects that arithmetic. Lumentum’s roughly 15% after-hours jump priced in an expected premium for its holders; Marvell’s after-hours slide suggests its shareholders are weighing the cost, dilution or debt load a deal of this size would bring.

    Who Gains and Who Has Reason to Hesitate

    Lumentum shareholders are the clearest beneficiaries if a premium deal is signed. Marvell would gain a far larger share of spending on each AI network link, a strategy consistent with its earlier purchase of Inphi, which built its optical electronics business.

    The more uncertain group is Lumentum’s other customers. Chip and systems companies that compete with Marvell may be reluctant to depend on a rival’s subsidiary for critical optical parts, and could shift volume to other suppliers. That dynamic could limit how much of Lumentum’s existing business Marvell actually keeps, and it gives competing optics makers a potential opening. Regulators are likely to examine precisely this question: whether the combined company could disadvantage rivals by controlling access to components they need.

    Background

    Marvell Technology designs semiconductors for data infrastructure: networking, storage and custom processors built for large cloud operators. Its 2021 acquisition of Inphi made it a leading supplier of the electronic chips used inside high-speed optical modules, positioning it squarely in the plumbing of AI data centers.

    Lumentum Holdings, spun out of JDS Uniphase in 2015, makes optical and photonic products, including the lasers and modules that carry data as light over fiber. As AI clusters have grown to thousands of interconnected accelerators, demand for these optical links has become a significant part of data center spending, drawing chipmakers toward the photonics business.

    Sources

    Source: Marvell nears $83B deal for Nvidia-backed optics maker Lumentum (Dealroom), reporting the value of Marvell’s talks to acquire Lumentum. Additional reporting: Marvell Is Said to Near Megadeal for Lumentum In AI Push (Bloomberg) and Lumentum Holdings (LITE) Surges 15% After-Hours Amid Acquisition Talks with Marvell Technology (GuruFocus).

    Primary sources: Marvell Technology, Form 8-K filed October 23, 2026: Third Amended and Restated Revolving Credit Agreement.

  • Japan’s $1B Tower Grant Treats AI Data Center Optics as Strategic Chip Capacity

    Japan’s $1B Tower Grant Treats AI Data Center Optics as Strategic Chip Capacity

    TL;DR · 30-second read

    The Short Version

    Artificial intelligence runs in huge computer buildings, and the chips inside must constantly pass data to one another. More and more of that traffic travels as light through glass fiber.

    Tower Semiconductor, an Israeli firm that makes chips for other companies, plans to spend $3 billion expanding its factories in Japan to make the chips behind those connections. Japan’s government pledged $1 billion in grants.

    The twist: this public money is not going to the famous processors that do the thinking. It is going to the plumbing that links them.

    Tower Semiconductor (Nasdaq/TASE: TSEM), the Israel-based analog chip foundry, announced in July a $3 billion investment to expand 300-millimeter silicon photonics, silicon germanium and advanced semiconductor packaging production in Japan. Stocktwits reported that the Japanese government has pledged an additional $1 billion in grants for the expansion, that Tower raised its 2028 targets to $3.6 billion in revenue and $1.2 billion in net profit, and that the shares rose 17% in premarket trading on the day of the announcement.

    On October 13, Tower followed with a 6-K filing announcing an expanded multi-year agreement with Semtech. Under that agreement, Semtech will invest in securing dedicated silicon germanium capacity across Tower’s fabs in the United States, Israel and Japan to supply optical and copper interconnect chips for AI data centers.

    Executive Summary

    The Japan program targets three technologies: silicon photonics, which routes data as light on a silicon chip; silicon germanium (SiGe), a manufacturing process for very fast analog transistors that drive lasers and amplify incoming optical signals; and advanced packaging, which assembles several chips so they behave as one unit. All three serve one job in an AI data center: moving data between processors rather than computing it. Tower will build the expansion on the former Panasonic semiconductor operations in Japan, which it controls through a 51% holding in TPSCo, according to its October 6-K.

    The significance is in what the subsidy is paying for. A $1 billion grant, equal to a third of the stated $3 billion investment figure, has gone to a specialty analog foundry making connectivity chips rather than to a leading-edge logic fab. Tower’s chief executive, Russell Ellwanger, described Japan as having “selected Tower to lead the expansion of these strategically important technologies.” Semtech’s October capacity agreement shows at least one major chip designer is committing money to reserve the same kind of SiGe supply.

    Japan Is Paying for the Data-Moving Layer of AI

    Discussion of AI hardware tends to focus on accelerators, the GPUs and custom processors that do the computing. Japan’s grant is aimed at a different layer. Silicon photonics, silicon germanium and advanced packaging are the technologies that carry data between those processors, across a rack, a row, or an entire campus. Ellwanger’s statement that the government chose Tower to lead “strategically important technologies” is the clearest evidence of how Japan has categorised this capacity. It has treated the interconnect layer as worth public money in its own right, not as an afterthought to compute.

    Semtech’s October 13 announcement explains why this layer has become a capacity question. The company argues that as AI clusters grow, a rising share of their power budget goes to moving data rather than computing it. It adds that each step in transceiver speed, from 800G to 1.6T to 3.2T (modules moving 800 gigabits, 1.6 terabits and 3.2 terabits per second), raises the requirements on the high-speed analog front end that drives and receives the signal. Those front-end parts include transimpedance amplifiers, which convert a light detector’s tiny current into a usable signal, and laser and modulator drivers. Semtech builds them on Tower’s SiGe process. An optical module cannot ship without these parts, however many processors are waiting to be connected.

    The people affected are optical transceiver makers, fabless designers such as Semtech that depend on Tower’s SiGe and silicon photonics processes, and data center operators whose network buildouts depend on module supply. There is a limit to the claim: this is one government backing one foundry. It does not establish a coordinated global policy shift. It does show that at least one major chipmaking nation now ranks AI connectivity capacity alongside the better-known categories of strategic semiconductor spending.

    Semtech’s Deal Puts a Paying Customer Behind the Thesis

    A government grant shows intent, while a customer committing money shows demand. Tower’s October 6-K describes Semtech as investing to secure dedicated SiGe capacity under a multi-year agreement that builds on a 20-year relationship. The capacity spans Tower’s fabs in the United States, Israel and Japan. Semtech presented the agreement as supply assurance for growing design wins in optical and copper interconnects. Its chief executive, Hong Hou, said meeting that demand “requires long-term capacity planning and commitment with our manufacturing partners.”

    Fabless chip companies generally do not commit their own capital to reserve foundry capacity unless they expect supply to be contested. That makes the Semtech agreement the most concrete market signal around the Japan expansion. It also supports the strategic reading of Japan’s grant, because the capacity Japan is subsidising is capacity that customers are paying to secure.

    Some claims in the filing remain unquantified. Ellwanger referred to “a continual expansion of our already leading market share in optical transceiver connectivity,” but neither company gave a share figure, a dollar value for Semtech’s investment, or a volume of reserved wafers. The 224 Gbps-per-lane TIA family that won an award at ECOC 2026 is evidence of technical position, not of contracted volume.

    Why the Expansion Runs Through Existing Japanese Fabs

    According to its October 6-K, Tower owns two fabs in Japan, one 200mm and one 300mm, through its 51% holding in TPSCo. The July plan uses those former Panasonic semiconductor operations rather than a greenfield site. Working inside an existing operation can shorten the path to output, because cleanrooms, utilities and a trained workforce are already in place. The new spending can then go mainly to tooling and process transfer.

    The 300mm detail matters for cost. A 300mm wafer has about 2.25 times the surface area of a 200mm wafer, so each pass through the fab produces more chips. Moving silicon photonics to 300mm signals that Tower expects these optical chips to run at volumes that justify larger-wafer economics. Specialty analog products have historically run on smaller wafers. Japanese capacity also gives Tower’s customers a further production region, and Semtech cited exactly that kind of geographic diversification when it secured capacity across three countries.

    What the 2028 Targets Do and Do Not Establish

    Tower tied the expansion to higher 2028 targets of $3.6 billion in revenue and $1.2 billion in net profit. That implies a net margin of about one-third, which would be a high level of profitability for a foundry. The market reacted quickly. Shares rose 17% premarket on the announcement, after nearly doubling year to date and rising more than fivefold over the prior 12 months.

    Targets are not commitments. Reaching them depends on the Japanese capacity coming online on schedule, qualifying for customers’ products, and filling with orders at favourable pricing, and the timing of the $1 billion in grants also bears on returns. The Semtech agreement supports the demand side of that equation, but no single customer contract covers a $3.6 billion revenue plan. Tower’s third-quarter results on November 2 are the next point at which the company can show progress against these figures.

    Background

    Tower Semiconductor, headquartered in Migdal Haemek, Israel, and listed on Nasdaq and the Tel Aviv Stock Exchange, is a foundry for high-value analog semiconductors. It manufactures chips designed by other companies. Its process platforms include silicon photonics (SiPho), silicon germanium (SiGe), BiCMOS, mixed-signal CMOS, RF CMOS, image sensors and power management. According to its October 2026 filings, Tower operates one 200mm fab in Israel and two 200mm fabs in the United States. It also has two fabs in Japan, one 200mm and one 300mm, held through its 51% stake in TPSCo, and it shares a 300mm facility in Agrate, Italy, with STMicroelectronics.

    The market context is the buildout of AI data centers, where clusters of thousands of accelerators must exchange data at high speed. Optical transceivers and active copper cables carry that traffic. Their analog front-end chips, many of them built on SiGe, are a specialised supply chain separate from the leading-edge logic used in processors. Semtech, a fabless designer of AI data center networking chips, has relied on Tower’s SiGe process for 20 years.

    Sources

    Source: TSEM Stock Soars Premarket: Japan Backs Tower’s $3B AI Chip Manufacturing Push (Stocktwits), coverage of Tower Semiconductor’s $3 billion Japan expansion, $1 billion in government grants and raised 2028 targets.

    Primary sources: Tower Semiconductor Form 6-K, October 13, 2026: Tower Semiconductor and Semtech Expand SiGe Partnership to Scale Power-Efficient Analog Connectivity for Data Centers; Tower Semiconductor Form 6-K, October 13, 2026: Third Quarter 2026 Financial Results and Conference Call announcement; Tower Semiconductor Form 6-K, October 13, 2026: Notice of Extraordinary General Meeting of Shareholders; Exhibit 99.1 to Form 6-K: Proxy Statement, Extraordinary General Meeting of November 25, 2026; Exhibit 99.2 to Form 6-K: Proxy Card, Extraordinary General Meeting of November 25, 2026.

  • Huawei’s 98% Link-Use Claim Pitches the AI Network as a Way to Win Back GPU Time

    Huawei’s 98% Link-Use Claim Pitches the AI Network as a Way to Win Back GPU Time

    TL;DR · 30-second read

    The Short Version

    Huawei, the Chinese technology company, showed European customers an upgraded line of networking equipment in Florence this week.

    Its main pitch: the cables and switching boxes that link the chips training artificial intelligence often leave those chips waiting in traffic. Huawei says its new design keeps 98% of the connections busy, makes the computing more than 7% faster, and needs 30% less equipment.

    It is also selling security built into the hardware, at a time when European buyers are rethinking how much they rely on Chinese suppliers. The performance figures are Huawei’s own.

    Huawei unveiled an upgraded Stellar Intelligent Network portfolio for Europe at the Huawei Network Summit 2026 Europe in Florence on October 17, 2026, issuing the announcement in Italian, Dutch, Turkish, Slovak and Russian. The update covers three product lines: Stellar Intelligent Fabric for data center networks, Stellar Intelligent WAN for links between sites, and Stellar Intelligent Campus Network for offices and branches. It also adds new eKit networking products for small and mid-sized businesses.

    Huawei said its self-developed “100T” switches enable a two-tier, multi-plane data center architecture that quadruples network capacity while using 30% fewer switches and optical modules, and that its Network Packet Load Balancing (NPLB) algorithm lifts link utilization to 98% and AI training and inference performance by more than 7%. The same week, Huawei pressed the same AI-era networking message in Vienna. It announced a network-security partnership with Macau operator CTM at Network X 2026. At UBBF 2026 on October 13, it highlighted large-scale commercial use of its NG WAN solution, where data communication product line vice president John Cai called AI the decisive opportunity for networks.

    Executive Summary

    The Stellar upgrade is a broad enterprise networking refresh, but its most consequential claim concerns AI data centers. Huawei is arguing that the network connecting AI accelerators is not passive plumbing. In its account, the network determines how much useful work expensive compute actually does, and a better-balanced fabric recovers performance that would otherwise be lost to congestion.

    The second thread is security. Nearly every product line in the upgrade carries a built-in security feature: an intrusion-detection card, an integrated quantum key distribution card, post-quantum cryptography across wired and wireless links, and threat detection running directly on switches. Huawei is presenting security as a property of the hardware rather than an add-on. That is a pointed choice in a European market where buyers have been reassessing their exposure to Chinese network suppliers.

    For buyers, the announcement is a set of specific, testable numbers rather than vague promises. However, none of the figures comes with a stated baseline, cluster size or independent benchmark. The useful response is to treat them as a procurement checklist, not a conclusion.

    The Network as a Compute Multiplier

    The headline claim rests on a mechanism that AI cluster operators know well. Large training jobs split work across many GPUs or other accelerators, which must constantly exchange intermediate results in lockstep. When those exchanges hit a congested link, every chip in the job waits for the slowest transfer. Conventional data center load balancing assigns whole traffic flows to paths, typically by hashing. That approach can pile several large flows onto the same link while neighbouring links sit underused. Huawei’s NPLB algorithm, as its name indicates, balances at the level of individual packets, spreading traffic across all available paths. Huawei says this drives network-wide link utilization to 98% and raises training and inference performance by more than 7%.

    If that figure holds in a customer’s environment, the economics are straightforward. A 7% throughput gain on a fixed cluster is roughly equivalent to adding 7% more compute without buying more accelerators, power or floor space. Accelerators are the dominant cost in an AI cluster, so recovered utilization is worth far more than the network’s own price tag. That is the commercial logic Huawei is selling: spend on the fabric to get more out of the chips.

    The claim needs context before it can be priced. Huawei has not said what the 7% is measured against, on which models, at what cluster scale, or with which accelerators and network adapters. Packet-level spraying also typically requires endpoints that can handle packets arriving out of order, so the gain may depend on the rest of the stack. Operators should treat 98% and 7% as hypotheses to reproduce on their own workloads.

    Fewer Tiers, Fewer Optics

    The fabric’s other claim targets hardware count. Huawei says its self-developed 100T switches support a two-tier, multi-plane architecture that quadruples network capacity while cutting the number of switches and optical modules by 30%. In industry usage, the 100T label denotes switching capacity on the order of 100 terabits per second per device. Higher-capacity switches let a network reach the same number of servers with fewer layers.

    That matters because optical modules — the transceivers that convert electrical signals to light at each end of a fibre link — are a large and growing share of AI network cost and power. Every extra tier between servers adds more links, more transceivers and more hops. A flatter design reduces all three, which is why Huawei frames the 30% reduction as a total cost of ownership saving. The affected parties are clear: cluster builders weighing capex, and optics suppliers whose unit volumes shrink when topologies flatten.

    As with the performance claim, the comparison point is unstated. Quadrupled capacity and 30% fewer components relative to which previous design, at what scale, are the numbers a network architect would need. Huawei describes a three-level energy-saving framework covering devices, network and operations aimed at European environmental and carbon requirements, but it attached no energy figures to it.

    Security Built Into the Box, Sold Into a Cautious Market

    Security runs through every line of the upgrade. On the WAN side, Huawei cites an intrinsic security card that it says detects and blocks intrusions in real time with over 95% precision. It also cites what it calls the first integrated quantum key distribution (QKD) card, which uses quantum physics to exchange encryption keys and which Huawei says cuts deployment cost by more than 60%, plus APN6-based isolation of sensitive data. On campus networks, Huawei offers post-quantum cryptography (PQC), meaning encryption designed to resist future quantum computers, across wired and wireless connections. It also says local inference on switches cuts threat-detection time from 5 minutes to under 10 seconds, and that more than 20 inline algorithms in branch gateways detect 95% of unknown threats. For distributed AI, Huawei says layered model partitioning keeps raw data on premises.

    Embedding security in network hardware has real operational appeal. It can mean fewer separate appliances, fewer integration points and faster response at the edge. It also concentrates trust in a single supplier. For any vendor, European buyers increasingly ask who can verify the security functions, which cryptographic standards they implement, and whether independent certification is available. That question carries particular weight for Huawei, given the scrutiny Chinese network suppliers have faced in Europe. Huawei’s Vienna announcements point the same way: the CTM partnership is framed around strengthening network security in the AI era. The company is making security its lead argument rather than a defensive footnote.

    WAN and Campus Claims Aimed at the Budget Line

    Beyond the data center, the claims target recurring costs. Huawei says its lossless Starnet algorithm delivers more than 95% cross-region compute efficiency and cuts dedicated line costs by 60%, an argument aimed at enterprises running AI workloads across distant sites. On campus networks, it claims an 84% reduction in wireless interference, double the industry-average performance, a 30%-plus cut in total cost of ownership for 10 Gbps campus networks via multi-GE switches and flexible Right-to-Use licensing, and mean time to repair under 5 minutes.

    The small-business lineup is more concrete. It comprises Aura10/AR580 converged gateways with antivirus and intrusion prevention, S630 10GE core switches that can manage up to 1,000 access points for hotels, and AP362H access points with 1.6x antenna gain for retail. These are product specifications rather than benchmark claims, and they show Huawei pushing the same security-first message down-market. Its NG WAN commercial momentum, presented at UBBF 2026, suggests the WAN story is being sold to operators and enterprises alike.

    Background

    Huawei is a Chinese technology company that supplies telecom operator equipment and enterprise networking gear. Stellar Intelligent Network is its umbrella for enterprise networking, spanning the data center fabric that links servers and AI accelerators, the wide-area networks that connect distant sites, and the campus networks inside offices, hotels and shops.

    AI data centers have made networking a first-order design problem. Training large models requires thousands of chips to exchange data constantly, so congestion, hardware count and optical transceiver costs now weigh on cluster economics alongside the chips themselves. Huawei presented the Stellar upgrade at its Florence summit days after related Vienna events, Network X 2026 and UBBF 2026, where it framed AI as the defining opportunity for network vendors and operators.

    Sources

    Source: Huawei presenta la soluzione Stellar Intelligent Network aggiornata per l’Europa, basata sulla filosofia della “connettività sicura e intelligente” — Huawei’s announcement of its upgraded Stellar fabric, WAN and campus networking portfolio for Europe, also issued in Dutch, Turkish, Slovak and Russian.

    Primary sources: Безопасность на основе искуственного интеллекта для Макао — CTM и Huawei заключают партнерское соглашение для усиления безопасности сети в эпоху ИИ (CTM and Huawei network-security partnership, Network X 2026, October 17, 2026); Решение Huawei NG WAN для крупномасштабного коммерческого использования стимулирует рост операторских сетей (Huawei NG WAN large-scale commercial use, UBBF 2026, October 15, 2026).

  • AI Optics Planning Reaches Analog Fabs: Semtech Pays to Reserve SiGe for 1.6T and 3.2T

    AI Optics Planning Reaches Analog Fabs: Semtech Pays to Reserve SiGe for 1.6T and 3.2T

    TL;DR · 30-second read

    The Short Version

    The giant computers behind artificial intelligence spend a growing share of their electricity just moving data between chips, not doing the math.

    Semtech, a California chipmaker, builds the small parts that push those signals down cables and turn them into light and back. It is now paying its long-time contract factory, Tower Semiconductor, to set aside production space for it in the United States, Israel and Japan.

    The surprise: the artificial intelligence boom is not only a scramble for famous high-end processors. Even the plainer chips that connect them are being booked under multi-year deals.

    Tower Semiconductor and Semtech announced on October 13, 2026 an expanded multi-year agreement under which Semtech will invest in securing dedicated silicon germanium (SiGe) manufacturing capacity at Tower’s fabs in the United States, Israel and Japan. The capacity supports Semtech’s signal-integrity chips, including amplifiers, drivers and equalizers used in optical and copper links inside AI data centers, at link speeds from 800G and 1.6T today toward 3.2T.

    Tower filed the joint announcement with US regulators in a 6-K the same day. The companies did not disclose the size of Semtech’s investment, the wafer volumes reserved or the agreement’s length beyond “multi-year.” In a separate 6-K, Tower said it will report third-quarter 2026 results on November 2.

    Executive Summary

    Semtech and Tower have worked together for 20 years, and Semtech’s high-speed analog parts already run on Tower’s SiGe process. What changes now is the commercial structure. Semtech is no longer only buying wafers as it needs them. Under the expanded agreement it will invest to secure capacity dedicated to its products, spread across three regions for what both companies describe as supply resilience and geographic diversification.

    The announcement matters because it reaches a part of the AI supply chain that gets little attention. Most discussion of AI chip supply focuses on leading-edge logic: accelerators and the digital signal processors built at the most advanced nodes. Every optical or copper link between those accelerators also needs an analog front end, the circuitry that physically drives and receives the signal. Semtech makes those parts on a specialty process at a specialty foundry, and it is now committing money to guarantee its share of that foundry’s output.

    The release frames the demand case in terms of power. It says AI’s next constraint is power, that a growing share of it is spent moving data rather than computing it, and that each step from 800G to 1.6T to 3.2T raises the requirements for bandwidth, reach and power per bit.

    Why a Link Chip Needs Its Own Fab

    An optical transceiver, the pluggable module that converts electrical signals to light and back, contains more than a laser and a digital chip. It also holds a transimpedance amplifier (TIA), which turns the faint current from a photodetector into a usable voltage, and a modulator or laser driver, which pushes the outgoing signal hard and cleanly enough to survive the trip. These are analog circuits. Their performance depends on transistor speed and noise rather than on how many logic gates fit on a die. Silicon germanium, which adds germanium to silicon transistors to make them faster, is a long-standing process choice for this kind of high-frequency analog work. It is also a different manufacturing line from the leading-edge digital processes that dominate AI supply headlines.

    The speed roadmap raises the stakes for these parts. Tower’s 6-K describes each step from 800G to 1.6T to 3.2T as raising the requirements for bandwidth, reach and power per bit, and as increasing the importance of the high-speed analog front end. Semtech’s recent ECOC 2026 award was for a 224 Gbps-per-lane TIA family built on Tower’s SiGe. That per-lane rate shows how much performance these analog parts must now deliver. Several of the architectures Semtech lists, including linear pluggable optics (LPO) and linear receive optics (LRO), are designed to remove or reduce the power-hungry digital signal processor in the module. That design shifts more of the signal-conditioning job onto the analog components Tower manufactures.

    Paying for Dedicated Capacity Is the Operational Signal

    The core of this announcement is the commitment. According to the 6-K, Semtech “will invest in securing dedicated silicon germanium (SiGe) capacity at Tower.” Semtech CEO Hong Hou tied that directly to supply planning: “Meeting that demand requires long-term capacity planning and commitment with our manufacturing partners.” A chip designer does not usually spend its own money to reserve capacity at a 20-year supplier unless it judges that buying wafers as needed carries too much supply risk for the volumes it expects.

    That is the sense in which AI interconnect planning has reached the analog fab. Neither company says SiGe capacity is short today, and the release should not be read as evidence of a shortage. It points instead to “strong design wins” and demand “expected to grow across a broadening base of customer programs.” The operational conclusion is narrower and better supported. For at least one major supplier of AI link components, securing specialty analog wafer capacity has become a deliberate investment and a multi-year commitment, the kind of arrangement more commonly associated with leading-edge logic.

    The people affected are those downstream of Semtech’s parts: transceiver and active-copper-cable makers, and the data center operators deploying 1.6T networks. For them, the availability of the analog front end now depends partly on how much capacity Semtech has locked in at Tower. The question for other Tower customers in data center optics is whether they will need similar commitments to secure their own supply.

    What Each Side Gets, and What Is Still a Claim

    For Semtech, the main benefits are assurance and geography. The capacity spans Tower’s fabs in the United States, Israel and Japan, so a disruption at one site need not halt supply of a product line. Semtech also positions its portfolio across both media in the data center, with copper links within the rack and optical links across the facility. Its active copper cable and linear equalizer products give it exposure if short-reach copper persists longer than some optics-first forecasts assume.

    For Tower, a customer-funded capacity commitment adds visibility in a high-growth segment and ties a key customer to its multi-generation technology roadmap. Tower CEO Russell Ellwanger described the partnership as ensuring “a continual expansion of our already leading market share in optical transceiver connectivity.” The release offers no market-share figures to support that statement, so it should be read as the company’s characterization. Tower’s third-quarter results on November 2 are the first chance to see whether data center demand appears in the numbers.

    Background

    Tower Semiconductor, based in Migdal Haemek, Israel, describes itself as the leading foundry of high-value analog semiconductor solutions. It manufactures chips designed by other companies on specialty processes, including SiGe, silicon photonics (SiPho), BiCMOS and power management. It owns one 200mm fab in Israel and two in the US, operates two fabs in Japan (200mm and 300mm) through its 51% stake in TPSCo, and shares a 300mm facility in Agrate, Italy with STMicroelectronics.

    Semtech, based in Camarillo, California, supplies high-performance semiconductors for AI data center networking and Internet of Things devices. Its data center products cover both optical links across a facility and copper links within a rack, a category growing as AI clusters move toward 1.6T and 3.2T network speeds.

    Sources

    Source: Tower Semiconductor, Semtech Expand Partnership to Secure AI Chip Capacity (MarketScreener), coverage of the expanded SiGe capacity agreement for AI data center interconnects.

    Primary sources: Tower Semiconductor Form 6-K, October 13, 2026: Tower Semiconductor and Semtech Expand SiGe Partnership to Scale Power-Efficient Analog Connectivity for Data Centers; Tower Semiconductor Form 6-K, October 13, 2026: Tower Semiconductor Announces Third Quarter 2026 Financial Results and Conference Call.

  • Lumentum’s Sellout to 2029 Turns AI Network Optics Into a Long-Lead Item

    Lumentum’s Sellout to 2029 Turns AI Network Optics Into a Long-Lead Item

    TL;DR · 30-second read

    The Short Version

    Lumentum, a company most people have never heard of, makes tiny lasers. They turn computer data into flashes of light so it can travel through glass fiber between the thousands of chips inside the huge buildings that run artificial intelligence.

    Its chief executive says its factories are fully booked into early 2029. For some products, about seven in ten of the parts customers want will still be missing by the end of next year.

    The race to build artificial intelligence is usually told as a story about chips. This shows a small, unglamorous part can set the pace too.

    Lumentum Holdings, the San Jose-based maker of lasers and optical components that carry data inside AI data centers, says its optical component production is fully booked into early 2029. Bloomberg reported on October 8, 2026 that the company, which it described as Nvidia-backed, sees its opto-parts capacity sold out to 2029. GuruFocus reported that chief executive Michael Hurlston said roughly 70% of demand for certain products will remain unmet by the end of next year, and about 30% of demand for other products will go unfilled until 2028. GuruFocus put the company’s market value at about $95 billion.

    The comments came two days after Lumentum filed its definitive proxy statement with the SEC on October 6. In that filing, the company called fiscal 2026, the year ended June 27, 2026, a “landmark year.” It reported record shipments of high-speed transceivers, transport products and laser chips, and said it had started production of next-generation 1.6T modules. It also said its optical circuit switch business had passed a $100 million quarterly revenue run rate.

    Executive Summary

    Lumentum is telling the market that it cannot make enough of the lasers and optical modules that link AI accelerators together. Its production is committed into early 2029. For some product lines, most of the demand it sees for next year will go unserved. Those are the parts that convert electrical signals to light and back again, so that thousands of GPUs can behave like one machine.

    The significance for infrastructure builders is timing. AI capacity plans have mostly been framed around securing GPUs, power and buildings. A multi-year sellout at a major laser supplier adds optical interconnect to the list of items that have to be reserved well before a cluster is commissioned. It also matters because Lumentum’s October 6 proxy statement argues that optics will move deeper into the server rack, which would raise the amount of optics each AI system needs.

    What is substantiated is narrower than the headline figure suggests. The company has stated a booking horizon and percentages of unmet demand. It has not disclosed which products are affected, who holds the committed capacity, or whether those commitments are binding orders or allocations.

    Why a Laser Sellout Becomes a Cluster Schedule

    Modern AI training and inference clusters are networks first. Every server in a scale-out cluster, meaning the fabric that connects many servers across a data center hall, reaches its switches through optical transceivers. A transceiver is a plug-in module that turns electrical data into light for the fiber run and back again. Each one depends on laser chips. Lumentum makes both the chips, including EML (electro-absorption modulated) and CW (continuous-wave) lasers grown on indium phosphide wafers, and finished modules such as the 1.6T (1.6 terabits per second) parts it began producing in fiscal 2026. A rack of GPUs that arrives without its optics cannot join the cluster, however many accelerators it holds.

    That is the mechanism behind the long-lead label. If Lumentum’s production is booked into early 2029, a buyer placing a new order with Lumentum today is effectively scheduling delivery more than two years out. Hurlston’s figures sharpen the point. Roughly 70% of demand for certain products will still be unmet by the end of next year, and about 30% of demand for other products will be unfilled until 2028. For hyperscalers, neocloud operators and the developers whose tenants need a working network fabric on a fixed go-live date, optics now has to be secured early, the way developers already secure grid capacity and electrical equipment.

    The claim has limits that matter. Lumentum is one supplier, and it has not said what share of the market its booked capacity represents. Buyers with diversified supply may be able to source modules or lasers elsewhere, if other vendors have room. The defensible conclusion is that optical interconnect has become a dated procurement constraint for anyone depending on Lumentum’s capacity. It is not proof that every AI cluster will be delayed, and it does not mean GPUs are no longer scarce.

    Scale-Up Could Raise the Optics Count Per Rack

    The proxy statement explains where Lumentum thinks demand goes next. Its results to date came from scale-out networks and scale-across networks, which link separate data center sites. The company says it is most excited about scale-up, the tight connections between GPUs inside a single rack or system. Today those links are largely electrical, over copper. Lumentum describes a shift toward optical connectivity within the rack that “has the potential” to drive its performance for years.

    If that shift happens at scale, the number of lasers needed per GPU rises. And it would do so at a moment when Lumentum says its existing capacity is already spoken for. The company also points to co-packaged optics (CPO) and near-packaged optics (NPO), designs that place the optical engine on or next to the switch or processor package. Lumentum says these designs eliminate the need for a digital signal processor (DSP), cutting power use and cost. They also depend on the ultra-high-power lasers Lumentum is developing. The scale-up thesis is the company’s forecast rather than a disclosed order book. Still, it explains why a supplier might describe demand in years rather than quarters.

    Optical Circuit Switches: A Reported Number and a Forecast

    Lumentum’s optical circuit switch (OCS) line is the clearest growth figure in the filing. An optical circuit switch redirects beams of light between fibers without converting the signal to electrical form, which can save power in large AI networks. The company says the platform has passed a $100 million quarterly revenue run rate. That is a reported operating figure.

    The next claim is of a different kind. Lumentum says it expects the product line to generate “revenue in the billions in upcoming years,” citing technological differentiation and a first-mover advantage. That is a management expectation with no stated timeline or customer base attached. Readers should weigh it as guidance, not as booked business. The same distinction applies to the 2029 sellout. A booking horizon is a strong demand signal, but its value to the company depends on the pricing and terms attached to it, and Lumentum has not disclosed either.

    Who Gains and Who Waits

    A supplier with sold-out capacity usually has leverage on price and terms. That can support margins unless long-term agreements locked in earlier pricing. Customers who committed early hold an advantage over those arriving now. In practice, AI operators that reserved optics alongside their GPU allocations are better placed than late entrants or smaller cloud providers. Module assemblers that buy laser chips rather than make them are also exposed to whatever allocation decisions upstream laser suppliers make.

    For data center developers, the practical effect is coordination. A powered, cooled hall delivers no revenue until the tenant’s network is live, and that network now has its own multi-year lead time. Lumentum says it has expanded wafer capacity across its global indium phosphide fabs and qualified both CW and EML process flows. It has not said how much capacity that adds or when the backlog eases.

    Background

    Lumentum was spun off from JDS Uniphase in 2015 and is headquartered in San Jose, California. It makes lasers, optical transceivers, transport products and, more recently, optical circuit switches. Its cloud and networking business supplies the components that move data between servers and switches in large data centers. The company’s fiscal year ends in late June, and fiscal 2026 closed on June 27, 2026.

    Demand for optics has risen with the size of AI clusters. Training a large model means synchronizing thousands of accelerators, and every link between them needs bandwidth. Network speeds have stepped up from 400-gigabit to 800-gigabit and now 1.6-terabit modules. Many of those modules rely on laser chips made from indium phosphide, a semiconductor material produced in specialized fabs that take time and capital to expand.

    Sources

    Source: Lumentum Holdings Inc. Faces Soaring Demand for AI Data Center Components (GuruFocus), which reported CEO Michael Hurlston’s remarks that optical component production is booked into early 2029. Also: Nvidia-Backed Lumentum Sees Opto-Parts Capacity Sold Out to 2029 (Bloomberg).

    Primary sources: Lumentum Holdings Inc., DEF 14A Definitive Proxy Statement, filed October 6, 2026 (SEC EDGAR).

  • Nokia’s 76.6x P/E Hinges on Cloud and AI Margins, Not 2028 Sovereign Satellites

    Nokia’s 76.6x P/E Hinges on Cloud and AI Margins, Not 2028 Sovereign Satellites

    TL;DR · 30-second read

    The Short Version

    Nokia, the Finnish company behind much of the equipment that runs phone and internet networks, is teaming up with satellite company ICEYE. Together they plan to build communication satellites that governments control themselves. The goal is to keep soldiers, border guards and emergency crews connected when ground networks are knocked out. The first satellites are planned for 2028.

    The satellites will not change Nokia’s finances soon. Investors currently pay about 77 times the company’s yearly profit for its shares. That price leans on Nokia more than tripling the profit it keeps from each euro of sales, mostly through software, cloud and artificial-intelligence networking, not satellites.

    Simply Wall St reported that ICEYE and Nokia Oyj have announced a partnership to build secure, sovereign broadband satellite communication systems in low Earth orbit (LEO) for governments. The systems target defense, border security, emergency services and disaster response, and the first satellites are planned for launch in 2028. The pairing combines Nokia’s secure networking and defense-focused portfolio with ICEYE’s sovereign satellite mission expertise.

    The partnership lands against a demanding set of expectations for Nokia. Consensus estimates assume 4.8% annual revenue growth over three years, an operating profitability shift from 3.4% to 11.2%, and earnings rising from €702.0 million today to €2.6 billion by 2029. The shares trade at 76.6 times current earnings.

    Executive Summary

    Nokia and ICEYE are building sovereign broadband satellite systems for governments: networks a state can own and control rather than rent from a commercial provider. These systems are meant to keep working when traditional terrestrial networks are unavailable or compromised. The first launches are planned for 2028, which puts the partnership squarely in the category of long-dated strategic positioning rather than near-term revenue.

    That timing is what makes the deal analytically interesting. Nokia’s valuation already assumes a large earnings recovery by 2029. The satellites arrive too late to do much of that work. Whatever justifies a 76.6x price-to-earnings multiple (the share price divided by a year of per-share profit) has to come from margin repair in Nokia’s existing network, cloud and software businesses. This includes the cloud and AI domains where Nokia competes as a challenger.

    For government buyers and network planners, the more durable signal is architectural. Sovereignty requirements that have shaped terrestrial fiber and mobile procurement are now being written into the satellite layer as well.

    Sovereignty Moves From the Ground Into Orbit

    Governments have spent years tightening control over the networks that carry sensitive traffic. That has meant specifying who builds them, who operates them and where the data travels. The Nokia and ICEYE partnership extends that logic upward. A sovereign LEO system is a constellation of satellites orbiting relatively close to Earth, operated under a government’s own control. It gives defense, border and emergency agencies a communications path that does not depend on ground infrastructure that may be damaged, congested or compromised.

    In practice, such a system works as a resilient alternate route. It carries traffic between field units and core networks when fiber or mobile backhaul (the links that connect local sites to the wider network) is unavailable. For planners, this means sovereign requirements no longer stop at the edge of the terrestrial network. A government that insists on controlled fiber and controlled mobile cores will increasingly expect the same of its satellite links. Vendors that can offer both layers under one security model gain an advantage in those procurements.

    Nokia’s contribution is its secure networking and defense portfolio. ICEYE brings sovereign satellite mission expertise. The combination is coherent on paper. The practical test will be whether the two can deliver an integrated system that governments buy as a single sovereign capability.

    Why 2028 Satellites Cannot Carry a 76.6x Multiple

    Consider the mechanism behind Nokia’s valuation. Current profit is €702.0 million on a 3.4% margin, and the shares trade at 76.6 times those earnings. The US Communications industry, by comparison, trades at 25.7x. Consensus forecasts close that gap through earnings growth, not satellites. They project revenue of €23.5 billion and earnings of €2.6 billion by 2029, with margins reaching 11.2%. On those forecasts, matching analyst scenarios would put Nokia on 22.0x earnings in 2029.

    That is a near-quadrupling of profit within roughly three years. The ICEYE satellites are not scheduled to launch until 2028. Even on schedule, an initial constellation entering service that late can contribute little to 2029 earnings. Viewed against the forecasts, the partnership is best read as an early option on government-grade connectivity, not a visible earnings driver in the next few years. Consensus fair-value work still points to a 21% potential downside from the current price. In other words, even the full forecast margin recovery does not obviously support today’s multiple.

    Per-share math makes the bar higher still. Forecasts assume the share count rises 3.42% a year for three years, which spreads any profit increase across more stock. The figures are discounted at a 7.87% rate to reach today’s implied value. For investors, the conclusion is straightforward. The satellite deal adds strategic breadth, but the multiple depends on margin expansion elsewhere.

    Where the Margin Has to Come From: Cloud, Software and AI

    To reach 11.2% margins, Nokia’s higher-margin segments and software-heavy offerings must carry more weight in the portfolio. Those include Network Infrastructure, Cloud and Network Services, and patent licensing. The recent Zain KSA rollout of Nokia Deepfield Cloud Intelligence illustrates the model. This network analytics software gives operators visibility and control when traffic spikes or incidents hit critical services. Recurring software and analytics revenue of this kind can lift the margin mix if execution holds.

    Nokia’s position in these markets cuts both ways. The company is a challenger in some cloud and AI domains. If competitors offer similar analytics at lower prices or with deeper hyperscaler partnerships (ties to the giant cloud operators such as those building AI data centers), Nokia may capture less of the benefit. The most optimistic published scenario leans hard on exactly this AI exposure. It projects revenue of €26.1 billion and earnings of €3.5 billion by 2029, assuming faster cloud and AI adoption. The gap between that case and the €2.6 billion consensus is, in effect, the AI networking question.

    The Mobile Drag Beneath the Story

    The near-term swing factor remains Mobile Networks, a segment facing competition and flat demand. Investors are watching for faster growth in cloud, fiber and private wireless to offset that pressure. Currency headwinds, pricing pressure, market share loss and earnings volatility are all live risks to the forecast path. None of these is addressed by a satellite program that begins launching in 2028. The ICEYE deal broadens what Nokia can sell to governments, but the investment case over the next three years still rests on the terrestrial and software businesses.

    Background

    Nokia Oyj is a Finland-based supplier of mobile, fixed and cloud network solutions, selling into North and Latin America, Greater China, India, Asia Pacific, Europe, the Middle East and Africa. Its business spans Mobile Networks, Network Infrastructure, Cloud and Network Services, and a patent licensing operation. Mobile Networks has faced competition and flat demand, so the company’s growth ambitions increasingly center on cloud, fiber, private wireless and software.

    Low Earth orbit constellations have become a major focus of the communications industry because they can deliver broadband-like service from space. Governments, meanwhile, have pushed for sovereign control over the networks that carry defense and emergency traffic. The Nokia and ICEYE partnership sits where these two trends meet.

    Sources

    Source: Is Government LEO Partnership Altering The Investment Case For Nokia (HLSE:NOKIA)? (Simply Wall St) — on the Nokia and ICEYE sovereign LEO satellite partnership and the analyst forecasts behind Nokia’s valuation.