DK Street Journal

Two end markets, one industry

On 29 July, Lennox International — which sells furnaces, air conditioners and heat pumps to American homeowners through a distribution network, alongside light-commercial rooftop equipment — told investors that the replacement wave it had expected in the back half of this year would not arrive until 2027. Management blamed affordability, elevated mortgage rates and weak consumer confidence. Residential revenue fell 7% in the June quarter, and unit volumes fell 12%. New-construction revenue dropped about 30% after the company declined low-margin work. Full-year adjusted earnings guidance came down to $23-$24 a share, and the shares fell about 19% that day.

That is not a company-specific accident. American residential cooling demand is running below the natural replacement rate — roughly 7.5m units against the 8m implied by replacing 130m installed systems every sixteen years — and the distributors' trade association expects only 2-4% volume growth this year. Lennox's own commercial division, meanwhile, grew 24% and had its guidance raised to about 20%.

A day later Trane Technologies, the Irish-domiciled maker of chillers, air handlers and building controls that also owns the Thermo King truck and container refrigeration business, reported organic bookings up 37% and a record $12.1bn backlog, up 70%, with about $6bn of it scheduled for 2027 and later. Bookings for applied systems — the large engineered chiller plants a data center needs — more than doubled for a fourth consecutive quarter, rising 130%. Four individual orders exceeded $100m.

The margin nobody expected

AAON, the Tulsa manufacturer whose BasX division builds purpose-built data-center cooling and pre-packaged outdoor mechanical rooms alongside its legacy rooftop units, is the name most often described as the small-cap way to own this build. Its June quarter was extraordinary: revenue doubled to $627m, BasX-branded sales rose 216% to $345m, and backlog reached $2.0bn, up 98%.

Gross margin nonetheless fell 230 basis points to 24.3%, a fourth straight quarterly decline from 34.9% two years ago, and full-year gross-margin guidance was cut 200 basis points to 25-26%. Part of the reason is the ramp of a new Memphis plant, which absorbed $18.1m of overhead against $3m a year earlier. But management also attributed the compression to faster BasX growth carrying lower margins than the Oklahoma rooftop business. The engineered data-center dollar dilutes the mix rather than enriching it — the reverse of the case usually made for it. Backlog also slipped sequentially from $2.13bn. The competitive backdrop explains the pricing: in liquid cooling no vendor is dominant, with Vertiv leading at just over 11% and the top five holding about 35%.

Three declines, three causes

Between 22 July and 21 August, AAON fell 25.9% and Lennox 24.4% while Trane lost only 4.3% and the broad market rose. Lennox's entire move was two sessions around its guidance cut. AAON's came in three legs: a 20.8% slide before it had reported anything, alongside Vertiv's 17% drop on a revenue miss; 8.8% on its own results day; and 8.9% in the week the 30-year Treasury yield hit 5.33%, a 19-year high, which marks down every long-dated equipment backlog. Trane had recovered its post-earnings dip in full by 5 August.

Over twelve months the split is by end market, not by product. Vertiv is up 105%, nVent 76% and Modine 45%; Lennox is down 29%, Watsco 22% and Carrier 7%. AAON's remaining non-residential market is genuinely shrinking: non-residential construction spending fell 7% in nominal dollars over the first five months of the year.

What the prices now assume

AAON trades at 13.2 times trailing gross profit, against roughly 26 times in late May and 22 times in February — gross profit grew 28% over that stretch while the shares fell. That is now level with Trane's 12.7 times, a multiple unchanged since May, despite AAON growing many times faster; AAON's free cash flow is negative, at -1.8% of market value on $102.6m of capital spending so far this year. Trane converts at 3.8%. Lennox, at 7.9 times trailing gross profit and 17.0 times forward earnings, is the cheapest of the three on every measure, and its 5.3% free-cash-flow yield is the highest — the price of a business waiting on the homeowner.

The setup

Where it stands — One industry, two end markets: commercial and data-center orders at records, residential volumes contracting, and the shares punished together. Would confirm — AAON's September-quarter gross margin lands inside the new 25-26% guidance while BasX backlog resumes sequential growth. Would invalidate — Trane's applied bookings growth falls below 100% year on year, ending four straight quarters above it. Watch next — Lennox's third-quarter report in late October, the first test of the $23-$24 full-year earnings guidance. Valuation — AAON 13.2x trailing and 11.1x forward gross profit versus about 26x in May; Trane 12.7x; Lennox 7.9x.

The Trade Desk, which runs a self-service platform where ad agencies buy display, video, audio and connected-television inventory across the open web, spent its August earnings call explaining a market rather than a company. Roughly a quarter of the spending on its platform comes from consumer-goods and automotive advertisers, and those advertisers are pulling back: carmakers and retailers are both cutting budgets as tariffs squeeze margins. Management added that some large brands had moved to cheaper programmatic-guaranteed buying instead of paying for premium decisioning, and acknowledged execution gaps of its own. Second-quarter revenue grew 3.0% to $715.1m. The third quarter was guided to at least $650m — about a 12% fall from a year earlier, and some 19% below what analysts expected.

Three smaller companies selling into the same marketing budget reported the opposite quarter.

The quarter the challengers had

Viant Technology, an independent buying platform worth $847m — roughly a seventh of The Trade Desk — grew second-quarter revenue 33.9% to $104.3m, with customer spending on connected TV up about half. Its Direct Access product, which buys inventory straight from publishers rather than through exchanges, went from about 50% of connected-TV spending on the platform to more than 80%, and Viant says the route cut the cost per thousand impressions by 35%. Management said friction between agencies and its largest competitor was pushing budgets its way.

The more telling number sits elsewhere in Viant's disclosure. Outcomes, an autonomous buying tool launched six months ago and positioned explicitly against Google's Performance Max and Meta's Advantage+, already accounts for 5% of all advertising spend on the platform. Model-native buying is taking share inside the open internet, not only outside it.

Digital Turbine, which places apps and in-app advertising directly onto Android handsets through carrier and manufacturer deals, grew fiscal first-quarter revenue 26.8% to $166.0m. Gross margin improved to 49.4% from 47.3%, full-year guidance went up to $650m-$670m, and net leverage fell to 2.5x from above 5.0x. Its account of where the money is going is blunt: open-web traffic is down roughly 10% year on year as AI answers displace clicks, with some categories off 20% to 40%, while app releases grew 60% and the average person now spends about five hours a day inside apps.

Zeta Global, a marketing cloud that predicts consumer intent from opted-in identity data, grew revenue 43.5% to $442.8m — 28% of that organic — at a 59.1% gross margin, and raised full-year guidance to $1.818bn. It is the group's follower rather than its argument: Zeta sells software into marketing departments and makes no autonomous-bidding claim, and its growth suggests marketing budgets are being redirected rather than cancelled.

What The Trade Desk still has

The deceleration is not one bad quarter. Revenue growth ran 17.7%, then 14.3%, then 11.8%, then 3.0% across four quarters, and second-quarter operating income of $101.6m was 13% lower than a year earlier. Customer retention is not the problem — it has stayed above 95% for more than a decade — but revenue per retained customer is. Connected TV still grows double digits and is now more than half of revenue; audio has been the fastest-growing channel for four straight quarters; joint business plans with large advertisers rose 38%.

The pressure is on price. The Trade Desk's take rate has sat between 19% and 21% since 2017, and Amazon's demand-side platform charges nothing on its own inventory and about 1% on some streaming deals. Where the reallocated dollars land is visible in the largest sellers' results: Amazon's advertising segment grew 26% to $19.8bn and Meta's grew 27% to $59.4bn in the same quarter, against a forecast 5.0% for total ad spending in 2026. Five times the market is not a cycle.

The price

The Trade Desk's 15.7x trailing earnings looks cheap and is an artifact: the forward multiple is higher, at 32.8x, because consensus has 2026 earnings per share at $0.40 against $0.91 last year, and 2027 lower again at $0.24. Price paid per dollar of trailing gross profit has fallen from about 5.3x in February to 2.50x, and free cash flow now runs at 13.9% of the market value. Among the challengers, Digital Turbine trades at 16.7x forward earnings — its trailing figure is negative — and Zeta at 29.2x forward, or 7.3x trailing gross profit.

The shares have followed the results. The Trade Desk's 50-day average has sat below its 200-day since 13 November 2025, an unbroken nine-month downtrend, and the stock lost 21.9% on 7 August on nine times its normal volume. Over the past month Zeta rose 38.9%, Digital Turbine 25.4% and Viant 11.0%, while Criteo fell 17.0% — advertising technology is sorting by result, not moving as a block. Two apparent gainers in the group are not re-ratings at all: Semrush is frozen at $12.00 after Adobe's cash takeout closed in April, and DoubleVerify, an ad-verification business that grew 2.5% last quarter, jumped because Nielsen agreed on 6 August to buy it for $13.60 a share.

The setup

Where it stands — The Trade Desk guides to its first revenue decline while three smaller platforms selling the same budget grew 27% to 44%. Would confirm — Viant's third-quarter revenue growth holds above 25% while The Trade Desk lands at or below its $650m guide. Would invalidate — The Trade Desk returning to double-digit growth in the fourth quarter with consumer-goods and auto spending recovering. Watch next — Third-quarter results in early November from The Trade Desk, Viant, Zeta and Digital Turbine. Valuation — The Trade Desk: 15.7x trailing earnings but 32.8x forward; 2.50x trailing gross profit against 5.3x in February.

The line between copper and fiber is dissolving

Credo Technology, a San Jose designer of high-speed connectivity silicon whose active electrical cables (AECs) carry data between accelerators inside an AI rack, has spent two years being described as the company that keeps optics out of the rack. Copper reaches about seven metres and draws roughly half the power of an optical link, which is why hyperscalers bought it, and why the company is the de-facto standard for intra-rack and rack-to-rack hops.

Its own guidance points somewhere else. On the June call covering the quarter ended 2 May, management guided fiscal 2027 to more than 80% revenue growth and said optical digital signal processors (DSPs), a silicon-photonics interposer and its ZeroFlap optics line would contribute more than $600m of it. Against fiscal 2026 revenue of $1.335bn, that is not a side project. Management also said scale-up copper — the accelerator-to-accelerator links the bull case rests on — contributes little in fiscal 2027 and becomes material only in fiscal 2028.

Astera Labs is travelling the same road from the other end. The Santa Clara firm sells PCI Express retimers, smart cable modules and, since this year, the Scorpio fabric switch that stitches accelerators together inside a rack. On its 4 August call it laid out a three-phase optical plan: high-density fiber with a tier-one artificial-intelligence provider in 2027, near-package-optics chipsets the same year, and co-packaged optics built into Scorpio from 2028. An optics stumble, in other words, is not a copper windfall for either of them.

What the two businesses did

Credo's fourth-quarter revenue was $437.0m, up 157.0% year on year. Gross margin reached 68.2%, the fifth straight quarter above the mid-sixties band the company was long assumed to live in, and operating margin widened from 20.4% a year earlier to 35.7%. The crack is sequential rather than annual: quarter-on-quarter growth ran 51.9% in the January quarter and 7.4% in the May quarter. Concentration is unresolved — the top four customers were 34%, 27%, 16% and 10% of revenue, about 87% between them.

Astera Labs is accelerating where Credo decelerated. June-quarter revenue of $392.4m grew 104.5% year on year, up from 93.4% growth in March, and management guided the September quarter to $540-560m, roughly 40% sequential growth. Gross margin is the cost: 76.3% in March, 73.3% in June, guided near 72%, with a long-run target of 70%. Management attributes the step-down to Scorpio's board-level mix rather than price, and says Scorpio margins converge on the corporate average at volume. Scorpio X alone is targeted above $1,000 of content per accelerator.

The two also compete directly. Astera's Taurus line is shipping pre-production 100G-per-lane silicon for 800G active electrical cables, with a ramp expected in the second half — the socket Credo owns.

The dislocation is next door

MACOM Technology Solutions makes analog and radio-frequency-to-lightwave chips for optical networking, data centers and defense radar. Its July-quarter revenue was $342.2m, up 35.8% year on year, with gross margin expanding to 58.3% from 53.7% and a record 1.6-to-1 book-to-bill; guidance calls for $415-425m. The shares are down 29.9% over three months. At 30.9x trailing and 27.2x forward gross profit, MACOM is the one name here whose numbers accelerated while its price did the opposite.

Coherent, a $55.6bn laser and photonics manufacturer with more than 51,000 employees, is the cheap end of the same rack: June-quarter revenue of $2.046bn grew 33.7% with a fourth consecutive quarter of margin expansion, to 38.5%, and it trades at 20.8x trailing gross profit.

Against those anchors, Credo at 47.3x trailing gross profit has barely moved — it was 49.8x on 21 May — while trailing gross profit grew 25.4% over the same three months. Its 37.3x forward earnings embeds the guided acceleration. Astera Labs at 54.1x trailing gross profit has come down from 67.1x in May but still costs 57% more per dollar of gross profit than the 34.4x of 20 February, on 72.1x forward earnings.

The week the discount rate moved

Between 17 and 21 August the whole rung sold off together, and copper got no shelter: Credo fell 18.5% and Astera Labs 11.0%, against Lumentum's 10.6% and Coherent's 17.6%. Fabrinet, which assembles optical modules, fell about a quarter after reporting record revenue of $1.316bn and guiding higher. The S&P 500 tracker lost 1.0% over the same four days. The mechanism was rates and demand narrative, not order books: the 30-year Treasury yield hit a 19-year high above 5.33% as a Treasury buyback plan failed to hold it, and a Wall Street Journal report of disappointing OpenAI second-quarter results hit AI infrastructure names the same session. Astera Labs' uptrend weakened in early August, its shorter moving average flattening toward the longer one, and has not recovered since.

The setup

Where it stands — Both interconnect leaders are building optical businesses, so the copper-versus-fiber split no longer separates them; MACOM's numbers and price point opposite ways.

Would confirm — Credo's fiscal first quarter shows optical DSP and ZeroFlap revenue disclosed as a growing line with gross margin held near 68%.

Would invalidate — Credo's top-customer share climbs above 34% while sequential growth stays below 10%, showing one program, not a broadening base.

Watch next — Credo reports fiscal first-quarter results for the quarter ended 1 August after the close on 1 September 2026.

Valuation — Credo 47.3x trailing and 25.7x forward gross profit; Astera Labs 54.1x and 34.4x; MACOM 30.9x; Coherent 20.8x.

The work of physically building an artificial-intelligence data center — bolting accelerators into racks, assembling 800-gigabit switches, wiring power shelves — is done by companies whose names appear on nothing. Three of them are US-listed, and in the past year each has taken a very different route into the business.

Celestica, a Toronto-based builder of switches, servers and data-center interconnects to hyperscalers' own designs, priced 9,677,419 shares at $310 in a $3bn offering against a prior close of $362.76, adding about 8% to its share count to fund working capital and plant. Sanmina, a San Jose maker of circuit boards, backplanes, enclosures and full systems, bought ZT Systems' data-center manufacturing operation from AMD for $2.55bn in cash and stock plus contingent consideration. Jabil, the Florida group that also makes medical devices, car electronics and packaging, bought neither, and instead shrank its diluted share count from 116.7m to 106.5m in under two years.

The accounting that sets the margin

Whether rack assembly is a real profit pool depends on who owns the accelerators. Under buy-and-sell terms the assembler purchases the graphics processors and memory, resells them inside the finished rack, and books enormous revenue at almost no incremental margin. Under consignment the customer owns the silicon and the assembler charges a conversion fee on a far smaller revenue base.

The damage is already visible in Taiwan, where Foxconn, Quanta and Wistron dominate rack assembly. Quanta's gross margin fell to 4.78% in the March quarter from 7.92% a year earlier precisely because of buy-and-sell accounting, and it is now negotiating customers toward consignment, with first orders expected in the second half. Celestica's gross margin in its June quarter was 12.29%; Sanmina's was 10.49%. The commoditization has not reached them.

What each one actually earns

Celestica has the strongest numbers in the group and the most concentrated ones. Revenue growth has accelerated four quarters running, to 62.4%. Its Connectivity and Cloud Solutions segment did $3.81bn, up 84%, at an 8.7% margin versus 8.3%, and the smaller Advanced Technology Solutions unit widened to 6.3% from 5.3%. Hardware Platform Solutions, the higher-value line where Celestica supplies its own designs, reached roughly $1.9bn. Management guided the year to $20.5bn of revenue and $11.30 of adjusted earnings per share. The offset is disclosed in the quarterly filing: three customers were 32%, 17% and 14% of revenue. Capital spending ran at 5.6% of revenue against 1.1% a year earlier, and trailing free-cash-flow yield is 1.52%, the lowest of the group.

Sanmina's headline growth of 69.7% is mostly purchased. ZT contributed about $1.1bn of the quarter's $3.464bn; core Sanmina's manufacturing revenue grew 14.1%. What the acquisition did deliver is margin: group gross margin has risen three quarters running, from 7.56% to 10.49%, and the manufacturing segment's margin gained 270 basis points to 10.2%. Cloud and AI is now 62% of revenue. The concentration sits inside the acquired unit rather than at group level — the 10-Q shows no customer at 10% of sales, while ZT's output is entirely AMD-based accelerated compute.

Jabil is the laggard on growth and the leader on cash. Revenue growth halved to 11.8% from 23.1%, and operating income grew more slowly still, at 10.4% — the leverage ran backwards even as gross margin improved to 9.46%. It expects about $13.6bn of AI-related revenue this fiscal year, up 50%, inside an Intelligent Infrastructure segment guided to roughly $16.5bn. Its trailing free-cash-flow yield, 4.58%, is three times Celestica's.

What the market pays

Because reported revenue at all three is inflated by pass-through component cost, the honest denominator is gross profit. Celestica now trades at 18.8 times trailing gross profit, down from about 26.5 times in May, while those profit dollars grew 26% since February. Sanmina is at 8.8 times, against roughly 14.1 times in May, and 15.6 times forward earnings versus 33.1 trailing. Jabil's multiple has round-tripped to 10.6 times, exactly its February level, on half the growth — the only one of the three where the decline is doing obvious work. Celestica still commands more than double Sanmina's multiple for broadly similar conversion work; it is cheap against its own history, not against its peers.

The selling itself was indiscriminate. Between 17 and 21 August all six US-listed assemblers fell between 12.8% and 16.3%, with no company news at any of them, after the 30-year Treasury yield topped 5.33% on 18 August, a 19-year high. Flex, whose cloud and power unit is guided to grow 65-75%, fell hardest at 16.3%; Plexus and Benchmark, which have the least data-center content, fell about as much as Celestica. Celestica's 50-day average crossed below its 200-day on 12 August; the other five have not. All six remain up more than 49% over twelve months.

The setup

Where it stands — Two of the three assemblers are expanding gross margin on AI rack work; the third is growing slower and generating more cash. Would confirm — Celestica's gross margin holding above 12% as 1.6-terabit programs ramp in the second half. Would invalidate — Gross margin at Celestica or Sanmina falling toward Quanta's 4.78% as buy-and-sell rack volume scales. Watch next — Jabil reports fiscal fourth-quarter results on 24 September, with fiscal 2027 guidance at a September briefing. Valuation — Price to trailing gross profit: Celestica 18.8x versus 26.5x in May, Sanmina 8.8x versus 14.1x, Jabil 10.6x versus 10.6x in February.

Both companies reported in the first week of August, and both beat. Entegris posted quarterly revenue of $883.2m, up 11.5% from a year earlier, and the shares jumped 15.5% on 4 August. Ultra Clean posted a record $644.9m, up 24.3%, beating on both lines and rising 7% in the session. Three weeks later Ultra Clean had given the entire move back and then some. Entegris had not.

What gets consumed and what gets ordered

Entegris supplies microcontamination-control filtration, ultra-pure specialty chemicals and gases, and the sealed carriers, called front-opening unified pods, that move wafers between tools. Management disclosed on its August call that about three-quarters of revenue is tied to wafer-start unit growth and only a quarter to customer capital spending, and raised its 2026 industry wafer-start growth assumption to 7-8% from mid-single digits.

The mechanism behind the growth is content per wafer, not tool count. Purity requirements at sub-5nm and 2nm nodes raise the number of filtration steps and the specification of the carriers, so the Advanced Purity Solutions segment reached $515m, up 17%, with liquid filtration at a fourth consecutive record and segment margin of 30.3%. Materials Solutions grew 5% to $371m, held back by mainstream logic, where foundry utilization is running at 80-85%. Memory is about 30% of revenue, and molybdenum precursor volume — used in the metal layers of newer memory stacks — roughly doubled year on year. Entegris holds an estimated 22-25% of the chemical-mechanical planarization consumables market after its 2022 CMC Materials purchase, competing with Fujimi, DuPont, AGC and Resonac for sockets that are hard to switch once qualified into a customer's process flow.

Operating income grew 55.1% on that 11.5% revenue line. The risk sits in geography: Asia-Pacific was about 79% of 2025 sales, and management has said expanded export controls reduced what it can sell into China.

The echo, one to two quarters behind

Ultra Clean builds precision gas and chemical delivery subsystems, weldments and process modules to its customers' designs, plus a smaller parts-cleaning and coating service. That is a different business from selling a qualified consumable: the volume arrives after the equipment makers book orders, and the margin reflects who owns the design. Gross margin was 16.1% last quarter, up from 15.3% a year earlier but structurally below the 19.6% of 2022, because the mix keeps shifting toward products — $572.7m against $72.2m of higher-margin services. Applied Materials and Lam Research together were 58.7% of 2025 revenue. Operating cash flow was minus $41.1m on a deliberate inventory build ahead of customer ramps, and management is building toward $3.5bn of annualized capacity by year-end.

None of that deteriorated in the quarter. The selling came from outside. A 1 July Bloomberg report that Meta would resell excess artificial-intelligence compute was read as evidence of overbuilding, and in three weeks Applied Materials fell about 23%, Lam Research 27% and KLA 28%. Applied Materials then reported revenue of $9.12bn, up 25%, and fell anyway; the group lost another 4-5% on 19 August on doubts about how long AI infrastructure orders last. What is being marked down is the 2027-28 order book, and Ultra Clean's revenue is that order book with a lag.

What each price now assumes

Entegris trades at 14.5x trailing gross profit, down from 16.4x in mid-May, and 13.5x forward — the advance since spring has come from profit, not from a higher multiple. The absolute level is still full: 29.2x trailing EV/EBITDA, 36.7x forward earnings and a 2.6% free-cash-flow yield. Ultra Clean sits at 9.9x trailing gross profit, essentially the 10.0x of early May, because gross profit rose roughly 7.5% while the shares fell about a quarter; forward, it is 7.9x gross profit and 24.2x earnings against consensus for 2027 revenue of $3.87bn.

Both stocks held a strong uptrend into late July and slipped out of it in early August, their 50-day averages still above the 200-day. From the 1 July peak Ultra Clean is down 41.0% and Entegris 13.0%. Over the past month Entegris was the only gainer among six suppliers to the fabs, while MKS fell 19.5% and Applied Materials 12.0%. The businesses agree on direction and disagree entirely on quality of revenue.

The setup

Where it stands — Entegris' consumable content is compounding at a cheapening multiple; Ultra Clean's faster growth is thinner, later and concentrated in two buyers. Would confirm — Entegris Q3 revenue of $905-935m with gross margin at 47.5-48.5%, and Ultra Clean printing $700-750m. Would invalidate — Materials Solutions failing to reach double-digit growth in the second half, or Ultra Clean guiding Q4 below Q3. Watch next — Third-quarter results from both in early November, with Applied Materials' order commentary before them. Valuation — Entegris 14.5x trailing gross profit and 13.5x forward, against 16.4x in mid-May; Ultra Clean 9.9x and 7.9x.

On 10 August, AECOM — an infrastructure consultancy that plans and designs highways, water systems, defense facilities and, increasingly, data-center campuses — published a record order book and a quarterly loss in the same release. A construction-management job awarded in 2019 took a $337m pre-tax charge for higher projected cost to complete, on terms the company now says it would not accept. The shares fell 19% over two sessions.

The charge is the noise. The signal is what the quarter, and those of its peers, reveal about where money from the artificial-intelligence build-out is actually landing — and it is not landing first on the drawing board, as the standard theory holds.

The theory, and the order books

The theory is sound. Roughly 2,200 to 2,600 gigawatts of generation sit in US interconnection queues, and the wait from application to commercial operation has stretched from under two years in 2008 to more than eight years by 2025. PJM, the grid operator for 13 states, has signed interconnection agreements for 103 GW since 2020; only 23 GW has entered service. In December 2025 federal regulators ordered PJM to rewrite its rules for generation co-located with large loads — more studies, more engineering hours.

Scarce capacity should show up as booked work. It has, modestly. AECOM's backlog rose 13% to $27.82bn, on quarterly wins of $4.2bn and a book-to-burn ratio of 1.6x. Stantec, the Edmonton-based engineering and architecture firm, reported a record C$9.2bn, up 17.5% — but only 7% excluding acquisitions, and equal to 13 months of work. Tetra Tech, a water and environmental consultancy, ended at $4.49bn, up 5% in the quarter.

Now the builders. MYR Group, an electrical contractor that erects transmission lines and substations, lifted backlog 20% to a record $3.16bn. Quanta Services, the largest US transmission contractor, grew its book about 50% to $53.4bn. The designers are adding work at a third to a half the rate of the firms that follow them.

More telling: not one of AECOM, Stantec or Tetra Tech discloses power, transmission, interconnection or data-center work as a dollar line. AECOM's management called data centers its fastest-growing business and said its water and Department of Defense pipelines each expanded about 30% in the quarter, without quantifying any of it. Stantec's named win is the design of Meta's $13bn Alberta campus, inside a Buildings unit where backlog grew over 40% — with nearly eight of its 17.5 points of growth coming from 2025 acquisitions.

Quanta, meanwhile, bought Percheron, a firm doing permitting, routing and land acquisition, saying it should compress interconnection timelines. Its largest high-voltage corridors are, in management's words, mostly in engineering now, with backlog entry from late 2026 and field work in the second half of 2027. The front-end work is happening. The contractor is doing it in-house.

Tetra Tech's growth is not fee revenue

Tetra Tech has posted three straight quarters of accelerating revenue: up 1.1%, then 10.6%, then 13.5%. Gross profit fell in every one of them, most recently by 3.3%. Gross margin dropped 323 basis points to 18.6%. The growth is subcontractor pass-through, not billable engineering hours.

Its stated headwinds are almost entirely unrelated to AI: federal contracting offices short-staffed, offshore-wind cancellations, and the collapse of the US Agency for International Development, which had been 10.6% of fiscal 2025 revenue and whose shutdown struck roughly $1.1bn of awards from backlog against a $92m impairment. US commercial revenue grew 1% as wind cancellations offset data-center and mining gains.

The shares rose 35.8% over three months anyway. Price per dollar of trailing gross profit is now about 10.1x, against roughly 7.4x in late May — a 37% re-rating on gross profit that is roughly flat year on year. The forward price/earnings ratio of 23.5x sits above the trailing 22.2x, because consensus models fiscal 2026 revenue 3.5% below fiscal 2025.

AECOM and Stantec: the opposite shape

AECOM's operating leverage predates the charge — fiscal 2025 operating income rose 24.1% on flat revenue, lifting operating margin to 6.36% from 5.14%. The full-year adjusted EBITDA margin target went up to 17.4%. The shares are down 46.7% over twelve months. The offset is cash: management cut adjusted earnings guidance to $3.95-$4.15 against a consensus near $5.97, and the problem projects will consume roughly $500m of cash through the first half of fiscal 2027, with buybacks deferred.

Stantec earned a record 18.7% quarterly EBITDA margin and lifted trailing earnings to C$4.44 a share from C$3.83, with operating margin at 14.4% versus 9.4%. Its shares fell 31.5%. The blemish is organic net revenue growth of 3.7%, with US organic growth flat.

One technical note does argumentative work. Between 17 and 21 August, as the 30-year Treasury yield touched 5.33%, a 19-year high, the contractors fell about 9%. Stantec was flat and Tetra Tech rose. Whatever repriced the builders that week did not touch the designers — their year-long de-rating has company-specific causes, one per company.

A customer class that did not exist at scale

American pipelines and electric utilities have spent the past year signing up a customer that barely existed five years ago: the hyperscale computing campus, which draws power like a mid-sized city and wants it contracted for a decade or more. Kinder Morgan, the Houston owner of about 83,000 miles of pipe that earns fees moving other companies' gas, refined products and crude, now carries a sanctioned project backlog of $9.6bn. Some 92% of it is natural gas, and more than 60% is aimed at power generation and data-center load. The backlog shrank from $10.1bn in the quarter only because $650m of finished projects entered service; management described a shadow list behind it worth over $10bn, with at least another $1bn expected to be sanctioned in the second half.

The operating numbers are moving the same way. Second-quarter adjusted EBITDA was a record $2.199bn, up 12%, and the company raised full-year 2026 adjusted EBITDA guidance to at least 5% above its $8.6bn budget, with adjusted earnings per share guided at least 12% above budget. Transport volumes rose 7% year on year; gathering volumes rose 26%, led by a 54% jump on its KinderHawk system in the Haynesville shale. Net debt fell to 3.6 times EBITDA from 3.8x at the start of the year, leaving roughly $3.4bn of borrowing capacity before the 4.0x middle of its target range. On 31 July federal regulators certificated two Southeast expansions — about 500 miles of new pipe, some $5.2bn of capital and 3.8m dekatherms a day of firm capacity.

The reason that matters is deliverability, not gas resource. Only one large interstate line, Mountain Valley, has been completed in a decade, and a new system takes five to eight years to permit and build. Working gas storage capacity grew 7% between 2010 and 2025 while demand grew far faster, cutting days of cover from 71.9 to 36.9. Incumbency on existing high-deliverability corridors is what makes Kinder Morgan the default counterparty for a new power or liquefied-natural-gas interconnect — and the same five-to-eight-year cycle is its main constraint on capturing the 2028-2030 squeeze with new-build capacity. Management also flagged tightening compression equipment, because turbine manufacturing capacity is being pulled toward power generation.

Against that, the shares fetch 7.00x trailing gross profit, down from 8.29x in mid-May, and 12.54x trailing EV/EBITDA against Williams at 15.72x. Kinder Morgan's trailing free-cash-flow yield is 5.61%; Williams' is minus 0.25%. The caveat sits in consensus, not the company: analysts model 2027 revenue growth of 2.0% and earnings of $1.53 against $1.52 this year, so the forward price/earnings ratio of 20.37x sits slightly above the trailing 19.86x. The backlog is not in the estimates.

The opposite balance sheet

NextEra Energy pairs Florida Power & Light, a regulated utility serving about 5.7m accounts over roughly 77,000 circuit miles, with Energy Resources, the country's largest developer of contracted wind, solar and storage. Second-quarter revenue rose 12.4% to $7.534bn and gross profit rose 43.4%. The signed development backlog reached 35.1 gigawatts after 3.6 GW of additions, and the company lifted its Florida large-load expectation to 8 GW by 2032 from 6 GW, each gigawatt worth roughly $2bn of rate-base investment. Florida regulators approved a four-year rate settlement in November 2025 with an authorized return on equity of 10.95% and base-revenue increases of $945m this year and $705m more in 2027. Earnings guidance of $3.92-$4.02 was left unchanged.

What differs is funding. NextEra's trailing free-cash-flow yield is minus 5.83% — a developer outspending its operating cash flow — and on 15 May it agreed to buy Dominion Energy at 0.8138 NextEra shares plus cash per Dominion share, a roughly $67bn transaction. That requires issuing about 716m new shares against 2,089m diluted, and raising authorized shares to 5.0bn from 3.2bn, at a special meeting on 3 September. The stock fell 4.63% on announcement and the two companies' combined market value dropped $5.0bn. The balance sheet is the asserted moat — the thing that lets NextEra underwrite multi-gigawatt contracts with hyperscalers — and this deal is what tests it. Dominion already has about 12 GW under executed electric service agreements; NextEra's Florida book remains 21 GW of interest with no announced transaction.

NextEra now trades at 8.37x trailing gross profit against 11.73x in mid-May, and near 21x its own adjusted 2026 guidance against the 24.44x price/earnings ratio it carried in May.

One session, five companies, no news

The rest of the delivery complex rolled over in the same week. Williams, the Tulsa operator of the Transco and Northwest systems and Kinder Morgan's closest rival for the same interstate revenue, raised its EBITDA growth target to 11%-plus a year and is moving into generation itself, putting a 200-megawatt plant in service in 18 months and closing a $5.34bn venture with Blackstone; its shares still fell 7.7% over three months. American Electric Power lifted contracted large-load additions to 69 GW from 63 and traded like a bond. On 21 August, Kinder Morgan fell 2.94%, Williams 3.87%, NextEra 1.88% and AEP 1.54% — while the S&P 500 gained 0.4% and the Treasury doubled buybacks of long-dated debt, pushing yields down from the 30-year's 19-year high of 5.33% three sessions earlier. Henry Hub gas rose 1.1% that day. No issuer news was discoverable for any of them.

So the usual explanations fail in sequence: not the discount rate, not the commodity, not guidance. Kinder Morgan was still in a clean uptrend on 10 August, its 50-day average above its 200-day, and had fully reversed by 21 August — a nine-session turn, not a slow de-rating, and it remains up 18.1% over twelve months. NextEra's deterioration began in June, and its 21 August close of $83.65 was its lowest since at least February. What is left is positioning: the same group of stocks sold by the same holders, at once.

The setup

Where it stands — Both businesses accelerated through the summer; both multiples compressed since May, Kinder Morgan's by 16% and NextEra's by 29%. Would confirm — Kinder Morgan sanctioning the flagged $1bn-plus of new projects, lifting backlog above $10bn, with power and data-center share holding above 60%. Would invalidate — A cut to Kinder Morgan's raised 2026 EBITDA guidance, or NextEra trimming its $3.92-$4.02 earnings range. Watch next — NextEra's special shareholder vote on the Dominion share issuance, 3 September 2026. Valuation — Kinder Morgan: 7.00x trailing gross profit, 6.86x forward, versus 8.29x in May. NextEra: 8.37x and 7.78x, versus 11.73x.

An agent business with a dollar figure

Salesforce, which sells the cloud software companies use to run sales, service and marketing organizations, reports fiscal second-quarter results on Wednesday 26 August. At its last print it did something it had spent a year avoiding: it attached a dollar figure to its AI agent product. Agentforce annual recurring revenue was $1.2bn, up 205% year over year, inside roughly $3.4bn once the Data 360 data platform and Informatica Cloud are included. The long-standing complaint — that Salesforce could report agent deals closed but not agent revenue earned — is out of date.

That matters because the entire question hanging over this corner of software is arithmetic. If an agent does the work of a licensed user, the licensed user goes away, and the vendor needs consumption-priced agent revenue to arrive faster than seat revenue leaves.

The numbers behind the accelerating line

Salesforce's growth has gone the right way for three consecutive quarters: 8.6% year over year, then 12.1%, then 13.3% to $11.13bn. Subscription and support revenue rose 14%, with the Data 360 and platform line up 25% to $3.68bn. Current remaining performance obligation — contracted revenue due within a year, the closest thing to a forward order book — reached $33.6bn, up 14%, with total obligations of $67.9bn.

The cleanest tell is gross margin. Inference costs land in cost of revenue, so a vendor selling AI badly bleeds there. Salesforce's held at 76.92% against 76.96% a year earlier. Operating margin widened to 21.8% from 19.8%, and operating income grew 25% on 13% revenue growth. Beneath that, $27.1bn of buybacks including a $25bn accelerated repurchase cut diluted shares 9.5%, to 871m.

So the share price rose 23% over the past month and the stock got cheaper against what it earns. At 5.48x trailing gross profit per diluted share it sits below February's 5.68x, because gross profit per share grew 17% over those six months. Forward earnings are 14.8x against 24.1x trailing, on consensus of $46.11bn of revenue and $14.16 of earnings per share for the January-2027 year.

What could still be true for the bears

Salesforce is a system of engagement, not infrastructure, and rivals are learning to read its data without paying it. Atlassian's latest shareholder letter lists Salesforce records and Slack conversations as context ingested into its own graph — Salesforce as a data source rather than a customer. Zeta Global, which sells a marketing cloud that scores consumer intent from opted-in data, told investors on 4 August that Gap had made Zeta its system of record, replacing Salesforce and three other vendors. Klarna's public departure last year was the second such exit after Workday. If the coordination layer consolidates above individual applications, the per-seat base erodes regardless of how fast Agentforce compounds.

The seat-priced control group

HubSpot, which sells an integrated marketing and sales suite to mid-market businesses, is where the thesis shows damage: revenue grew 19.8%, but net customer additions came in at 7,000 against 9,000–10,000 expected, second-half guidance was cut to 5,000–6,000 a quarter, and gross margin fell 159 basis points. The shares dropped 18% in a session on 6 August and have not recovered a bullish trend since.

The others diverge from that. Braze, which orchestrates cross-channel messaging for consumer brands, grew 30.2% — the fastest in the group — while giving up 290 basis points of gross margin and running a −13% operating margin. Freshworks, selling service software to companies under 20,000 employees, grew 16% with margin flat at 84.8% and cut diluted shares 7% this year. Zeta grew 43.5% but surrendered 295 basis points of margin on channel mix, and trades at 7.33x gross profit against Salesforce's 5.15x. Pegasystems is the deteriorating case: contract-value growth halved to 7% and management blamed AI-driven deal delays.

Twilio, whose application programming interfaces embed messaging and voice in other companies' software, is the consumption-priced counter-case and the most stretched. Revenue accelerated to 22%, net expansion is 116% — about five points of that is new US carrier pass-through fees — yet its price per dollar of gross profit per share went from 6.95x in February to 13.30x now, against 4.1% growth in that gross profit. Its whole month is one session: a 26.6% jump on 7 August, and down since.

The rally's actual cause

Every one of the seven rose over the past 30 sessions, an equal-weighted 19.6%, so this was not two names carrying a group. The trigger was a rotation out of AI hardware into beaten-down software after Anthropic's $65bn run rate landed below expectations and a Wall Street Journal report tallied some $3tn of off-balance-sheet AI commitments at nine large tech firms. Salesforce added a JPMorgan upgrade to a $250 target on 19 August. These stocks were down 44% to 49% year to date before it started; Salesforce's 50-day average crossed above its 200-day on 19 August, its first bullish reading of the year.

The setup

Where it stands — Salesforce's agent revenue is now disclosed in dollars and growing, while its multiple against gross profit is below February's.

Would confirm — Q2 current remaining performance obligation growth holding at or above 14%, with gross margin near 77%.

Would invalidate — Agentforce ARR growth decelerating sharply, or subscription growth slipping back below 10%.

Watch next — Second-quarter results after the close on Wednesday 26 August; the $25bn repurchase settles in the third quarter.

Valuation — 5.48x trailing gross profit per share against 5.68x in February; 14.8x forward earnings versus 24.1x trailing.

On 10 August, Everpure — the Mountain View maker of all-flash storage arrays that traded as Pure Storage until a February rebrand — said a second top-five hyperscaler had chosen its DirectFlash drives under a new design win and supply agreement. It named no customer and disclosed no financial terms, and said the deal is expected to contribute revenue only from fiscal 2028. Susquehanna and Morgan Stanley upgraded the stock the same day, at targets of $120 and $108. Within four sessions the shares had passed both.

What a shipped hyperscaler deal looks like is already on the record. Everpure's first, a March agreement to become Meta Platforms' primary storage provider, began recognizing revenue at roughly $30m in a quarter at gross margins above 90%. Management has said hyperscale product revenue was minimal last quarter as planned, that the significant majority arrives in the second half of fiscal 2027, and that the category should carry 75-85% gross margins. The August move discounted a logo, not a shipment.

The business under the logo

Everpure's operating numbers are the strongest in enterprise storage. Fiscal first-quarter revenue reached $1.053bn, up 35% year over year, with product revenue up 55%. Annual recurring revenue from its Evergreen//One consumption subscriptions passed $2bn, and remaining performance obligations — contracted work not yet billed — grew 41% to $3.8bn.

The costs are equally visible. GAAP operating margin was 1.89%, on $19.9m of operating income against more than a billion dollars of revenue. Product gross margin of 65.5% was up year over year but down 180 basis points sequentially. And the diluted share count rose 5.2% in twelve months, to 343.5m, as stock-based compensation quietly claims part of whatever the hyperscalers eventually pay.

NetApp is the opposite case

NetApp, the San Jose vendor whose ONTAP software runs the arrays it sells to banks, hospitals and governments, has the record Everpure lacks. Revenue growth accelerated through every quarter of fiscal 2026, from 1.2% to 12.5%. Operating margin widened from 19.8% to 27.3%. All-flash array revenue hit a record $1.2bn in the fourth quarter and $4.2bn for the year. Free cash flow rose 40% to $1.869bn, and the share count has fallen 6.1% in eight quarters.

What NetApp will not do is put a dollar figure on artificial intelligence. It reports roughly 500 AI and data-preparation wins in the quarter and more than 1,100 for the year — a count, never a revenue line.

The flash bill is arriving

Both companies buy the same inflating input. Contract prices for NAND flash memory are forecast to rise a further 10-15% quarter over quarter as suppliers reallocate capacity to enterprise solid-state drives for Nvidia's next server platform. NetApp has already conceded the arithmetic, guiding fiscal 2027 gross margin to 68.5-69.5% from 70.74% delivered, naming NAND and DRAM, with product margin bottoming in the fiscal fourth quarter.

That blended margin is not earned on the box. NetApp's fourth-quarter product gross margin was 56.1%, against 93% on support contracts and 85.7% on public-cloud services. The software attach carries the company; the hardware is increasingly a pass-through of someone else's price increase.

What the price already assumes

NetApp now costs 7.70x its trailing gross profit, against roughly 5.9x in late May and 4.4x in February. Trailing gross profit grew 5.1% across that stretch. Put differently, about 94% of the increase in market value is re-rating rather than earned profit. Everpure's expansion has been milder — near 10.1x in February to about 13.1x today — against trailing gross profit that grew 13.7%, nearly three times NetApp's rate.

NetApp's shares have held an uptrend since 21 May, the 50-day average above the 200-day for 92 straight days, and the name appears nowhere in the market's 90-day leaderboards. But 22.4 percentage points of its 38% three-month gain came in a single session, 29 May, the day after earnings, on 15.9m shares against a 3m norm. Strip that day and the advance is about 12.8%. Everpure, by contrast, has whipsawed all year and has traded without trend since late July. Both peaked on 14 August and both fell exactly 7.2% into 21 August, after a Wall Street Journal analysis of roughly $3trn in off-balance-sheet AI commitments knocked the memory complex down; Dell fell 12.0% and Hewlett Packard Enterprise 11.3% over the same stretch.

Everpure reports on 26 August. NetApp reports on 2 September. One of them has to show the hyperscaler revenue actually landing; the other has to show what NAND does to the box.

The setup

Where it stands — Everpure's price embeds a design win that ships in fiscal 2028; NetApp's embeds a multiple that rose 75% on 5% gross-profit growth. Would confirm — Everpure booking material second-half hyperscaler product revenue at the promised 75-85% gross margin. Would invalidate — NetApp cutting its 68.5-69.5% fiscal 2027 gross-margin guide again on flash costs, or Everpure's product margin falling below 65%. Watch next — Everpure's fiscal Q2 report on 26 August and analyst meeting on 23 September; NetApp's fiscal Q1 on 2 September. Valuation — NetApp at 7.70x trailing and 7.08x forward gross profit, versus 4.4x in February; Everpure near 13.1x trailing, from 10.1x.