DK Street Journal

SanDisk spent two years losing money making NAND flash memory, the chips that store data in phones, laptops and data-center drives. In the June quarter it earned gross profit of $7.58bn on revenue of $8.97bn, a gross margin of 84.6% against 26.2% a year earlier.

The composition is unusually pure. Revenue rose $7.06bn; gross profit rose $7.08bn. Cost of goods sold did not rise with the business — it fell, from $1.403bn to $1.383bn. Volume contributed nothing to the gain. It was all price. SanDisk's own results release attributes the quarter primarily to higher pricing, with data-center revenue doubling sequentially to $2.98bn and next-quarter revenue guided to $10.3–10.8bn.

Who actually paid

Dell Technologies, the largest seller of AI servers by volume at roughly a fifth of the market, grew April-quarter revenue 87.5% to $43.8bn. Gross margin came in at 17.75% against 21.12% a year earlier. Held at the prior year's rate, gross profit would have been about $1.48bn higher. NetApp, which sells flash storage arrays and the ONTAP software that runs them, guided fiscal 2027 gross margin to 68.5–69.5% from 71.3%, naming memory and component costs in its prepared remarks. On consensus revenue near $7.5bn, that is $135–210m for a full year.

SanDisk's single-quarter gain is about 4.8 times Dell's whole shortfall and about 35 times NetApp's annual one. The money is coming from handset makers, PC builders and hyperscalers buying drives direct — not chiefly from these two.

The alibi mostly holds

At Dell, the memory line is doing less work than the headline margin suggests. Infrastructure Solutions Group revenue rose 181% to $29bn and segment operating margin expanded 80 basis points to 10.5%, while management said that excluding AI-server mix the gross margin rate was up year over year, per Futurum's review of the quarter. Total operating income rose 197%. Dell booked $24.4bn of AI orders and ended with a record $51.3bn backlog. Memory at Dell reads as a gate on how much it can ship, not a destroyer of what it sells.

At NetApp the cost has been named but has not arrived. Reported April-quarter gross margin was 70.07%, up from 68.88%, with operating margin at 27.26%. The finance chief told analysts product margin should trough in the July quarter and improve as price increases flow through — recovery by pricing, not a permanent transfer.

Why the collector stopped rising

SanDisk's durable asset is the Yokkaichi and Kitakami joint venture with Kioxia, extended in January to December 2034, the largest NAND manufacturing footprint in the world. But it does not set the price: Samsung led the market with 29% share in the first quarter. SanDisk collects an industry cycle rather than authoring one — and it is pure NAND, with no DRAM or high-bandwidth memory.

That matters, because the two are separating. TrendForce sees NAND contract prices up 10–15% in the third quarter but moderating as consumer demand weakens, and forecasts NAND supply tilting toward balance and surplus in the second half of 2027 while DRAM stays short into 2028. Micron, which has both, is up 34% over three months against SanDisk's 13%.

The shares registered it early. SanDisk's uptrend broke on 29 July after 227 unbroken sessions, one day before that forecast published; it then fell 5.4% and 6.8% in the two sessions after record August results, rallied 14.1% on 13 August, and gave back 9.8% on 18 August. Dell and NetApp have held rising trends since 31 March and 21 May. All four names, including Seagate, fell together on 19 August as the 30-year Treasury yield topped 5.33%, a 19-year high.

What each price now assumes

Dell trades at 11.37x trailing gross profit, against 6.99x in mid-May — the multiple roughly doubled while the margin fell — and 23.3x forward earnings on consensus of $18.80 versus $8.68 delivered. NetApp's price-to-gross-profit went from 4.55x in early May to 7.79x, with a forward multiple of 21.6x on 12.5% growth. SanDisk is the cheapest on forward numbers, at 7.47x against 20.17x trailing — but only because consensus carries fiscal 2027 earnings of $210 a share against $73.76 delivered, which requires peak NAND pricing to survive the supply flip.

The setup

Where it stands — The memory windfall is overwhelmingly price on a flat cost base, and it dwarfs what Dell and NetApp have given up. Would confirm — Dell's 1 September quarter showing ISG operating margin again flat-to-higher despite further memory inflation. Would invalidate — NAND contract prices falling quarter over quarter, or SanDisk gross margin retreating below 70%. Watch next — Dell fiscal second-quarter results on Tuesday, 1 September, with second-half gross-margin guidance. Valuation — SanDisk 20.17x trailing earnings, 7.47x forward; Dell 34.18x and 23.28x; NetApp 30.25x and 21.60x.

On 13 August, MongoDB — which sells the document database Atlas as a managed cloud service — launched a hosted way for AI coding agents such as Claude Code, Codex and Devin to read and write live application data, alongside automated embeddings built on its Voyage AI acquisition. Brokers followed: Guggenheim went to $560 on expectations of 29% Atlas growth, with Citizens at $519, RBC at $515 and Stifel at $475. What has not happened yet is the earnings report. MongoDB reports its fiscal second quarter on 1 September.

MongoDB is grouped with four other companies under a data-management heading, and the grouping implies they share a demand cycle. They do not share a billing unit, and over the past month they did not share a direction. MongoDB gained 36.2% and Dropbox 11.3%; LiveRamp was flat, Teradata fell 8.8% and Varonis 10.9%. Equal-weighted the five are up 5.5%, all of it and more from one name.

The one business that is compounding

MongoDB's case is real. April-quarter revenue reached $687.6m, up 25.2% year on year, and gross profit grew faster still, up 26.9%. Atlas revenue rose 29.4% and is now about three-quarters of the total, against 72% a year earlier, while the older licence line grew 13%. Because Atlas is metered on workloads, that line is a direct read on how much customer software is actually running. The moat is switching cost: MongoDB's storage format and query language make a migration a data transformation and application rewrite, typically a six- to eighteen-month project, which is what holds off Oracle, Postgres and the hyperscalers' bundled databases.

The price is the harder question. Price-to-gross-profit stands at 18.9x, against 13.9x on 20 July, 15.2x in May, 17.2x in February and 11.5x a year ago — above every anchor of the past twelve months. Forward earnings are 72.0x consensus fiscal-2027 EPS of $6.12, and consensus has revenue growth slowing from 21.6% this year to 18.0% next. The multiple expanded while the forward growth curve flattened. Six sessions in early August — 3, 4, 7, 10, 11 and 13 — compound to 38%, more than the entire month's gain.

The three that are not

Teradata sells Vantage, an enterprise analytics platform still largely running in customers' own data centers. Its annual revenue has fallen from $1.833bn in 2023 to $1.663bn last year. In the June quarter total annual recurring revenue grew 1%, cloud recurring revenue 8% off a small base, and consulting fell 24%. The shares lost 23.7% on 5 August after management guided third-quarter revenue down 6% to 4%. Earnings are being carried by capital allocation — $40m of stock repurchased in the quarter, a $450m term loan retired — while management conceded that its new AI features drive utilization of capacity customers already bought rather than new recurring revenue. At 3.28x EV/EBITDA and a 28.4% trailing free-cash-flow yield it is optically cheap; the deterioration is doing the work.

Varonis, which maps and polices access to sensitive corporate data for chief information security officers, is the interesting middle case. It is converting from perpetual licences to subscription, which mechanically depresses reported revenue: recurring revenue excluding conversions grew 25% to $598.1m against reported revenue growth of 18.3%. The shares still fell 12.9% on 29 July because billings shrank 1.5% and gross margin narrowed 419 basis points to 75.3%. It trades at 9.16x price-to-gross-profit, above February's 6.0x but below 13.8x a year ago.

Dropbox grew revenue 0.9% last quarter; its earnings growth comes from an 18% reduction in share count, which is why its forward price/earnings of 11.1x sits below its trailing 18.7x. Its Claude and ChatGPT integrations reached over 150,000 users, but Dash revenue is not in guidance. LiveRamp is not a re-rating at all: Publicis agreed in May to buy it for $38.50 a share in cash, and it has traded within a fraction of that price ever since.

The backdrop matters: chip stocks shed more than $1trn in late July on fears AI infrastructure spending was peaking, and money rotated into software. Only one of these five has, so far, produced the consumption numbers to justify being on the receiving end.

The setup

Where it stands — MongoDB carries the entire data-layer advance at a multiple above every anchor of the past year, before reporting. Would confirm — Atlas revenue growth at or above 29% in the 1 September print, with full-year guidance raised. Would invalidate — Atlas growth below the mid-20s, or a guidance raise smaller than the 200 basis points brokers assume. Watch next — MongoDB fiscal second-quarter results after the close on 1 September 2026. Valuation — 18.9x trailing price-to-gross-profit and 16.6x forward, against 13.9x in July and 11.5x a year ago.

Three semiconductor companies get grouped together as the machinery of "edge inference" — the silicon that runs artificial-intelligence models inside a security camera, a car or a doorbell instead of in a data center, and gets paid per shipped device rather than per rack. Over the past twelve months the three moved apart violently, and the reasons had almost nothing to do with per-device AI content.

The bid, not the business

Ambarella designs low-power system-on-chip processors that fuse video compression, image processing and computer vision, sold to camera makers, automotive driver-assistance suppliers and robotics firms. On 31 July its shares jumped 16.1%, from $74.09 to $86.00, on 9.2m shares against typical daily volume of 1-2m. The cause was a Financial Times report that NXP Semiconductors was in talks to acquire it for more than $3bn — no agreement reached, no comment from either side. NXP's own shares fell more than 6% on the report, erasing roughly $4bn of value. Remove that one session and the rest of the month nets to about -2%, leaving Ambarella below where it traded a year ago.

The business underneath is real but slowing. Revenue growth has decelerated for four consecutive quarters — 49.9%, then 31.2%, 20.1% and 16.9% in the April quarter, on revenue of $100.4m. Gross margin was 59.1%. Inventory days rose to 145 from 99 on a deliberate pre-build ahead of memory price increases; that caution looks well judged, with TrendForce now modelling a 105-110% quarter-on-quarter surge in PC DRAM contract prices. Edge-AI chips were 80% of last fiscal year's $390.7m of revenue, and a ten-year agreement with South Korea's Hanwha Group is worth more than $800m.

What it is not is cheap. Ambarella trades at 94.4x forward earnings and 12.7x forward gross profit against consensus revenue growth of 13.2%. Its trailing price-to-gross-profit is 13.9x, against 13.7x a year ago — no re-rating at all, just price and gross profit advancing together from an already extreme level. Its scarcity value is genuine: Mobileye is majority-owned by Intel, Hailo is private, and the fast-scaling rivals — Horizon Robotics, Black Sesame — are Chinese and largely closed to Western OEMs. In driver-assistance silicon, though, it is a challenger to Mobileye, NVIDIA and Qualcomm, not an incumbent, and research and development absorbs 50-70% of revenue. One distributor, WT Microelectronics, was about 61% of last quarter's sales.

The royalty that didn't arrive

CEVA licenses digital signal processor cores and Bluetooth, Wi-Fi and cellular connectivity designs to chipmakers, collecting royalties on shipped devices. It ships no silicon itself. Second-quarter revenue rose 13.1% to $29.0m — a fourth straight quarterly acceleration — at an 87.4% gross margin, and management raised full-year growth guidance to 13-15% from 12%. The shares fell 13.4% that day and a further 10.8% on 18 August.

The split explains it. Licensing revenue rose 21% to $18.2m, a three-year high and 63% of the total, while royalties — the per-device stream the whole edge-inference case rests on — were flat at $10.8m. Customers shipped 567m units, up 16%, but Bluetooth shipments fell 16% to 295m. A new deal placing its NeuPro-M neural accelerator inside a major computing platform's custom silicon carries a 1.5-2 year path to production, meaning no royalties before roughly 2028. At 8.2x trailing gross profit, CEVA has given back part of a re-rating — it stood near 10.4x in mid-May — but remains about 40% above the 5.9x of a year ago.

The one that isn't an edge story

Lattice Semiconductor sells low-power field-programmable gate arrays into communications, computing, industrial and automotive equipment, and is up 91% over twelve months. Its record quarter — revenue of $201.1m, up 62%, at a 70.3% gross margin — came from Compute & Communications, up 83% on AI data-center capital spending and rising FPGA attach per server. Industrial and embedded, the actual edge segment, grew 36%. Management expects AI-related revenue to reach a quarter of this year's total and says 2027 is "pretty much booked." The 12.5% drop on 28 July was deal risk: the $1.65bn AMI acquisition closed on $1bn of cash and a $950m delayed-draw term loan. AMI is boot firmware and server manageability software — data-center plumbing, not edge silicon. Lattice's trailing price-to-gross-profit has expanded to 36.6x from 26.0x a year ago.

Two controls settle it. indie Semiconductor, the purest automotive edge-AI name, grew revenue 24% to $64.0m and is down 6.0% over the year, with gross margin compressed to 36.1% from 40.6%. Synaptics, which sells touch, display and ultra-low-power wireless edge chips, rose 47.9% and trades at 19.2x forward earnings — a quarter of Ambarella's multiple on comparable positioning. Exposure to per-device AI content was not what got paid; data-center attach was.

The past week's decline is shared and macro. The 30-year Treasury yield topped 5.33% on 18 August, a 19-year high, and the Philadelphia Semiconductor Index fell 4.98%. On 20 August Lattice fell 5.3%, indie 7.7% and Synaptics 3.7% with no company news between them.

The setup

Where it stands — Ambarella's twelve-month advance rests on one takeover session; CEVA's growth is licensing, not royalties; Lattice's is servers, not the edge.

Would confirm — CEVA royalty revenue breaking above $12m a quarter, or Ambarella guiding October-quarter revenue above $120m.

Would invalidate — A signed NXP-Ambarella agreement, which would make the July session a floor rather than a rumor premium.

Watch next — Ambarella reports fiscal second-quarter results on 3 September 2026.

Valuation — Ambarella: 13.9x trailing gross profit and 12.7x forward, against 13.7x a year ago and Synaptics' 6.8x forward.

The electrical guts of an aircraft — wire harnesses, in-seat power units, circuit-card assemblies, radar racks — are sold against certified part numbers, not brands, and the companies that make them almost never appear in a headline. Three of the largest American suppliers of that hardware reported June-quarter results this month, and all three reported records. Two of their stocks have roughly doubled in a year. The question is what was actually bought.

Ducommun: the arithmetic behind the margin

Ducommun, a Santa Ana, California manufacturer that builds other companies' designs — interconnects, printed circuit-board assemblies, lightning-diversion systems and contoured aluminum and titanium airframe structures — reported revenue of $224.5m, up 11.0% from a year earlier. Gross margin reached 28.0%, up 160 basis points, and operating income rose 59.4%. The company reported all-time-high remaining performance obligations of $1.2bn, quarterly bookings of $309.7m and a book-to-bill of 1.4x, with missile revenue up 68% on PAC-3, THAAD, Tomahawk and AMRAAM work.

The margin, though, has a mundane source. Management's multi-year performance-center consolidation hit its full $13m annual run rate in the quarter. On fiscal 2025 revenue of $824.8m, $13m is about 158 basis points — arithmetically almost the entire gross-margin gain. The parallel story, that engineered products have grown from 15% of revenue in 2022 to 23% today, has yet to show up as a separate contribution. And the missile narrative is not what is driving the top line: commercial aerospace revenue rose 16% to $89m, faster than defense's 7% rise to $124m.

That commercial leg is a real mechanism. The Federal Aviation Administration lifted the 737 MAX production cap in March, and Boeing has moved to 47 aircraft a month from 42, with 52 targeted for early next year; Airbus is pushing toward 75 A320-family jets a month by mid-2027. More shipsets, more content.

What the price has already taken is the harder point. A fiscal 2025 charge left trailing earnings meaningless, so gross profit is the usable lens: Ducommun trades at 12.86x trailing and 12.54x forward gross profit, against roughly 6.5x a year ago, while trailing gross profit grew 14.9%. On consensus 2026 earnings of $4.16 a share, that is 48.1x forward. L3Harris trades at 23.2x. Management unveils a successor to its VISION 2027 plan at an investor day in New York on 17 September; adjusted EBITDA margin is 17.1% against the existing 18% target.

Astronics: the record is in the small segment

Astronics, of East Aurora, New York, is not really an avionics vendor — it makes cabin and airframe electrical power, lighting and connectivity systems, plus automated test equipment for defense and semiconductor customers. Revenue rose 27.0% to $260.0m, gross margin expanded 760 basis points to 33.4%, and operating margin went from 2.3% to 15.6%. The company raised full-year guidance past $1bn for the first time on record bookings of $306m and record backlog of $780.5m.

The concentration matters. Book-to-bill was 2.78x in Test Systems, a unit that turned $0.6m of operating profit on $22.7m of sales; in the Aerospace segment tied to narrowbody build rates, it was 1.02x. About 70 basis points of the gross margin came from a one-time $2m tariff refund. Debt stands at $310.3m, roughly 1.9x consensus 2026 EBITDA, and the Lufthansa Technik patent fight continues in France and Germany even after a UK appellate court dismissed Lufthansa's damages appeal in its entirety in July. Still, at 10.24x trailing gross profit Astronics is the cheapest of the three, and its multiple rose 72% over a year in which gross profit grew 48.4% — more of that move was earned than at Ducommun.

HEICO: the aftermarket that isn't compounding

HEICO, a $50.6bn Florida group whose FAA-approved replacement parts undercut engine makers' own by 30% to 50%, grew revenue 25.3% to $1.376bn with operating margin of 25.5%. But gross margin compressed 110 basis points to 41.4%, and the parts franchise geared to airline flying grew just 2% organically — against industry traffic forecast to expand 4.9% this year. Growth came from acquisitions: four year to date, net debt at 1.74x EBITDA, and a $1.2bn senior notes offering closed in July. At 25.73x trailing gross profit versus roughly 24.6x a year ago, HEICO has not re-rated at all; its 18.3% gain was paid for in profit. Fiscal third-quarter results are due after the close on 25 August.

L3Harris is the counter-case. It beat on the quarter and raised full-year guidance to as much as $23.7bn on a $42bn backlog — and the shares fell, then kept falling. It has been in a sustained downtrend since 11 June, sits 26.8% below its 12-month high, and trades at 8.83x gross profit with a 5.4% free-cash-flow yield. Investors fixed on capital intensity: research and capital spending each up over 20%, and the missile-business listing pushed to mid-2027.

What the month actually was

None of this was a grind. Astronics' entire month was one session — a 16.7% jump on 12 August — and it has since surrendered roughly half of it, including a 5.2% drop on 18 August attributed to a broad industrials selloff rather than anything the company said. Three-quarters of Ducommun's month came in two sessions; 72% of HEICO's came in one. Only Ducommun has held an unbroken uptrend, its 50-day average above its 200-day since 24 December.

The setup

Where it stands — Two small suppliers have re-rated sharply on quarters whose margin gains have identifiable one-time components; the large-cap aftermarket compounder has not re-rated at all. Would confirm — Ducommun holding gross margin at or above 28% in the September quarter, with the consolidation program already at full run rate. Would invalidate — Astronics' Aerospace book-to-bill falling below 1.0x, or Ducommun's remaining performance obligations declining from $1.2bn. Watch next — HEICO's fiscal third quarter after the close on 25 August; Ducommun's investor day in New York on 17 September. Valuation — Ducommun at 12.86x trailing and 12.54x forward gross profit, against roughly 6.5x a year ago; Astronics 10.24x, HEICO 25.73x.

Liberty Energy, a Denver-based pressure pumper that fractures oil and gas wells across North American shale and owns two Permian sand mines, told investors on 22 July that its 2026 capital budget had climbed to roughly $1.5bn from about $1.2bn. The extra money does not buy pumping equipment. It buys long-lead deposits on gas-fired generators — $71m of them in the second quarter alone, according to the company. The shares fell about 22% the following day despite beating on both revenue and adjusted earnings.

That single decision reframed an entire rung of the oilfield-services chain. The question in this business is no longer how many wells get fractured. It is who pays for the generators, and whether anyone has signed for the electricity.

The core is not shrinking

Liberty's second-quarter revenue was $1.189bn, up 14% from a year earlier and 16% sequentially. Gross margin recovered to 17.5%, from 6.2% in the first quarter and 9.7% a year ago. Below that line the recovery thins out: operating income was $12.7m, down 70% year on year, for an operating margin near 1%. The shares have fallen 38% in three months and have traded in a downtrend since early August, their 50-day average below the 200-day.

The multiple has compressed with it. Liberty changes hands at 0.78 times trailing sales, against 1.36 times in mid-May, and at 7.0 times trailing enterprise value to EBITDA — the cheapest of the group, below Halliburton at 8.5, SLB at 12.5 and Baker Hughes at 13.6. Trailing free cash flow is negative, a yield of -9.6%. Some of the discount is therefore rational: shareholders are funding a build whose flagship, the PowerBridge joint venture serving a 2-gigawatt West Texas campus, expects first power only in the fourth quarter of 2027 and has no customer leases signed. The one named counterparty, Vantage Data Centers, reserved 400 megawatts of 2027 capacity in January. Liberty has locked in 3 gigawatts of equipment through 2029 from Bergen Engines and Wärtsilä, financed partly by $1.3bn of convertible notes and project-level debt.

ProPetro, a Midland-only pumper of roughly $1.5bn market value, is further along on the thing Liberty is promising. Its contracted power book rose to about 350 megawatts from 240 in the quarter, including a 60-megawatt hyperscaler project running live and already profitable. Its fracturing business is expanding too — a twelfth fleet activated in the second quarter, a thirteenth due by the end of the third, with Permian capacity management describes as basically spoken for. Yet revenue of $305.8m was down 6.2% year on year and the company still posted an operating loss of $4.8m. It is the only loss-maker here and carries the highest multiple among the North American pumpers, 9.15 times EBITDA. Brokers cut targets in early August — Citi to $16, Piper Sandler to $18 — while keeping buy ratings.

What the market will pay for

Baker Hughes shows the price of contracted demand. The Houston company's industrial and energy technology arm — liquefaction turbomachinery, where it holds roughly 90% of the market, plus gas turbines — booked record orders of $7.1bn, double last year, a 2.2 times book-to-bill, with obligations yet to be delivered up 19% to $37.1bn. Strip out data centers and orders would still have matched the prior record. None of it is in the income statement yet: reported revenue fell 2.4%. The shares rose 17% in a month to the richest valuation in the group, a 59% premium to Halliburton.

The macro reading is counterintuitive. Brent sits near $93, up more than 37% on the year on Middle East supply disruption the Energy Information Administration expects to persist into 2027. North American completions de-rated anyway, because the binding constraint is producer capital discipline and Henry Hub gas at $2.79, down from $3.00 in May. Halliburton is not defending share at all, pricing its domestic fleet hard while shifting equipment to Argentina and the Middle East; its international revenue grew 6%. SLB, Liberty's announced power partner, is also its competitor — its data-center unit grew 80% year on year and has delivered 1.3 gigawatts. And Select Water Solutions, which recycles and disposes of produced water, signed a seven-year Permian contract with a 128m-barrel minimum commitment; its infrastructure segment earned a record $102m at a 58% gross margin. Its shares, on the longest uptrend in the group, are up 2.8% in three months.

The completions dollar is migrating to water and power. The market is paying for it only where someone has signed.

The setup

Where it stands — Liberty's pumping margins are recovering while its equity prices a self-funded, uncontracted power build; ProPetro's power book is contracted but its operations still lose money. Would confirm — Liberty signing a named lease at the PowerBridge campus, or ProPetro turning an operating profit in the third quarter. Would invalidate — Liberty cutting 2026 capex back toward $1.2bn, or ProPetro's contracted megawatts stalling near 350. Watch next — Third-quarter results in late October, plus ProPetro's promised thirteenth fleet activation by 30 September. Valuation — Liberty at 7.0x trailing EV/EBITDA and 0.72x forward sales, against 1.36x sales in mid-May; Baker Hughes at 13.6x.

Palo Alto Networks, which sells firewalls and a widening stack of subscription security services to large enterprises and governments, closed its purchase of CyberArk in February for $21.1bn — the biggest deal in its history, and one aimed squarely at privileged access and machine identity. Chief executive Nikesh Arora calls the strategy "platformization": persuade a customer to buy six products instead of one. The permanent objection is that this is discounting dressed as strategy — deals pulled forward, revenue booked once, margin surrendered permanently.

The fiscal third-quarter numbers argue otherwise, and the test is not revenue. Revenue of $3.002bn included $388m from CyberArk and Chronosphere, an observability company acquired alongside it, leaving organic growth near 14%. But next-generation security annual recurring revenue reached $8.13bn, up 60%, of which $1.63bn was acquired — roughly 28% organic. Remaining performance obligations, the contracted revenue not yet recognized, grew 36% to $18.4bn; net of $1.8bn acquired, about 23%. Obligations running at double the organic revenue rate is what a company looks like when it is signing longer contracts, not cheaper ones. Management said on the earnings call that CyberArk synergy capture is running three to six months ahead of schedule. Reported gross margin fell to 67.6% from 74.2% two quarters earlier, but that is purchase accounting flowing through cost of revenue — the same distortion that makes the 302x trailing price-to-earnings ratio meaningless here.

What actually moved the prices

Almost none of it was Palo Alto. Over the 30 sessions to 19 August the seven largest listed cybersecurity names averaged about +7.7%, but that average is two stocks: Rubrik, which sells backup and recovery software repositioned as ransomware recovery, up 27.8%, and Zscaler, whose cloud gateway replaces corporate network security appliances, up 24.3%. The other five — Cloudflare +7.1%, Palo Alto +3.2%, CrowdStrike +1.1%, Fortinet -4.5%, Okta -4.9% — average +0.4%.

Nor was it drift. Two clusters did the work. On 3-4 August a broad enterprise-software bid ignited by Palantir's results lifted the entire group. Then on 10 August, after the Black Hat conference, CrowdStrike and Palo Alto hit record highs as BTIG told clients that AI agents had "fundamentally changed the threat landscape," citing a 2.5-fold rise in agent-triggered detection leads. All seven names peaked on 13 August. In the four sessions since they have given back 7.8% on average against a 1% decline in the S&P 500, as money moved back toward AI hardware — the software-and-services index fell 2.8% on 17 August while the semiconductor index rose 1.6%.

Three billing models, three different answers

Cloudflare is not really a security vendor. It runs a global edge network — 335-plus points of presence interconnecting with roughly 13,000 carriers — and bills by consumption, selling firewalls and Zero Trust access alongside content delivery and its Workers developer platform. Its growth is accelerating: second-quarter revenue of $696.1m, up 36%, after 33.5% and 33.6% in the prior two quarters. Dollar-based net retention reached 120%, six points better than a year ago; customers spending above $100,000 rose 27% to 4,698. Non-GAAP gross margin ticked up to 73.1%, its first sequential gain in eight quarters, while capitalized network spending runs at a guided 14-15% of revenue. Matthew Prince told investors more than half of network traffic is now AI agents rather than humans — for a company billing by the request, that is added units, not lost seats.

Okta is the mirror image. It sells single sign-on and identity governance priced per user, which is precisely the unit that AI-driven headcount restraint erodes. Revenue growth has slowed for four straight quarters, to 11.2%, and management guided the quarter reporting after the close on 26 August to $790-794m, about 9%. Current remaining performance obligations grew 12% last quarter and net retention has stabilized at 107%, but margins, not growth, are doing the work: operating margin reached 7.3%. Its biggest threat is Microsoft's Entra ID, bundled into enterprise Office subscriptions at roughly $6-9 per user a month — and Palo Alto now owns CyberArk's machine-identity depth.

Among the rest, CrowdStrike, the endpoint-detection platform, reports the same evening as Okta and fell 5.3% on 19 August into it; Fortinet, the appliance-led firewall vendor serving mid-market buyers, is the only name down over the month; Zscaler and Rubrik are diverging upward on momentum rather than on any disclosed change in their contracted backlog.

The price of being right

Every one of these businesses is more expensive than in the spring, measured against gross profit — the fairest lens when Cloudflare is loss-making and Palo Alto's earnings are wrecked by acquisition accounting. Palo Alto has gone from 20.2x trailing gross profit on 3 May to 38.4x, a near-doubling in fifteen weeks, and sits at 29.5x forward. Cloudflare moved from 45.4x to 56.8x, or 49.8x forward against consensus revenue growth near 29% in 2027. Okta went from 6.0x to 10.1x while its growth rate fell — the widest gap between business and price of the three. Cloudflare's shares have held an uptrend since 5 May, roughly 71 sessions with the 50-day average above the 200-day; CrowdStrike's turned decisively down on 6 August. The group is not moving as one.

The setup

Where it stands — Palo Alto's obligations are outgrowing its organic revenue by nine points, but its multiple of gross profit has nearly doubled since May. Would confirm — Fiscal fourth-quarter next-generation security ARR landing at or above the guided $8.90-8.95bn on 1 September. Would invalidate — Full-year remaining performance obligation guidance of $20.9-21.0bn cut, or organic ARR growth slipping below the mid-20s. Watch next — Okta and CrowdStrike both report after the close on 26 August; Palo Alto follows on 1 September. Valuation — Palo Alto at 38.4x trailing and 29.5x forward gross profit, against 20.2x on 3 May.

Almost every enterprise software company sells by the head: so many users, so many dollars a month. The Descartes Systems Group, a Waterloo, Ontario business whose network sits between shippers, carriers and customs authorities, sells by the event. A customs declaration lodged, a sanctioned-party screen run, a shipment routed, a carrier message delivered — each is a billable transaction. That makes its revenue line an unusually direct read on the paperwork of cross-border trade, and it is why the debate about artificial intelligence deleting software seats mostly misses it.

Paid by the filing

In the quarter to April, reported 3 June, revenue reached $193.6m, up 15%. Of that, $180.5m — 93% of the total — was the recurring services line; licences contributed $1.6m. Adjusted EBITDA was $89.8m, or 46% of revenue, against 45% a year earlier.

The operating leverage is the part worth pausing on. Operating income rose 40.8% on revenue up 15.7%, lifting operating margin to 33.3% from 27.4%. Revenue growth has now accelerated for four consecutive quarters, from 10.0%. Chief executive Edward Ryan runs a company of about 2,400 people and a $6.7bn market value; it is compounding its earnings roughly two and a half times as fast as its sales.

Fewer parcels, more paperwork

The mechanism behind that acceleration is two-sided, and one side is hostile. The American ban on duty-free treatment for low-value Chinese shipments remains the primary reason trade volumes are falling, and from 1 July the European Union began applying a €3 (about $3.50) customs duty to consignments under €150, pushing Shein and Temu toward local warehousing and consolidated hub shipments. That means fewer, larger filings — fewer billable events.

What offsets it is complexity. A 25% tariff on covered Brazilian products took effect 22 July, and certain Canadian imports face a new 50% duty from 19 August. Every rate change, exclusion and origin rule is content Descartes maintains and sells through its Global Trade Intelligence line, which management named as its main growth theme while describing the underlying freight market as weak. Its defensible position is regulatory data and carrier connectivity rather than execution software — the reason it competes with, rather than inside, WiseTech Global and Manhattan Associates.

The rally it did not join

Software shares surged over three sessions in late July after Huang rejected the AI-disruption thesis, saying customers will rent agents that use software tools and the industry will end up larger. Sixteen comparable application-software names rose an average 21.8% over the following month. Descartes managed 5.3%, and by 19 August sat slightly below its 29 July close.

The multiple reflects that. Descartes trades at 11.4 times forward gross profit — the fitting lens, since gross margins across these businesses range from 53% to 82% — against roughly 14.5 times a year ago. Manhattan Associates, the warehouse-software company whose gap higher powered much of the group's month, is back at 18.9 times forward gross profit, recovering nearly all of its twelve-month de-rating even though its operating income fell 10.2% last quarter on 9.3% revenue growth. The market has paid up for the business with negative operating leverage and left the one with widening margins where it was.

Two things argue for caution. Consensus models Descartes at $807m of revenue this fiscal year, growth of 9.7%, well below the 15.7% just delivered — a deceleration the weak freight backdrop partly justifies. And growth is part-bought: Idelic, a driver-safety platform, closed in April, and Drivin followed in July, with no disclosed split between organic and acquired revenue. Descartes has not held an earnings call since 3 June, so nothing in its share price since reflects new company information.

The setup

Where it stands — A transaction-billed trade-compliance network accelerating into a weak freight market, priced below where it was a year ago. Would confirm — Q2 revenue growth holding above 13% with adjusted EBITDA margin at or above 46%. Would invalidate — Growth falling toward the 9.7% consensus with acquisitions supplying most of it. Watch next — Fiscal second-quarter results, due early September. Valuation — 11.4x forward gross profit and 12.2x trailing, against roughly 14.5x a year ago; 33.7x forward earnings.

BETA Technologies, a South Burlington, Vermont builder of piloted electric aircraft, booked revenue of $14.7m in the second quarter — and some of its best customers are the companies trying to beat it to market. It supplies motors to Eve Holding, flight control computers to the Canadian developer Horizon Aircraft, and charging hardware to operators who have not yet chosen an airframe. Management expects 40-60% gross margins on that component line. Revenue rose 146% from a year earlier, gross margin recovered to 54.7% from -3.1% in the first quarter, and full-year guidance went to $42-50m from $39-43m.

That is the entire delivered top line of the American electric-aircraft business right now, and it belongs to the one company that decided to arm its rivals.

The month was three days, and it was about defense

Archer Aviation, BETA and Eve Holding rose an average of 22% over the month to 19 August. The obvious reading — that air taxis are being repriced — does not survive the comparison. Joby Aviation, the closest peer and the only one of the four with a nine-figure revenue guide, rose 3.3%. Kratos Defense & Security Solutions, which builds jet-powered target and attack drones, rose 31.8%. Red Cat, a maker of small military reconnaissance drones, rose 27.4%. Buyers wanted autonomy and defense exposure; the eVTOL leader was left out.

The gains were also concentrated. Three of Archer's sessions compound to +35.9%, more than its entire +21.5% month, while the other nineteen sessions net roughly -10%. Eve's single best day, 4 August at +19.2%, exceeded its whole 10.3% gain, and the shares have since round-tripped from $2.89 to $2.67. That 4 August session was not eVTOL news at all: the Nasdaq rose nearly 3% and small caps hit their first new highs since 2021 after the Federal Reserve held rates steady.

Archer bought its next act with stock

Archer, of Palo Alto, is developing the four-passenger Midnight aircraft and has yet to sell a seat. Second-quarter revenue was $5.0m; the operating loss widened 58.5% year on year to $279.2m. The share count is the story: 397.5m diluted shares a year ago, 781.7m now. On 9 August the company agreed to acquire Boeing's Wisk Aero, Insitu and SkyGrid units, paying in stock worth about 19.75% of the company plus warrants over a further $200m.

It buys something real. Insitu is profitable on more than $200m of annual revenue and has built over 4,000 unmanned aircraft. It is also not in consensus, has not closed, and is subject to antitrust review. Archer holds $1.6bn of liquidity against an adjusted EBITDA loss of $177m in the quarter, so the runway is bought — with equity. On certification, Archer says it is the only manufacturer in the fourth and final FAA stage with fully accepted means of compliance, while Joby has flown a conforming aircraft for Type Inspection Authorization, the step that puts FAA pilots in the cockpit. Archer's near-term cash is coming from Abu Dhabi, where it has begun receiving payments under a launch programme and is targeting a restricted UAE type certificate ahead of full FAA approval.

Eve's backlog is mostly a wish list

Eve Holding, based in Melbourne, Florida, with 198 employees, is less a standalone company than an Embraer programme with a listing. It has never reported revenue; the second-quarter operating loss was $37.2m. Its headline order book of about 2,700 aircraft, valued near $13.5bn at list prices, rests on non-binding letters of intent — only 100 units, from Revo and AirX, are firm. The engineering prototype has flown 66 times for 2 hours 46 minutes in total and needs 30 to 40 more flights before full wing-borne transition. Cash of $403m and $531m of liquidity are guided to last "through 2028" — the same year service is meant to begin. Consensus revenue for 2026 has been cut to $222k.

Why revenue exists before a certificate

The mechanism is the eVTOL Integration Pilot Program: under a June 2025 executive order, the Department of Transportation and the FAA selected eight projects across 26 states on 9 March 2026, letting uncertified aircraft fly paying missions. BETA flew the first such flight on 12 July, moving a transplant organ for United Therapeutics between Maryland and Virginia — commercial operations pulled forward more than a year against the certification clock.

What the prices already assume

Because gross margins differ wildly across the three, price-to-gross-profit is the honest lens for BETA: 282x trailing against 270x forward — essentially no compression, since 2026 revenue steps only from roughly $36m to $47m, and consensus does not show a profitable year until 2030. Archer trades at 710x trailing and 322x forward sales on $15.2m of 2026 consensus revenue. Eve has no sales to divide into, leaving book value: 41.1x, against 2.66x at Archer and 3.79x at BETA.

The setup

Where it stands — One of the three sells something today; all three carry multiples that assume certification arrives on schedule. Would confirm — BETA delivers within its raised $42-50m guidance and holds gross margin above 50% in the third quarter. Would invalidate — Archer's Boeing acquisition fails antitrust review or slips past year end, removing the revenue it prices. Watch next — Eve's third-quarter report, and whether the prototype logs the 30-40 flights to full transition. Valuation — BETA 282x trailing, 270x forward gross profit; Archer 710x/322x sales; Eve 41.1x book versus Archer's 2.66x.

Accenture, the Dublin-headquartered firm that sells strategy, systems integration and outsourced operations to the world's largest companies, produced roughly $91,500 of revenue per employee over the past twelve months. A year earlier the figure was about $86,600, and the workforce — 799,000 people — grew barely 1% in between. For a business whose unit of sale is a person's hour, that is the single cleanest test of whether artificial intelligence is compressing what an hour is worth. At Accenture it currently says no.

At Cognizant, the Teaneck, New Jersey outsourcer that runs offshore delivery for banks, insurers, healthcare payers and consumer companies, the same measure rose 1.8%, to about $60,700. Headcount ended June at 356,700, up 12,900 from a year earlier. Revenue grew 4.5%. That is growth bought with people rather than with price — the shape you would expect if rates were flat and volume were doing the work.

The number that would settle it is gone

Accenture's third quarter, ended 31 May, showed new bookings of $19.32bn, down 2% in dollars and 3% in local currency, against revenue of $18.72bn — a book-to-bill near 1.03, with orders falling while revenue rose 5.6%. Underneath, the mix is barbelling: 104 client engagements of $100m or more year to date, up 13%. Accenture is winning more of the very large transformation programs and losing short-cycle discretionary work.

Whether the AI work inside those programs is incremental revenue or the same maintenance revenue rebooked at a different rate is now unanswerable from disclosure. Management has stopped breaking out advanced-AI bookings, on the grounds that AI is pervasive across the firm. That is a defensible accounting position and an unhelpful one for anyone trying to size the transition. Full-year guidance stayed at 3–4% local-currency growth.

The deflation risk is not hypothetical. Infosys has told investors that clients are demanding effort reductions at renewal and mid-contract, across telecom, financial services, retail and utilities, and has named the phenomenon AI-led deflation without quantifying it.

Cognizant's margin is coming from the wrong line

Cognizant is the operational bright spot and the caveat at once. Operating income grew 7.0% on 4.5% revenue, a sixth straight quarter of adjusted operating-margin expansion, and full-year adjusted earnings guidance went up to $5.70–$5.82. But gross margin fell to 33.37% from 33.67%. The gains are landing in overhead and restructuring, not in delivery economics. Utilization excluding trainees ran 85% and voluntary attrition in technology services rose to 13.0%. Seven deals above $100m of total contract value closed in the quarter, three of them new logos, and clients are beginning to ask for rate cards that embed model inference costs — the billing unit itself is drifting away from the hour.

EPAM's shares and EPAM's numbers disagree

EPAM, a premium engineering shop that builds software platforms for banks, travel firms and life-sciences companies, earns about $89,400 per head from just 62,850 staff — nearly Accenture's productivity at a twentieth of its market value, $5.6bn. Margins are expanding smartly: operating income rose 20.4% on 4.5% revenue. Everything else is deteriorating. Growth has halved twice in three quarters, from 12.75%. The Americas, 57% of revenue, grew 0.5%. Free cash flow was minus $18m against plus $43m a year earlier. Full-year guidance was cut to 3.2–4.2%, with organic constant-currency growth of 2–3%, and the pipeline of AI-led managed-services deals has closed nothing, with contribution pushed to 2027. The shares fell 14.5% on 6 August and still finished the month up about 20%.

What the month was

The advance is dated, not gradual. Almost all of it landed in four sessions at the end of July and again on 19 August, when six of these names rose together without company news, funded by a rotation out of AI hardware into beaten-down software and services. Strip each name's two best days and the group is flat, with five of eight negative. Accenture is the exception: it still gains 11.8% across the period, on 12 up days out of 22.

Price levels leave room but less than they did. Accenture trades at 14.5x trailing and 13.2x forward earnings, and 4.76x forward gross profit — the fitter lens for labour businesses with different margin structures — against roughly 6.6x scaled to its share price a year ago, with an 11.2% trailing free-cash-flow yield. The trailing-to-forward spread is negligible: consensus assumes no inflection. Cognizant is at 10.6x forward and 7.2x EV/EBITDA. EPAM's 8.2x forward looks cheapest until you notice what it embeds — consensus 2026 earnings of $13.14 a share, up 15.1%, on revenue growth of 3.9% against management's 2–3% organic guide.

The setup

Where it stands — Accenture's revenue per head is rising without hiring; Cognizant's is not; EPAM's price recovered while its guidance was cut.

Would confirm — Accenture's next quarter showing revenue growth at or above 5% with headcount growth under 2%.

Would invalidate — Accenture revenue per employee flat or lower year over year, or new bookings falling a third straight quarter.

Watch next — Accenture's fourth-quarter and full-year fiscal 2026 results, due late September, with fiscal 2027 guidance.

Valuation — Accenture 14.5x trailing and 13.2x forward earnings, 4.76x forward gross profit versus roughly 6.6x price-scaled a year ago.

None of them had news. Over two sessions on 18 and 19 August, every listed supplier of the parts that go inside chipmaking tools — gas-delivery panels, probe cards, ion implanters, test handlers — dropped double digits at once. FormFactor fell 17.1%, Ichor 16.6%, MKS 13.0%, Veeco 12.8%, Axcelis 12.3% and Cohu 11.9%. Lam Research, which buys from several of them, fell 10.7%. Nvidia fell 2.1%.

The trigger sat in the bond market. The 30-year Treasury yield topped 5.33% on 18 August, a 19-year high, on sticky core inflation and deficit concerns, and the Philadelphia Semiconductor Index fell 4.96% that day. Two sessions were not the whole story, though: this rung has been bleeding since midsummer. Ichor is 45.5% below its 30 June close, Veeco 42.9%, MKS 36.7%, Axcelis 34.0% from its June high and FormFactor 28.1%.

Why this layer is the most rate-sensitive rung

These firms are paid a slice of someone else's tool, and the slice is booked late. Ichor's customers are now placing purchase orders six months out — a pattern management called abnormal. FormFactor's new Farmers Branch plant does not turn accretive until early 2028. Veeco's $200m advanced-packaging order, won in the June quarter, is mostly for 2027 delivery. Earnings power parked three years out is exactly what a 5.33% long bond discounts hardest.

What did not happen was a cut to the spending that feeds them. Lam Research raised its 2026 wafer fab equipment forecast to the low $150bn range from $140bn. Applied Materials guided its next quarter to $10.25bn, up 51% year on year on "unprecedented" demand, and said leading-edge logic, DRAM and advanced packaging would be about 80% of equipment-market growth in 2026 and 2027.

FormFactor: the numbers went the other way

FormFactor makes probe cards — the consumable interfaces that test chips while they are still on the wafer, designed in-house and sold against Japanese rivals MJC and JEM. June-quarter revenue rose 31.9% to $258.2m. Gross margin reached 50.7% against 37.2% a year earlier, and gross profit rose 80%; operating income rose more than fourfold. DRAM probe cards set a record on HBM4, the next generation of stacked memory, with high-bandwidth memory roughly two-thirds of DRAM revenue — and advanced HBM cards carry five to ten times the selling price of legacy ones. Foundry and logic did not stall either: a probe card for Nvidia's Rubin GPU is qualified and ships for revenue this quarter. Capacity, not demand, is the constraint until Farmers Branch ramps.

Against that, the shares trade at 21.8x trailing gross profit, versus 35.3x at the 30 June peak and 23.1x in February — the cheapest in six months, on a trailing gross-profit base 38.6% larger than a year ago. The honest caution is mix: September DRAM revenue is guided flat, shifting toward commodity DDR, which carries far less test content per wafer than an HBM stack.

Ichor: better business, still the expensive way to own it

Ichor assembles gas and fluid-delivery modules to other firms' drawings; Lam and Applied were 76% of its 2025 revenue. The recovery is real — revenue up 22.7% to $294.8m, gross margin back to 13.9% from a 4.6% trough, full-year guidance raised on 3 August to at least 30% growth, with internal content rising toward 35% of the bill of materials. But at 20.8x trailing gross profit it is still above its 19.0x February level, and 2.1x Ultra Clean's 9.7x for the same contract gas-panel work at a fatter 16.1% margin. Its $200m share sale was struck at $80.70; the stock closed at $61.21. The shares fell 12.8% after the print as negative operating cash flow overshadowed the raise.

Axcelis: the one where selling and fundamentals agree

Axcelis sells ion-implantation systems into mature-node, power-device and Chinese fabs — not the leading-edge logic, DRAM and packaging that Applied says will carry 80% of equipment growth. June-quarter revenue rose 10.6% to $215.2m, but gross margin slipped to 42.4% and operating income fell 30%. Trailing gross profit is 7.4% lower than a year ago; annual revenue has fallen two straight years, from $1.13bn in 2023 to $839m in 2025. Bookings roughly matched systems revenue, leaving backlog at $451.6m. China was 46% of the quarter. It is the only name here whose forward multiple of gross profit, 10.71x, sits above its trailing 10.43x — consensus expects the gross-profit dollars to shrink. It is also 59% more expensive than in February, and Applied holds about 63% of the implant market to Axcelis's 21%. Its $4.4bn all-stock merger with Veeco still awaits China's antitrust regulator, which moved the case onto a longer standard review.

The other three fall on the growing side. MKS, which has content on roughly 85% of wafer fab equipment, grew revenue 28.3% to $1.25bn and saw its multiple of trailing gross profit compress from 16.9x to 9.9x while that gross profit rose to $1.92bn. Cohu raised full-year guidance to about 35% growth with computing orders up 150%. Veeco raised revenue guidance but cut earnings guidance to fund 2027 capacity — the clearest statement that this layer is spending now for revenue later.

The setup

Where it stands — A rate shock repriced the whole sub-tool layer uniformly, but only Axcelis's shrinking gross profit matches the selling. Would confirm — FormFactor's September quarter landing near $270m with non-GAAP gross margin at the guided 54%. Would invalidate — Lam or Applied cutting 2027 equipment-spending guidance, turning a discount-rate move into a demand call. Watch next — Axcelis's next systems bookings print and China's SAMR ruling on the Veeco merger, expected before year-end. Valuation — FormFactor 21.8x trailing and 19.0x forward gross profit, against 35.3x on 30 June; Axcelis 10.43x trailing, 10.71x forward.