DK Street Journal

Circle issued more stablecoin dollars in the third quarter than it did a year earlier and made less money on them.

Reserve income, the interest Circle collects on the short-dated Treasuries and cash backing USDC, its dollar-pegged token, fell 8% to $658m as the return rate on those reserves dropped 61 basis points to 3.5%. Average USDC in circulation over the quarter grew 8%, to $73.3bn. Period-end circulation of $73.9bn was flat against a year ago.

The GENIUS Act, the federal stablecoin regime, takes effect in January 2027, and this layer has been bought on the premise that regulation turns it into financial infrastructure whose economics compound with adoption. The quarter tests that premise at the unit level. What a dollar of USDC earns Circle is set by the policy rate and then split with the exchange that distributes it. What a dollar of loan volume earns Figure Technology Solutions, which runs a blockchain-recorded marketplace for consumer lending out of Reno, Nevada, is set in a contract, and that number has not moved.

Coinbase took about half the interest

Of $409m in distribution and transaction costs, $331m went to Coinbase, roughly half of all reserve income earned in the quarter. Coinbase receives all of the reserve income on USDC held on its own platform and half the residual on USDC circulating elsewhere, under terms renewed through 2029. On its own October call Coinbase said it holds about 30% of USDC on-platform and captures roughly half the economics, describing the arrangement as "coopetition". Circle's distribution bill rises with its distributors' success.

That leg nonetheless held up. Net reserve margin rose 70 basis points year on year to 37.8%, slipping 80 points sequentially on the Hyperliquid arrangement, under which Coinbase passes about 90% of the reserve income onward. Management said the fourth quarter absorbs a full quarter of both Hyperliquid and a renewed Binance deal, and guided margins lower.

Combine the two disclosures and Circle kept roughly 1.3 cents a year on each dollar of USDC, against about 1.5 cents a year earlier, a figure derived from the reported rate and margin rather than disclosed. The rate did that, not the distribution terms.

The $182m that made the quarter a beat

Headline revenue of $864.9m was up 16.9% and earnings of $0.84 a share doubled the $0.42 consensus, but $182m of the $207m in other revenue was a one-time presale of tokens for Arc, Circle's new chain. Full-year other revenue is guided to $310–330m, of which core products account for $130–150m: about 4–5% of the $3.01bn consensus top line. Circle also disclosed a 5% workforce reduction, roughly 55 positions and $9.1m of restructuring expense, in a footnote rather than in its commentary. JP Morgan's Mike Colonnese cut his target to $100 from $104, Needham's John Todaro to $103 from $134.

Chief executive Jeremy Allaire was explicit that he is not defending revenue per dollar moved. "When this achieves scale, it's going to have enormous velocity and the unit costs are going to be much, much lower," he told investors on 4 November. "We are not shooting to let's go repeat the margin structure of the existing payments industry." On-chain USDC transaction volume grew 49% to $13.9 trillion in the quarter.

At $67.88 Circle trades at 58.5x this year's consensus earnings of $1.16, above its 36.5x trailing: consensus has reported earnings falling. On trailing gross profit it is at 24.2x against Coinbase's 8.1x, three times its closest listed peer after a 74% fall from its high.

Figure's take rate stayed inside the band it guided

Figure's third-quarter marketplace volume was $5.12bn, up 107% year on year and 20% sequentially, at the high end of its own $4.8–5.2bn range, before any contribution from Kiavi. The per-unit economics did not narrow as volume compounded: net take rate was 3.6% in the second quarter, inside the 3.5–4.0% guided range. What changed is the mix. Ecosystem and technology fees became the largest contributor to adjusted net revenue for the first time, and partner-branded gain-on-sale revenue grew 35.0% while partner-branded volume rose 151%, because partners are moving onto fee-based terms. Gross margin has fallen from 93.5% to 80.1% over four quarters on that shift, with revenue close to double.

The caveats are in the balance sheet. Figure bought Kiavi on 1 September for about $590m net of cash, funded with $600m of 8.5% senior notes, and trailing free cash flow runs deeply negative because loans held for sale consume cash before they are sold. Figure trades at 19.8x this year's consensus earnings against 25.8x trailing, with consensus earnings growth of 74%, and at 15.5x trailing enterprise value to EBITDA, the measure that now has to carry the notes.

Coin prices and a dead bill, not unit economics

Circle's results account for part of its mark-down: the core line is shrinking on a float that stopped growing, and the beat came from a token sale that does not recur. For Figure they account for none of it. The shares closed at an all-time low of $20.90 on 5 November, 69% below their high, with volume accelerating and results a week away.

The better explanation is the asset class and the statute book. Bitcoin fell about 6% to $75,700 and Ether to $2,247 on 5 November, triggering $1.09bn of liquidations, and the CLARITY Act, the market-structure bill meant to re-rate this layer, failed a Senate procedural vote 49 to 50 on 15 September. Over 30 days Circle fell 15.5% and Figure 18.8%, both less than Coinbase's 24.8%, and the levered coin-treasury vehicles in the same segment averaged 21.7%: the operating businesses were marked down at least as hard as the holding companies.

The rate input, meanwhile, has turned. The Federal Reserve raised its target range 25 basis points to 3.75–4.00% on 16 September, its first increase since 2023, and held there in October with two governors dissenting for more. Circle's 3.5% reserve rate is last year's cuts washing through a short book. Its constraint is share: USDC's slice of the stablecoin market has fallen from 34.9% to 23.1% in two years while Tether's USDT sits at $183.4bn.

Figure reports on 12 November, with Kiavi's first partial month and the interest on the notes that bought it both landing in the same statement. The line to read is the one Circle cannot set for itself.

The two companies that do most of the world's outsourced chip packaging and test reported the best quarters in their histories inside four days in late October. By the first week of November their shares had lost roughly a third of their value, and the explanation sits in election returns rather than in either income statement.

The shelf was supposed to work the other way. The standing belief about outsourced assembly and test is that the rent from the AI buildout's tightest physical step stays upstream with Taiwan Semiconductor Manufacturing's in-house chip-on-wafer-on-substrate lines, while the listed packagers get paid out of smartphone and automotive volumes at thin, mainstream rates. Both September quarters say otherwise. Gross profit per revenue dollar rose at each, the advanced work is now a disclosed dollar line, and the compression the theory predicts shows up in only one place: contract manufacturing.

The repricing arrived from the ballot box

The 3 November midterms produced a Democratic sweep of Congress, and voters in at least ten states backed restrictions on data-center expansion over electricity costs and land use. Two days later the Nasdaq Composite fell a record 1,421.52 points, or 5.52%, with the Philadelphia Semiconductor Index down more than 12% as the buildout became a political variable rather than an engineering one.

Almost half the month's decline landed in that single session: ASE Technology's American depositary receipts fell 20.1% and Amkor 14.5%, and both rebounded the next day. Over the 30 days to 6 November Amkor was down 38.5% and ASE 32.4%. The cleanest check that this was not a verdict on packaging is United Microelectronics, a mature-node foundry with roughly $300m of AI-related revenue in 2026 and no advanced-packaging socket at all: it fell 40.3% over the same stretch, more than either packager.

Amkor's phone book shrank as its computing book set a record

Amkor, which sells wafer bumping, flip-chip, fan-out and wirebond assembly and test to fabless designers and foundries from Tempe, Arizona, grew third-quarter revenue just 0.18% year-on-year to $1,990.6m, after 25.6% growth the quarter before. Gross profit rose 27.8% on that flat top line. Margin reached 18.26% against 14.32% a year earlier, a fourth consecutive quarterly expansion and above the 13.99% the company earned for all of 2025.

The mix did it. Computing — data center, infrastructure, storage — hit a record 26% of net sales against 19% a year earlier, while Communications fell to 37% from 51%. Advanced products, meaning flip chip, memory and wafer-level processing with related test, were $1,632m of the total, about 82%. The stalled top line is the phone side: AI data centers are absorbing memory, and IDC now expects global smartphone shipments to fall 16.7% in 2026.

"Customers are not only planning further ahead, but are also making long-term commitments and providing advance payments tied to future capacity requirements," chief executive Kevin Engel told investors on 26 October. Roughly $3bn of such prepayments have arrived since 2025, including $1.5bn from Nvidia for the Peoria, Arizona campus, which carries a finalized $407m CHIPS and Science Act award and begins volume production in early 2028. It earns nothing before then, and consensus has 2027 earnings per share at $2.76 against $2.53 this year — 9.1% growth, the depreciation eating the leverage. Amkor costs 13.1x trailing earnings and 6.58x trailing gross profit, against 10.8x gross profit on 21 August.

ASE put a number on AI packaging and raised it mid-call

ASE, the largest outsourced assembly-and-test provider, posted record consolidated revenue of TWD 244.03bn, up 44.8%, with growth accelerating through the year from 17.4% in the first quarter. Its assembly-and-test arm earned a 30.6% gross margin, eight points better than a year earlier, and delivered 91% of group operating profit on 62% of revenue. The electronic manufacturing services arm is where input costs bite: its margin slipped 0.6 points to 8.3% on component pass-throughs, for 9% of group operating profit.

"The so-called structural margin ceiling has now become the base for us," chief financial officer Joseph Tung said on 29 October. Management lifted 2026 leading-edge advanced packaging revenue guidance to over $4bn from $3.7bn during the call — roughly 15% of expected group sales — disclosed $300m of full-process chip-on-wafer-on-substrate work this year ramping above $1bn in 2027, and guided co-packaged optics past $0.5bn next year. Quotes for advanced packaging went up more than 20% in July.

The bill is cash. Capital spending was raised to $11.5bn, about 43% of expected revenue; net debt stands at 66% of equity and trailing free cash flow is negative 3.1% of market value. At 30.0x trailing earnings, down from roughly 42x in August, the cheap reading rests on 2027 consensus, which implies 60.5% net income growth.

Why the overflow reaches them at all

Neither company won these accounts from TSMC, which still earns 66.41% gross margin and says customers prefer its own flow. The packagers' share is set by TSMC's shortfall: non-TSMC players are estimated to add capacity equivalent to 110,000 advanced-packaging wafers a month by end-2028, and TSMC has transferred work to ASE and Amkor rather than see orders reach Samsung. Tool orders corroborate the ramp: ASMPT booked a record $1.13bn in the September quarter, up 138%, at a 1.69x book-to-bill, including thermocompression bonders ordered by packaging partners.

So the quarters account for none of the fall, and the two names were never one tape. Amkor is down 6.6% over twelve months and has been in a downtrend since 10 August, because 37% of its sales sit in a phone market losing units; ASE is up 105.8% and held an uptrend until 4 November, because its assembly arm is sold out and priced up. What the election changed is the terminal demand both are building into, and that is a question about permits and power, not about packaging yields.

Amkor's Arizona phase one completes late 2027 on a two-phase program management now sizes at $12bn. Every ton of that concrete was poured on the assumption that someone will be allowed to plug the finished chips in.

Ouster shipped more than 27,000 sensors in its third quarter, the most in its history, and collected less gross profit from them than from the 17,000 it shipped three months earlier. Gross profit fell 5.3% sequentially, to $25.3m from $26.7m, on 59% more units out the door.

The lidar maker's unit count has been the headline number in every quarterly release, and the "physical AI" case for sensor vendors is that deployment volume compounds into dollars. This quarter the count and the dollars came apart. Ouster disclosed over 7,000 lidar units against more than 9,000 in the prior quarter, while its Stereolabs ZED stereo cameras went to 20,000 from more than 8,000. Dividing reported revenue by those company-supplied counts gives roughly $2,197 of revenue per sensor, down about a third sequentially, and roughly $935 of gross profit per sensor, down about 40%.

The extra 10,000 units were the cheap ones

Revenue itself was fine: $59.3m, up 50.1% year on year and 9% sequentially, with product revenue a record $56m and a fifteenth consecutive quarter of growth. Gross margin printed 42.6% against 48.9% in the prior quarter, but that comparison is distorted by a one-time $5m tariff refund; excluding it, the earlier quarter was about 39.7%, so the underlying rate improved. The deterioration sits in the dollars each unit carries, which is a mix outcome rather than a pricing one.

The lidar line is where the company is placing its bet. "As the world's first native color LiDAR, REV8 is a paradigm shift that is proving to be an undeniable competitive differentiator," chief executive Angus Pacala told investors on November 3. Chief financial officer Kenneth Gianella said demand is "continuing to see that demand rotate more towards REV8." Against Hesai and RoboSense, which supply geometric lidar data and which held, in Hesai's case, 44% of China's long-range advanced driver-assistance lidar market in June 2026, Ouster's durable edge is regulatory: compliance with the National Defense Authorization Act and Buy-American rules bars its Chinese rivals from federally funded US projects. Third-quarter wins ran through the Utah, Washington, Florida and Ontario departments of transportation, plus defense customers Forterra, Seneca and GOQ.

Margin upside is going into tooling

Management holds a long-term gross-margin target of 35-40%, says it is currently beating that band, and is spending the difference on retooled lines and precision automation under a Benchmark partnership meant to support more than 100,000 lidar units of annual capacity by the end of 2027. No volume commitment underwrites that. The adjusted EBITDA loss narrowed about 60% year on year to roughly -$3m, the reported operating loss was $19.3m, and cash stood at $421m including about $191m raised in a July equity offering. The shares carry 22.8x trailing price-to-gross-profit, more than twice Cognex's 10.2x.

Cognex earned more per dollar of sales and got marked down

Cognex, which sells barcode readers and integrated vision systems to logistics, electronics and automotive manufacturers, ran the reverse arithmetic. Revenue of $313.4m rose 13.2% to a company record, gross margin expanded 312 basis points to 70.8%, gross profit rose 18.4% and operating income rose 74.5% to $100.8m. Full-year guidance went up, to revenue of $1.14-1.16bn and adjusted earnings of $1.68-1.72 a share. "At the midpoint of our guidance, we expect approximately 93% flow-through of incremental revenue to adjusted EBITDA, compared with roughly 70% in 2025," chief financial officer Dennis Fehr said on November 5.

The stock fell 18.3% across the two sessions into that print. Contemporaneous coverage put the company-specific reaction at about -6% on a slight revenue miss; the remainder landed in the November 5 washout, when the Philadelphia Semiconductor Index fell 12.2%, the session the paper has already described as repricing data-center demand expectations. Trailing price-to-gross-profit is now 10.2x against 13.4x in May, on a materially larger gross-profit base. The cautions are internal: automotive was flat year-to-date on European weakness, with Europe revenue down 15%, and management guided 2027 incremental flow-through down to about 60%.

A unit count has stopped predicting a gross-profit line

One quarter of results explains almost none of what happened to these shares, and the per-sensor mechanism that was supposed to bind them explains less. Ouster's dollars per unit fell because the marginal sensor is a camera rather than a lidar; Cognex does not disclose units at all and grew gross profit by price, mix and margin execution. Ceva, which licenses digital-signal-processor and connectivity designs and is paid a royalty on every device shipped, shows the same squeeze one layer down: licensee shipments of 567m units rose 16% in the June quarter while royalty revenue rose 1%, about 1.9 cents a device from 2.2 cents. It trades at 7.6x trailing gross profit, the cheapest of the three, and reports on November 9.

What survives is three separate businesses billing three different ways, all of them paid out of factory, logistics, infrastructure and defense budgets that have nothing to do with hyperscaler capital spending.

Ouster is building for a lidar volume nobody has promised it: lines sized for more than 100,000 units a year by the end of 2027. Until the orders arrive, the sensor it ships most often is the one that earns it least.

Between late October and the first week of November the four listed operators of outsourced customer contact and back-office work all reported, and the quarters pulled them apart. ExlService raised full-year guidance for the second time this year. Concentrix, which staffs and runs customer-contact operations plus front- and back-office automation for consumer-electronics, e-commerce, health-insurance and banking clients, guided its November quarter down. All four shares are higher over the month, each by double digits.

For two years these companies have been read as the clearest casualties of agentic artificial intelligence, and the month's advance was read as the turn: deflection arriving as revenue through contracts that charge for a resolved problem instead of an agent's hour. The filings carry the first half of that and almost none of the second. Work is moving between providers. The unit they bill has barely changed.

A seat at ExlService bills near twice a contact seat

ExlService, which sells insurance and healthcare analytics, claims processing, premium audit and payment accuracy, grew September-quarter revenue 19.7% to $634.0m, a third consecutive acceleration, with gross profit up 18.2%. Organic headcount rose 11% against 17% organic revenue growth; employee cost falling as a share of revenue is the mechanism management names. Divide the quarter's revenue by its 65,000 staff and annualize and the result is roughly $39,000 per employee, a derived figure, against about $21,600 on the same arithmetic at Concentrix and its payroll of 455,000.

"Data and AI-led revenue grew 36% on a reported basis and 31% organically in the third quarter, now representing almost 2/3 of EXL's total revenue," chairman and chief executive Rohit Kapoor said on the 28 October call. Digital Operations revenue fell 1.7% by design as AI was embedded in work the company already ran. Guidance went to $2.430bn-$2.440bn with adjusted earnings of $2.34-$2.36 a share. It is also the dearest name here, at 7.24x trailing gross profit and 6.95x forward, against 3.52x for Genpact, the nearest listed peer.

Concentrix's shrinkage was ordered by two clients

Concentrix's revenue declined 1.2%, or 0.5% in constant currency, after 5.4% growth in the February quarter. The November quarter is guided to $2.41bn-$2.46bn, a 3% to 5% constant-currency fall, because two hyperscaler clients are completing a reduction in support coverage, pulled forward from 2027. Beneath that sits a roughly three-point annual headwind from clients moving seats to cheaper countries. Gross margin held at 34.6%, and non-GAAP operating margin widened 30 basis points to 12.6%.

The mix story is real and small. "This quarter, we crossed the mark that 50% of our revenue now comes from businesses that we generated in the last 3 years," chief executive Christopher Caldwell told investors on 29 September, putting that pool's growth near 30% for the year. The software meter inside it, iX Suite licensing, carries about $120m of annual recurring revenue, some 1.2% of the $9.85bn consensus expects for fiscal 2026. Retention across the whole client base is 98% and the top 25 clients average more than 16 years, so capability is not the problem; price is. The company also acquired CastleHill Managed Risk Solutions for governance, third-party risk and AI-governance work.

Its shares rose 15.7% in the 22 October session, the largest single day on the list, with no earnings in the window; a dividend raise and that acquisition were the only datable news, so positioning is the likelier reading. At 0.57x trailing gross profit, a 31% trailing free-cash-flow yield and 2.74x forward earnings, the advance is closing a valuation gap. Trailing price-to-earnings is unusable: a $1.05bn non-cash goodwill impairment, triggered by the share price's own trading range, followed a $1.48bn write-down a year earlier.

TaskUs grew the line that costs most to deliver

TaskUs, which provides digital customer support, trust-and-safety moderation and data labelling for model training, grew revenue 5.4% to a record $314.7m while gross profit fell 9.2% to $108.9m, costing it 553 basis points of margin. "The current growth that we are seeing in AI services seems to be heavy on the onshore side, which comes at lower margins than what we make in our offshore business," chief financial officer Rishabh Khemka said on 4 November. AI services reached $75.0m, up 27.7%; trust and safety fell 15.8%. The largest client dropped to 17% of revenue from 27% on its own automation. Management's answer to the cost line is conventional price increases. The stock trades at 6.8x forward earnings.

ibex was handed the calls it automated

ibex, which runs customer care and technical support across voice, chat and SMS for banking, healthtech, retail and travel clients, is the clean case. "By having all contacts immediately reaching an AI agent, we were able to reduce abandoned calls by more than 60% without the need for our client to add additional headcount," chief executive Robert Dechant said on 5 November. The fibre provider in question then awarded ibex its traditional call-centre work. Revenue was a record $175.2m, up 16%, adjusted EBITDA rose 29%, and guidance went to $720m-$735m. Onshore revenue, the high-margin end, grew 37% and reached 29% of the mix. The shares cost 13.4x forward earnings against 13.5x trailing, which is no growth premium at all.

Only Genpact has moved off the headcount meter

Two of the four earned their re-ratings with results. ExlService because analytics and data engineering bill at nearly twice a contact seat, ibex because a six-week deployment bought it an incumbent's volume. TaskUs's advance rests on a guidance raise and a cheap multiple while the income statement deteriorates. Concentrix's rests on cash: $218m of adjusted free cash flow in the quarter and $550m or more of annual gross debt paydown targeted through 2027, against a top line consensus has flat to 2029.

The pricing migration the whole thesis needs has one listed example on the shelf's edge. "Non-FTE revenue exceeded 55% of total revenue in Q3, and our revenue growth continues to decouple from headcount," Genpact chief executive BK Kalra said on 5 November, with gross margin at 38.0% against 36.4% a year earlier. Everest Group describes the new billing units, per journey, per bot-hour, customer-experience-as-a-service, as beginning to emerge. Emerging is the right word: ExlService puts its own payment-integrity shift a couple of years out and calls the timing "unclear and unsettled".

So in this corner of services AI has become a question of which provider holds the account when volumes move, while the hour remains the thing counted on the invoice. The companies whose revenue per employee has pulled away are the ones that found something else to sell, and the next three quarters will show whether anyone can charge for a resolved problem at scale.

Three software companies that bill against something physical rather than a logged-in user reported within eight weeks of each other, and the profit each extracted from near-identical growth went three ways. Descartes Systems, which sells customs-filing, route-optimization and sanctioned-party screening software to carriers, brokers and freight forwarders, grew revenue 11.8% in its July quarter while operating income rose 38.5%. Bentley Systems, whose MicroStation and OpenRoads applications design roads, rail, bridges and power distribution, grew September-quarter revenue 12.2% and operating income 45.1%. Manhattan Associates, which runs warehouse and omni-channel execution software for retailers and third-party logistics operators, grew revenue 11.0% and operating income 6.8%.

They reported into a market that had just decided seat-priced software was the problem. ServiceNow, the workflow-automation vendor, beat on subscription revenue and raised its artificial-intelligence booking target again, then fell 25.9% from close to close on 29 October after guiding fourth-quarter contracted revenue to 18.5% constant-currency growth; commentary tied the fall to the fear that AI agents shrink demand for software sold by the seat. On a reported basis its operating income fell 32.7% to $385m on revenue up 20.5%, margin down to 9.4% from 16.8%; the company's own non-GAAP operating margin was 32.5%.

The shelf went up without checking the meter

The 13-name specialized-enterprise software group these three sit in rose as a block: the median member is up 22% over 30 days and 11 of 13 are up double digits, against a median twelve-month change of -3.8%. But the meter does not sort the winners. PAR Technology, in restaurant software, leads at 41% and Procore, which sells construction project management, at 39%. DocuSign, priced per user and per envelope, rose 13.2%. PTC, whose asset-design billing most resembles Bentley's, rose 0.8%.

The timing is worse for the mechanical story. Descartes ran from $80.43 on 8 October to $94.00 on 22 October with no company disclosure in the window; its last report was 10 September, and the discoverable events were a Loop Capital reiteration at $96 on 8 October and BMO Capital Markets lifting its target to $90 on 13 October. Manhattan gained 19% between 9 and 22 October, after a Morgan Stanley upgrade, and added about 4% after its 27 October results. Bentley's shares turned up on 19 October, two weeks before it printed.

What Descartes is paid for when volumes fall

The unit Descartes bills is the obligation, not the box. Chief executive Edward Ryan told investors on the 10 September call that overall shipment volumes fell in the quarter, with trucking down 4% year on year, and named rapidly changing tariffs, stepped-up customs and export-control enforcement and a heavier audit burden as the growth drivers. Services revenue was $188.6m, 94% of the total and up 13%, with licence revenue at $0.1m, so nothing is diluting the recurring line. Organic services growth, though, was just north of 9%: roughly four points came from about $250m of acquisitions, and revenue growth has decelerated from 16.2% two quarters ago. At 13.4x trailing gross profit, against 11.3x in mid-September, and 40.2x forward earnings, the market is paying now for leverage already delivered.

Bentley's margin, and the engineers it cannot hire

Bentley's annualized recurring revenue reached $1,582.8m, growing 13% in constant currency, with dollar-based net retention at 110% and operating margin at 29.1% against 22.5%. Executive chair Gregory Bentley called the retention figure "a historical high of 110%" on the 5 November call. It is not AI revenue: chief executive Nicholas Cumins said design-focused monetization starts in 2027 and will not be material that year, and that "Asset Analytics is already monetizing AI, Q3 was a great quarter for Asset Analytics," naming tower-inspection and road-monitoring deals with existing accounts. The demand behind it is a physical bottleneck: of 12–16 GW of data-center capacity projected for 2026 delivery, only about 5 GW is under construction, and nearly half of US engineers are 50 or older. The 13 October divestiture of Cohesive, its IBM Maximo consulting arm, lifts the mix to roughly 95% subscriptions and costs about $12m of fourth-quarter services revenue. Bentley is the cheapest of the three at 9.3x forward gross profit, and 31.8x forward earnings against 43.3x trailing.

Manhattan is buying the transition

Manhattan's cloud subscription revenue hit $134.4m, up 28%, exceeding services revenue for the first time, with contracted revenue yet to be recognized at $2.59bn, up 25%. Services grew 1%, and the maintenance annuity trailing retired perpetual licences fell 12.9% to $30.5m. New logos were only about 35% of new cloud bookings, so the growth is existing accounts buying more. Chief executive Eric Clark told investors on 27 October: "We are maintaining our 100% conversion success from AI pilot to AI subscription." What it has not yet bought back is margin — 26.5% against 27.5% a year earlier, though up from 22.2% in the June quarter. At 23.1x trailing gross profit, from 19.6x seven weeks earlier, and 45.3x forward earnings, it is the dearest of the three on the lowest gross margin, 56.9%.

Compounding, and what the re-rating actually bought

The results account for the direction and almost none of the timing. Descartes and Bentley widen margin because an extra filing or an extra consumption unit rides a platform already built; Manhattan is funding a conversion in which services revenue grows 1% while cloud capacity has to be sold, so its incremental dollar costs more. That split is a pay-structure difference inside one shelf, and the shelf rose without it. The sharpest test sits on the other side: after its crash ServiceNow trades at 8.9x forward gross profit, below Descartes at 12.5x and far below Manhattan at 22.7x. Whatever October paid for, it was not a premium for non-seat meters earned on numbers.

Descartes' own analysis of the overturned tariffs expects brokers to file more post-summary corrections, not fewer. The obligation is symmetrical: taking a duty away generates the same paperwork as putting one on, which is the one thing in this chain that no agent can delete.

Applied Digital, which builds data-center shells and leases them to artificial-intelligence computing tenants, spent more servicing its debt last quarter than it collected in rent. Interest expense reached $77.4m in the three months to August 31, up from $8.0m a year earlier, against $65.8m of base rent. Group revenue more than quadrupled to $341.9m, but $183.5m of that was low-margin tenant fit-out construction, which is why reported gross margin fell to 15.3% from 42.5% two quarters earlier.

That inversion is what the signed-lease story leaves out. Fifteen-year leases with creditworthy tenants have been read as annuities — megawatts contracted, rent fixed at signing, financing a detail of the plumbing. The sequence actually runs the other way: interest accrues from the day capital is drawn, rent starts only when a hall is energized, and just 250 of Applied Digital's 1,410 contracted megawatts were in service at the quarter's close.

The window between the draw and the rent is about to widen

"Over the next 12 months, we expect to place over 600 megawatts into service versus 250 megawatts over the past 12 months," chairman and chief executive Wes Cummins told investors on October 7. At the company's guided capex of $11m to $13m per megawatt, that step-up has to be funded before a dollar of it bills. It is being funded off the common equity line: $562.5m drawn from Macquarie Asset Management under a preferred-equity facility of up to $5.0bn, with a further $337.5m after quarter end, alongside $1.59bn of project notes placed at 7%. "Our first Polaris Forge 1 related notes placed last year carry a 9.25% coupon. This tranche priced at 7%," chief financial officer Mohammad Saidal L. Mohmand said on the same call, crediting anchor tenant CoreWeave's credit quality.

CoreWeave's own 2032 bond has since yielded above 16%, and convertibles, the channel these developers lean on, had their worst October since 2008 at -6.4%. The repricing of data-center risk after the midterms, already visible in power equipment, lands here as a cost of future capital.

Cipher had to buy its yield back

Cipher Mining, a bitcoin miner converting its powered Texas sites into leased AI capacity, disclosed on November 2 that cost per leasable information-technology megawatt at Barber Lake has risen to $13m from the $10.4m it originally underwrote, after $409m of tenant change orders — and that it put in additional capital to restore yield on cost to roughly 15% from the 13% the project would otherwise have earned. The rent is not movable: the lease was extended to a 20-year base term and about $9.1bn of contracted revenue. Of $6.7bn of debt, $4.5bn is non-recourse at the project level, and management says every contracted project is funded through completion. "We are not a neocloud," chief executive Tyler Page said. "Lenders and investors are placing a premium on predictable, contracted cash flows." Reported revenue fell 62.8% to $26.7m because mining was switched off at Black Pearl to hand the building over.

HIVE owns the chips the landlords refuse to buy

HIVE Digital Technologies, a miner that also sells graphics-processor cloud capacity, has $600m of contracted off-take but borrowed about $129m at a 9.0% fixed rate at 80% loan-to-cost to buy a $165m cluster itself. Revenue grew 69.1% to $77.1m at a gross margin of -37.3%, growth having decelerated from 285% four quarters earlier, and diluted shares rose 38.1%.

How much of the fall the numbers justify

For two of the three, they do. Cipher trades at 20.4x trailing sales on revenue that is falling by design; the bull anchor is an enterprise value near $6bn against roughly $864m of projected average annualized net operating income, derived from its own disclosures. HIVE's 1.12x book is a book of depreciating hardware earning a negative margin.

Applied Digital is the odd one. Revenue rose 322%, adjusted earnings before interest, taxes, depreciation and amortization went from $0.5m to $64.4m, and its project coupon tightened 225 basis points, while the shares fell 32.6% over the 21 sessions to November 6. At 27.2x trailing gross profit against 60.74x in late August, the decline has finally done the valuation work a 43% fall three months ago did not. The single line that genuinely deteriorated is the financing line, and it is the one that decides whether the next 600 megawatts earn their rent.

The nearest dated test belongs to the smallest of them: HIVE reports on November 9, with the call the following morning. It is the one company here whose hardware, not its landlord's balance sheet, has to hold its value.

The change in rack power architecture that is meant to multiply the dollars of power silicon sitting beside every accelerator has so far paid its loudest advocates almost nothing. Monolithic Power Systems, which designs the direct-current-to-direct-current conversion chips that step voltages down for processors, booked $506.7m of Enterprise Data revenue in its September quarter, and by its own account none of that is 800-volt content: its gallium-nitride parts are sampling, with revenue dated to late 2027 and a material ramp to 2028.

The move from 12-volt to 48-volt in-rack distribution and now to 800-volt direct current was read as the reason to own the gallium-nitride specialists and to fear for the incumbent's position. What the last two prints show is the reverse order of payment. The incumbent's dollars are accelerating at today's architecture; the architectural pure-play is shipping single-digit millions; and the credible threat to the final conversion stage is being bought and litigated rather than engineered in 800 volts.

The socket that pays today

Monolithic Power's Enterprise Data line ran $262.8m, $380.6m and $506.7m through the first three quarters of 2026, reaching 42.1% of revenue in the September quarter, up 164.6% year on year — roughly a $2.0bn annualized rate out of the quarter just filed. Total revenue growth did not decelerate into that mix: 26.1%, then 47.6%, then 63.1%. Gross margin moved 20 basis points across four quarters, ending at 55.35%, and management raised its Enterprise Data growth floor to 150% from 130%.

The margin stability has a stated mechanism. Asked on 29 October about rising foundry wafer costs, chief executive Michael Hsing said: "We definitely absorb some of the cost increase. And not that we don't increase the price, okay? And we some -- we share the cost with -- the cost increase with our customers." He added: "During the capacity constraint, we don't gouge our customers." The company says existing capacity supports a $6bn annual rate, targets $8bn in the second half of 2027, and needs no significant capital spending to get there.

It is still the expensive name in its own complex: 40.9x forward earnings and 65.4x trailing, against 85.5x trailing in mid-August and 117.7x in May. onsemi, by contrast, trades at 19.3x forward earnings.

The 800-volt bill is not due yet

Industry work on the transition sizes the prize precisely: power-semiconductor content per rack is estimated to grow more than sixfold by 2027, reaching roughly $140,000 at today's prices from about $15,000, as rack power climbs from around 120kW toward 600kW and above.

Navitas, which sells gallium-nitride power chips and silicon-carbide devices and employs 190 people, is the purest listed claim on that chain. Its quarter inflected: revenue of $13.99m, up 38.4%, with reported gross margin swinging to 39.3% from -9.5% three months earlier, an operating loss narrowing to $10.9m excluding stock compensation and acquisition items against a reported loss of $34.4m. December revenue is guided to $16.0m with more than 40% from AI infrastructure, about $6.4m. Operating expense is guided to $25.0m.

Chief executive Chris Allexandre dated the compounding himself on 3 November: "it's not inconceivable to see the quarter-over-quarter compound growth to start to accelerate late '27 into '28 into '29." Breakeven, as management frames it, needs quarterly revenue in the low $60m range. The company ended September with $531.5m of cash and no debt, near $405m after the Claros acquisition, and has 45.9x forward sales to defend.

Cash and patents, not volts

The pressure on the last conversion stage is coming from buyers. onsemi agreed to pay $600m in cash plus $175m of earnouts for Ferric's integrated voltage regulators, and lifted its 2030 AI data-center target by $500m to above $3bn; its AI data-center revenue was $180m last quarter, roughly $620m expected for the year. "Ferric blows away everybody's z-height," chief executive Hassane El-Khoury said on 2 November.

Vicor is charging for permission instead of parts: $66.6m of its $197.1m quarter was patent royalties, gross margin reached 68.3%, and an International Trade Commission investigation seeks to exclude unlicensed vertical-power systems. Infineon, the German power-chip maker guiding to about €16.3bn for its fiscal year, announced a gallium-nitride intermediate-bus collaboration with Nvidia on 21 October.

What the selloff did and did not know

Monolithic Power fell 17.0% in the 30 sessions to 6 November and sits 29.6% below its 3 June close; Navitas fell 25.2% and is 72.9% below 26 May. Both lost most of it on 5 November, when the Philadelphia Semiconductor Index dropped 12.25% on Qualcomm's handset guidance, with incoming senators calling for brakes on data-center construction. No company-specific development for Monolithic Power surfaced that day, and it was the mildest faller in the group: Vicor lost 48.0%, Infineon 25.0%, onsemi 24.8%.

So the two markdowns do not mean the same thing. Monolithic Power's quarter accelerated and its margin held, which leaves sector-wide repricing and policy risk around AI capital spending as the likelier reading of its fall. Navitas's quarter also improved, but the decline is unwinding a May spike in a name still priced on 2029 volumes.

The open question is disclosure. Neither Monolithic Power nor its largest customer publishes per-platform power allocation; management quantifies the ramp only as content gains as new platforms arrive. Investors have been burned on exactly that gap before: in November 2024 a research note that Monolithic Power's Blackwell allocation was "at risk", with Renesas said to have taken part of it, cost the shares about a quarter of their value in a session.

When the rack power bill really arrives, it will be large enough that nobody will have to guess who won it. Until then the only names collecting are the ones selling into the architecture already installed.

Dycom Industries, which engineers, buries and splices fiber for carriers including AT&T, Verizon and Lumen, told investors in August that its next quarter would bring in less revenue than the one it had just reported. The guide was $1.90–1.98bn against the $2.006bn delivered in the quarter to 1 August, with an adjusted earnings midpoint of $4.56 a share below the $4.79 consensus. ITG, a smaller contractor doing the same outside-plant work for regional fiber builders, pre-announced a third quarter above its own outlook on 22 October. Over the month to 6 November Dycom lost roughly a quarter of its market value and ITG two-fifths.

Both were bought as the ground crew of the AI interconnect buildout, on the belief that hyperscaler dark fiber and intercity routes were already landing in contractor revenue. Dycom's latest reported quarter says otherwise: it carried only "initial revenue contribution from fiber infrastructure programs for hyperscalers," while AT&T and Verizon each exceeded 10% of total revenues. Carrier access spending still sets the revenue line, and the cost of standing up crews sets the margin.

What a mobilized crew costs before it bills

Dycom's headline growth of 45.6% was flattered by purchases — Power Solutions in December 2025 and National Technology Integrators on 2 July 2026 for $275m, which contributed $22.9m in the quarter. Organic growth was 16.7%, down from 24.7% three months earlier. Gross margin fell to 16.20% from 17.90%, and the Communications segment margin dropped 134 basis points to 13.6% on expansion investment, deferred wireless work and fuel. The order book moved the other way: total backlog reached a record $12.242bn, up 53.2%, with $6.472bn due inside twelve months. Days sales outstanding improved to 101 from 108, so this is not cash trapped between digging and being paid.

ITG puts the same cost in plainer language. Its second-quarter adjusted EBITDA margin was 12.9% against 14.8% a year earlier, attributed primarily to "startup costs associated with onboarding and mobilizing employees and subcontractors, deploying fleet and equipment, establishing local operating infrastructure and preparing new programs for customer launch." Dycom's chief executive Dan Peyovich framed the payoff on the 26 August call: "We are comfortable scaling that business, and that is really where you see the margin raise, the high teens to low 20s as we go forward."

The carrier number behind the guide

MasTec supplied the read-across on 29 October. Its Communications revenue fell 12% to $807m and segment adjusted EBITDA fell 41% to $61m, because customers deferred second-half spending amid speculation about a potential fourth nationwide wireless carrier and announced reorganizations at major customers. MasTec expects a flattish Communications year in 2027, with carrier weakness offsetting hyperscaler and federally funded rural broadband growth. That last programme is not yet paying contractors either: more than $22bn of the $42.45bn Broadband Equity, Access and Deployment allocation sat at the Commerce Department as of June 2026, Texas paused disbursements in August, and most funded construction reaches end users in 2027 or 2028.

ITG's drop is not in its filings

ITG priced its initial public offering at $16.00 on 1 July, below its $19–22 range, and now trades at $4.575. Its next-twelve-months backlog rose to $1,517m from $1,259m on awards from Ziply Fiber and Intrepid Fiber Networks, and preliminary third-quarter adjusted EBITDA of about $64.5m beat its own $63.0m outlook. Its worst session was 28 September, a 32.5% fall with no discoverable company disclosure; short interest rose 209% between the 15 and 30 September settlement dates, to 1.76m shares. The likelier reading is a seller rather than a business break.

Price per dollar of Dycom's trailing gross profit is 4.80x, against 9.6x in late August and 12.05x in mid-May; forward earnings are 12.19x against 18.71x trailing, EV/EBITDA 7.68x and the trailing free-cash-flow yield 7.49%. ITG's trailing earnings are unusable on a small loss, leaving 8.47x forward earnings and 2.42x forward gross profit against 5.08x trailing — with a trailing free-cash-flow yield of minus 11.4%.

So the de-rating splits on cause. Dycom's August-to-October leg is explained by its own disclosures: halved organic growth, a compressed segment margin, a sequential guide down. Its last two sessions before 6 November, down 7.2% and 9.3% to new lows, carried no company news and coincided with voters approving data-center restrictions in 40 of 41 jurisdictions — the political risk that repriced power equipment this week. On ITG's side there is no reported deterioration at all, only mobilization costs the company has quantified and a guide it has now exceeded twice.

ITG publishes full third-quarter numbers after the close on 10 November; Dycom reports in mid-to-late November. One of them will show whether the crews hired this year have started to bill.

Vertex, which sells indirect-tax determination and compliance software to corporations in retail, communications, leasing and manufacturing, added barely any customers last quarter and still raised its full-year guidance. Annual recurring revenue grew 11.7% to $724.3m while average revenue per customer rose 9.2% to $145,793 — arithmetic that leaves implied customer growth near 2.3%. Almost all of the growth came from accounts already signed, putting more invoices through the meter.

Three vendors on the same financial-services software shelf reported inside four days: Vertex and Fair Isaac on 4 November, BlackLine on 5 November. They did so after twelve months in which the nine-name group lost 23% equal-weighted, a de-rating widely read as artificial intelligence hollowing out software priced by the seat. The prints point somewhere else. What decided these three quarters was who sets the billable unit — Vertex's is set by European legislatures, Fair Isaac's was set by Fair Isaac until a federal housing regulator intervened, and BlackLine's seat meter was dismantled by BlackLine.

A legislature sets the invoice count

Vertex's revenue reached $213.3m in the quarter, up 11.0%, with cloud revenue growth re-accelerating above 20% against a full-year cloud guide that had been cut to 18% in August. Net revenue retention improved a point sequentially to 106%, gross retention held at 95%, and full-year guidance went up to revenue of $828.9–830.9m with the adjusted EBITDA midpoint lifted $10.9m.

The driver is statutory. "E-invoicing ARR up over 100% year-on-year," chief executive Christopher Young told investors on 4 November. "Growth in e-invoicing contributed meaningfully to the re-acceleration in cloud revenue… This reflects strong execution, the successful launch of the French mandate and the contribution from Brinta." Germany, Norway and Slovakia begin mandates in January 2027, with Spain and phased UK obligations running into 2029. One offset was disclosed: the loss of a payroll original-equipment customer worth about $1m a quarter, enough on its own to take retention down a point in the fourth quarter.

GAAP profit is still nominal — operating income of $0.3m and a net loss of $46.6m — so the stock is judged on gross profit, at 5.46x trailing against roughly 5.09x on 30 October and 3.85x at the end of September. The re-rating is largely done.

The price Fair Isaac can no longer raise

Fair Isaac's fiscal year, ended 30 September, was the better of the two reported records: revenue of $2.535bn, up 27.3%, and Scores revenue of $1.652bn, up 41%, with business-to-business Scores up 39% in the fourth quarter. That came almost entirely from lifting the mortgage score price from $4.95 to $10 while score-pull volumes fell high single digits.

That lever is now pinned. Fair Isaac has held the wholesale tri-merge bundle at $10 for 2027 for resellers whose lenders commit to pulling a FICO score on every mortgage application. The substitute costs about a tenth as much: VantageScore 4.0 sells near $0.99, with TransUnion's pricing extended through 2028, and Rocket Mortgage has made VantageScore its preferred provider on eligible loans. Management says it has seen no meaningful volume loss to competition. It also withheld fiscal 2027 guidance entirely, with mortgage programs unsettled and the pricing year resetting on 1 January while the fiscal year began 1 October.

"Transparency and detail on our programs and pricing have not been particularly helpful for FICO," chief executive William Lansing said on 4 November. "We're not going to share every pricing program. We're not going to share what the prices are."

The shares fell 29% on 29 September, the day after the Federal Housing Finance Agency's director, Bill Pulte, said Fannie Mae and Freddie Mac would adopt a pricing grid incorporating VantageScore, and more than 13% on 26 October when Pulte pledged further action and the flat $10 price was read as defensive. Twelve-month decline: 59.5%. The stock sits at 19.3x trailing and 15.6x forward earnings, and that forward figure rests on a $45.25 consensus the company declined to endorse. The part growing independently is software: platform annual recurring revenue of $453m, up 72%, now 53% of total.

BlackLine swapped meters and the new one is late

BlackLine, whose software automates the corporate accounting close, cut full-year revenue guidance to $759.5–761.5m from $765–769m, guided the fourth quarter about 5.7% below consensus, and lost 25.1% on 6 November. Growth has stepped down three quarters running, 9.7% to 9.2% to 8.5%. Retention rose to 103.5% on platform conversions carrying 10–40% pricing uplift, which the company said was offset by lower user expansion. Paid agentic artificial intelligence moved from the first half of 2027 to mid-to-end 2027, waiting on customers' auditors and governance committees.

The cash went the other way: non-GAAP operating margin expanded 450 basis points to 25.9% on under 1% headcount growth, the trailing free-cash-flow yield is 11.4%, all of it going to buybacks, and at 2.66x trailing gross profit it is the cheapest of these names. "If customers were replacing us, it would show up first in our largest accounts as churn, smaller contracts and/or shorter commitments," chief executive Owen Ryan said on 5 November. "We are seeing the opposite, larger deals, longer commitments and strong enterprise renewals."

Four names made a nine-name turn

The shelf's 11.6% equal-weighted gain over the month to 6 November came from Vertex at 38.4%, Q2 Holdings at 27.8%, Tyler Technologies at 26.7% — a vendor of court and municipal software to state and local governments, not a financial-services meter at all — and nCino at 18.6%. The median member moved 0.25%; BlackLine fell 8.2%. The window opens almost exactly at the 5 October peak in the 10-year Treasury yield, which eased back below 5% by 30 October.

Only one of those gains traces to a change in the company's own unit economics. Q2 Holdings, which sells digital-banking platforms to regional banks and credit unions, rose on cost: its first 30% adjusted EBITDA margin, up 580 basis points, on revenue growth of 11.9% — and at 8.99x trailing gross profit it is the dearest name measured here. Fair Isaac's reported economics accelerated while its equity halved, because the market is pricing the lever that is frozen rather than the year that was banked. BlackLine's growth and its cash generation now disagree, and the thing meant to reconcile them has slipped a year.

Fair Isaac's lender clients have until 31 January to choose which 2027 pricing path they take. The company has already said it will not tell them, or anyone else, what the prices are.

FuelCell Energy has sold more stationary power than it can presently build. The company, which makes, installs and services fuel-cell power plants for utilities, hospitals, wastewater plants and lately data centers, reported $1.3bn of committed backlog and $2.4bn of awarded capacity as of 31 July. Against that $3.6bn book sits a factory that ran at an annualized rate of about 37.1 megawatts during the quarter, with a stated ambition of up to 500 megawatts annualized at its Torrington works.

The order book had been read as proof that the data-center pivot was working: a first 75 MW capacity reservation with a major data center, a 30 MW deposit under an agreement with Fit Energy for up to 380 MW, and an 11% jump on 6 October when Fit Energy unveiled a Pennsylvania campus plan. What the filing showed was a conversion problem. The company booked $17.0m of charges on inventory and firm purchase commitments under Phase 0 of the Fit Energy capital equipment agreement, and a class action filed 11 September alleges it misled investors between 24 June and 1 September about whether its manufacturing capacity could meet the production rate the deal required, disclosing the shortfall only in its 2 September quarterly report. The lead-plaintiff deadline is 10 November.

Five fiscal years of shipping plants below cost

Revenue for the quarter ended 31 July was $33.0m, down 29.4% year on year. The gross loss of $24.5m works out to a -74.2% gross margin, against -11.0% a year earlier, and the operating margin was -141.4%. The negative unit economics are structural: gross margin has been negative in each of the last five fiscal years, from -22.5% in 2021 to -16.7% in 2025. Funding that gap has cost shareholders their share of the company, with diluted shares up 188% year on year to 70.4m. An upsized offering of 10,714,286 shares in July raised about $225m gross at $21.00; the 6 November close of $9.955 is 53% below that price.

The shares fell 45.5% between 7 October and 6 November and are still up 29.0% over twelve months. The fall has not made them cheaper on sales. Trailing price-to-sales is 5.16x, above the 4.26x of early May, because revenue contracted faster than market value. The forward 3.18x rests on consensus revenue of $250.1m for fiscal 2027, a 69.2% step up, with earnings per share still at -$1.18. Book value is the one undemanding anchor, at 0.72x.

Conversion, as practiced down the same industry code

Bloom Energy engineers solid-oxide systems that make on-site electricity from natural gas, biogas or hydrogen without combustion. Its September-quarter revenue was $1,235.8m, up 138.1%, at a 33.8% gross margin, with product revenue of $1,073.6m carrying a 38.4% product margin. Operating income reached $225.4m against $7.8m a year earlier, and on 22 October the company raised full-year guidance for the third time this year, to $4.4-4.6bn.

"Operating income in this quarter alone exceeded operating income for all of last year. That is the operating leverage embedded in our model now showing up in the results," chief financial officer Simon Edwards told investors on 22 October.

Bloom shares fell 14.1% on 5 November, the morning it announced a 900 MW deployment agreement with NTT Global Data Centers for initial shipments in 2027, and 40.3% across the month. Trailing price-to-sales is 13.37x, down from 31.81x in early May while revenue tripled. Its own file carries two cautions. Oracle has invoked force majeure on Project Jupiter, which carries Bloom's largest single deployment at up to 2.45 GW, roughly a tenth of its identified 25 GW pipeline; and the $2.2bn of zero-coupon notes priced last October convert at about $194.97, above the 6 November close of $173.24.

Two markdowns, two causes

One company's results account for its de-rating and the other's account for none of it. What moved Bloom is siting risk for the campuses whose interconnect queues create behind-the-meter demand, after a Democrat won the Texas governorship on 3 November promising a pause on new data centers in a state holding 27% of planned data-center gigawatts on Wells Fargo's count. The two group members without that exposure barely registered it: Plug Power, whose gross margin reached roughly break-even at -0.9% in the June quarter, fell 19.7%, and Ballard Power, which sells stacks for buses, trucks and rail and holds the group's only positive gross margin at 19.8%, fell 24.0%. The month repriced data-center power, not fuel cells.

That leaves FuelCell Energy holding the one problem an election can neither cause nor cure. Torrington would have to run roughly thirteen times faster than it did last quarter before the backlog begins to pay.