DK Street Journal

Two cloud-communications companies reported June quarters days apart, and both put up revenue growth of roughly 22%. Only one of them grew gross profit at anything close to that rate. The difference is not execution in the ordinary sense. It is a decision by American mobile carriers, made in January, that flows through both income statements the same way and lands very differently depending on what the business underneath is actually doing.

The fee that inflates revenue and nothing else

Twilio, whose programming interfaces let software developers embed text messages, phone calls and email inside their own applications and bills them by the message and the minute, reported revenue of $1.5bn, up 22% reported and 17% organic. Bandwidth, a smaller rival that owns the carrier network its traffic runs across and sells voice and messaging to large enterprises and contact centers, reported $220m, up 22%. Twilio's gross profit rose 20.4% to $725.9m. Bandwidth's rose 9.6% to $78.5m.

The mechanism is application-to-person (A2P) messaging surcharges — per-message fees the carriers levy on automated text traffic. T-Mobile raised them effective 19 January 2026 across short code, 10-digit long code and toll-free routes. Providers collect the money and hand it on. Revenue rises, dollar gross profit does not, and reported gross margin falls. Bandwidth's quarterly filing puts the arithmetic plainly: first-half cost of revenue rose $61m, of which $41m was higher pass-through messaging surcharges. Twilio's carrier fees were $71m in the quarter and are running toward $250m for the year.

At Twilio, optical. At Bandwidth, structural.

Strip the fees out and Twilio's non-GAAP gross margin was up 60 basis points year on year rather than down 160. Its reported GAAP gross margin slipped only 65 basis points. Dollar-based net expansion — what existing customers spend this year against last — reached 116%, and customers spending over $1m grew more than 20%. Operating income more than doubled, to $84.5m from $37.0m. Management raised full-year organic growth guidance to 13-13.5% from 9.5-10.5%.

Bandwidth's core did not keep pace. Cloud communications revenue grew 12% while the surcharges grew 54%. Voice, its largest line, was $121m and grew 9%; management attributed the slowdown to customer deployment timing. Gross margin fell to 35.7% from 39.8%, and has declined every quarter since 41.0% in early 2025. The company remains unprofitable on a GAAP basis, with a $4.6m operating loss in the quarter. Consensus has 2026 revenue growing 20% and 2027 growing 4% — the Street is modeling this year's surge as largely non-repeating.

What each one owns

Twilio held 26.2% of global revenue in communications-platform-as-a-service (CPaaS) in late 2025, ahead of Sinch at 12%. Its defensible asset is direct carrier interconnection across 180-plus countries, built over fifteen years, and no single customer exceeds 10% of revenue. It has positioned frontier model vendors as distribution rather than threat, shipping a native integration with OpenAI's Realtime API. Management cited one artificial-intelligence customer that went from $200,000 to $9m of annual recurring revenue in five quarters.

Bandwidth's asset is the owned network, which won it the communications layer inside Salesforce's Agentforce Contact Center. Its software services ARR rose to $25m from $15m at year-end — real, growing fast, and under 3% of the $900m of revenue guided for this year. Net retention was 107%, with record revenue per customer of $256,000.

The shares agreed, violently

Bandwidth fell 18.8% on 28 July and a further 29.3% the next day, 42.6% across the two sessions around its report. Remove those two days and it would be up roughly 32% over the past month. Twilio gapped 26.6% higher on 7 August; remove that one session and it would be down about 15%. Neither chart is a gradual drift — both are single verdicts on guidance. Twilio's 50-day average has sat above its 200-day since mid-April; Bandwidth's crossed back down on 6 August.

Gross profit is the honest yardstick here, because Twilio's 29.5x trailing earnings multiple is distorted by a $1.07bn one-off gain and Bandwidth barely earns a GAAP profit at all. Twilio trades at 12.45x trailing and 11.91x forward gross profit, against roughly 6.8x in February — the multiple nearly doubled in six months while the business improved considerably less than that. Bandwidth sits at 5.46x trailing and 4.99x forward, down from about 7.2x at its July high but still more than three times its February level, on gross profit growing under 10%. Twilio guides third-quarter organic growth down to 11-12%.

The setup

Where it stands — Twilio is converting its revenue growth into gross profit; Bandwidth's June quarter showed it is not, and its shares repriced accordingly.

Would confirm — Bandwidth's cloud communications growth staying near 12% or below while surcharge revenue keeps outpacing it.

Would invalidate — Bandwidth gross margin recovering toward 39% with voice reaccelerating above 15% in the September quarter.

Watch next — Third-quarter results: Twilio has guided organic growth to 11-12%, Bandwidth to second-half growth slightly below the first half.

Valuation — Twilio 12.45x trailing / 11.91x forward gross profit versus about 6.8x in February; Bandwidth 5.46x / 4.99x versus about 1.6x.

Analog Devices told investors on Wednesday that the analog chip market it has spent a decade describing as a 5-7% grower should now compound at double digits for several years. The reasoning is arithmetic, not enthusiasm: roughly 100 gigawatts of new data-center capacity between 2026 and 2031, with each gigawatt consuming $1bn to $1.5bn of the power-management, signal-conditioning and data-conversion parts the company sells. ADI supplies the analog layer around digital compute — the chips that turn physical current, voltage and radio signals into numbers — for factory automation, test equipment, cars and telecom networks.

The quarter underneath that forecast was the best in company history. Revenue reached $4.02bn, up 39.6% year on year, beating consensus and lifting October-quarter guidance to $4.2-4.4bn against $3.55 of expected earnings per share. GAAP gross margin was 67.3%, against 62.1% a year earlier. Industrial, at 49% of revenue, grew 53%. Communications grew 84%, four-fifths of it data center. The shares closed down slightly on the day and sit 16% below their 22 June high.

Three chipmakers, one peak date

That date matters, because NXP Semiconductors and Microchip Technology peaked in the same session. NXP, the Eindhoven-based supplier of automotive microcontrollers, radar and secure connectivity chips, has fallen 30.1% from its 22 June close; Microchip, the Arizona broad-line vendor of 8-, 16- and 32-bit microcontrollers alongside analog and timing parts, is down 25.0%. Both crossed from an uptrend into a downtrend in mid-August, their 50-day averages slipping under their 200-day.

Nothing in either company's results supports the move. NXP grew revenue 19.5% to $3.496bn in the June quarter, its fourth consecutive quarter of acceleration, with gross margin at 57.3% and operating income up 55.9%. On the 28 July call management said book-to-bill is above 1.0 and rising, backlog visibility now runs 18 months, and customer escalations doubled. Crucially, it attributed the growth to content per vehicle rather than restocking: Western tier-one inventories show no replenishment and ordering remains hand-to-mouth. NXP guided September revenue to $3.75bn. The stock fell 7.7% anyway.

Microchip's print was the sharpest reversal of the bear case. June-quarter revenue rose 38% to $1.485bn. Gross margin has now expanded for five straight quarters, from 51.6% to 63.2%, and the under-utilization charges from idled fabs are shrinking — about $30.5m guided for September, down from $38.5m. Distributor inventory is down to 25 days, the low end of its historical range; bookings were the strongest in four years. September revenue was guided to $1.59-1.62bn against roughly $1.55bn expected. Its competitive edge in this cycle is counterintuitive: some $450m of idle equipment and slack internal fabs at a moment when AI has crowded out external foundry, packaging and test capacity industry-wide.

What actually moved the prices

Strip each stock's two worst sessions from the last 90 days and the picture separates. Microchip's 18.0% decline becomes a 1.4% gain; ADI's 6.2% decline becomes a 10.8% gain. Both hinge on 23 and 26 June — days when the whole semiconductor complex unwound, with the VanEck semiconductor ETF down 6.5% and Micron down 11.4%. July was then the worst month for chip stocks in more than a decade, erasing over $1trn. The August leg was a discount-rate event: the 30-year Treasury yield touched 5.33% on 18 August as the Philadelphia Semiconductor index dropped 5.4%.

NXP is the exception. Its 27.1% 90-day fall only improves to 14.5% without its two worst days — a genuine, sustained drift. The company-specific news does not explain it: the succession to Rafael Sotomayor completed in October 2025, and the reported approach to Ambarella, the edge-AI vision chip maker valued near $3.3bn, is an outbound purchase.

Where the price leaves room

Against mid-May, all three multiples have compressed: NXP from about 19.5x forward earnings to 15.0x, Microchip from 28.8x to 24.1x, ADI from 33.4x to 30.1x. NXP trades below its 19.2x trailing multiple because estimates are rising — consensus has FY2026 earnings up 28% — and carries a 5.2% free-cash-flow yield. Microchip's 106.8x trailing multiple prices trough profits and says nothing; 24.1x forward sits on earnings consensus expects to double. ADI is the one where the fall looks like a re-rating of a rich stock rather than a mispricing: 30.1x forward and 18.7x forward gross profit, after gaining 62% over twelve months. NXP, over the same twelve months, is flat.

The same macro, two opposite reports

Pegasystems, the Cambridge, Massachusetts vendor of workflow and customer-engagement software sold to banks, insurers, healthcare payers and governments, told investors on 22 July that its customers had stopped signing. Annual contract value (ACV) — the only clean growth measure at a company whose reported revenue lurches on upfront license accounting — reached $1.62bn, short of the $1.658bn expected. That is 7% growth, against 12% the prior quarter and 17% two quarters before. Founder and chief executive Alan Trefler called it a "max confusion moment", with enterprises whipsawing between maximizing AI spend and controlling it. Management had already back-loaded two-thirds of the year's net-new ACV into the second half; it then said the first half "significantly underachieved" and that recovery "will be very difficult". Citizens cut the stock to Market Perform, and a Sell downgrade followed on 13 August citing execution and competition.

Two weeks later Appian, the McLean, Virginia low-code vendor whose platform generates the forms, workflows and data models that would otherwise be hand-coded, reported the same environment with the opposite result. Cloud subscription revenue rose 23% to $131.7m, and net expansion inside existing cloud accounts held at 115%. Adjusted EBITDA of $16.2m beat the company's own $5-8m guide. Full-year revenue guidance went to $845-853m — above where analysts sat. Management said it saw no lengthening of deal cycles and pulled 2027 hiring forward into this year.

Whether AI eats the workflow tool

This is the rung of enterprise software with the most direct exposure to agents that write a process rather than execute a designed one. So far the evidence at Appian runs the other way: 85% of second-quarter new logos bought AI-enabled tiers, and it cited a health insurer expecting more than $10m of savings over three years from automated medical-records processing. Its pitch is to be the deterministic layer underneath the models — governance, data access, an auditable process.

Pegasystems is making a nearly identical argument and getting a worse hearing. It shipped Infinity 26 in July, extended its Blueprint design tool across the full build-and-operate lifecycle, and exposed the platform so outside agents can invoke Pega workflows — priced per case rather than per token, explicitly to spare customers runaway token costs. Pega Cloud ACV grew 22% in constant currency and is now 57% of total, but that decelerated from 27%, and the mix shift keeps pressuring legacy license and maintenance lines on the way to a targeted 75%.

What each price already pays for

Appian's re-rating has been fast: price-to-gross-profit stood at 3.07x in early May and 3.63x in late July, and is 4.62x on forward gross profit now. That is roughly 38x forward earnings on consensus of $0.99, paid against a curve consensus expects to slow to 10.7% revenue growth in 2027. Gross margin compressed to 71.2% from 74.3% a year earlier on AI cost mix, and the operating line swung to -2.7% with an $11.8m net loss.

Pegasystems is priced as the broken one. It trades at 13.8x forward earnings against 17.8x trailing, and 3.90x forward gross profit against the 4.72x it carried in May — a month of gains has not restored its own three-month multiple. The cash is real: $288m of first-half free cash flow, full-year guidance of $575m reaffirmed, and $360m spent buying back 9m shares, more than the cash generated. Net income still fell 55.7% to $13.3m.

The month was one stock, and one session

Over the past 22 sessions Appian rose 44% while Pegasystems added 2.8%. Remove each name's two best days and Appian is up 4.7% and Pegasystems is down 15.3%. Pegasystems' single best day, +13.6% on 28 July, was the market-wide rotation out of semiconductors into discounted software, not company news; the same bid lifted ServiceNow and monday.com and returned again on 18 August. Appian's 50-day average crossed above its 200-day in mid-August. Pegasystems' has sat below since 30 January without a break.

One shared overhang remains. The trade-secrets case Appian won — a $2.04bn jury award in 2022 — was reversed, and Virginia's high court upheld that reversal on 8 January 2026, sending it back for a new trial on both liability and damages.

The setup

Where it stands — Appian is executing and richly priced; Pegasystems is cheap with contract-value growth halved and its downtrend unbroken since January. Would confirm — Pegasystems' third-quarter total ACV growth reaccelerating above 7% in constant currency. Would invalidate — Appian's cloud net ARR expansion falling below 110%, or full-year guidance trimmed from $845-853m. Watch next — Pegasystems' third-quarter results, where management itself called the season a "lousy time for turnaround". Valuation — Pegasystems 13.8x forward earnings vs 17.8x trailing; Appian 4.62x forward gross profit vs 3.07x in May.

Dell Technologies recognized $16.1bn of artificial-intelligence server revenue in the quarter to 1 May, and the more of those machines it shipped, the less of each dollar it kept. Revenue for the quarter rose 87.5% to $43.8bn. Gross profit rose 57.6%. The gap is the whole story: gross margin fell to 17.75% from 21.12% a year earlier, a compression of 337 basis points, and it has now declined in every fiscal year of the AI build — 23.83%, then 22.24%, then 20.00%.

Dell sells servers, storage and networking through its Infrastructure Solutions Group and PCs through its Client Solutions Group. It is the largest branded seller of AI servers by volume, and in a box built around someone else's accelerator, memory is the single biggest input Dell buys itself.

Who pays for the memory

Contract prices for server DRAM (dynamic random-access memory) are expected to rise 13-18% quarter on quarter in the third quarter of 2026, with the market undersupplied. TrendForce expects the shortage to persist through 2027, RDIMM bit supply growing only 15-20% against faster server CPU shipments. Crucially, several large US cloud providers have locked multi-year agreements capping their increases — so the rises land on buyers without such contracts and on incremental volume outside them. Dell told investors in May that demand exceeds supply and named memory, CPUs and hard drives as the constraints, warning of shortages in the second half of its fiscal year. Amazon has attributed a $20bn rise in capital spending to higher memory prices rather than added capacity.

NetApp, which sells all-flash enterprise storage arrays and the ONTAP data-management software that runs them, has guided the same mechanism explicitly: fiscal 2027 gross margin of 68.5-69.5% against 70.74% delivered in fiscal 2026, citing higher memory and NAND flash costs. That comes with the best revenue growth in three years — up 12.5% in the April quarter — and record all-flash array revenue of $1.2bn.

Celestica, the Toronto contract manufacturer that builds switches, interconnect and full server racks for hyperscalers, sits on the other side of the trade. Its gross margin is 12.29%, roughly one-sixth of NetApp's, because it is paid to convert components into finished racks rather than to sell a product at a price. June-quarter revenue rose 62.4% to $4.70bn and operating margin expanded, to 9.75% from 9.42%. Management guided gross margin to stay in the mid-11% range through 2026 and said material availability, not factory floor space, is the bottleneck, with customers placing non-cancellable orders for long-lead silicon — a structure that hands component risk back to the buyer.

What the prices already assume

Dell's shares have risen 220% in twelve months. Its trailing twelve-month gross profit rose 20.7% over the same period, to $25.55bn. Price-to-gross-profit — the only comparable lens when margins in one group range from 12% to 70% — stands at 11.25x, against 6.99x in mid-May and roughly 4.5x a year ago. NetApp shows the same shape in miniature: gross profit up 6.2%, shares up 80%, the multiple moving from about 4.8x to 7.86x.

Celestica is the exception, and it is the surprising one. Its trailing gross profit grew 52% year on year, yet its price-to-gross-profit fell from 26.26x in mid-May to 19.19x. The break traces to a single dated event rather than to demand: on 5 August it priced 9.68m shares at $310.00 for $3bn of gross proceeds, a 14.5% discount to the prior close, to fund capacity. The stock fell about 15% the next session and now trades below the offer price — this after the company raised full-year guidance to $20.5bn of revenue and disclosed custom-rack programs with OpenAI and AMD.

One caution on Dell's advance: it is concentrated, not gradual. The 13.3% gain over the past 30 sessions becomes a 5.4% loss if the two best days are removed, and four sessions in that window fell 5% or more.

The setup

Where it stands — AI revenue is arriving at Dell and NetApp with lower gross margins, while Celestica's fee-based model passes the cost through.

Would confirm — Dell's fiscal Q2 gross margin printing below the 17.75% of the prior quarter as AI mix rises again.

Would invalidate — Dell holding gross margin near 20% while AI server revenue grows, showing memory costs are contracted forward.

Watch next — Dell reports fiscal Q2 on 1 September; NetApp reports fiscal Q1 after the close on 2 September.

Valuation — Dell at 33.8x trailing and 23.4x forward earnings; 11.25x trailing gross profit against 6.99x in mid-May.

On July 30, Roblox told investors that bookings — the cash players spend on Robux, the currency used inside experiences that other people build — would fall between 14% and 18% this quarter. It would be the first decline in the company's history. Management also withdrew full-year guidance, citing "increasing variability" from safety, discovery and new-product changes.

Roblox runs a user-generated 3D platform: creators build the games, players buy currency to spend inside them, and Roblox takes a cut. That model depends on two inputs — how many people show up, and how much each one spends. Both are now going the wrong way at once. Daily active users have fallen for three consecutive quarters, from 152m in the third quarter of 2025 to 123m in the second quarter of 2026. Hours engaged dropped from a peak of 40bn to 29bn. Average bookings per daily user slid from $15.97 in late 2024 to $12.66.

The deceleration is startling in its speed. Bookings grew 43% year over year in the first quarter of 2026. One quarter later they grew 8%. The quarter after that, they are guided to shrink — and that is despite Russia being reinstated late in the second quarter, giving the third a full-quarter benefit from a reopened market.

Two causes, one self-inflicted

The first cause is regulation, and it is compounding rather than one-off. Roblox began voluntary age estimation — a video selfie processed by a third-party vendor into age bands — in November 2025, made it mandatory in Australia, New Zealand and the Netherlands from December, and went global in January. That bracket lines up exactly with the quarters in which users began leaving. The company settled with Alabama for $12m and Nevada for $12.5m over child safety, agreeing in both cases to mandate face scans or government-ID checks; Oklahoma became at least the tenth state to sue. Meanwhile Texas's App Store Accountability Act is in force after the Fifth Circuit stayed an injunction against it in June, and Utah's law requiring explicit parental permission for minors on user-generated-content apps took effect July 1. Age-check penetration is already 57% globally and roughly 70% in the United States.

The second cause Roblox chose. Management attributed the miss to a collapse in monetization among under-13s plus a deliberate overhaul of its discovery algorithm, which now surfaces games that keep players engaged longer but extract less money from them. Asked when that trade pays off, management declined to give a timeline.

Costs are moving the other way. Roblox told investors third-quarter margin pressure splits roughly evenly between fixed-cost deleverage and artificial-intelligence infrastructure spending, and that infrastructure cost is shifting from variable per-hour to fixed. Cost per hour rises as hours fall. Gross margin is 80.1%, but the operating margin in the second quarter was -15.6% on revenue of $1.469bn.

What is left is cash

The bull case is the cash statement. Free cash flow rose 66% year over year in the quarter, to $294m, and reached $1.64bn over twelve months against a market value of $28.49bn. But bookings are recognized as revenue over time, so today's cash is the tail of yesterday's spending — the part already sold. The multiple has done the adjusting: 5.01x trailing sales and 3.98x forward, against roughly 21x a year ago. Consensus still models $7.16bn of 2026 revenue and no positive net income before 2030, estimates that predate the withdrawn guidance. Benchmark, cutting to a sell rating, called it "platform lifecycle decline".

The hours are contested, too. Epic Games' creator-built Fortnite islands took 47% of that game's playtime in May, up from 38% a year earlier. Roblox's answers are diversification — the top ten games now account for 20% of hours versus 30% three years ago — and demographics, with users aged 18 to 34 in the United States up 40% and a 50% payout premium offered to creators of adult content. Its advertising pivot, a partnership routing rewarded video and immersive ads through Google's ad systems, is real but unquantified in the company's own disclosure; the widely cited $1.2bn 2026 figure is a Morgan Stanley estimate, not a Roblox one.

Nobody else is falling

If this were engagement monetization de-rating as a category, the neighbors would show it. They don't. Unity, which sells both the game engine developers build on and the advertising network they use to acquire and monetize players, is up about 25% over twelve months and 61% in the past month. Its second quarter was Roblox's mirror image: strategic revenue of $486m, up 38%, with its Vector advertising platform crossing a $1bn annual run rate two quarters early and free cash flow of $202m. Electronic Arts rose 23% over the year, Take-Two 6%, NetEase was roughly flat. Roblox is alone.

Bilibili, the Chinese video community whose users skew young and whose content centers on animation, comics and games, is down about a third over the year — but for unrelated reasons and without a single violent session. Its problem is growth, not profitability: revenue rose 6.0% in the first quarter, to about $1.04bn, while gross margin reached 37.1% and 2025 became its first profitable year. Advertising grew 30%, offset by mobile games falling 12% against a strong prior-year title. Its games business is gated on Beijing's licence approvals, not American child-safety law. It trades at 34.4x trailing earnings that only just crossed zero, and reports second-quarter results on August 27.

The setup

Where it stands — Roblox is guided to its first bookings decline ever, with users, hours and spending per user all falling together. Would confirm — Third-quarter bookings land at or below the $1.58bn low end, with daily users below 123m. Would invalidate — Daily users and hours stabilize sequentially and bookings per user turns up from $12.66. Watch next — Roblox third-quarter results, late October; Bilibili second-quarter results on August 27. Valuation — Roblox at 5.01x trailing sales and 3.98x forward, against roughly 21x trailing a year ago; free cash flow yield 5.76%.

Okta reports fiscal second-quarter results after the close on 26 August, seven days from now, and management has already told the market what to expect: revenue of $790m-794m for the quarter ended 31 July. That is about 9% growth, a step down from the 11.2% it reported in the quarter before. The share price has moved in the opposite direction from the growth rate all year.

The login layer, priced per head

Okta sells the corporate front door — single sign-on, multi-factor authentication, a central user directory — plus Auth0, the developer-facing customer-identity business it bought in 2021. The moat is switching cost rather than technology: moving off Okta means re-plumbing authentication for every application and migrating a user directory with no downtime, a one-to-two-year project most chief information officers will not start. Its neutrality across clouds is what keeps it alive against Microsoft's Entra ID, which arrives bundled inside Microsoft 365 enterprise licences at roughly $6-9 per user per month and is the single largest threat to the franchise.

The first quarter was a good one. Revenue reached $765m and current remaining performance obligations rose 12% to $2.499bn, one point above reported growth — real, but not an inflection. Dollar-based net retention ticked from 106% to 107%. Newer products — identity governance, privileged access, and identity for autonomous AI agents — reached roughly a quarter of bookings and lift deal size about 40% when attached. Chief executive Todd McKinnon called demand for the agent products "bigger than anything we've ever seen". Margins agree: GAAP operating margin went from 5.7% to 7.3% year over year, and free cash flow is guided to $855-885m for the year.

What has not happened is reacceleration. Annual growth has fallen four years running — 42.9%, 21.8%, 15.3%, 11.8% — and consensus models 10.0% this year and 9.5% next. The re-rating happened anyway. Okta fetched 5.97 times trailing gross profit on 3 May and 10.29 times by 29 July; the forward figure is 9.47 times, and forward earnings 36.7 times against roughly 11% earnings growth. Of the 55% twelve-month gain, two sessions around the May print — up 30.1% and 13.4% — supply about 86%. The stock has gained 1.1% since 1 June and slipped 4.8% in the past month. Thirty-one analysts rate it buy, and their mean target of $120 sits 15% under the quote.

The one that got cheaper

SailPoint governs who may access what — the audit-and-entitlement layer above the login — and is the leader in that niche. Its first quarter put annual recurring revenue at $1.163bn, up 26%, with the SaaS portion up 36% and adjusted operating margin 3.3 points wider. Customers above $1m of recurring revenue grew 32%. The stock still fell 11.5% on 9 June, because full-year guidance stepped growth down to 21-22% from 28% delivered.

It is not levered — the term loans were repaid at the IPO — but Thoma Bravo still owns the large majority of it, which caps how much any re-rating can run. On about $11bn of market value, SailPoint trades near 9.4 times recurring revenue, against roughly 14 times at the February 2025 listing on $813m of it. The business is 43% bigger and the multiple a third smaller. Over twelve months the shares are down 1.6%; the 27% gain of the past month came across fifteen unremarkable sessions after it shipped a merged human-and-agent control plane, with second-quarter results due 9 September.

The third name isn't in this business

BlackBerry is grouped with them and sells no identity software. Its revenue is QNX, the real-time operating system embedded in cars, plus secure government communications; royalties recognise as vehicles ship, and the QNX backlog is near $950m. The turn is real — revenue up 25.6% and operating margin from 1.6% to 10.1% — but it comes off a base that shrank 29.5% two years ago, and at 43.9 times forward earnings it is dearer than Okta on every line. Its 131% year rests on three gap sessions, and the uptrend broke on 6 August.

The force underneath all three is a change in the billing unit. Per-seat identity is capped by headcount; governance, privileged access and agent identity are billed per identity, and machine identities already outnumber humans by dozens to one. That is why Palo Alto Networks closed its $25bn purchase of CyberArk in February. The prize is real. Okta's price now assumes it wins the prize; SailPoint's assumes it does not.

The setup

Where it stands — Okta has been re-rated about 70% on gross profit without a change in growth; SailPoint's multiple compressed while its recurring revenue grew 43%. Would confirm — Okta guiding fiscal Q3 revenue growth near 9% on 26 August with net retention flat at 107%. Would invalidate — Okta's cRPO growth stepping into the mid-teens, running several points clear of reported revenue. Watch next — Okta reports 26 August; SailPoint reports fiscal Q2 on 9 September with annual recurring revenue guided to $1.22bn. Valuation — Okta: 36.7x forward earnings, 9.47x forward gross profit, versus 5.97x gross profit on 3 May. SailPoint: 9.4x recurring revenue versus 14x at listing.

Range Resources, a Fort Worth company drilling the wet southwestern corner of Pennsylvania's Marcellus shale, does not sell the same product as its neighbors. About a third of what it lifts is ethane, propane, butane and condensate — molecules that leave the United States by ship rather than by pipeline. Its pre-hedge natural gas liquids (NGL) price in the June quarter was $29.10 a barrel, roughly $3.49 above the Mont Belvieu benchmark that normally sets Gulf Coast pricing, according to its second-quarter release. Management then raised full-year guidance to a $2.50 premium over that benchmark.

The premium is not a trading fluke. Waterborne ethane exports from the US averaged 658,000 barrels a day in the quarter, up 40% year over year, with a record near 750,000 barrels a day in June; propane and butane exports rose 30% to 2.6 million barrels a day. Dock capacity, not the domestic gas market, sets Range's marginal barrel, and its East Coast position prices into Europe. A further 360,000 barrels a day of liquefied petroleum gas export capacity is due in early 2027, alongside Range's own Repauno terminal capacity.

What the mix does to the income statement

Range's revenue grew 19.1% to $833.6m and operating income rose 73.9%, lifting operating margin to 39.1% from 26.8%. One caveat belongs beside that: net income fell 17.8% to $195.3m, because derivative marks below the operating line went the other way. The improvement is operational; the bottom line is not clean.

EQT, which produces from about 2.0 million gross acres and is the largest of the three at a $33.5bn market value, went the opposite way. Revenue fell 29.2% to $1.81bn and operating margin collapsed to 25.1% from 44.3%. Its execution was not the problem: it sold 634 billion cubic feet equivalent, above guidance, on capital spending of $666m, held unit operating costs to $1.03 per thousand cubic feet equivalent, and cut net debt to $5.5bn from $7.7bn at year-end. The Equitrans midstream acquisition shows up as that cost advantage rather than as a fee stream insulated from gas prices. Its one genuinely non-gas revenue mechanic — a 10-year, 325,000 dekatherm-a-day supply deal with CPV for a 2-gigawatt West Virginia data-center plant, priced off PJM power markets — does not start until as early as 2031.

Expand Energy, formed from the Chesapeake–Southwestern merger and holding the largest position in Louisiana's Haynesville shale that feeds Gulf Coast liquefied natural gas plants, realized $2.42 per thousand cubic feet, down 17.4%. Revenue fell 19.7% to $2.96bn. It carries the least debt relative to earnings, near 0.5 times, bought back $850m of stock in the quarter, and is adding a $1.25bn gas marketing business, Twin Eagle, to shave $0.15–$0.25 off a roughly $2.70 breakeven. It is also six months into a search for a permanent chief executive.

Who actually owns the winter

The suppressant is shared: working gas in storage reached 3,153 billion cubic feet in early August, 198 billion above the five-year average, and Appalachian basis widened, with the September TETCO M2 contract at a $0.95 discount. The December contract nonetheless trades above $4 per million British thermal units against a summer prompt near $2.80–$3.00, on an Energy Information Administration 2026 average estimate of about $3.67.

That premium accrues unevenly. Expand has hedged 41% of 2027 volumes. Range's 2027 book is 270,000 million British thermal units a day of swaps at $4.05 plus 80,000 in collars floored at $4.00 — about a fifth of gas volumes, struck above today's curve. EQT's disclosed 2027 position is summer collars alone, $3.00 puts against $4.51 calls. The most exposed name to a cold winter is the one whose quarter was worst.

All three fell over the three months to 19 August, bottomed on 21 July and have recovered roughly 8% since, without changing trend. Valuation does not follow the operating results. On trailing enterprise value to EBITDA, Expand is cheapest at 3.81 times, EQT sits at 6.30 and Range — the only one growing — is dearest at 7.33. Range is also the only one whose forward price/earnings multiple, 9.85 times, sits below its trailing 11.12, meaning analysts model earnings up; EQT's forward 12.66 against trailing 11.78 says the reverse. Free cash flow yields are within a point of each other, near 11–12%. The market has paid for part of Range's liquids premium and none of EQT's unhedged optionality.

The setup

Where it stands — Range's realizations and margins are improving on export-linked liquids while its shares track two shrinking dry-gas producers. Would confirm — Range holding an NGL premium at or above $2.50 a barrel over Mont Belvieu in the third quarter. Would invalidate — That premium turning negative as new export capacity slips, collapsing the realization gap toward EQT's $2.65. Watch next — Third-quarter results in late October, plus weekly storage prints against the 198 billion cubic foot surplus. Valuation — Range 7.33x trailing EV/EBITDA and 9.85x forward earnings against 11.12x trailing; EQT 6.30x and 12.66x forward.

Adeia told investors on 3 August that it now expects its semiconductor licensing arm to reach $200m of annual revenue, twice the target it had carried. That arm collected roughly $26m across all of 2025. It booked $48m in the first half of this year.

Adeia is a 150-employee San Jose company, spun out of Xperi, that manufactures nothing. It owns patents and bills two very different customer sets. One is media: pay-TV distributors, streamers and consumer-electronics makers pay a fee tied to subscribers and devices. The other is semiconductors, and it is the one that has changed.

What Adeia is actually paid for

The technology is hybrid bonding — joining two silicon dies directly, copper pad to copper pad, with no solder bump between them. Below roughly a ten-micron pad pitch, nothing else works. The density gap is the whole argument: about 14,000 signal connections per square millimeter against roughly 1,500 for conventional micro-bump stacking, at three times the interconnect energy efficiency. That is why high-bandwidth memory, the stacked DRAM sold alongside artificial-intelligence accelerators, is migrating to it. Samsung is procuring around 50 die-to-wafer hybrid bonding systems for a Pyeongtaek production line, with BESI as preferred tool supplier, applying the technique from 16-layer HBM4E before going fully hybrid at HBM5.

Adeia is positioned to tax that. A 2025 patent-influence ranking placed it first in foundational hybrid-bonding patents, ahead of TSMC, Samsung, Micron and IBM — TSMC holds more patents, Adeia the more foundational ones. Signed licensees include Samsung, SK Hynix, Micron, Kioxia, Sony, AMD and Nvidia. The holes are named and large: TSMC, which built its own stacking IP, and Intel, with Foveros Direct, are both unlicensed, and TSMC may simply never sign.

Underneath, the media base is eroding on purpose-built decline. Pay-TV is roughly 35-40% of revenue with subscribers falling more than 7% a year. Non-pay-TV recurring revenue grew 54% year on year in the June quarter and is now nearly double the pay-TV recurring line. Establishing the floor got harder: DISH's Chapter 11 filing on 30 June automatically stayed Adeia's April infringement suit in Colorado, and Adeia sued Fubo in Delaware on 1 July after failing to agree terms.

June-quarter revenue was $96.1m, up 12.1%. Operating margin slipped to 26.4% from 27.5% as enforcement spending — $30-35m budgeted for 2026 against roughly $7m a quarter in 2025 — absorbed the growth. Management reiterated full-year guidance of $395-435m while conceding it is pacing toward the low end. The shares fell 9% after hours on the print because revenue missed consensus by less than $1m.

The other licensor

InterDigital, a Wilmington, Delaware company that licenses cellular, Wi-Fi and video-codec patents to handset and streaming firms, gapped 16% on 30 July. Its June quarter contained $103.7m of catch-up revenue from an Amazon agreement whose final value binding arbitration will set over the next 18-24 months. Reported revenue still fell 13.4% year on year and operating margin dropped to 53.5% from 68.3%. Third-quarter guidance of $154-158m, struck on existing contracts, sits about 40% below the June quarter. Consensus has revenue falling 3.6% in 2026 and 5.3% in 2027. The recurring leg is genuinely growing — annualized recurring revenue hit a record $625.7m, up 13% — just far slower than the price. A jury trial against Disney over Disney+, Hulu and ESPN+ codec use opens 8 September in the Central District of California, after InterDigital won a second pan-European injunction on 23 July.

Where the prices sit

InterDigital's enterprise value is 18.4x trailing EBITDA, up from roughly 12x in May at a $6bn market capitalization; it is now $8.5bn. Its forward price/earnings of 38.1x exceeds trailing 28.2x, the arithmetic of falling earnings. Adeia's forward 18.8x sits under trailing 23.7x, EV/EBITDA is 12.6x, and free cash flow yield 6.1%. The counterweight: at $26.78 Adeia still trades 54% above its two-year average close of $17.36. The average of ten analyst targets is $33.66.

The trend has not endorsed either story. InterDigital's 50-day average has been below its 200-day since early May, through the entire 28% month; strip its two best sessions and the month is +5.3%. Adeia is +2.2% over 30 days, and −17% without its own two best days, having fallen 10.2% in three sessions to 19 August as semiconductor shares broadly retreated. The forward moat question is dated, not rhetorical: the foundational Ziptronix bonding patent expired 4 May 2026, leaving roughly 1,100 untested continuations to carry royalties into an HBM5 ramp arriving around 2029.

The setup

Where it stands — Adeia's semiconductor royalties are inflecting while its shares sit a fifth below the high; InterDigital's re-rating rests on non-recurring money. Would confirm — Adeia's second-half semiconductor revenue exceeding the $48m booked in the first half. Would invalidate — Full-year revenue landing below the $395m floor of guidance, or semiconductor royalties stalling near $15m a quarter. Watch next — InterDigital's jury trial against Disney opens 8 September; Adeia reports the September quarter in early November. Valuation — Adeia at 18.8x forward against 23.7x trailing; InterDigital at 38.1x forward against 28.2x trailing.

The layer nobody sees

Before a graphics processor can be packaged, the silicon has to be mounted on something. That something is a build-up substrate: a laminated carrier of resin film and copper that routes tens of thousands of connections out of the chip stack and into the board. AI packages are getting physically larger and adding memory stacks, which means more layers, bigger panels and lower yields — and the industry that supplies them has spent a decade being treated as a commodity laminate business.

It is not being treated that way now. Ibiden, the Japanese group founded in 1912 that makes package substrates in its Electronics division alongside a ceramics business supplying diesel particulate filters, grew revenue 26.4% in the June quarter to ¥123.2bn — its fourth consecutive acceleration. Operating margin reached 21.1%, from 18.1% a year earlier. On 4 August it lifted its net profit forecast for the year to March 2027 to ¥84bn from ¥58bn, ahead of the ¥73.3bn analysts expected.

The Taiwanese carrier makers moved further. Unimicron, the Taoyuan maker of printed circuit boards and flip-chip carriers, expanded gross margin by 11.6 points in four quarters to 24.7%, with operating income up 342%. Nan Ya PCB, the Nan Ya Plastics subsidiary that builds substrates for processors, graphics chips and memory modules, expanded gross margin by 15.2 points and swung from a small quarterly loss to NT$2.90bn of operating income. Kinsus, a smaller Taoyuan carrier specialist, grew revenue 30.7% with margin up five points.

Why the pricing is holding

The mechanism is unusually legible. Demand for flip-chip ball-grid-array (FCBGA) substrates exceeds available capacity by more than 50%. Ajinomoto, the Japanese food group that controls roughly 98% of the intellectual property behind the build-up film these carriers are made from, raised film prices 30% for the third quarter, and substrate prices themselves are up 5-10% for the second half. TrendForce models the market moving from rough balance this year to a 22% shortfall in 2027 and 29% in 2028, despite $19.1bn of announced expansion.

Everyone is spending into it. Ibiden's board approved about ¥500bn of electronics capital spending through March 2029, targeting 2.5x its current substrate capacity, with its Ono plant — which started AI server output in October 2025 — still using only half its building. Unimicron raised 2026 capex a third time, to NT$53.7bn, with 80-85% aimed at build-up substrate. Ibiden and Unimicron between them hold about 74% of the FCBGA market, but Ibiden's grip on the very highest end is loosening as Unimicron qualifies. One caution on pricing: Ibiden itself told analysts last October it was not optimistic about product prices, even while describing its high-end capacity as sold out.

What the shares already assume

Gross margins here range from 10% at LG Innotek to 36% at Ibiden, so price against gross profit is the comparable lens. On that basis Ibiden has gone from 7.9x trailing gross profit a year ago to 17.8x in February to 40.0x now. Unimicron went from 12.7x to 65.3x, Kinsus from 6.2x to 45.0x, and Nan Ya PCB stands at 100.8x — dearest in the group on every measure.

The prior cycle is the uncomfortable comparison. Unimicron's trailing gross profit of NT$27.3bn is 54% of the NT$50.4bn it earned in 2022, when the last substrate shortage pushed its gross margin to 35.9%. Nan Ya PCB's is 29% of its 2022 figure; Kinsus remains 37% below. Only Ibiden has cleared the old peak, at ¥137.7bn against ¥127.5bn. Consensus closes the gap by assuming acceleration from here: Unimicron revenue up 48.7% in 2027 and Nan Ya PCB up 61.4%, faster than either is currently reporting. Nan Ya PCB's 76x forward earnings only becomes 29x if that lands, and analyst targets for it span NT$875 to NT$2,444.

The one going the other way

AT&S, the Austrian board and substrate maker in Leoben and the smallest of the group at €6.1bn, has fallen 34% from its June high. In that stretch it swung from a €55.9m quarterly loss to €40.9m of profit, gross margin rising from 5.5% to 22.2%, and raised full-year revenue growth guidance to 45-55% with margin guidance lifted and Malaysian expansion covered by long-term commitments from AMD and one other customer. It trades at 14.7x forward gross profit.

The two Korean names show the re-rating is specific. Samsung Electro-Mechanics earns only about 22% of revenue from substrates, the rest from ceramic capacitors and optics, yet carries Ibiden-like multiples. LG Innotek, mostly camera modules, sits at 5.5x forward gross profit. Strip both out and the remaining five average a 636% twelve-month gain, against 614% for all seven.

Recent trading has been driven by index moves as much as by substrates. Six of the seven had their best session on 31 July, when the KOSPI rose 17.9%, its largest one-day gain on record. Remove each company's two best days from the past month and the group's 10.9% average gain becomes a 14.8% decline. Unimicron and Kinsus made fresh highs on 17-18 August; Ibiden is 25% below its June peak.

The setup

Where it stands — Substrate profits are inflecting hard, but multiples have expanded roughly five to seven times over twelve months against far smaller gross-profit growth. Would confirm — Unimicron and Nan Ya PCB posting third-quarter gross margins above 25%, extending the four-quarter climb. Would invalidate — Taiwanese monthly revenue flattening sequentially, or 2027 consensus growth of 49% and 61% being cut. Watch next — Monthly Taiwanese revenue disclosures in early September, then Ibiden's September-quarter results in late October. Valuation — Unimicron 65.3x trailing and 52.3x forward gross profit; Ibiden 40.0x and 33.4x; AT&S 20.1x and 14.7x.

Corning beat on both sales and earnings for its June quarter. Then it told investors to expect $4.9bn to $5.0bn of core sales in the September quarter. Wall Street had been modeling the top of that range. The shares fell 19.3% the next session, leading a rout across optical stocks and marking the company's worst day since March 2020.

Corning is a 175-year-old glass and materials maker that now sells four largely unrelated things: optical fiber, cable and connectors to telecom carriers and data-center operators; glass substrates for televisions and phone screens; ceramic emissions parts for vehicles; and polysilicon for solar. Only the first has anything to do with artificial intelligence, and the market spent the first half of 2026 pricing the whole company as though it were the only one.

The engine is accelerating

Optical Communications grew sales 32% year on year to $2.07bn in the June quarter, with the enterprise networks line — the part that sells into hyperscaler campuses — up 65% and generative-AI product sales close to double. Segment net income rose 77% to $438m, a record 21% margin, and management's description of the constraint was blunt: if it could make more, it could sell more.

The mechanism is content per graphics processing unit (GPU), and it is unusually concrete. Clusters above roughly 130,000 GPUs require a third optical layer, which adds about 50% more content. Each bandwidth step doubles fiber counts; the move from NVIDIA's Hopper generation to Blackwell took fiber per GPU from eight to sixteen. Should scale-up interconnect inside the rack shift from copper to optical, that figure could reach 160. Management expects 1.3x to 1.5x optical content per GPU by 2028 on the conservative path alone.

That is why three of the largest technology spenders have tied up capacity in advance. Meta committed up to $6bn in January, NVIDIA followed in May with capital attached — pre-funded warrants and a commitment to expand US optical connectivity capacity tenfold, including new plants in North Carolina and Texas — and Amazon signed a multiyear fiber agreement in June. Corning's internal Springboard plan was upgraded in May, not trimmed: a $20bn sales run-rate by end-2026, $30bn by 2028, $40bn by 2030.

The drag is the other half

The guidance that cost the stock a fifth of its value came from the consumer side. DRAM and NAND memory prices are inflating the bill of materials in a smartphone, and Corning now expects handset units to fall by a mid-teens percentage in 2026 — a demand problem it did not create and cannot fix. In the same quarter it folded Display and Specialty Materials into a single Glass Innovations segment and carved out a new Solar segment. Solar sales grew 90% but the segment lost $7m at the margin line on a $30m extended maintenance shutdown at its wafer plant. Automotive grew 2% into a market down 2%.

The honest caveat is that consolidated growth has slowed for four consecutive quarters, from 20.9% to 16.6%. But the mix improved: operating income grew 21.8% on that 16.6%, gross margin held at 36.1% against 36.0% a year earlier, and free cash flow was $1.42bn.

What the price now assumes

Because Lumentum and Applied Optoelectronics both carry losses that make price-to-earnings meaningless, the comparable lens across this group is price-to-gross-profit. Corning trades at 22.4x trailing and 19.7x forward, back below the 23.2x it commanded on 3 May and far under the roughly 35.7x implied at its 29 June closing peak of $255.69. Coherent, a $69bn maker of lasers and datacom optics whose June quarter grew 33.7% to $2.05bn, is cheaper at 18.9x forward. Lumentum, whose quarterly revenue doubled to $1.01bn before a $7.16bn charge, sits at 31.9x. Applied Optoelectronics, a $12bn transceiver supplier growing 86% but loss-making at a 27.7% gross margin, is at 41.3x.

On earnings, 48.9x forward is not cheap in absolute terms — though it is marginally below the roughly 52x premium that looked like the central risk in this name back in May. Consensus has not moved against it: 2027 revenue of $22.8bn and earnings of $4.30 a share imply 31.7% profit growth.

The trend signal only turned negative on 14 August, when the 50-day average crossed below the 200-day — six weeks after the fall that caused it. On 18 August the whole complex reset again after Fabrinet dropped sharply on a beat; Corning fell 7.7% while the broad market gave up 0.7%. It remains up roughly 148% over twelve months, and down only over the three-month window.

The setup

Where it stands — Corning's fiber segment is sold out and growing 32%, while its consumer-glass and solar halves are absorbing a memory-driven handset downturn. Would confirm — September-quarter core sales landing at or above the $5.0bn top of guidance, with optical margin holding near 21%. Would invalidate — Optical Communications growth falling below 20% year on year, or 2027 consensus earnings dropping from $4.30. Watch next — Third-quarter results in late October, plus any restatement of the $20bn end-2026 Springboard run-rate. Valuation — 19.7x forward price-to-gross-profit versus roughly 35.7x at the June peak; 48.9x forward earnings against Coherent's 42.0x.