DK Street Journal

TD Synnex buys hardware, software and cloud services from vendors and resells them to the value-added resellers, integrators and managed service providers who serve corporate and government buyers. That is the boring half. The other half is Hyve, which designs and builds data-center racks to hyperscaler specification, and in the quarter ended 31 May it booked $5.5bn of gross billings there, up 117% year over year after 95% growth the quarter before. Hyve now runs programs with each of the five largest American hyperscalers, including Amazon Web Services, Microsoft Azure and Google Cloud.

The assumption attached to that kind of work is that it is a pass-through: you buy the graphics processors and the memory at whatever the market demands, add a sliver, and ship. The company's own segment disclosure says otherwise. Hyve's non-GAAP operating margin was 3.3% in the quarter, against 2.0% in distribution — and the blended margin expanded 22 basis points precisely because the hyperscaler book grew as a share of the mix. Hyve's own margin did fall 50 basis points on ramp inefficiency, while distribution's rose 34.

At the consolidated level the same thing is visible. Revenue reached $19.57bn, up 31.0%, and gross profit grew faster still, up 33.3% to $1.339bn. Gross margin widened to 6.84% from 6.73%. Gross billings of $28.9bn rose 33.4%, dead level with gross profit growth — the channel is holding its take rate while the dollar value of what passes through it inflates. That inflation is real and mechanical: conventional memory contract prices rose 90-95% in a single quarter early this year because AI data centers are expected to absorb roughly 70% of high-end DRAM in 2026 and manufacturers are steering wafers toward higher-margin high-bandwidth memory.

What the ramp costs

The bill arrives on the balance sheet. Net working capital went from $4.2bn to $4.9bn in one quarter and free cash flow turned negative, on accelerated investment in Hyve capacity. The gross cash conversion cycle barely moved, at 17 days, so this is scale rather than deterioration — but scale has to be funded. TD Synnex extended its trade receivables securitization facility to January 2028 and launched a new €650m European receivables facility in June. The cost of that funding is going the wrong way: the 30-year Treasury yield topped 5.33% on 18 August, a 19-year high.

The trailing free-cash-flow yield is 1.87%. And Hyve does not compete with other distributors for this revenue — it bids against Taiwanese contract manufacturers Quanta, Wiwynn, Wistron and Foxconn, plus Celestica and Supermicro, which hold a structural cost advantage in build-to-spec racks.

The price of the same growth, three ways

Because gross margins run 6.84% at TD Synnex, 11.26% at Arrow and 21.74% at Insight, price per dollar of gross profit is the only figure that compares them. TD Synnex trades at 4.20x trailing, up from roughly 3.4x twelve months ago. It is the most expensive of the three on that measure and the thinnest on cash return.

Arrow Electronics, which distributes semiconductors and passive components and separately sells data-center and security infrastructure through resellers, grew revenue 31.8% to $9.99bn and nearly doubled operating income to $377.3m. Its enterprise computing backlog is up more than 75% to a record. It trades at 9.67x forward earnings against 13.35x trailing — but consensus FY2026 earnings of $21.81 a share sit almost exactly on the 2022 cyclical peak of $21.80, which is why the shares fell 8.45% the day after that quarter.

Insight Enterprises, which sources, configures and manages technology for enterprises across 25 countries, is the one name that got cheaper on its own growth: price per gross-profit dollar fell to 2.35x from about 2.50x while trailing gross profit rose roughly 10%. Cloud gross profit grew 39% to $171m and services 21% to $95m; hardware gross margin fell 110 basis points, so the 60-basis-point blended expansion came entirely from the non-hardware side.

The chart and the guide

TD Synnex shares did nothing from August 2025 into March, then repriced violently — the monthly average close went from $157.40 in March to $277.08 in June, with a 16% two-session jump around the March quarter. They peaked at $290.14 on 23 June and have fallen 13.8% since, roughly flat over the past month. The turn came with the guide: management pointed to about $18.6bn of revenue and $4.50 of adjusted earnings for the current quarter, both below the $19.6bn and $4.85 just delivered. Consensus has revenue growth dropping from 20.5% this year to 7.0% next.

The setup

Where it stands — Hyve is the higher-margin segment and doubling, but funding it has taken free cash flow negative and the multiple has already expanded. Would confirm — Hyve gross billings growing above 50% year over year in the October-quarter report with segment operating margin holding near 3%. Would invalidate — A second consecutive quarter of negative free cash flow without a corresponding step-up in Hyve billings. Watch next — TD Synnex reports fiscal third-quarter results in late September, against guidance of $18.6bn and $4.50. Valuation — 4.20x trailing and 3.92x forward gross profit, against roughly 3.4x a year ago and 2.68x at Arrow.

PPL Corporation, which owns transmission and distribution wires serving about 3.6 million customers in Pennsylvania, Kentucky and Rhode Island, told investors this month that signed electric service agreements for data centers in its Pennsylvania territory had reached 32 gigawatts. Those are contracts, not inquiries. Pennsylvania's utility regulator approved a dedicated large-load rate class in a March settlement requiring a ten-year load and financial commitment, and PPL's version demands guaranteed payment on at least 80% of reserved capacity plus termination fees. The shares have fallen anyway — down 6.4% over twelve months while the S&P 500 gained 20.3%.

Requests are not a backlog

Set that against Oncor, the pure wires business inside Sempra, the San Diego holding company that also owns California's SDG&E and SoCalGas. Oncor's headline number is far larger: an active large-load interconnection queue of roughly 271 gigawatts of data centers plus 18 gigawatts of other industrial load. But only about 16 gigawatts is attached to projects with in-service dates through 2034. The rest is optionality, and Texas regulators are now testing it: Governor Greg Abbott ordered a verification audit of every data-center project in the interconnection queue on 3 August, and ERCOT has said it will not finish before December.

CenterPoint Energy, whose electric arm owns the wires around Houston and whose gas arm serves eight states, sits between the two. Its 14 gigawatts of large-load projects would lift system peak by more than 65%, and management says $900m of customer cash and security is already in hand against it — against system upgrade costs under $60m per gigawatt. Texas's Senate Bill 6 framework charges $50,000 per megawatt in non-refundable interconnection fees, which is what separates a deposit from a phone call.

The businesses improved

All three reported operating leverage in the June quarter. CenterPoint grew revenue 10.7% and operating income 28.1%, and raised its ten-year capital plan to $66.7bn while stating no additional equity is required to fund it. Sempra's revenue was flat, yet operating income rose 42.5% and adjusted earnings per share 30%, with 2026 guidance of $4.80–5.30 reaffirmed. PPL's operating income rose 17.0%, guidance of $1.90–1.98 held, and it deployed $2.3bn of capital in the first half, some 30% more than a year earlier, toward rate-base growth above 10% a year.

On forward earnings, Sempra now trades at 16.2x, down from about 17.9x three months ago and against a utility sector near 18.8x; its price-to-book of 1.66x is the lowest of the three. PPL is at 17.6x against roughly 18.8x a year ago. CenterPoint, at 20.3x, is the outlier: still dearer than the 19.9x its price implied twelve months ago, so its fall has only partly worked off a premium rather than created a discount.

Rates, or Texas?

The convenient explanation is duration. The 30-year Treasury yield reached its highest level since 2007 in mid-August, and utilities were among the weakest sectors into Jackson Hole. But the 21 August session discriminates by geography, not by yield sensitivity: the S&P 500 rose 0.20% while Sempra fell 5.42%, American Electric Power 4.05% and CenterPoint 3.45% — and the longest-duration, slowest-growing regulated names barely moved, Consolidated Edison down 0.65%, WEC 0.27%, National Grid 0.12%. Rates set the level; the Texas audit set that day's ranking. Merchant generators took more: Vistra fell 16.1% over the month and NRG 14.1%, against 9.3% at CenterPoint and 2.9% at PPL.

The complications are real. Sempra has launched a $3bn at-the-market equity program, so its rate-base growth is partly funded by new shares; CenterPoint says its is not. Kentucky regulators set LG&E and KU's allowed return on equity at 9.775%, below the 9.90% proposed in settlement, a reminder that the return on data-center capital is granted, not earned. Sempra and PPL have both traded below their 200-day averages since midsummer; CenterPoint turned lower on 13 August.

The setup

Where it stands — Three wires utilities raised or held guidance and de-rated; only PPL's demand book is largely contracted. Would confirm — CenterPoint converting its 14 gigawatts into ERCOT Batch Zero approvals with capital-tracker recovery intact. Would invalidate — Oncor's queue shrinking materially after the audit, or PPL's signed Pennsylvania agreements falling below 32 gigawatts. Watch next — ERCOT's data-center audit, due to conclude in December, covering 250–300 large consumers. Valuation — Sempra 23.9x trailing and 16.2x forward; PPL 17.6x forward; CenterPoint 20.3x, above its year-ago 19.9x.

Fabrinet reported the largest quarter in its history after the close on 17 August. One line undid it. Datacom revenue — the part that carries artificial-intelligence traffic — slipped 1% from the prior quarter because sales to NVIDIA declined.

Fabrinet is a Thailand-based contract manufacturer that assembles optical transceivers, tunable lasers and active optical cables for the equipment makers whose names appear on them. Revenue for the June quarter was $1.316bn, up 45% year on year and above its own guidance; full-year revenue rose 36% to $4.641bn. Telecom grew 70%, and two new datacom customers — Amazon and a merchant transceiver vendor — largely offset the NVIDIA decline. Needham's Ryan Koontz nonetheless read the datacom line as share loss at NVIDIA, "likely to Chinese competition", while keeping a $650 target. Guidance for the September quarter put revenue at $1.375–1.425bn but earnings at $4.10–4.25 a share, flat sequentially. Four customers each supplied a tenth or more of fiscal 2026 revenue: Cisco at 20%, NVIDIA 16%, Nokia 11% and Amazon 11%. The shares fell 17.5% the next session and 25.4% over five.

Three businesses, one price move

What separates these companies is what each keeps from a transmitted bit. Fabrinet's gross margin was 12.3% in the June quarter, and its annual figure has sat between 12.0% and 12.7% every year since 2022. Thirty-six percent revenue growth produced no expansion at the gross line at all; the leverage shows up lower down, where operating income rose 42.7% and operating margin reached 9.97%. That is the arithmetic of a contract assembler whose moat is process qualification measured in years, not pricing power.

Ciena sells the systems, not the assembly. Its coherent transport platforms carry traffic between data centers and across carrier networks, and in the quarter ended 2 May revenue grew 39.5% while gross margin expanded to 44.0% from 40.2%. Operating margin went from 3.1% to 15.1%. Cloud providers were 46% of sales and grew 70%, and backlog rose more than $600m sequentially to a record $7.7bn. Its claim is WaveLogic 6 Extreme, a 1.6-terabit-per-wavelength modem management says no rival matches in the wide-area network — a lead Nokia, which bought Infinera to add coherent silicon, is spending against.

Corning is the passive control: fiber, cable and connectors, alongside display glass. Group revenue grew 16.6% in the second quarter on a flat 36% gross margin, but Optical Communications, now 49% of sales, grew 32% to $2.07bn with segment net income up 77%. Enterprise sales rose 65% and data-center sales nearly doubled. The mechanism is physical: clusters above 130,000 graphics processors need three optical layers rather than two, adding half again as much content, and a fully optical scale-up would need 160 fibers per processor against about 16 today. Demand for the high-density products already exceeds Corning's capacity, which is why Meta, NVIDIA and Amazon have signed multiyear supply agreements.

The day none of it mattered

On 18 August nine optical names fell together and only one had reported. Fabrinet lost 17.5%, Coherent 12.8%, MaxLinear 16.0%, Ciena 8.9%, Corning 7.7%, Amphenol 5.0%. Two disclosures explain the rest: Anthropic told investors annualized revenue reached $65bn at the end of July, below the figures circulating privately, and a Wall Street Journal tally put roughly $3trn of off-balance-sheet commitments on nine tech balance sheets. Fabrinet's own downtrend had begun earlier: its 50-day average crossed below its 200-day on 6 August, twelve sessions before the print. Corning's did the same after its 28 July call, having already fallen 18.2% in late July and fully recovered by mid-August.

What the market now pays

Fabrinet trades at 28.1x trailing gross profit, down from about 50x in May and 42x in February, and 21.4x forward — on gross profit that grew 34.6% to $556.5m over the same stretch. It is cheaper than at any point this year and only marginally above the 24x it carried a year ago, with consensus 2027 earnings of $18.20 a share against $13.05 delivered.

Ciena and Corning are different. Ciena's 23.4x is a third off May's 39.6x but more than triple the 7.1x of a year ago, and its trailing price-to-earnings ratio is 128x. Corning's 20.9x has given back the entire 2026 re-rating yet remains nearly double the 11.4x of last August. Those declines are doing rational work against multiples that had tripled.

The anomaly is Amphenol, which makes electrical and fiber-optic connectors and interconnect systems. Its June-quarter revenue grew 55% with gross margin up more than four points to 40.5%; information-technology datacom was 43% of sales and rose 89%, orders hit a record $10.7bn for a book-to-bill of 1.23x, and management reported no material supply bottlenecks. It is the cheapest of the four on trailing gross profit at 17.3x — and the only one whose uptrend has held unbroken since mid-June.

The threat that would justify a permanent de-rating of Fabrinet is co-packaged optics, which moves the light source onto the switch package and cuts the pluggable module out. That is not yet happening at scale: co-packaged optics is about 0.5% of AI data-center optical modules, with NVIDIA's Rubin Ultra generation targeted for late 2027. Corning, meanwhile, mapped $10bn of photonics revenue by 2030 while calling scale-up adoption timing the biggest uncertainty in its plan and declining to guide 2027.

The setup

Where it stands — Three optical suppliers sold off together since May while revenue accelerated and margins held or expanded at each. Would confirm — Fabrinet's September-quarter datacom revenue returns to sequential growth as Amazon volumes ramp. Would invalidate — A second consecutive quarter of declining NVIDIA revenue, or gross margin below 12%. Watch next — Fabrinet reports fiscal first-quarter results in November; Ciena's fiscal third quarter lands in early September. Valuation — Fabrinet at 28.1x trailing and 21.4x forward gross profit, against 50.4x in May and 24.3x a year ago.

The suppliers to the nuclear industry do not share an income statement, and this summer they stopped sharing a share price. In the month to 21 August the businesses that dig and enrich uranium climbed, while the two closest to finished reactors fell. Nobody in the group disputes the demand story — data centers want firm, carbon-free power. What separates them is what each one earns today, and what its buyers are paying for a dollar of it.

The one that raised guidance and hit a low

BWX Technologies has built the reactors, fuel and missile launch tubes for the US Navy's submarines and carriers for decades under a security-cleared, sole-source arrangement — an unusually protected revenue base for an industrial company. Second-quarter revenue was $901.6m, up 18% year on year. Backlog reached $8.4bn, up 40%, on a trailing book-to-bill of 1.7 times. On 3 August the company raised every 2026 line it guides: revenue to roughly $3.8bn, adjusted earnings to $4.70-4.80 a share, free cash flow to $345-360m.

The shares then set a 52-week low of $156.95 on 20 August.

The single soft number sits in the smaller half of the company. Commercial Operations — steam generators, small-modular-reactor pressure vessels, TRISO fuel pellets, medical isotopes — grew revenue 72%, and its margin guidance slipped to about 13% from about 14% as BWXT hires and adds capacity. Government margin guidance went the other way, up to about 20.5%. The de-rating, in other words, is being charged against the start-up costs of the growth business, not the franchise. Management is conspicuously refusing to fund the speculative end: it licensed its mPower reactor design to Applied Atomics rather than absorb more than $600m of licensing spend, and is holding back full commercial fuel capex until a federal award lands.

Price per dollar of trailing gross profit tells the story cleanly. BWXT costs 18.5 times today, against 24.3 times three months ago, 25.9 times six months ago and 21.8 times a year ago — while trailing gross profit itself grew 11.5%. The forward price-to-earnings multiple has fallen to 33.1 times from roughly 46 times in May, on estimates that rose in between.

The one whose earnings collapsed and multiple expanded

Cameco mines and mills uranium in Saskatchewan, converts and fabricates fuel, and owns 49% of Westinghouse, whose AP1000 is the only large Western reactor with a completed US reference plant. Second-quarter revenue fell 7.2% to $814.1m. Gross profit fell a third. Net income fell to $25.2m, down 92%, which management tied to a non-recurring Westinghouse payment on the Czech Dukovany project booked a year earlier.

The contracted-book insulation did not show up. Cameco's average realised price was $67.79 a pound last quarter — 28% below the $94 long-term contract price, which has held since June at an 18-year high — because deliveries come off contracts signed when uranium was cheap. Yet the stock is up 43.5% over twelve months and costs 46.1 times trailing gross profit, against 32.8 times a year ago. Its forward multiple, 66.4 times, is essentially where it stood in May despite a 15.5% fall since February: estimates came down with the price.

What the price carries is Westinghouse, and that is dated. The unit filed a confidential draft prospectus on 31 July, the Department of Energy committed $17.5bn of conditional financing for AP1000 long-lead items in June, and Cameco tracks 91 AP1000 opportunities worldwide. An AP1000 ordered now produces electricity around 2033.

The one with no earnings at all

Oklo is designing the Aurora, a sodium-cooled fast powerhouse of 15-75 megawatts, and booked the first revenue in its history last quarter: $1.21m, against a $73.2m operating loss. Its dates moved forward, not right — its Groves isotope reactor reached first criticality less than 11 months after groundbreaking, the DOE approved the safety analysis for the Idaho unit, and the 2028 start-up stands. That unit proceeds under DOE authorisation rather than standard NRC licensing, which takes the regulator off the near-term critical path. What moved was the cost of those dates: 2026 operating cash use guidance rose to $120-150m from $80-100m, capital spending to $400-500m, funded by $1.9bn of at-the-market issuance that lifted diluted shares 25.8%. Of the $42.09 share price, roughly $17 is cash; the remaining $4.3bn of market value is an option on reactors not yet built, at 2.27 times book.

Rates explain one day, not the month

On 18 August the 30-year Treasury yield touched 5.323%, its highest since 2007, and all six nuclear names fell together, Oklo worst at -6.5%. It was the only such session of the month. A discount-rate story cannot explain a month in which the miners rallied through rising yields — and the fuel side has a mechanism of its own: the waivers that let US utilities buy Russian enriched uranium expire on 1 January 2028, removing about a quarter of US supply against Western capacity that cannot be expanded in time.

The setup

Where it stands — The nuclear supply chain split this summer: uranium names rallied on scarce enrichment, while the two closest to built reactors de-rated. Would confirm — BWXT's third-quarter revenue growth holding in the high teens with 2026 adjusted earnings guidance of $4.70-4.80 intact. Would invalidate — A cut to BWXT's Government Operations margin guidance of about 20.5%, or a Cameco realised price falling below last quarter's $67.79. Watch next — Westinghouse's public S-1 filing, and BWXT's federal fuel-programme award, both expected during 2026. Valuation — BWXT: 18.5x trailing gross profit, 17.1x forward, against 25.9x in February. Cameco: 46.1x trailing, against 32.8x a year ago.

Datadog's largest customer, an artificial-intelligence company running 17 of its products under a nine-figure contract, renewed this summer — and then cut how much it consumed.

That single decision reset the company's forward line. Datadog, which sells cloud monitoring software — infrastructure metrics, application traces, logs and security alerts — priced by volume consumed rather than per user, had just reported its fastest growth in two years. June-quarter revenue of $1.1215bn was up 35.6%, the fourth straight quarterly acceleration. Management then guided the September quarter to $1.135–1.145bn, or 28–29%, and the shares fell 19% in a session on 6 August, with analysts widely identifying the customer as OpenAI.

Volume arrives, margin leaves

The usage cut is the visible problem. The quieter one is that metered AI traffic is the least profitable revenue Datadog books. Gross profit grew 33.4% in the June quarter against revenue's 35.6% — the wrong way round for a business whose case is that inference and agent telemetry compound into recurring revenue. Gross margin has fallen about 180 basis points over three quarters, to 78.6% from 80.4%, because the compute and hosting that ingest all that telemetry sit in cost of revenue. Agent activity is real: tool calls through Datadog's model-context-protocol server rose fourfold in a quarter. Each one costs something to store.

Demand-side friction is arriving too. Management said finance chiefs are pushing back on bills and that customer priorities have moved from validating AI to optimizing its cost — which is why Datadog launched pricing that eliminates unpredictable charges for high-cardinality metrics.

What the selloff obscured is that the non-AI business is in its best shape in years. Growth from customers with no AI story reached the high 20s, a fifth consecutive quarter of acceleration from 18% a year ago. Net revenue retention sits in the low 120s. Customers spending over $100,000 a year number 4,720 and supply 91% of recurring revenue, and 13% of all customers now run ten or more products, up from 7%. Free cash flow was $279m, a 25% margin.

Palo Alto's growth has a purchase price

Palo Alto Networks — firewalls plus subscription threat prevention, cloud and identity security sold to large enterprises and governments — is the seat-and-platform counter-case, and its headline growth is partly acquired. In its April quarter the company reported next-generation security annual recurring revenue (ARR) of $8.1bn, up 60% — of which $1.6bn came with CyberArk and Chronosphere, leaving roughly 28% organic. Remaining performance obligation of $18.4bn grew 36%, with $1.8bn acquired, or about 22% organic.

The purchase shows up elsewhere. Gross margin fell 539 basis points to 67.6%. GAAP operating income swung to a $183m loss. Diluted shares rose 5.2% to 744m, because CyberArk was bought in February for $21.1bn, mostly in stock. Adjusted free cash flow, at $910m against $578m, held up.

The move was the sector's, not theirs

Both stocks entered uptrends in the last week of May — Palo Alto's a full week before it published anything, on 2 June. May was the best month for software stocks since October 2001, and since 6 May the peers moved in step: Fortinet up 71%, Okta 73%, Snowflake 130%, against Datadog's 64% and Palo Alto's 95%. Fortinet is the counter-evidence on quality: it grew 25.6% with a record 33.7% operating margin and free cash flow tripled to $966m, and its shares fell over the past month.

That leaves the price. Datadog costs 26.59x trailing gross profit and 23.61x forward, against 16.98x in early May — expansion, though gross profit largely kept pace. Palo Alto is at 38.22x trailing and 29.30x forward, against 20.18x in early May and roughly 16.7x in February, an 89% re-rating since spring on trailing gross profit that grew 4.9%. It has published no financial statement since 2 June.

The setup

Where it stands — Datadog is 18% below its August peak on accelerating non-AI growth; Palo Alto is near highs on three-month-old numbers.

Would confirm — Datadog's September-quarter gross margin printing below 78% while revenue growth holds near the guided 28-29%.

Would invalidate — Gross margin recovering toward 80% with growth above 30%, or Palo Alto disclosing organic next-generation ARR growth above 35%.

Watch next — Palo Alto reports fiscal fourth-quarter and full-year results on 1 September, its first new information since 2 June.

Valuation — Datadog 26.59x trailing and 23.61x forward gross profit versus 16.98x in May; Palo Alto 38.22x and 29.30x versus 20.18x.

Four of the companies that build the machinery of automation reported June-quarter results in the first week of August. All four beat expectations. All four fell — Rockwell Automation by 7.4% the session after it reported, Symbotic by 16.2% in a single day, Cognex by about 15% over the following fortnight. Between them they cover the physical layer that AI is supposed to be moving into: the cameras that inspect parts, the controllers that run a production line, the robots that pick cases in a distribution center.

The quarter that didn't count

Cognex sells machine-vision systems — image-based barcode readers, smart cameras and deep-learning software that let a factory or warehouse locate, identify and inspect parts without a person looking at them. June-quarter revenue rose 16.9% to $291m. Gross margin reached 70.6%, against 67.4% a year earlier, and operating income nearly doubled. Management raised full-year adjusted earnings guidance to roughly $1.66, up about 63% at the midpoint, and lifted its outlook for semiconductors, electronics and packaging to double-digit growth on AI infrastructure spending. Logistics, a quarter of 2025 revenue, grew double digits for a tenth consecutive quarter.

The objections raised against the print were small. Revenue landed fractionally under consensus, European revenue fell 15% in constant currency, and the shares had nearly doubled off a $34.60 low. Management also flagged about 75 basis points of gross-margin pressure in the second half from memory-chip price inflation — the same AI-driven memory shortage that is inflating its customers' bills — and said price increases should fully offset it by 2027.

What Cognex is not yet is an AI hardware story in dollars. Its data-center supply-chain business, inspecting components and server racks, is a low-single-digit share of revenue, though growing more than 30%. The durable asset is the largest installed base in industrial machine vision, which supplies proprietary training data for edge models; the threat is Keyence, the Japanese sensor maker that has consistently grown faster, and Zebra Technologies, whose machine-vision unit is compounding at strong double digits. Cognex added 4,500 new customers so far this year.

At 57.7x trailing and 36.0x forward earnings, against a consensus that has 2026 earnings per share reaching $1.68 versus $0.68 reported for 2025, the stock is expensive in absolute terms and cheaper than it has been all year relative to its own profits.

Where the de-rating is earned

Rockwell Automation, the largest pure-play industrial automation supplier and owner of the Logix and FactoryTalk control ecosystem that factories rarely switch away from, grew 10% organically and raised guidance for a second straight quarter. Warehouse and e-commerce orders rose 30%. But Lifecycle Services shrank 2% organically, its book-to-bill was 0.97, and management conceded the business is only now back to pre-COVID controller volumes. At 33.1x forward earnings against consensus 2027 revenue growth of 5.5%, the multiple was carrying more than the orders book.

Symbotic, which builds complete robotic warehouse systems for retail distribution centers, delivered the cleanest operational improvement: revenue of $721m, gross margin up to 22.3% from 16.6% six quarters ago, and a first meaningful GAAP profit. Yet growth is decelerating — 29.4%, then 23.1%, then 21.7% year on year — backlog is flat at $22.5bn, and more than 84% of fiscal 2025 revenue came from Walmart. The back-of-store rollout that would broaden it will not touch results until early 2028. At 47x trailing gross profit it is the most expensive name in the group on every measure.

Teradyne is the reason this group looks like a winner at all. Its shares are up 236% in twelve months on semiconductor test demand, where compute revenue rose sixfold; the other three average 22%. Its robot arms and mobile-robot business — the independent read on cobot demand — grew 33% to $100m, still only about 7.5% of sales.

What actually moved them

The premise of an automation trough is wrong: July's ISM Manufacturing index registered 55.6%, the highest since May 2022, with factory employment expanding for the first time in 33 months. The mid-August leg down came from the bond market. The 30-year Treasury yield hit a 19-year high on 18 August, and the Philadelphia semiconductor index fell 5% as higher discount rates cut what investors would pay for profits arriving later. Teradyne fell 8.8% that day alongside Aehr Test and FormFactor.

The result is uniform. Measured against trailing four-quarter gross profit — the least distorted lens where three of the four have volatile tax and stock-compensation lines — every name got cheaper over three months while its gross profit grew: Cognex from 15.0x to 13.4x, Rockwell 10.7x to 9.9x, Teradyne 25.0x to 21.7x, Symbotic 63x to 47x. Earnings are catching up to price. In late July and early August, Teradyne, Cognex and Rockwell each rolled from a steep uptrend into a shallower one, in that order, while Symbotic's spring downtrend eased on 19 August.

The setup

Where it stands — Cognex's profits are accelerating while its multiple compresses; Rockwell and Symbotic are de-rating against outlooks that justify it. Would confirm — Cognex September-quarter gross margin holding near 70% despite the flagged memory-cost headwind. Would invalidate — Cognex logistics revenue breaking its double-digit growth streak, or full-year guidance trimmed below $1.64. Watch next — Cognex reports third-quarter results in late October; Rockwell closes fiscal 2026 in September. Valuation — Cognex 57.7x trailing and 36.0x forward earnings, against 78.7x trailing in early May.

Brookfield Renewable Partners — a Toronto-based partnership that owns hydroelectric, wind, solar and storage plants across the Americas and Europe, and controls 51% of the reactor builder Westinghouse — told investors on 31 July it had earned record funds from operations of $421m in the second quarter, up 13% from a year earlier. Management confirmed that about $175m of that figure was "other income" in the hydro segment: gains on developed assets and non-core disposals. Roughly two-fifths of the record, in other words, came from selling plants rather than running them.

The accounting statements tell the plainer version. Revenue fell 1.8% to $1.66bn. Gross margin dropped to 21.3% from 58.7% a year earlier, operating income declined 15%, and the partnership posted a net loss of $101m — its fifth negative quarter in eight. That is why the 71.2x trailing price/earnings ratio is unusable and why unitholders are asked to look at funds from operations instead. At $2.14 per unit over the last twelve months, the units change hands at about 15.3 times a measure that is currently part non-recurring.

The recycling is not hidden or improper — Brookfield's model is to build, sell high and redeploy, and it agreed or closed $2.2bn of disposals in the quarter against $5bn deployed or committed to growth. It commissioned 1.3 gigawatts and signed power purchase agreements on 2.6 GW. But the growth is being funded expensively: the partnership priced C$750m of green bonds on 20 August, the ten-year tranche at 4.949%.

Clearway is the inverse trade

Clearway Energy owns roughly 5,000 net megawatts of installed US wind and solar plus about 2,500 MW of gas-fired capacity, all sold forward under long-term contracts, and is run with 60 employees. Its second-quarter revenue rose 22.7% to $481m, the third straight quarter of accelerating growth, while operating income climbed 36.5% to $116m. In June it restructured the contracts on all three of its Texas wind projects, extending them beyond 2040 at better pricing. The one real deterioration is a weather-driven cut to 2026 cash available for distribution, to $430-470m from $470-510m, which management attributes to a poor first-half wind resource and calls transitory; it reaffirmed a target of $2.70 or better per share for 2027.

The shares are down 16.5% over three months and sit at 0.72 times book value — below the carrying value of a fleet that is mostly contracted, operating generation. Clearway's trend broke on 20 July, a fortnight before the guidance cut, and the 5 August cut session does not rank among its 25 worst days since February. Its five worst sessions of the period contain no earnings date at all.

The variable is the 30-year Treasury

That is the tell. The long bond reached about 5.31% on 17 August, its highest since 2007, and topped 5.33% the next day on deficit issuance and sticky inflation. Clearway's declared quarterly dividend annualizes to $1.8408, a yield near 5.7% — leaving a spread over the 30-year of well under a percentage point. Brookfield's second-worst session since February, a 4.76% drop, was 18 August itself.

The controls confirm this is not a developer problem. NextEra Energy, which pairs the Florida Power & Light utility with the largest US renewables developer, grew second-quarter revenue 12.4%, lifted its backlog to 35.1 GW and is recontracting output at roughly a $20/MWh premium — and its uptrend broke on 18 August. Constellation Energy, the biggest unregulated nuclear operator in the country, raised full-year guidance to $11.50-12.50 a share and is down 11.9% over twelve months. Whatever is repricing these cash flows is not the demand line.

Policy adds a dated cliff. The One Big Beautiful Bill Act terminated the production and investment tax credits for wind and solar projects beginning construction after 4 July 2026 and placed in service after 2027, while a federal court vacated the Treasury guidance narrowing the 5% safe harbor in June — with appellate review likely to outlast the deadline it governs.

And the AI offtake everyone is underwriting remains mostly promissory. Brookfield's Microsoft agreement commits $10bn to more than 10.5 GW between 2026 and 2030, but it is a framework with no disclosed delivery split. Clearway's 17-plus GW of co-located data-center complexes sit at the sponsor, with its own first investment window around 2030. The demand does show up in prices — North American solar contracts hit $64.49/MWh in the first quarter, an index high — just not yet in either company's contracted cash.

On multiples, Brookfield is the cheapest of the four on enterprise value to EBITDA at 9.85x, against Clearway at 14.6x, Constellation at 14.7x and NextEra at 15.9x. The catch is that the support rests on a metric propped up by disposals. Clearway's is the cleaner arithmetic: about 12.1 times its reaffirmed 2027 distributable-cash target, on a trailing free-cash-flow yield of 10.0%.

The setup

Where it stands — Clearway's operations are improving into a falling share price; Brookfield's units hold up on earnings quality that is partly non-recurring. Would confirm — Brookfield's next quarter showing FFO per unit growth with disposal gains below 10% of the total. Would invalidate — A retreat in the 30-year Treasury below 4.75% that leaves both names still falling. Watch next — Brookfield unitholders vote on the one-for-one BEP/BEPC merger on 14 October; NextEra shareholders vote 3 September on the Dominion issuance. Valuation — Brookfield 9.85x trailing EV/EBITDA and ~15.3x last-twelve-month FFO per unit; Clearway 0.72x book, 12.1x its 2027 cash target.

For a contract manufacturer, buying a chip designer's captive assembly business is a wager that someone else's cost center can be your profit center. Sanmina, the San Jose group that builds circuit boards, backplanes, cable assemblies and enclosures for industrial, medical, defense and telecom customers, placed that wager on ZT Systems, acquiring the data-center infrastructure manufacturing arm from Advanced Micro Devices for $2.55bn in cash plus up to $450m of contingent consideration. AMD kept the systems-design engineers and sold the factories. The first full quarter of ownership is now reported.

The accretion test

Revenue in the quarter ended 27 June reached $3.46bn, up 69.7% from a year earlier. The important number is beneath it. Gross margin expanded 160 basis points to 10.49%, and operating margin rose to 6.43% from 4.70%. Gross profit doubled while revenue grew by two-thirds — the opposite of what pass-through volume does to a book.

Inside the Integrated Manufacturing Solutions segment, where ZT sits, revenue rose 79.4% to $2.96bn and gross margin widened 270 basis points to 10.2%. ZT contributed $1.1bn of that; the legacy Sanmina business grew 14.1% on its own. Management attributed the margin gain to mix including ZT, plus non-recurring engineering work. The acquired factories are earning more per dollar of revenue than the business that bought them, not less.

That matters beyond Sanmina, because it identifies where the value in an artificial-intelligence rack actually settles. Hyperscale buyers increasingly bypass the branded server vendors and contract directly with the manufacturers for full-rack builds — integration, power distribution, thermal work and test. Those are engineering services sold by the rack, not metal bent by the hour.

Two corroborations

Celestica, the Toronto manufacturer that designs and builds switches, interconnects, servers and storage for hyperscalers, reported second-quarter Connectivity and Cloud Solutions revenue of $3.81bn, up 84%, at an 8.7% segment margin against 8.3% a year earlier. Its smaller Advanced Technology Solutions unit added 100 basis points to reach 6.3%. Total revenue grew 62.4% and operating income 68.2% — leverage, not dilution. Celestica now sells its own 1.6-terabit switch designs rather than assembling someone else's, with ten such programs ramping.

Flex, the Singapore-founded group whose plants turn out power supplies, switchgear and busway alongside automotive and medical assemblies, is the diversified control. Its Cloud and Power Infrastructure segment grew 35% to $2.2bn at a 9.7% operating margin — roughly double the company's consolidated 4.94%. Flex raised fiscal 2027 guidance to $33.7-35.2bn of revenue and said it is on track to spin the segment out as a separate public company in the first quarter of calendar 2027. Three companies, three structures, one direction of margin.

What would break it

Sanmina's own management named the fragilities. The non-recurring engineering benefit continues into the fourth quarter and then ramps down. Legacy ZT programs are declining toward zero as the accelerated-compute work replaces them. Working capital will build as that ramp proceeds, pressuring near-term cash flow. And the concentration has migrated from customer to end market: cloud and artificial-intelligence infrastructure is now 62% of revenue, at $2.15bn, against a $1.32bn industrial, medical and defense book growing 4.8%. Sanmina guides fiscal 2026 to $14.0-14.3bn of revenue and diluted earnings of $11.90-12.20, with fiscal 2027 targeted above $16bn. Net leverage is 0.29x.

The shares have gone the other way. Sanmina fell 18.3% over three months, Celestica 14.4% and Flex 16.0%, a stretch that includes a 19-year high in the 30-year Treasury yield on 18 August — a discount-rate move that lands hardest on order books dated 2027 and 2028. Sanmina's price against a dollar of trailing gross profit fell from 13.2x in May to 8.8x. At 33.1x trailing and 15.6x forward earnings, it carries the widest gap between the two in this group; consensus has fiscal 2026 earnings at $12.10 against $4.46 delivered last year.

The setup

Where it stands — Sanmina's acquired factories lifted segment gross margin 270 basis points in their first full quarter, while the shares fell 18% in three months. Would confirm — Fourth-quarter IMS gross margin holding at or above 10% after the non-recurring engineering benefit fades. Would invalidate — Fiscal 2027 revenue guidance below the stated $16bn, or IMS gross margin returning toward last year's 7.5%. Watch next — Sanmina's fiscal fourth-quarter results, due late October, the first quarter without the flagged engineering tailwind. Valuation — 33.1x trailing and 15.6x forward earnings; 8.8x trailing gross profit against 13.2x three months ago.