DK Street Journal

Three companies that sell the electronics sitting at the edge of a broadband network reported June quarters within a month of one another, and the results had almost nothing in common. That is the finding. The industry label implies a shared cycle; in practice each vendor is paid by a different customer — a cable operator, a hyperscale data-center buyer, or a rural telephone company waiting on Washington — and only two of those three are writing checks today.

The one being paid

Harmonic sells CableOS, software that replaces the refrigerator-sized cable modem termination systems in an operator's headend, to companies such as Comcast and Charter. Since selling its video division to MediaKind on 16 June for $137.9m, it is a pure-play broadband vendor. June-quarter broadband revenue was $133.5m, up 54% year on year and above its own guidance. Backlog plus deferred revenue reached $587.6m, up 71%, with 73% due to convert inside a year.

The demand is not subsidy. Charter is spending roughly $11.4bn of capital this year and plans to upgrade 35% of its footprint to the DOCSIS 4.0 standard, work operators deferred in 2025 while waiting on new Broadcom silicon. Harmonic has been named market-share leader in virtual CMTS by Dell'Oro Group, with more than 60 operators deployed against CommScope, Casa Systems and Vecima. Concentration is easing too: revenue from customers outside its two largest operators reached almost $50m, up 44%.

Management raised full-year guidance twice, to $505-525m of revenue and $0.67-0.75 of earnings per share. Consensus still carries $490m and $0.642 — below the floor of the company's own guide. At 19.57x forward earnings and 5.42x trailing gross profit, Harmonic is the most expensive of the three, on the lowest estimates.

The one that is two businesses

Adtran, a Huntsville, Alabama vendor that also controls the German optical maker formerly known as ADVA, contains both cycles at once. In the June quarter optical networking grew 22% to $109.7m, and revenue from hyperscalers buying data-center interconnect gear rose 97%. Meanwhile access and aggregation fell 5% and subscriber solutions dropped 14% sequentially. Total revenue growth decelerated across four quarters from 22.7% to 6.1%.

The good half did not carry the bad. Gross margin slid from 39.5% to 37.0% as supply constraints pushed mix toward cheaper pluggable optics, and operating income swung to -$10.1m. Adtran had already pre-announced revenue of $280-282m against its own $283-303m guide, blaming one large customer's project delay; the shares fell 14% that day. Five sessions since late May account for essentially the entire 45% three-month decline. At 1.48x trailing gross profit it is the cheapest of the three, with $79.2m of cash, a new $350m revolving credit facility led by JPMorgan, and a standing German profit-transfer agreement carrying potential minority buyout obligations near $352m.

The one still waiting

Calix sells access systems and a subscription cloud platform to small and rural operators. Revenue grew 21.3% to $293.3m, the fourth straight quarter above 20%, with record contracted backlog of $386m and operating income of $21.8m against $0.4m a year earlier. Notably, the hardware line grew faster than the software line — appliance revenue $243m, up 23%; software and services $50m, up 16% — which inverts the usual platform narrative.

What has broken is margin. Gross margin has fallen three quarters running to 54.6%, and the September quarter is guided to a 52% midpoint. The cause is memory: prices for the DRAM used in broadband gear have risen sharply as AI servers absorb capacity, taking memory past 20% of a mid-range router's bill of materials from roughly 3%, and Gartner expects DRAM prices up 47% this year. Calix is moving to monthly surcharges and calls the third quarter the bottom. Harmonic bought its full-year memory early.

The federal Broadband Equity, Access and Deployment (BEAD) program, long treated as Calix's catalyst, was restructured: about 66% of eligible locations now go to fiber, 21% to low-earth-orbit satellite and 11% to fixed wireless, saving roughly $21bn. Construction largely runs 2026 to 2028, with first connections in 2027. Management calls satellite a 5-10% threat to its addressable market.

What the shares did

Calix spent 148 consecutive sessions from 13 January in a firmly downward trend, its 50-day average below its 200-day; that ended on 19 August, on volume many multiples of the prior day's. It is 15.2% above its June low and 28.1% lower over twelve months. Harmonic is up 39.6% over the year but 26.6% below its high. Adtran round-tripped an entire optical re-rating, from $8.59 to $19.40 and back to $8.01.

The setup

Where it stands — Cable capex and hyperscaler optics are funding orders now; subsidized rural fiber shipments sit in 2027.

Would confirm — Calix September gross margin at or above the 52% guided midpoint, with fourth-quarter margin guided higher.

Would invalidate — Harmonic's backlog falling below $500m, or full-year revenue landing under the $505m guided floor.

Watch next — Harmonic's investor day on 15 September 2026; Adtran's September quarter, guided to $275-295m.

Valuation — Calix 23.56x forward against 51.33x trailing; Harmonic 19.57x forward; Adtran 1.48x gross profit, 6.99x EV/EBITDA.

Ichor Holdings assembles precision gas and chemical delivery panels — welded fluid modules, machined blocks, finished subsystems — to other companies' engineering drawings, then ships them into the wafer-processing tools that Lam Research and Applied Materials sell to chipmakers. Those two customers were 76% of Ichor's 2025 revenue. Its order book is therefore those two firms' build plans, one quarter early — not an independent read on artificial-intelligence demand.

On 3 August Ichor raised full-year revenue growth guidance to at least 30%, guided the September quarter to $315-345m, and said customers were placing orders six months out — a pattern management called abnormal. June-quarter revenue rose 22.7% to $294.8m, and gross margin recovered to 13.9% from a 4.6% trough last autumn. The shares fell 12.8% on a revenue miss management traced to a flow-controller shortage that pushed shipments into July.

The comparison that matters

Ultra Clean Holdings does the identical thing. It contests the same gas-delivery, frame and subsystem-integration content, and Applied and Lam are the largest customers of both firms. In the June quarter Ultra Clean grew revenue 24.3% to $644.9m at a 16.1% gross margin — faster growth, fatter margin, same demand signal.

Ichor trades at 20.8x trailing and 17.0x forward gross profit. Ultra Clean trades at 9.7x and 7.8x. Gross profit is the honest lens here because Ichor lost money over the past twelve months, leaving its trailing price-to-earnings ratio meaningless. Ichor's own multiple has already come down hard: 43.6x at the 30 June peak, 28.5x in mid-May, 20.8x now. It was about 5.5x a year ago. The de-rating is genuine and it is not obviously finished.

MKS: the business improved, the multiple halved

MKS supplies vacuum and pressure control, radio-frequency power delivery, plasma and reactive-gas products and lasers — and, through its 2022 purchase of Atotech, electroplating chemistry for printed circuit boards and advanced packaging. Semiconductor was $554m of June-quarter revenue, or 44%; Electronics & Packaging contributed $381m, up 44%. Group revenue rose 28.3% to $1.248bn and operating margin reached 20.1% against 13.9% a year earlier.

The leveraged balance sheet from Atotech is shrinking rather than compounding. Net leverage fell to 3.0x adjusted EBITDA from 4.0x, helped by a $100m voluntary term-loan prepayment, and non-GAAP interest expense dropped to $33m from $46m. Management says chemistry demand is the strongest it has ever seen, with AI-grade chemistry now 15-20% of chemistry revenue against roughly 5% in 2024.

MKS now trades at 9.9x trailing gross profit, below the 12.1x of mid-May and far under 17.8x at the June peak. Its forward price-to-earnings ratio of 21.4x is half the 43.1x trailing figure — consensus has 2026 earnings of $13.13 a share rising to $17.50 in 2027.

Axcelis: the one where the numbers agree with the price

Axcelis builds ion implanters, which drive dopant atoms into wafers, sold largely into mature-node automotive, industrial and silicon-carbide power customers. Revenue rose 10.6% to $215.2m, but gross margin slipped to 42.4% from 47.0% in the December quarter and operating income fell 30%. Backlog was $452m on a book-to-bill of roughly 1x. China rose to 46% of revenue from 40%.

Its forward price-to-gross-profit of 10.7x sits above its trailing 10.4x — consensus expects gross profit to shrink. Axcelis is also the only one of the three with a binary event pending: its $4.4bn all-stock merger with Veeco has cleared US and European review and both shareholder votes, but China's competition regulator moved the case into an in-depth review.

What actually moved the shares

Nothing sub-fab-specific. On 28 July a Chinese state-backed firm began mass-producing immersion lithography machines and worries spread that AI infrastructure spending was peaking; chip stocks lost more than $1trn. On 18-19 August the 30-year Treasury yield hit a 19-year high of 5.33%, taking the sector down again. Remove each name's two best and two worst sessions from the past month and Ichor's 27% decline becomes 16%, MKS's becomes 4%, and Axcelis is flat. FormFactor, which sells probe cards for testing high-bandwidth memory, rose over the same stretch on record DRAM probe-card volumes.

The spending backdrop went the other way too: Lam raised its 2026 wafer-fab-equipment forecast to the low-$150bn range from $140bn. All three peaked in late June; Ichor is 45.5% below its $112.28 high, MKS 36.7% below, Axcelis 34.0%.

The setup

Where it stands — Two near-identical subsystem builders serve the same two customers, and Ichor's multiple is more than double Ultra Clean's. Would confirm — Ichor's gross margin rising roughly 100 basis points a quarter through Q4, as guided, while the discount to Ultra Clean persists. Would invalidate — Ichor's September quarter landing below the $315m guidance floor, or Lam and Applied cutting 2027 build plans. Watch next — Ichor's September-quarter report; China's competition ruling on the Axcelis-Veeco merger, targeted to close in 2026. Valuation — Ichor 20.8x trailing and 17.0x forward gross profit; Ultra Clean 9.7x and 7.8x; MKS 9.9x and 8.5x.

Celestica, the Toronto-based contract manufacturer that designs and assembles switches, storage and complete data-center racks for cloud operators, raised its full-year revenue target on 28 July to $20.5bn from $19bn. It lifted adjusted earnings guidance to $11.30 a share from $10.15, and named two new programs: custom racks for OpenAI, with mass production in 2027, and interconnect for AMD's Helios system, sampling at the end of this year.

Since the end of June the shares have lost 17%. Flex is down 30% over the same weeks, Sanmina 23% and Jabil 16%. These four companies do the physical work of the AI build — integrating accelerators, power shelves, busway and liquid cooling into racks that hyperscale operators once bought as branded servers from Dell or Hewlett Packard Enterprise. Their order books accelerated all summer. Their valuations went the other way.

The share-loss theory does not survive contact with the numbers

The cleanest bear case for a Western rack builder is that hyperscalers pull custom design toward the Taiwanese original design manufacturers, where the scale sits. That is testable monthly, because those companies report revenue every month. Hon Hai, better known as Foxconn, posted record July revenue of NT$946.5bn, up 54.2% year on year, crediting its cloud and networking division. Wistron reported July revenue up 60.8% and guided AI server growth to strengthen into the fourth quarter. Quanta booked its second-highest month on record.

Celestica's June quarter grew 62.4%. That is the same rate. Nobody is taking share from anybody; the market is growing faster than any of them can build.

Underneath the top line, Celestica's operating income rose 68.2% to $458.3m — faster than revenue, which is what operating leverage looks like. Its Connectivity and Cloud Solutions segment grew 84%, and the Hardware Platform Solutions line of custom rack and switch builds, now 41% of company revenue, grew 58%. Management said the binding constraint is material availability, with lead times past 52 weeks and customers placing non-cancellable orders for long-lead silicon.

The durable risk is concentration, and it is real. Celestica's annual report disclosed that three customers accounted for 32%, 14% and 12% of 2025 revenue. A single design loss resets a large share of the revenue line. But concentration is a standing condition, not a July event.

Flex is a power business wearing an assembler's clothes

Flex, the Singapore-founded manufacturer whose Anord Mardix unit makes switchgear and power distribution for data centers and whose JetCool line does microconvective chip cooling, grew its Cloud and Power Infrastructure segment 35% to $2.2bn at a 9.7% operating margin, with the power business up more than 70%. It guided that segment to 65-75% growth this fiscal year and says more than 90% of the next three quarters is booked. A tax-free spin-off of the segment is targeted for early 2027.

The premise that Flex's other end markets are dragging is wrong too: Reliability Solutions, its industrial, energy and medical arm, grew 12% to $2.7bn with operating margin up 130 basis points. Group gross margin expanded to 9.42% from 8.70%.

Sanmina is the genuinely ambiguous name. The San Jose company's heritage is defense, medical and industrial electronics; it bought ZT Systems' data-center manufacturing arm from AMD. Revenue rose 69.7% year on year — and fell 13.7% sequentially, as legacy ZT programs wind toward zero. Core Sanmina, excluding the acquisition, grew 14.1%. Gross margin went up rather than down, to 10.49% from 8.87%, but management attributes part of that to non-recurring engineering fees it expects to fade. Jabil, the slowest grower at 11.8%, is also the least de-rated.

What the price now assumes

For businesses converting 9-12% gross margins, price against gross profit is the honest lens; net income at Sanmina is distorted by acquisition amortization. Celestica trades at 19.15x trailing gross profit against 26.26x in mid-May, a 27% compression while gross profit grew 55.7%. Flex sits at 15.04x against 21.42x. Sanmina is at 9.07x against 11.55x in early May. On forward gross profit the three sit at 14.53x, 12.72x and 8.19x against Jabil's 11.99x — the group has converged toward its least AI-exposed member. Celestica's forward price/earnings ratio of 26.5x rests on consensus earnings of $11.37 this year and $19.87 next.

Two dated events carry most of the damage. Celestica priced a $3bn share offering at $310 against a $362.76 close, a 14.5% discount and roughly 8% dilution; the stock fell 14.8% the next session. Then on 18 August the 30-year Treasury yield hit a 19-year high of 5.32% and the Philadelphia Semiconductor Index fell 4.96%; all four assemblers dropped between 7.6% and 8.9% that day. Strip those three sessions and Celestica is up roughly 18% over a 90-day window that otherwise shows a decline. Over twelve months Flex is still up 127% and Celestica 63%.

The setup

Where it stands — Revenue and margins are accelerating at Celestica and Flex while their gross-profit multiples have compressed roughly 27-30% since mid-May. Would confirm — Celestica's third-quarter revenue printing inside the $5.25-5.55bn guided range, near 69% growth. Would invalidate — A named program loss, or Taiwanese monthly revenue growth pulling durably ahead of Celestica's. Watch next — Foxconn, Wistron and Quanta report August revenue in early September; Sanmina's fiscal fourth quarter follows in October. Valuation — Celestica 19.15x trailing gross profit, 14.53x forward, against 26.26x in mid-May and 28.71x on 7 May.

Three companies that charge for software the way a utility charges for electricity — by the request, the server, the gigabyte — reported June quarters this month. Two of them grew at almost exactly the same rate. What separated them was a single customer.

The same top line, two outcomes

Datadog sells cloud monitoring and analytics to developers and operations teams, metered by host, by gigabyte of logs ingested and by custom metric. Revenue reached $1.12bn in the June quarter, up 35.6% and the fifth consecutive quarter of acceleration. Dollar-based net retention sits in the low 120s, and customers spending over $100,000 a year rose to 4,720 from 3,850, generating 91% of annual recurring revenue.

None of that mattered on 6 August. Management disclosed that its largest customer — a leading AI company running 17 Datadog products on a nine-figure contract, widely believed on Wall Street to be OpenAI though the company has never named it — renewed on reduced usage from the September quarter. Guidance for that quarter implies 28-29% growth, down from 36%. The shares lost 20.4% that day.

The fact buried underneath is more interesting than the hole. Datadog's non-AI revenue is now growing at a high-20s percentage rate, up from 18% a year earlier — five straight quarters of acceleration in the ordinary enterprise base. More than 750 customers are AI-native, 31 of them spending over $1m and eight over $10m. Neither Datadog nor Cloudflare publishes AI revenue as a share of the total, so no outsider can size the concentration precisely; that opacity is itself the risk consumption billing carries.

Cloudflare, which operates an edge network of security, content delivery and per-request serverless compute across 335-plus locations, beat consensus revenue by 4.6% at $696.1m, up 35.9% and a fourth straight acceleration. Net retention improved six points year on year to 120%. Remaining performance obligations — contracted revenue not yet recognized — reached $2.73bn, up 38%. Customers above $100,000 rose 27% to 4,698.

It was bought, though. Non-GAAP gross margin of 73.1% is 320 basis points below last year, as free network traffic grew faster than paid, and the company booked $165m of severance and restructuring alongside a 20% workforce reduction. Management is deliberately refusing the capital-expenditure arms race, holding network capex at 14-15% of revenue.

Akamai is a different company wearing the same label

Akamai, the original content-delivery carrier, embeds servers inside 4,400-plus internet service providers. Delivery revenue fell 6% to $396m last quarter and is guided to a mid-single-digit decline for the year. Security grew 10% to $604m; the new cloud infrastructure line grew 39% to $99m. Total revenue rose 5.4% — but trailing gross profit fell 0.5% and operating income dropped 47%, with capital spending at 32% of revenue heading toward roughly 40% for the year, funded by convertibles now totaling about $7.6bn, with the buyback paused. The shares fell almost 7% after results.

The offset is $2.8bn of signed multi-year compute contracts, including an $1.8bn seven-year commitment from Anthropic — the largest deal in Akamai's history — and a $600m robotics deal. Revenue from them starts in the fourth quarter of 2026 at roughly $20-25m. Anthropic itself told investors its annualized revenue passed $65bn in July, from $9bn at the end of 2025, so the counterparty is not the question. Timing is.

What a dollar of gross profit costs

Gross margins here run 78.6%, 71.8% and 55.8%, so revenue multiples flatter Akamai and punish nobody usefully; price against gross profit is the comparable lens. Cloudflare trades at 56.8x trailing gross profit, up from roughly 43x in mid-May — a third of its 41% three-month gain is re-rating, not compounding. Datadog is at 26.4x, essentially where it stood in mid-May, so its quieter 8.5% three-month gain was earned; measured from early May, when it was near 17.7x, it remains about 49% dearer. Akamai has fallen to 6.7x from 8.4x, and its 16.9x forward earnings sit on consensus 2026 earnings per share of $6.69 — 5% below 2025.

The charts flatter all three. Datadog's 50-day average has held above its 200-day since 21 May and Cloudflare's since 5 May, while Akamai's crossed lower on 2 July and has not recovered. But over the month to 19 August, Datadog fell 11.3% and Akamai 8.4% as Cloudflare rose 7.1% — and strip Cloudflare's two best sessions and that becomes roughly -4.8%. On 18 and 19 August all three fell, as the 30-year Treasury yield touched about 5.33%, its highest since 2007. Long-duration software was sold, not bought.

The setup

Where it stands — Two usage-billed businesses grew alike last quarter and now carry multiples two-to-one apart; the third is de-rating on shrinking gross profit. Would confirm — Datadog's September quarter landing above the 28-29% guide with net retention still in the low 120s. Would invalidate — Cloudflare's net retention slipping below 118% or non-GAAP gross margin resuming its decline from 73.1%. Watch next — Akamai's fourth-quarter report, the first to include revenue from the $1.8bn Anthropic commitment. Valuation — Price to trailing gross profit: Cloudflare 56.8x (49.8x forward), Datadog 26.4x (23.9x), Akamai 6.7x.

Anterix has 64 employees, owns no network, and holds exclusive 900 MHz spectrum across the continental United States, Alaska, Hawaii and Puerto Rico. Its business is selling or licensing slices of that spectrum to electric utilities that want private LTE networks to reach substations, reclosers and meters — the low-capital alternative to trenching fiber to every point on the grid. In its fiscal year ended March 2026 it signed new spectrum agreements with CPS Energy, Texas-New Mexico Power and NorthWestern Energy worth $23.9m of contracted proceeds, then added Benton Public Utility District in April for $0.8m.

Over roughly that same stretch the equity went from about $0.65bn to $1.73bn.

The multiple that actually applies

Conventional multiples are useless here. Trailing revenue is about $7.0m, which makes price-to-sales of 246x a number without content. Reported fiscal 2026 net income of $90.6m sits on an operating loss of $41.6m: the profit came from a $105.4m gain on exchanging narrowband for broadband licences in 219 counties and a $34.8m gain on asset sales. In the June quarter, with no such gains, net income fell 99% to $0.24m.

The honest anchor is price per megahertz-POP, the unit spectrum trades in. Management's own arithmetic puts executed utility deals at an average $1.40 per MHz-POP and the Federal Communications Commission's AWS-3 auction at $2.50, implying about $7.5bn of gross asset value against a market capitalization that implies roughly $0.60. So the shares remain below transaction comparables — while sitting at about triple the implied value of six months ago. Roughly 3 billion MHz-POPs, 85% of the holding and including the largest metros, are still unmonetized.

The demand side is real. American investor-owned utilities have proposed more than $1.4trn of capital spending for 2026-2030, a rise of over a fifth on the prior five years, as data-center load forces grid rebuilds. Anterix describes about twelve live utility opportunities spanning seven- to nine-figure deals. On 11 August it also publicly backed a SpaceX proposal to let 896-901/935-940 MHz licences meet buildout obligations via satellite direct-to-device — a third monetization path alongside utility lease and sale.

The conversion, though, is slow. Anterix collected $16m from customers in the June quarter and expects only about $10m more for the rest of the fiscal year. It ended with $116m of cash, no debt, and quarterly operating costs near $13m plus spectrum clearing — call it two years. Diluted shares rose to 19.63m from 18.70m on option exercises; the count is growing, not shrinking, and $226m of buyback authorization expires next month.

The other two are not the same trade

Cogent Communications sells internet transit, private networks and colocation from 3,035 buildings, mostly to small enterprises and bandwidth-heavy customers. Its second-quarter service revenue was $235.6m, down 4.3% from a year earlier. The acquired Sprint Wireline book has fallen from a $118m run rate to $34m, masking 29% growth in the legacy Cogent business since the deal closed. Adjusted net leverage improved to 6.23x, helped by a $225m sale of ten former Sprint data centers, and margins expanded 90 basis points — but the board holds the quarterly dividend at $0.02 after a 98% cut, and free cash flow is positive only while T-Mobile subsidy payments run, about two years more. The genuine option inside it is wavelength services, up 63.8% to $14.8m — roughly 6% of revenue, and about 3% of a North American market where Zayo is spending $4bn on AI long-haul fiber and Lumen has held Frost Radar's top ranking three years running.

Sify Technologies, a Chennai group running enterprise networks, colocation and digital services, is the cleanest growth story of the three and the most heavily financed. June-quarter revenue rose 15.9% to about $141m and operating income doubled, producing its first positive net quarter in five. But fiscal 2026 interest expense of roughly $45m exceeded operating income of about $32m, and net debt reached some $381m. It has 134 MW of revenue-generating data-center capacity against 154 MW installed, and sold just 5 MW last quarter — the ramp is back-end loaded. The International Finance Corporation has committed $371m for two AI-ready sites totalling 103 MW, into a domestic market that added 258 MW in the first half of 2026, up 59%. Trailing enterprise value to EBITDA is 8.9x.

What the shares did

Anterix has held an uptrend since late January, its 50-day average above its 200-day every session — and unusually, it was earned gradually: strip its two best days and the year is still up 182%. It peaked on 4 August and has given back 16.6% since, including a double-digit drop the day after the June-quarter print. Cogent's fall came the opposite way, in violent single sessions of -34.9%, -29.3% and -25.9%. Sify has drifted: up 3.1% over thirty days, down 11.3% over three months, so its twelve-month gain is a late-2025 base effect rather than a live advance.

The setup

Where it stands — Anterix's re-rating rests on a $3bn pipeline, not the roughly $400m of proceeds contracted to date.

Would confirm — A signed utility agreement in the eight- or nine-figure range, at or above $1.40 per MHz-POP.

Would invalidate — Fiscal 2027 contracted proceeds finishing near the $10m of additional collections management guided to.

Watch next — The $226m buyback authorization expires in September 2026; management has said it may be renewed rather than spent.

Valuation — Roughly $0.60 per MHz-POP implied, versus $1.40 on executed deals and $2.50 at the AWS-3 auction.

The disruption is also the supplier

ADT, which sells professionally installed and continuously monitored alarm, fire and video systems to American homes and small businesses, launched a self-installed product called ADT Blu in the second quarter at an entry price of $10 a month. The platform underneath it belongs to Arlo Technologies, the Carlsbad, California maker of wireless cameras and video doorbells that also supplies the hardware and cloud service behind Comcast's Xfinity offering, Europe's Verisure and Samsung SmartThings (Security Systems News). The company most often cast as the threat to monitored contracts is a counterparty, not a rival.

The substitution itself is genuine. For the first time, more American alarm owners installed their own systems than hired a professional, and professional monitoring now ranks ninth of twelve factors buyers weigh; only 5% of non-users said they intended to buy monitoring within a year (SafeHome 2026 market report). Ring, Amazon's doorbell brand, is named by 43% of users as their primary system. ADT's fight is to defend the monitored socket, not to win the device.

So far it is holding. Second-quarter revenue of $1.31bn grew less than 2%, but gross profit rose 10.2% as margin widened to 53.4% from 49.4%. Ending recurring monthly revenue was $360m, and the share of that revenue lost to cancellations over twelve months was flat at 13.1% both sequentially and against last year — no deterioration in the metric the DIY thesis predicts. Adjusted free cash flow of $406m was up 48%, and management raised full-year guidance. The soft spot is below gross profit: operating income fell 7.9%. Net debt of $7.4bn sits at 2.8 times earnings before interest, taxes, depreciation and amortization, 98% fixed-rate, with the largest maturity not due until 2030 (Investing.com). Management flagged 2027 headwinds of roughly $50-100m each in cash taxes and cash interest as tax losses run out and swaps expire.

Losing money on cameras on purpose

Arlo's pro forma product gross margin was minus 11.6% last quarter. That is the strategy: every device sold below cost enters a household into a subscription funnel whose lifetime value rose 15% to $967, against service gross margin of 84.1%. Paid accounts reached 6.3m, up 23%, annual recurring revenue was $365m, and service is now 60% of the top line. Revenue grew 20.5% and gross profit 29.5% — hardware is expanding, not shrinking. Management raised full-year guidance to $580-600m of revenue and $0.90-1.00 of earnings per share.

That progress is more visible in the price. Arlo trades at 5.46 times trailing gross profit, roughly 2.5 times ADT's 2.23. Its forward multiple of 16.2 times earnings falls to about 13.6 on the top of its own guidance; ADT trades at 9.6 times trailing and 8.0 times forward earnings, and 4.9 times EBITDA — the fitting lens for a leveraged annuity.

Resideo's problem is a memory chip

Resideo makes thermostats, water controls and smoke and carbon-monoxide detectors under the Honeywell Home and First Alert brands. On 3 August it distributed one share of ADI Global Distribution for every two of its own in a tax-free separation; ADI, a $4.8bn-revenue low-voltage equipment distributor, was the larger half by sales (Commercial Integrator). The close went from $36.23 to $26.17 the next session — value moved, not lost. A second drop, 12.7% on 13 August, followed the standalone outlook of $2.90-2.95bn revenue and $605-625m of adjusted EBITDA (FinancialContent). Strip those two sessions and the month is down 3.9%; count the ADI stub near $21, worth about $10.60 per original share (TradingView), and a continuing holder is down roughly a tenth, not two-fifths.

The real news was the margin guide. Products gross margin hit 43.6%, a thirteenth straight quarter of expansion, and management said the streak ends this quarter: memory costs are up about fourfold and metals about 35%. Memory suppliers have reallocated capacity to servers and high-bandwidth chips, with AI data centers projected to absorb roughly 70% of high-end DRAM this year and prices up 63% in the second quarter alone (J.P. Morgan). A smoke-detector maker is being repriced by an AI shortage. Separately, one large original-equipment security customer moving production in-house removes $40-50m of second-half revenue. Resideo repaid a $900m term loan at the spin.

Both subscription names have recovered on the charts — ADT up 5.6% and Arlo 2.8% over 30 days, each moving back above its long-term moving average in the first week of August — while remaining down 16.3% and 19.1% over twelve months. Housing supports the caution: existing-home sales fell 4.2% in the first half and household mobility sits at a record-low 11.2% (Harvard Joint Center for Housing Studies), which starves new installations even as remodeling spending grows.

The setup

Where it stands — Two subscription businesses with improving unit economics; one manufacturer repriced by a spin-off and rising input costs. Would confirm — ADT attrition at or below 13.1% next quarter, and Arlo paid accounts above 6.5m. Would invalidate — ADT recurring monthly revenue falling below $360m, or Arlo service revenue growth slipping under 10%. Watch next — Resideo's third quarter, its first clean standalone period, guided to $705-730m of revenue. Valuation — ADT 9.6x trailing and 8.0x forward earnings; Arlo 5.46x trailing gross profit against ADT's 2.23x.

On 18 August the yield on the 30-year US Treasury bond topped 5.33%, its highest level in 19 years, on inflation and federal-deficit worries. The corner of the semiconductor industry that fell hardest that day was not the designers of artificial-intelligence processors. It was the back end of the chain — the firms that take finished silicon wafers and turn them into packaged, tested chips, plus the equipment makers who sell them the machines. Amkor fell 11.8%, Teradyne 10.5% and Kulicke and Soffa 11.7% in a single session.

That ordering is the story. The outsourced assembly and test industry — OSAT, in the trade's shorthand — is the layer of the artificial-intelligence buildout that is spending now and collecting later.

Why a bond yield reprices a packaging plant

Amkor, a Tempe, Arizona packager that does bumping, probe, assembly and final test for fabless designers and foundries, guided 2026 capital spending to $2.5-3.0bn. Consensus revenue for the year is $7.66bn. Roughly a third of the top line is going back into buildings and tools, part of it funded by a $1.15bn convertible issued at a zero coupon. ASE Technology Holding, the Kaohsiung-based world leader in packaging and test, raised its 2026 capital budget by $2bn to a record $10.5bn and told investors free cash flow stays negative into 2027. Powertech Technology, a Hsinchu packager weighted toward memory chips, runs a trailing free-cash-flow yield of minus 5.53%.

Capacity that produces nothing until 2027 or 2028 is a long-duration asset. When the discount rate on that stretch of time jumps to a two-decade high, the asset is worth less today even if every operating number improves. Over the same 21 sessions the demand side rose — the S&P 500 tracker gained 3.5%, Nvidia 8.6%, Micron 9.6%, TSMC 3.7% — while all six listed packagers fell between 6.6% and 23.8%. This was a financing shock, not a demand signal.

The operating numbers went the other way

Amkor's June quarter revenue was $1.898bn, up 25.6% year over year. Gross margin widened to 16.79% from 12.04%, and utilization climbed from the low 50s to the high 70s. Management guided the September quarter to $1.95-2.05bn at an 18.5-19.5% gross margin.

ASE grew 34.1% in the quarter, the fourth straight acceleration. Its assembly, test and materials arm — 66% of revenue but 94% of operating profit, the rest being a large, thin-margin contract-manufacturing business that flatters revenue and dilutes margin — set a record at 27.3% gross margin and is guided to break 30% in the December quarter. July revenue rose 43.2%, with that packaging segment up 49.5%. Powertech grew 28.0% with gross margin at 21.76%, and posted its highest quarterly earnings per share in four years.

The standing fear about this industry — that foundries and memory makers will keep advanced packaging for themselves — is going the other way. TSMC began delegating the chip-on-wafer step it had kept almost entirely in-house to outside packagers from the start of August, on top of 240,000-270,000 CoWoS wafers already outsourced this year, most of them to Amkor. Powertech, meanwhile, committed $400m to a Singapore packaging venture with Broadcom, with AMD as first customer.

Where the caution is real

One name deserves it. Amkor's consensus earnings per share nearly double this year, to $2.479, then rise 7.9% in 2027, to $2.675. The reason is arithmetic: construction of the first plant on its $7bn Arizona campus finishes in mid-2027, with production starting in early 2028, so depreciation lands a year ahead of the revenue. Its largest division, communications, was guided down by a high single-digit percentage this quarter, with headwinds into the first half of 2027.

That shows in the three-month record, where the damage is narrow rather than broad. Amkor is down 27.7% over that stretch; ASE is up 5.5% and Powertech up 15.6%.

Amkor now sits 45.6% below its 52-week high of $93.55. At that peak it carried 37.7x the same $2.479 estimate it trades at 20.53x today — the estimate never moved. It fetches 10.93x trailing gross profit, against 18.76x in mid-May, on a gross profit that grew 75%. Its 9.80x trailing enterprise value to EBITDA and 4.08% free-cash-flow yield are the only positive cash yield in the group. Powertech trades at 21.14x forward against 26.58x trailing, on consensus profit growth of 45.6% for 2027. ASE's trailing multiple has fallen to 42.42x from 52.55x in May while earnings nearly doubled; its US-listed receipt makes per-share forward figures unreliable, so only that direction is assertable.

The setup

Where it stands — Three packagers are compounding revenue at 25-34% while the market discounts capacity that starts earning in 2027 and 2028. Would confirm — Amkor's September quarter lands inside the guided $1.95-2.05bn with gross margin at or above 18.5%. Would invalidate — ASE's December-quarter packaging gross margin fails to exceed 30%, or advanced utilization slips below 80%. Watch next — ASE reports August monthly revenue in early September; Amkor reports third-quarter results in late October. Valuation — Amkor 20.53x forward and 22.72x trailing, against 37.7x forward at its high; Powertech 21.14x forward, 26.58x trailing.

Emerson Electric sells the instruments, valves and control software that keep refineries, liquefied natural gas (LNG) trains and generating stations running. For most of its 136 years that made it a bet on hydrocarbons. Its fiscal third quarter, reported on 4 August, described a different customer.

Underlying orders across the company grew 7%. Inside that number, what management calls its growth verticals grew 27%, led by power orders up 37% and semiconductor orders up 53%. Orders for Ovation, Emerson's distributed control system, rose 31%, and its backlog now stretches into late 2027 and early 2028.

Why the electricity build hits Emerson first

A control system is not a commodity purchase. Ovation and its sister platform DeltaV are embedded in gas plants, pipelines and much of the legacy American nuclear fleet, and replacing one means re-engineering and re-validating the plant. That entrenchment is why new generating capacity converts almost mechanically into Emerson content. Management pointed to fleet modernizations and behind-the-meter data-center opportunities as the source of the power growth, with cybersecurity rules pulling forward retrofit cycles at critical infrastructure. Named wins in the quarter included a 2.1 gigawatt power retrofit for Mexico's CFE and pressurizer valves for a 2.4 gigawatt reactor in China.

The project funnel — work identified but not yet booked — grew $1.2bn in three months to $12.4bn, of which roughly $3bn is power and $2.2bn LNG, most of it greenfield. Backlog reached $8.2bn, up 7%. Revenue growth accelerated through the year, from 4.1% in the first quarter to 7.0% in the third, and gross margin widened to 54.5% from 52.6%. Free cash flow was $1.3bn, up 36%. Emerson raised its full-year outlook to adjusted earnings of $6.55 a share.

Two cautions. The strength is American: sales rose 10% in the United States while Europe fell 1% and China fell 3%. And a Middle East disruption cost roughly $25m of sales in the quarter, with about $100m expected for the year.

The factory floor is only level again

Rockwell Automation sells the programmable controllers, drives and FactoryTalk software that run discrete manufacturing lines, through independent distributors. Its quarter was operationally excellent: revenue of $2.31bn, gross margin of 49.5% against 40.9% a year earlier, and operating income up 68.6%. It lifted full-year organic growth guidance to a 7.5%-9.5% range and earnings to a $13.15 midpoint.

The qualification came from management itself, which described the company as only now getting back to pre-COVID controller unit volumes. The destocking that followed the pandemic has unwound; a new upcycle has not obviously begun. Lifecycle Services, the project and services arm, shrank 2% organically with a book-to-bill of 0.97, and food and beverage capital spending has yet to inflect. E-commerce and warehouse orders, up 30%, are the exception rather than the base. The shares fell 7.4% on the day of the print.

Symbotic's business improved; its schedule slipped

Symbotic builds entire automated warehouse modules — robots, software and structure — for retail and wholesale distributors, with Walmart still dominant among them. Its numbers are improving fastest of the three: revenue of $720.8m grew 21.7% while gross profit grew 49.5%, gross margin has widened from 16.6% to 22.3% over six quarters, and the operating line swung from a $20.2m loss to $32.9m of income.

The shares fell 16.2% the morning after, on 6 August. The reason was time, not demand: conversion of the 400-store Walmart back-of-store contract is not expected until early 2028 and is not in backlog, and of the $22.5bn contracted backlog only about 15% converts to revenue within twelve months. The GreenBox joint venture's Lathrop site is built but still 60 to 90 days from going live.

What the prices already assume

Symbotic trades at 47.2x trailing gross profit — the fair lens here, because gross margins run from 22% to 55% across the three and its price-to-earnings ratio is a meaningless 438x. Rockwell sits at 9.85x that measure and Emerson at 8.91x. Symbotic's own multiple was near 68x in February, when trailing gross profit was about a tenth lower, so the multiple did the falling, not the business.

Emerson carries the widest gap between what it has earned and what it is expected to earn: 34.3x trailing against 24.1x forward, with enterprise value at 19.4x EBITDA and a trailing free cash flow yield of 3.9%. Rockwell's trailing multiple has come down from 45.3x in May to 40.5x, and to 32.8x forward — compression delivered by earnings growth near 25% rather than by a falling share price.

The shares have followed the order books, with a lag. Emerson's 50-day average crossed above its 200-day on its earnings day and it closed at a 52-week high of $164.40 on 12 August. Rockwell's uptrend weakened on 7 August. Symbotic's 50-day has sat below its 200-day since 21 May, though the stock is marginally higher over the past month, its damage done between February and July.

The setup

Where it stands — Emerson is being re-rated on power and semiconductor orders; Rockwell and Symbotic are executing well but waiting on customer schedules. Would confirm — Emerson's fourth-quarter power orders growing at a double-digit rate again, with the project funnel above $12.4bn. Would invalidate — Emerson book-to-bill falling below 1.0, or backlog slipping from $8.2bn. Watch next — Emerson's fiscal fourth-quarter results in early November, guiding fiscal 2027 against $6.55 of adjusted 2026 earnings. Valuation — Emerson at 34.3x trailing and 24.1x forward earnings, against Rockwell's 40.5x and 32.8x.

The chip that ships with the module

Rambus, a San Jose designer of memory-interface chips and a licensor of memory-architecture patents, sells a registering clock driver into the server dual in-line memory modules (DIMMs) that plug into data-center processors. Its unit volumes are set by how many modules ship — not by how much buyers spend on the memory inside them. With DRAM contract prices up sharply, that distinction has become the case against the stock.

TrendForce forecasts that registered-DIMM bit supply will grow only 15-20% year over year, well short of projected server CPU shipments, and already anticipates a server DRAM shortage in 2027. Baird downgraded Rambus to Neutral on precisely that mechanism, holding its $120 target while arguing memory scarcity could cap growth through 2027.

The results themselves point the other way. June-quarter revenue was a record $207.4m, up 20.4%, with gross margin of 79.8% against 74.9% a year earlier. Product revenue reached a record $99.2m, overtaking royalties as the largest line. The September guide makes the mix shift explicit: product of $110-116m against royalties falling to $69-75m from $84.2m. The silicon cycle is carrying the licensing annuity, not the reverse. Management framed the volume driver as channel count — Advanced Micro Devices at 12 memory channels, Intel moving from eight to 12 — with each channel adding companion-chip content. It also conceded there is no pricing opportunity in standards-based parts, where Renesas and Shanghai-listed Montage Technology compete for the same socket, and described the pooling standard the group is named for, Compute Express Link (CXL), as "a fragmented ASIC market" it addresses through licensed intellectual property rather than products.

Microchip already finished its correction

Microchip Technology, the Arizona maker of 8-, 16- and 32-bit microcontrollers and analog parts sold to carmakers and factory-equipment builders, is the one member with real memory-disaggregation revenue — and the one whose downturn is over. June-quarter revenue was $1.485bn, up 38.0%. Gross margin reached 63.2% against 53.6% a year earlier, a fifth consecutive quarterly expansion, while operating margin recovered to 22.7% from a 3.0% trough.

Distributor inventory has been drawn to 25 days, the low end of its historical range, on the strongest booking quarter in roughly four years. September is guided to about $1.603bn, up 40.6% year over year. Management warned the guided 66-67% gross margin is not repeatable, flattered by a lumpy licensing quarter billed at full margin. Internal fabs are still underloaded, with underutilization charges guided to $30.5m from $38.5m. Data-center sales are the growth engine: roughly $591m in 2025 heading to about $1bn in 2026, with 14 PCIe Generation 6 design wins entering production at end-June. Net debt of about $5.2bn sits at 2.85x EBITDA, expected below 2.5x this quarter.

The two smaller names

Everspin Technologies, an 85-person maker of magnetoresistive RAM (MRAM) — memory that retains data without power — posted record revenue of $18.7m, up 41.9% and above its own guidance. The bulk of the surprise is a $40m, two-and-a-half-year subcontract with a United States prime contractor to qualify MRAM for the defense industrial base. Its pooling work is real but pre-revenue: a CXL-attached proof of concept scheduled for the Storage Networking Industry Association developer conference on 28-30 September, and an evaluation agreement with MaxLinear on storage acceleration for artificial-intelligence servers. Roughly $4m a quarter of litigation cost keeps the September quarter at a reported loss.

Allegro MicroSystems, which makes magnetic sensor and power chips for automotive motion control and has no memory content whatever, is the group's laggard on operations: revenue grew 27.5% to $259.2m but operating margin was 9.8% on a 48.5% gross margin, the thinnest of the four, and it carries the highest forward earnings multiple at 38.2x.

All multiple, no earnings

Trailing price-to-earnings is unusable across this group — Everspin's is negative, Allegro's runs to 486x, Microchip's 107x on trough profits — so price against gross profit is the lens that compares them. Against 21 May, Rambus compressed from about 28.0x to 16.6x while trailing gross profit grew 6.6%; Microchip from 18.2x to 13.6x on gross profit up 13.3%; Everspin from 25.9x to 12.5x on gross profit up 11.4%. Forward multiples sit below trailing at all three, meaning consensus still expects gross profit to rise.

Rambus broke trend first, leaving its uptrend on 6 July and turning negative on 22 July, four weeks before the others. The recent leg was not gentle: Rambus fell 10.6% across 18-19 August, Everspin 12.9% from its 17 August close, Allegro 11.0% in the 19 August session alone. The context was rates — the Philadelphia Semiconductor Index fell 4.96% as 30-year Treasury yields hit 19-year highs, and chip stocks kept falling on 19 August even after yields eased, inside a sector drawdown that has erased more than $1 trillion of chip-stock value since late July. Tellingly, Astera Labs — the pure-play CXL and PCIe fabric vendor these names are valued against — is roughly flat over the same three months. The pooling premium did not leave the company that actually carries it.

The setup

Where it stands — Three memory-adjacent chipmakers reported accelerating revenue and wider margins into June, then de-rated 15-46% on multiple alone. Would confirm — Rambus September product revenue landing in or above the guided $110-116m range while royalties fall as guided. Would invalidate — Microchip December-quarter bookings turning below 1.0 book-to-bill, or distributor inventory rebuilding past 30 days. Watch next — Everspin's CXL-attached MRAM demonstration at the SNIA developer conference, 28-30 September 2026. Valuation — Rambus 16.6x trailing gross profit and 15.2x forward, against roughly 28.0x in May; Microchip 13.6x and 11.2x.

Three software companies whose customers are warehouse operators, civil engineers and apartment landlords each reported results in the past month, and each reported a business in better shape than a year ago. The market has since behaved as though only one of them had news.

The one that got paid

Manhattan Associates, which sells warehouse-management and order-management software to retailers, grocers and third-party logistics firms, has now re-accelerated for four straight quarters: revenue growth of 3.4%, then 5.7%, 7.4% and 9.3% year over year, the last on revenue of $297.8m. Cloud subscription revenue rose 26% and remaining performance obligations — contracted revenue not yet recognized — reached $2.47bn, up 23%. The optical drag is the shrinking perpetual-license line; excluding it, revenue rose 13%. Conversions of on-premise customers account for more than 40% of new cloud bookings, which matters in a warehouse-software market worth roughly $4.77bn where the top five vendors — Manhattan, Blue Yonder, Körber, Oracle and SAP — hold only 25–30% between them.

The shares went from $159.29 on July 27 to $211.74 on July 29. They have drifted 2.5% lower since. That leaves Manhattan 71% above its April low, 5.8% under its 52-week high, at 37.5x forward earnings against 58.5x trailing and 18.9x forward gross profit — the dearest of the three on every lens, while reported gross margin fell to 53.2% from 56.8% and GAAP operating income declined 10.2%. Consensus has revenue growing 8.2% this year and 8.5% next, below the 9.3% just posted. Nothing further is priced in the estimates; a good deal is priced in the stock.

The one that didn't

Bentley Systems sells the design and asset software behind roads, rail, water systems and electricity networks — MicroStation, OpenRoads, ProjectWise — competing with Autodesk, Trimble and Hexagon, with its defensible ground in large horizontal infrastructure. Annual recurring revenue reached $1.536bn, up 12% organically, with net revenue retention of 109% and account retention of 99%. Management named double-digit US backlogs in transportation, water, power and data centers, with the Power Line Systems unit driving grid work.

That end market is physical, not thematic. Large power transformers now carry lead times of two to five years and switchgear exceeds 60 weeks, which pulls engineering and design forward of equipment delivery. Bentley is a beneficiary of the AI buildout that is being valued as a victim of it: free cash flow of $498m over twelve months, up 15%, against 26.1x forward earnings and 7.5x forward gross profit, roughly a third of Manhattan's multiple on that measure. The stock closed at $35.84 on July 28 and $35.85 on August 19 — 37% below its 52-week high and down 32.9% over a year.

The one in between

AppFolio, a cloud platform for property managers that also earns fees on tenant payments, screening and insurance, grew revenue 19% to $281m with units under management up 8% to 9.6 million and value-added services up 22% to $219m. GAAP operating margin widened to 18.8% from 17.2%, and the company lifted full-year non-GAAP operating margin guidance to 26.5–28.0% from 24.7% in 2025. The standard worry — that interest on customer funds held in transit deflates as rates fall — has the direction wrong: the Federal Reserve has left the funds rate at 3.50–3.75% for five straight meetings and cut nothing in 2026, with three dissents favoring a hike. AppFolio trades at 31.2x forward earnings and 10.9x forward gross profit, 49% above its April low.

What moved them

All three gains belong to two rotation windows, not to a steady re-rating. The late-July window did most of the work; on August 19 a selloff in AI hardware pushed money back into enterprise software, with ServiceNow up 6.5% and a semiconductor ETF down 2.2%. It is a rotation, not a duration trade: the 30-year Treasury yield topped 5.33% on August 18, a 19-year high, which normally works against long-dated software cash flows. Bentley and AppFolio have both turned upward on trend, their 50-day averages crossing above their 200-day on August 14 and August 4; Manhattan's remains neutral.

The setup

Where it stands — Three improving vertical-software businesses carry very different multiples: Bentley at 26.1x forward earnings, Manhattan at 37.5x near its high. Would confirm — Bentley's Q3 annual recurring revenue growth holding at or above 12% with net revenue retention at 109% or better. Would invalidate — Manhattan's remaining performance obligations growing below the guided 18–20% range, or Bentley's retention slipping under 105%. Watch next — Third-quarter results: Manhattan in late October, Bentley and AppFolio in early November. Valuation — Bentley 7.5x forward gross profit versus 8.0x trailing; Manhattan 18.9x forward, 19.5x trailing; AppFolio 10.9x forward.