DK Street Journal

The Trade Desk is paid a percentage of the advertising money that passes through its buying platform. In 2025 that percentage went up. The money passing through went up more slowly — and that, not fee compression, is what the last year of trading has been about.

The distinction matters because the company reports only its own take, never the gross. It discloses client spend on its platform once a year, in the annual report, defined as spend on inventory, data and services plus its own fee — $13.4bn for 2025 against $2.896bn of revenue, or 21.6 cents on the dollar, versus $12bn and $2.445bn in 2024, or 20.3 cents. Revenue grew 18.5% that year while the spend behind it grew about 12%. Chief executive Jeff Green has said the take rate has been flat for a decade, through Kokai, identity products and new artificial-intelligence tools. Either way, the meter that decelerated first was volume routed through the seat.

The routing, quarter by quarter

What followed is visible without any modeling. Year-on-year revenue growth went 17.7%, 14.3%, 11.8%, then 3.0% in the June 2026 quarter, to $715.1m. Operating margin fell to 14.2% from 16.8%, with operating income down 13% on revenue up 3%. Guidance for the September quarter is at least $650m — about a 12% decline — and roughly $160m of adjusted earnings before interest, taxes, depreciation and amortization. "We underperformed our own expectations for two main reasons: macro conditions have made it more difficult for some of the world's largest brands to grow, and we didn't execute as well as we could have," Green told investors on the August 6 call. Consumer packaged goods and automotive, about a quarter of platform spend, were the stated drag: "Autos and CPG have both been set back by tariffs and oil prices."

The competitive half is dated and specific. Amazon charges roughly a 1% fee on open-web publisher ads and nothing for guaranteed deals on its own media, against a take reported at around 20% for The Trade Desk; its demand-side platform went from under 10% to just under 20% of programmatic share in about fifteen months. Digiday's survey work found the share of brands and agencies saying they had used The Trade Desk in the prior year fell from 55% to 39% between early 2024 and late 2025 — a routing datapoint that preceded the revenue break. Omnicom and Publicis have both audited its fee structures. On 4 September, 575 jobs — 15% of staff — ended, at a severance cost of $39m to $51m.

The same impressions, a different meter

Magnite runs the independent sell-side platform publishers use to monetize inventory, and it is paid per impression it clears rather than as a share of the buyer's budget. In the June quarter its contribution after traffic-acquisition costs — the revenue it keeps — rose 17% to $189.6m, with the connected-television line up 36% to $97.1m and the open-internet display line up 2%. It raised full-year guidance to 13-14% growth on 5 August, the same fortnight The Trade Desk guided down. "CTV represented 51% of total Contribution ex-TAC in the quarter," chief executive Michael Barrett said on the call, attributing the growth to same-store gains at Netflix, Roku and Warner Bros. Discovery rather than a single renewal. Jounce Media's supply-path work puts Magnite's coverage of US streaming supply at 99% on a dollar-weighted basis.

That is the mechanism in one line: streaming supply is expanding faster than streaming budgets, so the price per thousand impressions is falling — Netflix's average fell from about $42 to roughly $31 after Amazon's ad tier launched. More impressions at less money each is neutral-to-good for a per-impression fee and corrosive to a percentage-of-spend fee. Programmatic demand is not shrinking; the dollars are being counted differently on each side of it.

What the prices now say

The Trade Desk's shares have traded with the 50-day average below the 200-day without a break since mid-November 2025; Magnite crossed back the other way in June and has roughly doubled off its March low. At 2.3x trailing sales against roughly 7.6x a year ago, The Trade Desk looks cheap against its own history — but forward earnings multiples sit above trailing ones, near 37x, because consensus has 2026 revenue falling about 5% and earnings per share at $0.39 against $0.91. The cheapness is the denominator moving, not the price. Magnite, at about 21x forward earnings, now commands a higher price against gross profit than The Trade Desk does, an inversion of how these two have traded for most of their listed lives.

The verdict the evidence supports is narrow and uncomfortable: the de-rating is earned by the business, but it is earned by lost routing volume, and the fix — winning spend back from a competitor charging a twentieth of the fee — is not something a cost cut delivers. Against that sits the one fact the bear case does not absorb: about $1.5bn of cash, a trailing free-cash-flow yield near 13%, and $78m of stock repurchased in the June quarter against a market value under $7bn.

On Monday, The Trade Desk leaves the S&P 500 and drops into the small-cap index, a little over a year after joining at $80.21. The November print will show the take. The number that would actually settle the argument — how many dollars still route through the platform — is disclosed once a year, and not until the annual report.

Ascendis Pharma, a Danish rare-disease company whose three marketed drugs all run on one chemistry, sold €315m of product in the June quarter — double a year earlier. The operating line moved with it: a margin of 18.4%, against minus 33.5% in the same quarter of 2025, the fifth consecutive quarterly improvement.

What is being tested is not whether the science works but what a launch curve is worth once it stops being a promise. The market pays roughly 11 times forward sales for Ascendis and about 57 times for Cytokinetics, the South San Francisco company whose oral cardiac myosin inhibitor MYQORZO went on sale in January — two versions of the same specialist-launch archetype, priced five-fold apart.

One chemistry, three meters

TransCon is a prodrug: a parent hormone tethered to an inert carrier that releases it slowly, converting a daily injection into a weekly or monthly one. Ascendis sells it three ways and is paid differently by each. Skytrofa, a weekly growth hormone, is roughly flat at €55m a quarter and management guided it to stay stable — it is neither funding the story nor spoiling it.

Yorvipath is the engine. Quarterly revenue has compounded rather than stepped — €143.1m, €187m, €197m and then €252m in the June quarter — on roughly 1,000 unique new U.S. patients a quarter, with the drug available in more than 35 countries. Hypoparathyroidism had never had a replacement hormone, so the meter is not prescriber breadth but whether patients stay: the company cites 82–86% sustained multicomponent response and 95% retention at five years, and spends on patient-access staff rather than sales reach. "The majority of the drop-off is during that titration period in terms of when patients experience the most amount of change and where additional education and a higher touch support model makes sense," Jay Wu, president of Ascendis U.S., told investors on the 13 August call. AstraZeneca's eneboparatide hit its Phase 3 primary endpoint this year but was not approved as of September; GlobalData models Yorvipath at $1.66bn of 2030 sales against $492m for the challenger.

The third meter started this quarter. Yuviwel, for pediatric achondroplasia, booked €8m in its first partial quarter with more than 220 patient enrollments and over 65% of them reimbursement-approved. On that same 13 August call, chief executive Jan Mikkelsen dismissed BioMarin's patent suit: "whatever has happened, it will not have any material impact on Ascendis pathway. I can guarantee that." Eighteen days later Ascendis settled, agreeing to pay BioMarin 20% of U.S. Yuviwel net sales retroactive to first commercial sale, and 18% in Europe, Brazil and South Korea, until May 2030. Every dollar of the newest launch now carries a burden the June accounts did not.

What the tape has and has not paid for

Ascendis is up 3.9% over ninety days and 21.8% over twelve months, against revenue that nearly doubled, and its 50-day average crossed above its 200-day on 14 September. The next session the shares fell 9.49%, from $264.92 to $239.77 on more than three times average volume, immediately after Ascendis disclosed that all licences granted to Novo Nordisk in metabolic and cardiovascular disease — including once-monthly TransCon semaglutide — revert to Ascendis with no continuing obligations either way. Management had described that programme as advancing four and a half weeks earlier. No source attributes the move, but it is the only dated company event in the window.

The wider group of small- and mid-cap biotechs flagged as acquisition candidates has fallen about 10% in thirty days, and the decline sits almost entirely in the names with no product to count: Dyne Therapeutics down 30.2% after a Novartis muscle-wasting failure read across to myotonic dystrophy developers on 8 September, MoonLake down 25.6%, Denali down 16.5%. The commercial-stage members average a loss of 1.7%.

The mirror at 57 times sales

Cytokinetics' launch is inflecting too. MYQORZO went from $4.8m to $25.3m of net product revenue between the March and June quarters, with new-to-brand share above 40% and more than 700 prescribers, over half of them first-time or low-volume writers. "Just 5 months into the commercial launch of MYQORZO, we are demonstrating increased velocity that's exceeding expectations," chief executive Robert Blum said on 6 August. The friction is measurable: about 2,000 prescriptions in the quarter converted to 1,500 patients dispensed, the gap being risk-evaluation certification and prior authorization stretching to weeks.

The difference is what it costs to get there. The quarterly operating loss widened to $175.9m, combined research and selling expense guidance was raised to $860–890m, and the launch is funded by equity — a May offering raised about $760m net. Consensus shows no net profit until 2029. When the ACACIA-HCM trial hit both primary endpoints in non-obstructive disease, roughly doubling the treatable population, the shares fell about 7% on the day the data were presented in Munich.

The verdict

Cytokinetics' slide is a valuation doing its job: paying 57 times forward sales for revenue that arrives after three more years of losses leaves nothing for good news to buy. Ascendis is the opposite case — the doubling, the margin turn, the guided €500m-plus of 2026 operating cash flow and a debt-free balance sheet are all earned, and the share price has credited almost none of it. What the last month legitimately took back is narrower than the chart suggests: a fifth of Yuviwel's U.S. economics now belongs to BioMarin, and the metabolic option Novo Nordisk was carrying came home unfunded.

The guarantee is the tell. Ascendis's chemistry does what the label says; what each dollar of it earns is decided by counterparties — a court, a partner, a payer — and two of the three moved in three weeks.

Copart agreed on 10 September to acquire ACV Auctions for $10.50 a share in cash, valuing the company at roughly $1.9bn, according to the merger announcement. The buyer runs salvage auctions, mostly for insurers disposing of totaled cars. The seller runs a digital marketplace where franchised and independent dealers sell used vehicles to one another, taking a fee per transaction plus inspection, transport and short-term financing, and never owning the car. The premium was 45.4%; ACV shares gapped from 7.11 to 10.41 in a single session on 114.9m shares, against about 1.1m the day before.

The reason the price is the less interesting number: the channel Copart is buying into is shrinking in the only unit that pays it. Neither ACV nor its rivals are paid for used-car prices. They are paid per car that changes hands, and cars are changing hands less often.

The dislocation that stops a car from selling

When wholesale values slide, sellers anchor to last month's price and buyers bid this month's. ACV management put a figure on the gap in the June quarter — roughly 600 basis points between what sellers wanted and what buyers would pay, which knocked 300 to 350 basis points off auction conversion rates. The market-wide count went with it: dealer wholesale volumes, per the National Auto Auction Association, were down about 6% year on year in June and down 8% in July, a figure ACV's incoming chief financial officer Timothy Fox cited directly on the 10 August call.

Prices are not compensating. The Manheim Used Vehicle Value Index stood at 208.2 in August, up 0.4% on the year and down 0.9% on the month.

ACV grew revenue without growing transactions

ACV sold 211,472 marketplace units in the June quarter, essentially flat against roughly 210,000 a year earlier. Revenue still rose 10.4% to $213.9m — all of it from price per unit and attach. Auction revenue per unit rose 6% to $554; the transport, capital and data businesses grew 17%, with transport up 19% on 125,000 deliveries. Gross profit grew 4.1%, gross margin fell to 49.88% from 52.90%, and the quarter produced a $6.0m operating loss.

"We delivered record adjusted EBITDA per unit, increasing 11% year over year in Q2," outgoing chief financial officer Bill Zerella told investors on 10 August. Per-unit economics were the story precisely because units were not. A $50m accelerated buyback was in progress that day, with management calling the shares undervalued. One month later the board agreed to sell at $10.50, or about 2.2x consensus revenue for the year.

The name that did not move

OPENLANE, the former KAR Auction Services, runs the other listed digital wholesale marketplace plus a floorplan lending arm for independent dealers. Its only listed competitor being consolidated should have been worth something. It rose 1.1% on deal day and closed 16 September a shade below its pre-announcement price. Its June quarter shows the same squeeze from the other side: revenue up 15.1% to $554.6m, gross margin down to 40.12% from 42.41%, operating income down 19.4%. It trades at 9.9x enterprise value to earnings before interest, tax, depreciation and amortization, against 13.0x for Copart.

Copart itself fell 6.65% on the announcement and remains 4.8% below its pre-deal close. Its own growth had stalled — fiscal 2026 revenue of $4.666bn was up 0.4%, net income down 4.4% — at 19.8x trailing and 18.9x forward earnings. Buying the dealer channel is what a salvage auctioneer does when its own lane stops widening.

What is actually healing is credit, not volume

CarMax, the largest used-car retailer and the one that also lends, cut gross profit per retail used unit by $230 to $2,177 in the quarter to 31 May, and comparable used-unit sales still fell 0.8%. It trades at 22.0x forward earnings against 37.0x trailing, on consensus earnings per share of $2.708 for the year to February — about 0.5% growth, and below the $3.21 it earned in fiscal 2025. Meanwhile subprime 60-day-plus auto delinquencies, a record 6.90% in January, improved to 5.67% by June.

That is the honest split. The consumer-credit side of used vehicles is repairing off a record-bad peak; the transaction count is not. Strip out ACV's takeover pop and the group has been falling — CarMax down 1.0% over thirty days, OPENLANE down 1.8%, Carvana down 12.2%. ACV's quote, pinned within three cents of the deal price every session since, now carries no information about wholesale volumes at all.

The tender is expected to close by year-end. The next hard count of used-vehicle units, rather than of dollars per unit, arrives when CarMax reports before the open on 29 September.

No power-market datapoint landed on 14 September. Vistra's shares fell 5.16% that session and NRG's 4.40%, against less than half a percent for the S&P 500, while Constellation Energy, the largest US nuclear operator, dropped 7.09% and Talen Energy 8.38% — the same day chip shares sold off on worries about the pace of artificial-intelligence spending. Over the preceding week Oracle lost 13.9% from its 8 September close and Nvidia 6.0%. What moved the electricity companies was sentiment about their customer, not a price in any power market.

That is the problem with owning them as an artificial-intelligence trade. Vistra runs roughly 38,700 megawatts of nuclear, gas, coal, solar and battery capacity and sells power to about 4.3 million retail customers across 20 states; NRG serves some six million retail customers under the Reliant, Direct Energy and Green Mountain brands alongside its own fleet. Neither is paid for load growth. Each is paid by an auction clearing at an administrative cap, by a Texas spot market flooded with batteries, and by hedges struck years before delivery. Both reaffirmed guidance this year. Both have lost about a third of their value over twelve months.

The price is capped at one end and compressed at the other

On 22 July, with Texas supply tight, power cleared at $57 a megawatt-hour; Vistra's management said it could have cleared $400 to $500 had batteries not been competing into the same hours. Across the Electric Reliability Council of Texas grid, the intraday spreads merchant gas plants monetize fell about 50% year on year by June and ancillary-service prices 45%, with federal projections putting Texas battery capacity at 37 GW by the end of 2027.

At the other end sits PJM Interconnection, the mid-Atlantic grid operator, whose capacity auction for the 2028/29 delivery year cleared at the $325 per megawatt-day price cap for a third consecutive auction — 2.5% below the prior year's $333.44 — while PJM's own simulation put the uncapped clearing price at $555 and the auction still left the system roughly 6.8 GW short of its reserve-margin target. Scarcity is being rationed by rule rather than paid to generators. NRG says less than half of about 2 GW of upgrades it could make to its PJM plants would be economic even at $555.

Vistra sold the recovery already

Vistra had sold forward about 94% of expected 2027 generation and 72% of 2028. Second-quarter adjusted EBITDA rose about 30% to $1.77bn with the generation segment up 68%, and the 2027 opportunity range of $7.4bn-$7.8bn first set in November 2025 still stands, now trending to its lower end as the Texas forward curve fell. Downside is bounded: the section 45U nuclear production tax credit phases out only above $25 a megawatt-hour and disappears at $43.75, converting weak prices into a federal transfer.

Its data-center business is contracted rather than announced: 2,609 MW to Meta from PJM nuclear plants with deliveries starting in late 2026, and 1,200 MW to Amazon Web Services from Comanche Peak beginning late 2027. After Governor Greg Abbott ordered an audit of every data center in the Texas interconnection queue on 3 August — some 474 GW of requests, about 90% of them data centers — chief executive Jim Burke said that "baseload projects, because they've been studied, you'd expect those to be moving forward," and that "2026 wholesale power prices being soft is not that big of a surprise" to the company.

NRG earns more when power is cheap

NRG's second-quarter adjusted EBITDA rose 34% to $1.2bn, but $370m of that was the first full quarter of the acquired LS Power fleet; adjusted earnings per share fell to $1.49 from $1.73 and gross margin to 14.5% from 16.5%. Texas EBITDA dropped $131m as round-the-clock Houston power averaged $33 a megawatt-hour against a $52 planning assumption. The offset is subscription income — Smart Home customers grew 8% to 2.45 million — and the fact that roughly half of the Texas margin base is a retail spread that widens as wholesale prices fall. 2026 guidance of $5.325bn-$5.825bn was reaffirmed.

The growth story is later and less certain. Its 1.2 GW, $3.2bn hyperscaler plant has aligned commercial terms but no final investment decision, with commercial operation targeted for late 2029; more than 95% of its free cash flow would come from fixed capacity payments, and the company's leverage target has slipped to 2029 to fund it. "This is the model for how large load growth should work," said Robert Gaudette, its chief executive. "The customer supports the investment, with reliability and affordability protected for all." NRG closed at $105.99 on 15 September, a fresh 52-week low.

What the de-rating actually removed

Vistra trades near 10x trailing enterprise value to EBITDA and NRG marginally above it, against nearly 15x for Constellation. Vistra's reported earnings are distorted by hedge marks — 2025 diluted earnings were $2.21 a share against $7.00 in 2024 — which is why 16.5x forward earnings sits against 23.7x trailing; NRG is the cheapest of the three on forward earnings at 12.1x.

So the businesses did not deteriorate; the premium did. Both guidance sets held through the twelve months in which the shares lost a third, and the 14 September session shows the pair moving with Oracle and Nvidia rather than with any megawatt-hour. But the two are not one claim. Vistra has already sold 2027, so a firmer power curve cannot reach its margin before 2028, while contracted nuclear deliveries and the tax-credit floor limit how far the downside goes. NRG has been marked down for a price move that helps half its Texas book, and its compensation for the other half does not begin until late 2029. Meanwhile GE Vernova, which sells the turbines this same demand requires, is up 44.9% over twelve months: investors are paying for the equipment and marking down the electrons.

The first Meta megawatt-hours are due to leave Perry and Davis-Besse before this year is out. That delivery, not an interconnection queue, is the test.

SailPoint, whose Identity Security Cloud decides which employees, contractors and machines may open which applications, finally put a dollar figure on agent identity. On its September 9 results call the company said artificial-intelligence products now carry more than $70m of annual recurring revenue, better than 30% of the net new ARR it booked in the quarter, against a $100m target for this fiscal year and $800m by fiscal 2029.

That matters because identity software is sold on two incompatible meters. Okta, which runs corporate single sign-on and owns the Auth0 login service developers embed in their own apps, bills Workforce Identity per employee per month — the meter that shrinks if software agents replace headcount. SailPoint charges per identity governed, and sells machine identities, non-employee identities and agent identities as distinct governed types inside one platform. Which meter the agent era pays is the difference between identity as a steady low-double-digit business and something faster.

The meter that counts non-humans

SailPoint's booked numbers behave the way the per-identity theory predicts. ARR reached $1.231bn, up 25%, with the SaaS portion at $847m and up 36%. Net revenue retention was 113%, and average ARR per SaaS customer rose 17% to more than $400,000 on 16% growth in SaaS customers — existing accounts buying more per customer, not merely more customers. Remaining performance obligations reached $1.9bn, up 30%, which management attributed to larger deals rather than longer contracts. Full-year ARR guidance went up to $1.38bn.

The demand story management tells is regulatory as much as technical. "The EU AI Act contains specific human oversight and audit logging requirements, which are scheduled to become enforceable as early as 2027," chief executive Mark McClain told investors on September 9. "We believe regulators will increasingly ask who is accountable for this agent and its actions." The pending Entro Security acquisition, expected to close this quarter, extends discovery to more than 1,200 non-human identity types.

What the income statement says instead

None of that shows up cleanly in reported results. Revenue was $308.8m, up 16.8%, a deceleration from 21.6% the prior quarter, because 97% of net new ARR landed as SaaS: each $5m shifted out of term licence costs roughly $10m of recognised revenue, about a $5m headwind in the quarter alone. Gross margin fell to 59.0% from 67.3% — purchase-accounting amortisation and SaaS mix — so gross profit of $182.1m grew 2.4% while revenue grew nearly seven times that. The company remains GAAP loss-making, with a $57.3m operating loss.

Okta's problem is the opposite. Its per-seat Workforce half was 59% of annual contract value and grew 11%; the per-monthly-active-user and per-token Customer Identity half was 41% and grew 13%. Auth0's plans meter monthly active users and include only 1,000 machine-to-machine tokens a month before extra charges — exactly the counter agents should inflate — and the mix still did not move. Net retention is 107%; customers above $100,000 of contract value grew 6%. Chief financial officer Brett Tighe, on the August 26 call, called agent revenue "Still immaterial. Still very small. We're very early innings." What did move was profit: current remaining performance obligations rose 14% and total RPO 17% against 10.6% revenue growth, GAAP operating margin doubled to 13.3%, and $125m of stock was repurchased.

What the price has paid for

Okta has more than doubled in twelve months, gapping 33.7% the day after August results in a sector-wide cyber rally that also gave CrowdStrike its best session ever, then adding 12% on September 14 on a broad identity and AI-security bid with no company news. Its price-to-trailing-gross-profit has gone from 5.97x in early May to 13.06x, with forward gross profit at 12.45x offering almost no relief and forward earnings at 48x. SailPoint, up 44% over ninety days but still down 6% over twelve months, trades at 15.10x trailing gross profit — dearer than Okta on the measure that normalises for their 59% and 80% gross margins, and marginally cheaper on sales at 9.70x against 10.20x. No forward consensus is published for it.

So the honest split: Okta's re-rating is earned by operating leverage and one quarter of faster bookings, and management has guided the bookings acceleration to end immediately, with third-quarter current RPO growth back to 11-12%. Nothing in its disclosure yet shows agents paying. SailPoint has the evidence on the meter that should catch them first, and none of it in its reported gross profit. Chief executive Todd McKinnon described one customer evaluation where "we detected 50 instances of a Claude agent in the environment… we came back a few weeks and there was 1,500 Claude agents."

At Okta, agent single sign-on is still bundled into the standard product with a consumption cap built in and switched off — the meter exists in the software and not yet on the invoice. SailPoint is already sending the bill; what it has not yet shown is that anyone recognises the revenue this year.

An Alcoa executive spent part of 10 September explaining to investors the exact circumstances under which American smelters would stop making money without the price of aluminum falling at all. Molly Beerman, the company's chief financial officer, said that if Washington extends tariff relief beyond Canada to partners such as Japan, South Korea or Europe and the last million tonnes of imports is covered, "then you can expect the Midwest premium to reduce in response to essentially wipe out the tariff benefit." Century Aluminum, the listed company whose profit most nearly is that benefit, fell 15.8% over the five sessions to 16 September, more than twice Alcoa's decline.

The stake is a printed number due in weeks. Century guided third-quarter adjusted earnings before interest, tax, depreciation and amortization to $325–345m on an assumed US Midwest premium of $1.09 a pound. On 20 August the September contract slumped 8.2% to 95 cents after reports that Washington would halve the 50% tariff on Canadian metal. Those talks collapsed and the 50% rate stayed, but the premium never went back: duty-paid metal fetched more than $2,850 a tonne in early July and roughly $2,094 by late August. The metal itself did not move with it.

Where the dollar comes from

Century runs primary aluminum smelters in the United States and Iceland and owns a carbon-anode plant in the Netherlands. It realizes the London price plus the regional premium and pays for alumina, delivered power and anodes — three inputs set elsewhere. When both revenue numbers rose together in the June quarter, everything followed: revenue of $752.1m, up 19.7% year on year, gross margin of 30.3% against 5.8% a year earlier, and adjusted EBITDA of $327m on shipments of 130,632 tonnes. Price alone added $95m sequentially.

"Today, both plants are producing at full capacity into a market that needs every unit we can produce," chief executive Jesse Gary told investors on the 6 August call, after the Mt. Holly expansion completed in June and Grundartangi's second line restarted six months early. The cost stack is guided to worsen into the September quarter — $10–15m of summer energy, $5m of raw materials and $20–25m of hedge settlements against a $15–25m volume and mix tailwind. The offset is alumina, the biggest single input, which Beerman expects to stay oversupplied into 2027.

Century changed hands at 6.6x trailing and 4.0x forward earnings, against 140x in early May when the shares were priced for the tariff to hold forever. Alcoa — which produces around 900,000 tonnes a year in Canada and is paying more than $1bn in Section 232 duties to bring most of it south, the mirror of Century's position — trades at 9.5x trailing and 7.1x forward. Consensus still has Century earning $10.07 a share this year against 42 cents last, rising again in 2027.

Kaiser's month was one day

Kaiser Aluminum rolls, extrudes and draws mill products for aerospace, packaging and automotive customers, buying metal on contractual pass-through terms and getting paid per pound of conversion. It has no direct claim on the premium, and its drawdown is not about one. Second-quarter sales were a record $1.26bn, up 53%, net income more than quadrupled, and packaging conversion revenue rose 34% to $174m. Of its 17.3% loss over the month, 13.2% came on 18 August, the session it named Amcor's Fred Stephan to succeed Keith Harvey as chief executive from 1 November. Strip that day and the shares are down 4.8%. At 10.9x both trailing and forward earnings the market pays nothing for this year's step-up, and consensus puts 2027 earnings 17% lower.

Nor is aluminum alone in the selling: Carpenter Technology fell 22.9% and ATI 18.7% over the same 30 days while copper producers were flat to higher, and the dollar firmed as traders priced a Federal Reserve rate rise into 3.4% inflation.

The verdict

Part of Century's fall is honestly earned — the thing it sells did get cheaper, and a quarter budgeted on $1.09 a pound will not print $1.09. What nothing yet explains is the remainder: four consecutive quarters of widening gross margin, cash that exceeded total debt for the first time in July, and a forward multiple of four while the estimate under it still rises. The caveat that belongs beside that multiple is that the earnings are a spread between two administered numbers, and four times a policy decision is not four times an ordinary earnings stream. Kaiser's case is unrelated: record conversion economics, modelled as a peak, plus a leadership handover.

Century's own hedge against premium risk is more policy. A 20 July executive order lets companies investing in US primary capacity import approved volumes at 25% rather than 50%, and Century expects clearance for 300,000 tonnes a year from 2027 tied to its Oklahoma project — "quite material" to cash flow, Gary called it. The insurance against one signature is another signature.

Edgewater Research told clients on 14 September that Monolithic Power Systems' supplier allocation on Nvidia's next-generation accelerator platform was at risk, and the shares fell 7.39%. The session was ugly for chips generally — Nvidia fell 3.36% and a broad semiconductor exchange-traded fund 4.75% after Anthropic's Dario Amodei published an essay urging AI developers to slow down, with Elon Musk and Sam Altman echoing him — but Monolithic fell roughly half again as far as the index, and the excess belongs to the note.

The note hit the one thing Monolithic's income statement cannot show. The company sells power-management semiconductors — the controllers and modules that step rack voltage down to the sub-1-volt rails an accelerator's processor runs on — and it wins those designs years ahead, then gets paid per unit on a platform the customer can re-architect or second-source at the next generation. Content per accelerator is a decision made annually, at somebody else's desk. And since 7 July, a company eight times Monolithic's market value has been inside the same socket: Analog Devices closed a $1.5bn all-cash purchase of Empower Semiconductor, whose integrated voltage regulators put conversion in the processor package.

The incumbent is growing faster than ever

Monolithic's June quarter was the best in its history. Revenue rose 47.6% year on year to $980.6m, operating income rose 84.4%, and the operating margin went from 24.8% a year earlier to 31.0%. The enterprise-data end market — the accelerator business — produced $380.6m, up 164.3% year on year and 38.8% of the company. Management raised its full-year floor for enterprise-data growth from 85% to 130%. Founder-chief executive Michael Hsing told investors the company had become "the highest power density company in the world", with solutions for "not only top-of-rack switches but TPUs, NIC cards, and other things in the rack."

One line did not move: gross margin has sat between 55.1% and 55.4% in all eight reported quarters. The AI mix has brought volume, and it has not brought price.

The challenger is buying the socket

Analog Devices sells tens of thousands of catalog parts into industrial, automotive and communications customers, which is why it was supposed to be the analog-cycle name here. It isn't. In the quarter reported on 19 August, revenue rose 39.6% to $4.02bn, gross margin reached 67.3% from 63.1% a year earlier, and operating income nearly doubled. Communications grew 84%, with data center now about 80% of that segment, and data-center optical and data-center power each more than doubled as the industry moves to 800-volt distribution. Chief executive Vincent Roche called energy "the most persistent constraint to scaling next-generation systems". Management called intermediate-to-core conversion for 6,000-amp processors one of the fastest and largest growing analog opportunities, and said book-to-bill was "not unusually elevated" with customers running lean — so the channel restocking that usually explains an analog upswing is not the explanation here. The fourth quarter is guided to roughly $4.3bn.

Analog Devices was already among the main suppliers of voltage regulator modules — Monolithic's territory — before it bought Empower.

What fell, and what didn't

Over thirty days the drawdowns scale with AI-power content rather than with analog exposure: Vicor down 21.6%, Monolithic 18.5%, Analog Devices 7.2%, Texas Instruments 5.8%. Texas Instruments is the honest cycle control, and it is the one name of the group still in an uptrend, with revenue growth accelerating to 22.8% and operating margin at 42.3%. Monolithic's trend rolled over hardest and earliest, turning down in the first week of September; it closed Monday 32.4% below its 52-week high.

On forward gross profit, Monolithic is the most expensive of the four at 24.5x, against Texas Instruments at 18.8x, Vicor at 25.7x on consensus that has its revenue growing 55.6% next year, and Analog Devices the cheapest at 17.7x. Monolithic's trailing multiple of gross profit is 31.1x, down from 48.7x in May and 35.4x in mid-August. This is a de-rating that started four months before the Edgewater note.

The verdict

Monolithic earns the growth; what nothing in the reported numbers earns is the premium. A flat gross margin plus revenue concentrated in one accelerator program is a volume business dressed as a franchise, and the compression from 48.7x to 31.1x is the market pricing a risk the income statement will only ever show after the fact. Monolithic's management has said it has no concentrated customers, and Nvidia's published partner list for high-voltage racks names ten suppliers rather than one — the allocation risk is a claim about next generation, not a disclosed fact about this one.

Analog Devices is the harder one to square. Its business accelerated for four straight quarters, its margins expanded, it guided up, it is the cheapest of the group on gross profit, and it spent $1.5bn to move up the same power tree — and the shares turned down on 10 September anyway. If the market is marking down AI power content, it has marked down the company attacking the socket less than the one holding it, which is at least internally consistent, and it has marked both down while their customer's shares rose.

Vicor held the accelerator power socket inside Nvidia's H100 before Monolithic took it away, and now monetizes the patents behind that architecture, collecting $30.4m of royalties in the June quarter. It is the standing proof that this socket changes hands — and a fair picture of what a company is left with afterward.

Borr Drilling, which owns a fleet of modern jack-up rigs and rents them with crews and equipment to national oil companies and independents from Mexico to Southeast Asia, kept those rigs running 98.4% of their scheduled time in the June quarter and still earned almost nothing. Revenue of $232.3m produced an operating margin of 0.13%, down from 18.6% in the March quarter.

That is the whole argument for owning or avoiding a jackup contractor, compressed into one line of an income statement. The revenue is arithmetic — rigs contracted, times day rate, times the share of days actually paid for — and the cost per rig-day barely falls when a rig idles. Utilization held; price and cost did the damage. And the damage now meets a fixed bill. Borr refinanced substantially all its debt during the quarter, issuing $1.1bn of senior secured notes at 8.75% due 2032, $935m at 9.00% due 2034 and $300m of 3.5% convertibles, while extending its revolver to $250m. Near-term maturity risk is gone; a hard cash schedule replaced it. "The new notes amortized at 5% per annum, equating to $101.75 million on a full year basis," chief financial officer Magnus Vaaler told investors on the August 12 call. That starts in July 2027, against $473.6m of cash and undrawn revolver today. The refinancing also carried a $176.3m debt extinguishment charge, the largest piece of the quarter's $241.4m net loss.

One buyer sets the price

The shallow-water price cycle has one author. When Saudi Aramco suspended rig contracts in 2024 it released supply into a market that could not absorb it, and global jackup day rates fell about 17% through 2025 as more rigs chased fewer jobs. The rebound that was supposed to arrive this year did not: Westwood's rig data shows working utilization of the marketed Middle East jackup fleet falling from 83% in early February to around 69% by late April as hostilities across the Arabian Gulf delayed contract starts. Borr told investors that first-half additions to Middle East backlog were the lowest in more than 25 years. Meanwhile 22 of the 37 suspended rigs have mobilized out of Saudi Arabia to West Africa, South America and Asia — the very markets Borr sells into. The overhang did not disappear; it moved next door.

The counterintuitive part is what a crude rally does to this business. The paper has treated the Hormuz disruption as a cost shock rather than a revenue event for oilfield equipment owners; Brent has since risen above $107 a barrel after Saudi Arabia shut its East-West pipeline following drone attacks, and the reading holds harder. Borr is not paid for oil. It is paid a fixed day rate, and $7.3m of the quarter's sequential EBITDA decline was higher conflict-related insurance and fuel. "The region still has substantial underlying demand, which was close to materializing prior to the onset of the conflict, and we believe this delayed activity should reenter the market once conditions stabilize," chief executive Bruno Morand said on the same call, guiding to roughly 23 active rigs in the third quarter and EBITDA improving significantly from the June quarter's $43.8m.

The deepwater mirror

Noble, the Texas-based contractor whose earnings come mostly from ultra-deepwater drillships, is also in a trough — June-quarter revenue of $719.7m was down 15.2% and produced a $36.7m net loss, and full-year EBITDA guidance was cut to $850–925m, largely on two rigs in Brazil. But its forward book is firming: backlog of $6.8bn, leading-edge Tier-1 drillship rates in the mid-$400,000s a day, and utilization of its five ultra-harsh-environment North Sea jackups up to 80% from 66%. Chief executive Robert Eifler told the July call that 77 rig-years of ultra-deepwater work were contracted industry-wide in the first half, the most in over a decade.

What the market is actually marking down

The two trade at nearly the same trailing enterprise value to EBITDA — about 9.3x for Noble, 9.7x for Borr — which makes them look like the same bet. They are not. Borr's multiple sits on an EBITDA base consensus expects to fall 57.7% next year, to $183.4m, on flat revenue; the collapse is entirely margin. Its trailing free cash flow yield is minus 11.6%, against a positive 4% at Noble. And in 2027, the recovery year, consensus turns $427.8m of Borr EBITDA into $34.5m of net income, 8%, because depreciation and that 9% coupon consume the rest. Noble converts a third of its EBITDA at the same point.

So the six-month split — Borr down about 19%, Noble down roughly 5% — is earned on the shallow-water side and provisional on the deep. Borr's de-rating is the market pricing a day-rate cycle plus a leveraged balance sheet, and the business supports it. Noble is being discounted for 2026 white space in front of a backlog that is still growing, which is a calendar problem if the 2027 inflection arrives and something worse if it does not.

Borr has 73% of this year's rig-days sold at an average of roughly $134,000. Aramco's next jackup tenders are due at the end of September, from a country that just lost its main route around the Strait of Hormuz.

dLocal moved $17.7bn through its emerging-market payment rails in the June quarter, nearly double a year earlier, and earned less on each of those dollars than at any point in its public life. The Montevideo-based company, which collects and disburses money in Brazil, Argentina, Mexico, Nigeria and dozens of other markets for ride-hailing, streaming, e-commerce and remittance platforms, reported record gross profit of $127m. It also reported a gross margin of 31.8%, down 6.7 percentage points from a year earlier.

That gap is the whole story of what a payments company actually sells. Neither dLocal nor Shift4 is paid for transactions; both are paid a spread on the value that passes through them, and the spread is the only line that tells you whether winning more volume makes the owner richer. On that meter the two are diverging — and the divergence, not the volume growth, is what the market has been repricing.

The take that keeps shrinking

dLocal's gross profit per dollar of total payment volume has fallen from 1.16% in the June quarter of 2024 to 1.07% in 2025 to 0.718% now — a one-third decline in a single year. The annual pattern is older and unbroken: gross margin was 53.4% in 2021 and 36.8% last year. The mechanism is mix. Local-to-local flows, where dLocal never converts currency, reached 61% of volume, up six points; the cross-border business it is displacing carries the foreign-exchange spread on converting a merchant's collections out of reais or pesos, which is where much of the old margin lived. Volume is also concentrating in a handful of enormous customers who price accordingly.

"Were you to back out that one very large ride-hailing merchants mix gains at a lower take rate, take rate would have been relatively flat sequentially," chief executive Pedro Arnt told investors on the August 13 call. Revenue rose 55.8% to $400m; operating income rose 12.2%. Management raised full-year volume growth guidance to 60–70% and gross profit growth to 25–30%, but left operating profit growth at 27.5–32.5%. At the Goldman Sachs technology conference on September 10, Arnt said take rates are falling across the payments industry and faster in emerging markets, which started higher — a deceleration, not a reversal.

The spread that widened, and the debt that ate it

Shift4, the Allentown acquirer that sells card acceptance and its SkyTab point-of-sale system to restaurants, hotels and stadiums, has the opposite problem. Its reported revenue of $1.295bn is not what it earns — strip the interchange and network fees it passes to the card networks and $624m remains. On $61bn of end-to-end volume it earned a blended 65 basis points, or 65 cents per $100 processed, against 63 cents on $50bn a year earlier, with a mix shift toward smaller merchants offsetting enterprise pricing pressure.

What broke was below the spread. Gross profit rose 52.5%; operating income rose 14.3%; net income fell 35.3% to $22m. Organic growth in revenue less network fees was 11%, meaning most of the headline came from acquisitions — chiefly Global Blue, the $2.5bn tax-free-shopping refund business that takes a commission and a currency conversion on tourist spending, a demand curve with nothing in common with an American restaurant. Interest expense more than doubled to $65m on $4.52bn of debt, with net leverage at 3.7x against a self-imposed 3.75x cap. Guidance came down roughly $40m at the midpoint on that interest and $20m more on currency, to non-GAAP earnings of $5.15–5.35 a share.

"We can grow meaningfully without adding a single new customer," chief executive David Lauber said on the August 6 call, describing the conversion of merchants already running Shift4 software onto its own processing rails. The market's answer came the same week: the shares fell 22.9% in two sessions. They are down 48% over twelve months, and the recent 7% bounce starts from that floor. dLocal, by contrast, sits 6% below its 52-week high.

What each spread is worth

Shift4 trades at 2.3x forward gross profit and 9.1x trailing earnings before interest, taxes, depreciation and amortization, with a trailing free-cash-flow yield of 12.5%; dLocal at 8.1x forward gross profit and 13.8x EBITDA. For the same unit of economics — a dollar kept after the cost of moving money — dLocal costs three and a half times more. Shift4's 7.1x forward earnings rests on a consensus estimate of $6.36, above management's own guidance; against the company's midpoint it is nearer 8.6x. Analysts still average a $60.86 target while cutting it, Truist to $46 and Goldman Sachs to $49.

The verdict splits cleanly. Shift4's de-rating is earned, but by the balance sheet rather than the business: the spread widened, the volume grew 22%, and leverage taken on to buy a tourist-refund business converted all of it into interest. dLocal's re-rating is the harder one to defend — the dollars are real and rising, the guidance was raised twice, but every incremental dollar arrives thinner, and the market is paying a premium multiple for a company whose unit economics have halved since 2021.

dLocal's fastest volume growth since 2022 bought it 28.6% more gross profit. Whatever the next doubling buys is the number that settles this.

Snowflake used its September 2 earnings call to confirm it has agreed to acquire Observe, a company that builds monitoring software on the open Apache Iceberg standard rather than a proprietary data store. The price is reported at about $1bn, with roughly one percentage point of Snowflake's raised full-year product revenue forecast attributed to it.

That puts Snowflake — which sells credits consumed per second of virtual-warehouse compute, on a cloud data platform used for analytics and data sharing — into the market where Datadog meters a bill per monitored host, per gigabyte of logs ingested and per application trace. Both companies bill for volume. Both just reported four straight quarters of accelerating volume. Over six weeks the market has priced them apart: Datadog's share price against its trailing gross profit fell from 30.4x in late July to 25.0x, while Snowflake's rose from 27.7x to 31.3x.

What Snowflake bought the re-rating with

Product revenue reached $1.49bn in the July quarter, up 37% and accelerating for a third consecutive quarter. Net revenue retention was 126%, and remaining performance obligations reached $9.0bn, up 30%, with about 54% expected to convert inside twelve months. Management raised full-year product revenue guidance to $6.07bn from $5.84bn in May, and the shares rose more than 22% after hours. "AI is bringing new workloads onto the platform," chief executive Sridhar Ramaswamy told investors on September 2, describing first-party products Cortex Code and CoWork as a second leg and "AI activation" lifting overall platform consumption as a third.

The volume arrives at a cost. Snowflake cut full-year non-GAAP product gross margin guidance to 74% because the fast-growing AI workloads carry lower contribution margins today — the first reversal of a gross margin trend that had improved every year since fiscal 2022. The company remains unprofitable on a reported basis, with a $191.7m quarterly net loss, narrowed from $298.0m a year earlier.

What Datadog lost the re-rating with

On August 6 Datadog reported $1.12bn of revenue, up 35.6%, the fourth straight quarter of acceleration from 28.4% and a record $115m sequential addition. Then it disclosed that a significant AI customer using 17 of its products had renewed with a usage reduction starting in the third quarter. Guidance for that quarter implies 28-29% growth, and the shares fell about 19% in the session, with analysts widely assuming the account is OpenAI.

"If you back out our largest customer from our growth, you get pretty much the same growth rate as the rest of the business has been accelerating very steadily," chief executive Olivier Pomel said on the call. The supporting number is real: revenue from non-AI customers grew in the high 20s, a fifth consecutive quarter of acceleration off that base. But Datadog discloses no figure for what its AI customers or its largest account contribute. The most granular concentration data it gives is that 4,720 customers spending above $100,000 a year generate 91% of annual recurring revenue. The single relationship now setting the share price is not quantified in any public percentage.

The verdict

Datadog trades at 25.0x trailing and 22.1x forward gross profit; Snowflake at 31.3x and 26.9x, on a 67% gross margin against Datadog's 79%, and with reported operating losses. The market is paying more for consumption that management has told it will come at a lower margin than for consumption that has been growing faster and is disclosed as concentrated. MongoDB's 13.6% fall on September 2, after beating every estimate and raising guidance, says the judgment across metered-database software is about forward consumption rather than reported results.

Datadog's discount has a named, prospective cause, which is why it is not a pure mispricing — but nothing in the disclosure explains why the accelerating remainder of the base earns no credit at all. Snowflake's premium rests on guidance and a backlog it has yet to convert, and on a margin line it has already guided down.

Datadog's largest customer renewed and then consumed less. The next thing arriving in that market is a competitor that already holds the customer's data.