DK Street Journal

The line that grew was not the famous one

Dover, a diversified industrial group that sells pumps, connectors, refrigeration systems, marking equipment and fueling gear, told investors in July that its brazed-plate heat exchangers had just delivered their best quarter in the company's history. Those units, sold under the SWEP brand, sit between the water loop that touches an AI server's chips and the loop that dumps the heat outside. Demand is outrunning supply, so Dover is doubling capacity on that product line over the next twelve months, phased through the back half of 2026 and into 2027. The segment housing it, Climate & Sustainability Technologies, grew 8% organically.

The business investors actually associate with Dover and liquid cooling did nothing of the sort. Colder Products Company, whose UQD and Everis quick-disconnect couplings are the de facto standard for hyperscale direct-to-chip cooling, sits inside Pumps & Process Solutions — a segment that was essentially flat organically last quarter, with AI and energy-component strength offset by a weak polymer-processing comparison, though margin there hit a record 35%.

The arithmetic explains why. The entire global market for universal quick-disconnect couplings in liquid cooling was about $95m in 2024, on a path to roughly $1.9bn by 2030. Against Dover's approximately $8.7bn of 2026 consensus revenue, even total dominance of today's coupling market rounds to nothing. The engineering is genuine — CPC's new Everis DC full-flow connector cuts pressure drop by up to 90% versus a valved part of the same size — but the dollars are not there yet, and Dover discloses no thermal-connector revenue line at all.

The facility side discloses its number

SPX Technologies, which builds Marley evaporative cooling towers, dry coolers and heat exchangers for the building-side water system that feeds those chip-level loops, is the one company here willing to put a figure on it. It lifted its 2026 data-center revenue forecast to $430m from $300m, against $200m last year, and now sees $1.1bn of data-center capacity once lines are fully ramped, up from $750m. Backlog in heating and cooling reached $919m, up 59% organically, on a book-to-bill of 1.4 times, with hyperscaler agreements providing three-to-five-year visibility. Revenue grew 22.9% to $679m and operating income 34.6%.

Segment margin fell 260 basis points, and management attributed all of it to identified, temporary items: roughly 80 basis points of startup cost on new capacity, 80 of net tariff and 50 of inflation. That capacity is coming mostly from throughput work at existing Olathe and Springfield plants rather than construction — routine capital spending runs at 1.5–2% of revenue.

Both grew; both got cheaper

Dover's price per dollar of trailing gross profit has slid from 10.0x in February to 8.6x in May to 8.1x now, below the 8.0x of a year ago, while trailing gross profit rose 8.5% to $3.33bn. SPX's fell from 13.4x in February to 10.3x, roughly its level twelve months back, as its trailing gross profit grew 18.5% to $996m. In both cases the numerator did the work. Dover trades at 24.1x trailing and 18.8x forward earnings, SPX at 36.4x and 24.5x.

The integrator keeps the premium

Vertiv, which assembles the power and thermal chain — coolant distribution units, racks, Liebert cooling — guides to $14bn of 2026 revenue, up 37%, and expanded gross margin to 37.7% from 32.2%. It fetches 23.3x trailing gross profit, near three times Dover's, and it is buying the ends of the chain the component makers occupy: Strategic Thermal Labs for cold-plate design, with ThermoKey pending for heat rejection. Flex, a contract manufacturer earning a 9.4% gross margin, bought JetCool for the same reason; its cloud and power segment grew 35% and is guided to 65–75% for the year.

That matters for how the past year gets read. Grouped together, these four cooling-adjacent names returned about 45% over twelve months — but two-thirds of it came from Flex alone, up 120%. Dover added 13% and SPX 10%. Parker-Hannifin, the motion-control group whose quick-coupling division is CPC's closest rival, rose 37% on something else entirely: record fiscal fourth-quarter sales of $5.755bn, aerospace organic sales up 13.4% and a record $8.5bn aerospace backlog. Parker is the most expensive of the group at 15.6x gross profit, up from 12.9x a year ago — moving the opposite way to the two cooling names, and for aircraft reasons.

What actually moved them

Dover fell 9.7% on 23 July despite in-line results and a raised earnings outlook, then recovered the whole drop by 5 August — before grinding down another 4.9% over eleven sessions, on no company news, to $201.17. The last leg was not about cooling: the 30-year Treasury yield touched a 19-year high of 5.33% as a tally of $3trn in off-balance-sheet AI commitments circulated, and everything financed by debt repriced. Over four sessions to 21 August, Vertiv fell 13.3% and Flex 16.3%, against 5.8% at SPX and 2.2% at Dover.

The setup

Where it stands — Both cooling suppliers grew gross profit and shrank on the multiple; Dover's disclosed AI exposure is heat exchangers, not couplings.

Would confirm — SPX's second-half heating-and-cooling margin recovering the 260 basis points management called temporary.

Would invalidate — Dover's data-center-linked revenue share falling back below the 25% of 2026 sales it now cites.

Watch next — Dover's third-quarter results in late October, and whether SWEP capacity additions land on schedule.

Valuation — Dover 8.1x trailing gross profit versus 10.0x in February; SPX 10.3x versus 13.4x; Vertiv 23.3x.

A record order book and a falling share price

Vishay Intertechnology, the Malvern, Pennsylvania maker of the small power transistors, diodes and passive parts that sit on nearly every industrial circuit board, spent the summer being told by customers to build more capacity. Second-quarter orders ran at 1.32 times shipments, and backlog climbed to $1.9bn — about six months of coverage. Distributors, who had been holding 20 weeks of stock, cut it to 18 while selling through 20.5% more than a year earlier, the shape of real consumption rather than a restocking scramble. Adjusted revenue was $919m, up 20.5% from a year ago, and management attributed 30.1% industrial growth to smart-grid, AI power and high-voltage direct-current projects, with some customers ordering beyond 52 weeks out to lock in production slots.

The shares fell 16.9% over the following month.

The content arrives at the wrong margin

Here is the catch that the order book hides. The product doing the AI work — the MOSFET, the switch that converts rack voltage down to what a processor can use — is Vishay's least profitable line. MOSFETs generated $188.9m of revenue at a 13.8% gross margin last quarter, against $187.0m of diodes at 24.4%, and a company-wide 22.6%. Compare onsemi, the vertically integrated silicon-carbide supplier that owns its own crystal growth: its second-quarter gross margin reached 39.3%, guided to 40-42%, with data-center silicon-carbide revenue up nearly 60%. Same megawatt, roughly three times the margin. Volume growth at 13.8% buys far less earnings than the headline order book implies.

Vishay is also spending into it. Capital expenditure is guided to $400-440m for 2026, roughly half of it a German 12-inch fab that will not make non-automotive parts until mid-2027, and free cash flow stays negative this year. To fund it the company sold 15m shares at $50 on 29 June, raising about $830m net and adding roughly an eighth to the share count. At $31.04 the stock sits well below that price. Price per dollar of trailing gross profit has gone from 4.3x in February to 8.0x in late May to 6.3x now — a de-rating of a fifth while trailing gross profit itself compounded 17%, from about $595m to $695m. On forward earnings the stock is 36.9x against 120.7x trailing, but that gap is recovery, not growth: 2025 net income was -$9.0m.

Wolfspeed's fall needs no explaining

Wolfspeed, the Durham, North Carolina silicon-carbide maker that emerged from a prepackaged Chapter 11 last September having cut roughly $6.7bn of debt to about $2bn, is the opposite case. Fourth-quarter revenue of $149.6m missed the $223.6m consensus by a third, the fourth consecutive quarterly decline and 24.1% below a year earlier. Non-GAAP gross margin was still around -20%; adjusted EBITDA was -$62.4m against $1.1bn of cash. Consensus has fiscal 2027 revenue falling a further 15% to $642.8m.

The AI piece is real but unpriced. Data-center revenue more than doubled in fiscal 2026 and grew about 20% sequentially, with a design win at Taiwan's LITEON — disclosed only as growth rates, never dollars, inside roughly $106m of total power revenue. A named customer without a number is a slope, not a line item. The stock has fallen 63.4% from $73.50 on 26 May, and price-to-sales tracks the story rather than the business: 1.4x in February, 4.3x at the May peak of the 800-volt trade, 2.1x today, on a revenue base that shrank throughout.

Whose socket it is

NVIDIA's published 800-volt direct-current ecosystem names Infineon, onsemi, Navitas, STMicroelectronics, Texas Instruments, Analog Devices, Renesas, ROHM and Monolithic Power as its power-semiconductor partners. Neither Wolfspeed nor Vishay appears, and the infrastructure transition starts in 2027. Navitas, the gallium-nitride and silicon-carbide pure play, has $10.5m of quarterly revenue at a negative gross margin and a $3.3bn market value — larger than Wolfspeed's $1.39bn on a fourteenth of the sales. That is what the market is paying for optionality it believes in.

AXT, the Fremont maker of indium-phosphide and gallium-arsenide substrates for optical data-center links, is the one name here whose numbers already arrived: record revenue of $47.6m, up 164% year over year, gross margin of 45.0% against 8.2% a year ago, third quarter guided near $66m, with prepaid long-term supply deals from Coherent and Casella. It is priced accordingly at 50.6x forward gross profit — about eight times Vishay's multiple on the same measure — and remains hostage to Chinese export permits that capped its shipments in late 2025.

The last five sessions were a bond story

None of this explains the past week, in which all seven US-listed discrete and power semiconductor names fell between 8.9% and 15.4% together, alongside onsemi at -10.1%, Texas Instruments -5.7% and Analog Devices -4.0%. The 30-year Treasury yield hit a 19-year high of 5.33% on 18 August and more than $1trn came off global chip market values on fears that AI infrastructure spending is peaking. Estimates did not move; the multiple did.

The setup

Where it stands — Vishay's order book is inflecting on AI and grid demand while its shares de-rate; Wolfspeed's decline matches a still-shrinking, loss-making business. Would confirm — Vishay's third-quarter gross margin printing at the guided 24.0%, with MOSFET margin above 15%. Would invalidate — Book-to-bill falling below 1.0 or distributor weeks-on-hand rising back above 20. Watch next — Vishay's third-quarter results, guided to $945-975m of revenue; Wolfspeed's first quarter, guided to $140-160m. Valuation — Vishay: 6.3x trailing gross profit versus 8.0x in late May and 4.3x in February; 36.9x forward earnings against 120.7x trailing.

Three lenders to the artificial-intelligence build-out reported June-quarter results within four weeks of each other, and all three disclosed the same thing: the money they lend is earning a wider margin than it did a year ago. That is the opposite of what a capital glut looks like, and it happened while the companies borrowing that money were being marked down hard.

What the lenders disclosed

Apollo Global Management is a credit-first investment manager whose retirement arm, Athene, sells annuities and must find long-dated assets to back them. On 4 August it reported $74bn of new originations in the quarter, a record, and $150bn in the first half. Three-quarters carried an average rating of BBB+ and was priced at 280 basis points of excess spread over Treasuries; the sub-investment-grade remainder came at 440 basis points. The price matters more than the volume. Athene's net investment spread — what it earns on assets less what it owes policyholders — widened to 114 basis points from 97 the prior quarter, which management attributed to higher yields on Apollo's own originated paper. Fee-related earnings rose 25% to $785m. The $35bn financing package for Broadcom's AI compute, the largest private-credit deal ever announced, is not in the $74bn; it draws from late 2026 into 2027.

HA Sustainable Infrastructure Capital is the clean control. A $5.1bn balance-sheet lender to energy and climate projects with 170 employees, it funds itself in the same bond market it lends against, so a squeeze would show up immediately. It did not. Portfolio yield rose to 9.2% from 8.2% while its weighted-average interest cost moved only to 6.2% from 5.8%, and its own new-issue credit spread compressed below 200 basis points against 340 in 2021. New investments are being underwritten above 11%. The company raised guidance on 24% year-to-date adjusted earnings growth and a return on equity above 15%.

Blackstone, the largest alternative manager at $1.35tn of assets, is the only firm sitting on both sides — it owns data-center operator QTS as equity while originating credit against the sector. It carried that platform at $185bn in the second quarter, up from $130bn at the start of the year, and infrastructure assets reached $90bn, up 40%. Revenue grew 28.5% and operating margin reached 61.9% against 51.6% a year earlier. QTS is also a borrower: it refinanced ten data centers with a record $3.46bn commercial-mortgage-backed loan.

The borrowers are paying for it

The mechanism is visible in credit markets. Investment-grade data-center secured bonds have widened 40 basis points since mid-June and their high-yield counterparts 120, on Morgan Stanley's numbers, as supply floods in — data-center asset-backed issuance has gone from $4bn in 2020 to $61bn so far this year. CoreWeave, which rents AI computing capacity, now pays $640m of net interest a quarter against $267m a year ago. Over three months Applied Digital, which builds and leases AI data centers, fell 42.7% and CoreWeave 18.2%. None of this is a rates round-trip: the 30-year Treasury yield hit a 19-year high on 18 August, with markets now leaning toward increases rather than the cuts priced in January.

The month, though, was not about lending

On 10 August Nvidia signed memorandums of understanding with Apollo, Blackstone, Brookfield, BlackRock, Goldman Sachs and KKR to mobilize more than $500bn of third-party capital for compute infrastructure. No partner has disclosed a dollar commitment and the agreements await final documentation. The next session, Apollo rose 7.8%, KKR 7.2% and Blackstone 5.9%.

But the best 30-day performer in the group is TPG, at 26.7% — not an Nvidia partner, with no disclosed data-center credit book, and fee earnings up 43% to $315m all the same. Blue Owl added 27.0%, Ares 19.7%, Apollo 12.6%. Meanwhile the listed vehicles that actually hold these loans lagged badly: Ares Capital gained 5.8% and Blackstone Secured Lending 6.7%. The fee franchises re-rated; the loan books did not. And over twelve months the same managers remain deep underwater — Blue Owl down 37.3%, KKR 20.9%, Blackstone 11.5% — against a market up 20.3%. August was a recovery from a de-rating, not a premium for AI lending.

Where the prices sit

Apollo is the dislocation. Twelve-month return of exactly zero, revenue up 64.5%, and 28.8x trailing earnings collapsing to 15.2x forward against consensus 2027 earnings growth of 22.4%. The caution is that its quarterly accounting earnings are mark-driven — the March quarter carried a $1.9bn loss — and the overhang from March's redemption panic at its non-traded business development company has eased rather than cleared, with management flagging 36 competitors now chasing retail annuity money.

Blackstone tells the opposite story: 31.9x trailing, versus 30.8x in early May, while its market value grew from $146.5bn to $173.7bn. The advance was earnings, not enthusiasm. HASI's 61x trailing multiple is an accounting artifact; on 13.3x forward and 2.0x book it is the cheapest of the three, but a 48.1% twelve-month gain means most of that repricing has been collected.

The setup

Where it stands — Lending margins widened at all three firms, but August's share gains tracked the whole alternative-manager group regardless of data-center exposure. Would confirm — Athene's net investment spread holding above 110 basis points in the September quarter, with originations again above $70bn. Would invalidate — Excess spread on Apollo's investment-grade originations falling back toward 250 basis points as competing capital arrives. Watch next — Third-quarter results in late October, plus the first named project under Nvidia's financing memorandums. Valuation — Apollo 28.8x trailing and 15.2x forward; Blackstone 31.9x and 24.1x; HASI 13.3x forward and 2.0x book.

On 3 August, Texas Governor Greg Abbott directed the state's Public Utility Commission and its grid operator to verify every data center advancing through the interconnection process before another one is approved. The queue at ERCOT — the Electric Reliability Council of Texas, which runs the state's grid — holds roughly 474 gigawatts of requests, about 90% of them data centers, and the audit covers 250 to 300 projects. ERCOT then missed its 7 August deadline for "Batch Zero," the first tranche of large loads to be studied, and asked for extensions. The review is expected to run for months.

That lands on two companies more directly than on anyone building turbines or reactors.

The wires business gets paid on capital, not on power

CenterPoint Energy is a Houston holding company whose electric arm owns the poles, wires and 239 substations that deliver power in Houston but sells no generation; it also distributes natural gas across eight states. Sempra, of San Diego, owns two California utilities and, in Texas, Oncor — a pure transmission-and-distribution business serving 3.8m customers across 140,000 miles of line.

Neither profits from a megawatt sold. They profit from capital spent: a commission sets an allowed return on the substations and lines they build, and that return accrues whether the load behind it is an aluminum smelter or a training cluster. CenterPoint Houston Electric is authorized 9.65% on a 43.25% equity layer; Oncor won 9.75% in an April settlement worth about $560m of extra revenue. Roughly 85% of CenterPoint's spending is recovered through automatic trackers rather than full rate cases, which is how it keeps the lag between spending and earning short.

So the only thing that matters is whether the load is real. Texas Senate Bill 6 was written to answer that: loads of 75 megawatts or more must demonstrate site control and post security starting at $50,000 per megawatt, doubling at the next stage. CenterPoint says it already holds $900m of customer cash and security against its pipeline, expects system upgrades to cost under $60m per gigawatt, and earns about $6m per gigawatt each month in demand charges. That is a genuinely contracted business — for the projects that clear.

The results say one thing, the shares another

CenterPoint's second-quarter revenue grew 10.7% to $2.15bn and operating income 28.1% to $534m, after a first quarter that grew 1.9%. On 28 July it raised its ten-year capital plan to $66.7bn from $65.5bn and reiterated 2026 earnings guidance of $1.89 to $1.91 a share. Two-thirds of the increase is driven by large-load customers, and Houston Electric's rate base is guided to grow about 18% a year.

Sempra's quarter was stranger: revenue flat at $3.00bn, operating income up 42.5%, net income up 68.5%. The uplift arrives below the revenue line because it is a regulated return, not a volume. Its $65bn five-year plan is 95% regulated, and guidance was reaffirmed. Oncor received 44 gigawatts of Batch Zero requests, a 140% lift to its peak — none of it in the capital plan.

Both stocks fell through August, and gradually: CenterPoint's worst session was -2.65%, Sempra's -2.68%. There was no gap. Every large regulated utility fell together, which points at the discount rate. The 30-year Treasury yield touched 5.323% on 18 August, a 19-year high.

That matters more here than for most sectors, because rate-base growth is cash burn: CenterPoint's trailing free cash flow yield is -10.5%, Sempra's -10.7%. The gap is funded externally. Oncor's authorized cost of debt is 4.94%, while Ameren Missouri priced 30-year first mortgage bonds at 5.55% in February — before the long bond rose further. New money already costs more than the debt embedded in rates.

Which fall bought something

On unchanged consensus earnings, Sempra's forward multiple has gone from 17.9x in May to 16.5x, against a utility sector median near 18.8x, at 1.68x book — the cheapest per dollar of equity of the three. It is also selling 45% of its infrastructure arm for about $10bn, removing nearly $9bn of debt. CenterPoint's forward multiple fell from 22.2x to 20.7x, but it was 19.9x a year ago: the decline has unwound three months of expansion, not created a discount.

NextEra Energy, which pairs Florida Power & Light with the largest US renewables developer, is the useful contrast. Its Florida franchise is accelerating — large-load expectations raised from 6 to 8 gigawatts by 2032, a 35.1-gigawatt competitive backlog — yet it trades at 3.09x book and will issue roughly 716m shares to Dominion holders, with a shareholder vote on 3 September. Its de-rating is about issuance and funding cost, not demand. So is the rest of the group's.

The setup

Where it stands — Earnings and capital plans are rising at both Texas wires utilities while their shares fall on financing cost and a state-ordered queue freeze.

Would confirm — CenterPoint's September distribution cost recovery filing implemented on schedule, and Batch Zero studies resuming with its 14 gigawatts intact.

Would invalidate — A cut to CenterPoint's $1.89-$1.91 guidance, or large-load projects withdrawing rather than posting the $50,000-per-megawatt security.

Watch next — The ERCOT audit's completion, expected within months; NextEra's shareholder vote on the Dominion merger on 3 September.

Valuation — Sempra 16.5x forward and 24.3x trailing, versus 17.9x forward in May; CenterPoint 20.7x forward against 19.9x a year ago.

The price everyone quotes is not the price anyone earns

Cameco, the Saskatoon miner that also owns 49% of the reactor builder Westinghouse, sold its uranium last quarter at an average of US$67.79 a pound. The published spot price at the end of July was $86.36, and the price utilities pay under new long-term contracts has sat at $94 since June, the highest in 18 years. Cameco's realized price was up 18% from a year earlier and still 28% below that contract number, because deliveries come off a book signed when uranium was cheap.

That gap is the story of the whole nuclear fuel chain right now. The headline prices are at decade highs; the profit-and-loss statements are not. Over the past month shares in the fuel names rose together — Cameco 11.5%, the enricher Centrus 14.3%, the developers NexGen 21.5% and Uranium Energy Corp 26.8% — while the metal did nothing at all. Strip each name's two best sessions, clustered on a handful of shared news dates in late July and early August, and the month's gain becomes a loss. This was an event, not a re-rating.

What actually re-contracted

The events were real. Centrus, a Bethesda, Maryland company that sells separative work units — the measure of enrichment effort, or SWU — and no mined ore, disclosed a backlog of $4.5bn extending through 2040, against $2.3bn of commercial backlog three months earlier. Of that, $2.4bn of enrichment work is now under definitive agreements rather than contingent ones. It signed a firm high-assay low-enriched uranium off-take with the reactor developer X-energy and a letter of intent with Oklo to fuel up to five units from 2029, both carrying prepayments. Separately, the Department of Energy conditionally committed $17.5bn of loans on 23 June for long-lead equipment for up to ten Westinghouse AP1000 reactors, and Cameco disclosed that Westinghouse had confidentially filed for an initial public offering.

The mechanism underneath is a deadline. US reactor operators bought about 3.28 million SWU of Russian enrichment in 2025, nearly 26% of their purchases, and the waiver authority that permits it expires no later than January 2028. Urenco is adding roughly 700,000 SWU a year in New Mexico by early 2027, only a partial replacement. That shortfall is why enrichment quotes are elevated — SWU prices rose about 11% last year to $108.70 a unit.

Two income statements going the wrong way

Centrus is not capturing it yet. Enrichment-segment revenue rose 22% to $153.4m, but volumes fell 23% and realized SWU pricing rose 3% while unit costs rose 13%. Gross margin compressed to 28.3% from 34.9%, and operating income fell 69% to $10.4m. Consensus has 2026 earnings per share at $2.50, down 44% from last year. Management itself says utilities remain in wait-and-see mode until it demonstrates centrifuge installation. The balance sheet is improving — $1.9bn of cash, commercial Piketon production pulled forward to 2029 — while the earnings line deteriorates.

Cameco's quarter was worse on the surface and better underneath. Revenue of $814.1m fell 7.2%, gross margin dropped to 21.1% from 29.3%, and net income fell 92% to $25.2m, largely because its share of Westinghouse EBITDA halved to $163m against a prior year containing a one-off payment on the Czech Dukovany project. Production guidance of 19.5–21.5 million pounds held through weather disruptions at Key Lake and McArthur River and a Cigar Lake suspension after quarter-end. Its fuel-services realized price rose 13% to C$41.67 per kilogram of uranium. New contracts now carry floors in the high-$70s and ceilings near $160.

The valuation is the uncomfortable part. Cameco has fallen 17.9% in six months and become more expensive doing it: price per dollar of trailing gross profit has gone from roughly 33x a year ago to about 45x, because gross profit fell faster than the shares. Forward earnings multiple is 64.6x, up from roughly 56x in May, resting on a projected jump to $2.64 of 2027 earnings per share from $1.54 this year.

The controls

NexGen, a pre-revenue developer whose Rook I project sits in Saskatchewan's Athabasca Basin, trades at 5.27x book on a $7.02bn market value with first pounds around 2030 and financing for a C$2.2bn build not yet closed. It has 11.3 million pounds contracted, all spot-indexed, and deliberately leaves 96% of reserves uncommitted — the mirror image of Cameco's floors-and-ceilings book. Uranium Energy Corp, a Texas-based in-situ recovery producer, is guided to fiscal 2026 revenue of $26.9m, down 59%, with an $85.5m loss. Their outsized monthly gains are leverage to a metal that did not move.

One name went the other way: BWX Technologies, which builds naval reactors, fell 10.2% over the same month. Whatever lifted the fuel names, it was not indiscriminate enthusiasm for nuclear.

The setup

Where it stands — Backlogs and policy are re-contracting the Western fuel chain; realized prices and margins at both revenue-generating companies are still falling. Would confirm — Cameco's realized uranium price closing on spot in coming quarters, or Centrus posting SWU pricing gains above 10%. Would invalidate — Centrus funding Piketon through further equity sales without new priced volume, or Cameco cutting its 19.5–21.5m lb guidance. Watch next — Westinghouse's public IPO filing, and the January 2028 expiry of Russian enrichment import waivers. Valuation — Cameco 168.1x trailing and 64.6x forward earnings versus about 56x forward in May; Centrus 71.6x trailing against 72.0x forward.

Adobe's management did something in June that the disruption script did not anticipate. It deferred planned Creative Cloud price increases and leaned harder into a free tier that now carries more than 90m monthly active users — deliberately trading near-term subscription revenue for reach. Figma, reporting on 5 August, went further: its newest AI features, including an in-product agent, Code Layers, Motion, Shaders and the prompt-to-design tool Make, carry no price at all and are excluded from guidance entirely.

Both companies are absorbing the cost of generative AI rather than billing it. That is the opposite of what the bear case predicted, which was that AI would erode the per-seat subscription while the vendor had nothing to replace it with. What is happening instead is that the seat is holding and the AI is free.

Adobe: the price cut it chose

Adobe, which sells Photoshop, Illustrator and Premiere to professional creators and Acrobat to nearly everyone else, grew revenue 12.7% to $6.62bn in its May quarter — a record, and the fourth consecutive quarter of acceleration. Total annual recurring revenue reached $27.1bn, up 12.5%. Roughly $480m of that came from Semrush, the search-optimization vendor Adobe bought for about $1.9bn in April.

Strip the acquisition and the company's own ARR growth target falls to about 8.3%. Management attributed roughly half the reduction to deferring price increases and half to freemium expansion. Against that, AI-first ARR tripled year on year past $500m, with Firefly approaching $300m.

Generative compute is invisible in Adobe's cost of revenue: gross margin sat at 89.2%, versus 89.1% a year earlier. The strain is one line down. Operating income grew 6.1% against 12.7% revenue, and operating margin fell from 35.9% to 33.8%. Net income rose 1.2%; earnings per share held up mainly because the share count fell 7.2% in a year. Chief financial officer Dan Durn left four days after the print, and RBC cut its price target citing the exit and the ARR change.

Figma: AI as a cost of goods

Figma, the browser-based design canvas product teams use to draw and hand off interfaces, grew revenue 48.2% to $370.1m — its third straight acceleration. Net dollar retention was 136%, and customers spending over $100,000 a year rose 46%. More than 80% of its larger paid customers consume AI credits weekly, and about two-thirds of expansions come from adding full seats at renewal. The seat is not being deleted; it is being bought.

The bill arrives in cost of revenue. GAAP gross margin fell to 83.7% from 88.8%, gross profit grew 39.6% against 48.2% revenue, and the operating line swung from a $2.1m profit to a $117.3m loss. The number that has not been priced: gross margin rose 4.3 points sequentially, from 79.4%, while the new AI products remain free. Non-GAAP gross margin was 85%, up 2.5 points, with a 14% free-cash-flow margin and $1.7bn of cash. Management ascribes the volatility to unmonetized launches and routes inference across multiple model providers and its own first-party models to control it. Third-quarter guidance implies deceleration to about 36%.

The competitive field, meanwhile, widened: Google upgraded its free design canvas Stitch in March and Anthropic launched Claude Design in April, turning a near-monopoly into a crowded field. Figma also disclosed the departure of its chief product officer and chief marketing officer, with a chief technology officer search underway.

The group is not one business

Creative software as a category rose about a third over the past month, but Unity supplied 44% of that average by itself — and Unity is no longer a design-seat business. Its advertising arm grew 63% year on year and its Vector ad platform passed a $1bn run rate two quarters early; that is a mobile ad network, not a creative tool. Autodesk, which sells AutoCAD and Revit to engineering and construction firms, grew revenue 18.4% and trades at 35x trailing earnings — it was never de-rated in the first place.

Neither Adobe nor Autodesk reported inside the window. Adobe's last results were 11 June, and its 25% advance since late July is multiple repair on a rotation out of expensive semiconductor names into beaten-down application software. Its 50-day average crossed above its 200-day on 18 August, the first non-bearish reading in roughly a year. Excluding the concentrated late-July rotation sessions, Adobe is up only about 9.6% across 19 trading days, eight of them down — a grind, not a gap. Figma's month was stranger still: it fell 14.9% the session after its beat-and-raise, then rebuilt the gain over the following ten sessions.

On price per dollar of gross profit — the fair comparison, since the two carry margins five points apart — Adobe trades at 4.89x trailing and 4.65x forward, against 4.40x in late July and a five-year median trailing price/earnings ratio near 42. It sits at 11.3x forward earnings with a 9.7% free-cash-flow yield. Figma is at 13.26x trailing gross profit, 2.7 times Adobe's, though below its own 13.91x reading of 29 July: gross profit grew faster than the share price. Its final lockup expires on 31 August, freeing venture holders of more than half the company.

The setup

Where it stands — Both companies are absorbing AI cost while demand accelerates; neither has yet turned generative usage into a priced billing unit. Would confirm — Figma's gross margin extending its sequential recovery above 85% GAAP, and Adobe's September quarter showing net-new Digital Media ARR stabilizing. Would invalidate — Adobe's organic ARR growth falling below 8% or Figma's third-quarter growth undershooting the 36% implied guide. Watch next — Figma's final lockup expiry on 31 August; Adobe's fiscal third-quarter results in September. Valuation — Adobe 15.78x trailing and 11.32x forward earnings; Figma 13.26x trailing gross profit versus Adobe's 4.89x.

Datadog sells cloud monitoring software that watches customers' servers, applications and log files, and charges by the volume of data it swallows rather than by the seat. In the June quarter its revenue reached $1.12bn, up 35.6% — the fourth consecutive quarter in which its growth rate rose. Management told analysts on 6 August that the part of the business with nothing to do with artificial intelligence was growing in the high-20s, against 18% a year earlier, a fifth straight quarter of acceleration from that base.

Then it disclosed that its largest customer, described only as a significant AI company using 17 Datadog products, had renewed its contract at reduced usage starting in the September quarter. Guidance for that quarter came in at $1.135-1.145bn, growth of 28-29%. The shares fell 20.4% in a single session.

What is left is a price

That selloff did something unusual: it left Datadog costing exactly what it cost three months earlier. At $233.40 the shares are 26.34x trailing gross profit; on 21 May, at $218.04, they were roughly 26.45x. Gross profit grew as fast as the share price, so the advance since May contained no re-rating whatsoever. Stretch the window and the picture reverses — in late February the same measure was about 15x, so six months have added roughly 75% to the multiple. All of it landed before May.

Every other name in this corner of software did the opposite. Cloudflare, which runs a global edge network selling security, content delivery and serverless computing priced by usage, grew revenue 35.9% to $696.1m — statistically the same rate as Datadog. But its gross profit grew only 30.2%. Cost of revenue rose 53%, and reported gross margin fell to 71.8% from 74.9%, well under the high-seventies it held from 2021 through 2024. Management attributed the compression to a mix shift between paid and free network traffic, not to hardware, and noted the first sequential margin gain in eight quarters.

The shares are 55.30x trailing gross profit, against 44.2x in May and 38.9x in February — the group's most expensive claim, placed on its slower-growing line.

The demand underneath Cloudflare is not in doubt

Customers paying more than $100,000 a year reached 4,698, up 27%, with a record 986 net additions over twelve months; dollar-based net retention rose to 120%, six points higher than a year ago; contracted backlog reached $2.73bn, up 38%, growing faster than revenue. More than half of traffic across the network was non-human for the first time, and the developer count hit 7.4m after adding nearly 2m in one quarter, versus 1.5m in all of 2025.

What is absent is a revenue line attached to any of it. And chief executive Matthew Prince explicitly declined the inference-hardware land grab, arguing that selling commodity compute is unattractive. Network capital spending was 7% of revenue in the quarter, guided to 14-15% for the year — meaning the second half roughly doubles in intensity. Datadog's capital spending plus capitalized software runs 4-5% of revenue.

Two controls, pointing opposite ways

Palo Alto Networks, which sells firewalls and security subscriptions to large enterprises, has published no financial statements since early June; fourth-quarter results land 1 September. Its price per dollar of trailing gross profit went from about 27.0x in May to 37.50x — pure multiple, against an unchanged set of numbers. Those numbers show the cost of the $25bn CyberArk acquisition closed in February: April-quarter revenue up 31.1%, gross profit up only 21.5%, gross margin 67.6% against 72.9%, and a GAAP operating loss of $183m.

Dynatrace, Datadog's closest rival in monitoring software, grew just 16.2% and trades at 8.48x trailing gross profit with a 3.93% free-cash-flow yield. It also supplied the cleanest evidence that AI telemetry is a broad volume driver rather than one cohort's spending: more than 1,000 customers now monitor AI workloads in production, up from 850, and that group's consumption is growing 1.5 times as fast as everyone else's. CrowdStrike, meanwhile, is not in this trade — its 50-day average slipped below its 200-day in July and it has been in a clear downtrend since 6 August, despite 25.6% revenue growth and a 57.80x multiple.

The week the order became visible

All five peaked on 13-14 August. Into 21 August, CrowdStrike fell 12.3%, Cloudflare 10.1%, Palo Alto 8.8%, Datadog 8.6% — and Dynatrace, the cheapest, 0.5%. The drawdown sorted almost perfectly by price per dollar of gross profit, in the week the 30-year Treasury yield reached its highest level in nearly two decades. Higher long yields cut the present value of distant profits, and these are the most distant profits in software: Cloudflare's forward price/earnings ratio is 225x, Datadog's 92x.

The setup

Where it stands — Datadog's multiple has been flat since May while gross profit grew 33%; Cloudflare's expanded 25% on gross profit growing 30%. Would confirm — Datadog's September quarter landing at or above the 28-29% guide with non-AI growth still accelerating. Would invalidate — Cloudflare's gross margin falling below 71% as second-half capital spending doubles, with no disclosed AI revenue line. Watch next — Palo Alto reports fiscal fourth-quarter results on 1 September, its first numbers since early June. Valuation — Datadog 26.3x trailing and 23.4x forward gross profit; Cloudflare 55.3x and 48.4x, against 38.9x in February.

The record quarter nobody bought

MYR Group, a contractor founded in 1891 that strings high-voltage transmission line for utilities and wires data centers, airports and hospitals for developers, closed its June quarter with the best numbers it has ever reported. Revenue rose 20.1% to a record $1.08bn. Gross margin widened 170 basis points to 13.2%, net income nearly doubled, and backlog reached a record $3.16bn, a fifth higher than a year ago.

The composition is the story. Of that backlog, $1.89bn sits in the commercial and industrial division and $1.27bn in transmission and distribution. In the quarter itself, grid revenue grew 4% to $524m — nearly two-thirds of it routine work under master service agreements — while commercial and industrial revenue grew 42% to $558m, overtaking the grid division for the first time. The company most cleanly positioned as a play on interconnection is growing because of building wiring.

Booked, not burned

That is not a demand failure; it is a construction calendar. MYR Group's two Xcel Energy transmission awards, worth more than $200m combined, contribute no revenue until the second half of 2027 and then run for roughly eighteen months. Management told investors on its July call that most of the major projects it has won will be constructed between 2028 and 2030. A 500 kV Arizona substation and a 345 kV Texas rebuild sit in that queue.

The delay has a named cause. Under the Federal Energy Regulatory Commission's (FERC) Order 2023, utilities now study proposed grid connections in clusters rather than one at a time, with deadlines and penalties for late studies. Throughput is improving — PJM Interconnection has processed more than 170,000 MW of generation requests since 2023 — but against a national queue near 2,200 GW. Texas has paused new data-center interconnections in the face of an estimated 474 GW of requests, more than five times ERCOT's record peak. American Electric Power raised its five-year capital plan by $6bn to $77.9bn, including $33bn of transmission, and has threatened to leave PJM and SPP over how slowly connections clear. S&P Global puts US utility capital spending near $1.3 trillion for 2026-30. The money is committed. The billable hours are years out.

The de-rating is a multiple, not an estimate

MYR Group shares closed at $318.17 on 21 August, against $438.96 three months earlier. Consensus earnings for this year did not fall over that stretch; the multiple did, from 36x to 26x. Goldman Sachs analyst Neil Mehta made the mechanism explicit on 12 August, cutting his target to $422 from $469 by lowering the target enterprise-value-to-EBITDA multiple to 16.5x "to reflect broader AI and power market multiples." The stock trades at 16.4x trailing EV/EBITDA. Quanta Services, the tier-one electric-power contractor with a $100bn market value, trades at 35.1x — and grew revenue 41% last quarter, so the premium is not unearned, merely large.

The break was collective and fast. Between 17 and 21 August, Sterling Infrastructure fell 14.6%, Comfort Systems 10.9%, Quanta 8.8%, EMCOR 8.7%, Dycom 7.3% and MYR Group 6.4%; MYR Group's 50-day average crossed below its 200-day on 19 August, ending an uptrend that had held since spring. The proximate cause sits upstream: investors have been rotating away from AI infrastructure names whose capacity is increasingly debt-funded, and European Central Bank economists warned on 18 August that a valuation correction is likely.

The consultancies went the other way

AECOM, a fee-for-service design and program-management firm selling to transport agencies, water utilities and defense ministries, rose 4.0% in those same four sessions — after falling 19% in two days around its 11 August results. It took a $337m pre-tax charge on a single construction-management project awarded in 2019 on terms it says it would not accept today. Quarterly revenue fell 14.2% to $3.59bn and the company posted a $76m operating loss. Free cash flow guidance was cut to $300m from $400m, with roughly $500m of cash going out through the first half of 2027 against a $600-650m claims position management concedes needs years and litigation.

Underneath, orders are at a record: backlog up 13% to $27.8bn, with $1.60 booked for every dollar burned in the quarter. Its Department of War pipeline is up 30%, its US water pipeline the same, and president Lara Poloni called data centers among the fastest-growing parts of the business. The shares are down 46.7% over twelve months and sit at 10.4x the $6.22 consensus expects in fiscal 2027.

Tetra Tech, a water and environmental consultancy whose federal work was disrupted by 2025 agency cuts, is the recovery trade. Revenue grew 13.5% to $1.31bn, backlog reached a record $4.5bn and guidance was raised to $1.56-1.59 per share. But gross margin fell to 18.6% from 21.8%, operating income declined 4.3%, and its data-center practice runs at roughly $60m a year — about 1.4% of guided revenue. Whatever is re-rating Tetra Tech, from 17.3x forward earnings in May to 23.5x now, it is not megawatts.

The setup

Where it stands — MYR Group's grid backlog is at a record while grid revenue grows 4%; the earnings from it arrive in 2028-2030.

Would confirm — Transmission and distribution revenue growth accelerating above 10% year on year in the September or December quarter.

Would invalidate — Backlog falling below $3bn, or full-year organic growth guidance cut beneath the 13-15% management set in July.

Watch next — MYR Group's third-quarter results in late October, the first to include Valley Electric and Comet Electric.

Valuation — 29.7x trailing and 25.9x forward earnings, against 36x forward in May and Quanta's 42.9x forward today.

A capped price for scarce power

On 14 July, PJM Interconnection — the grid operator for 13 states from Illinois to Virginia, plus Washington DC — published the auction that sets what power plants are paid simply to be available in the 2028-29 delivery year. It cleared at $325 per megawatt-day, the legislated ceiling and 2.5% below the previous auction, the third consecutive clear at the maximum. PJM still procured 6.8 GW less than its reserve-margin target required.

Absent the price collar, PJM's own model shows the auction would have cleared at $555. That gap is the mechanism the market is now pricing: roughly $230 a megawatt-day of scarcity rent that a tightening grid is generating but not paying out. Demand outran supply, and the clearing price went down.

Vistra, which sells retail electricity and gas to about 4.3m customers across 20 states while operating some 38,700 MW of nuclear, gas, coal, solar and battery generation, cleared 10,924 MW into that auction. NRG, which serves roughly six million retail customers under the Reliant, Direct Energy and Green Mountain brands from a mostly gas fleet concentrated in Texas, holds about 2 GW of upgrade projects in PJM — and told investors less than half of them clear an economic hurdle even at the higher $555 reference, so it is negotiating directly with buyers instead.