DK Street Journal

Arm Holdings, the Cambridge company whose processor designs sit inside nearly every smartphone and which collects a fee on each chip that ships, told investors in late July that its royalties from data-center silicon had more than doubled in a year. In the same presentation, the finance chief cut the company's full-year royalty growth guidance to the high teens from the roughly 20% described three months earlier, and guided the September quarter's royalty growth to the low teens.

That is the whole argument about the chip intellectual-property business in one quarter. Arm earns a rising rate per chip — Armv9 cores and pre-integrated compute subsystems command a higher royalty than the older designs they replace — on a unit base that is going backwards, because handset makers are raising prices across mid and premium tiers as memory costs climb. Revenue reached $1.289bn, up 22.4%. Royalties were $715m and licensing $574m, up 23%. Cumulative shipments of Neoverse, its server core, passed 1.5 billion; the most recent 500 million took nine months, against six years for the first billion.

Two cents a device

CEVA licenses digital-signal-processor cores, neural-network accelerators and wireless connectivity blocks — Bluetooth, Wi-Fi, ultra-wideband, cellular internet-of-things — to chipmakers who put them in high-volume, low-priced devices. Its June quarter shows what that rung earns. Revenue rose 13.1% to $29.0m, licensing rose 21% to $18.2m — its best in three years — and royalties rose 1% to $10.8m, even though customers shipped 567m units, up 16%. That is roughly 1.9 cents a device, and falling per unit. Bluetooth chips, at 295m units, were down 16%; cellular IoT hit a record 68m.

The company raised full-year growth guidance to 13-15% and doubled its non-GAAP operating margin to 11%. But the growth is licence-fee growth, which is lumpy by nature. Its edge-artificial-intelligence royalty line does not yet exist: the NeuPro-M neural accelerator deal it signed with a large computing-platform company is on the usual one-and-a-half to two-year path to production. Management's stated mechanism for lifting the rate — selling complete subsystems, such as a full Wi-Fi 6 plus Bluetooth chip, rather than component blocks — raises deal size and royalty per unit together. It has not shown up in the royalty line.

What Arm is paying to move up the stack

Arm's other change is that it has started selling finished silicon. Demand for its AI server processor has risen above $2bn from the $1bn indicated a quarter earlier — and management guides that product's gross margin to the high 30s or low 40s, against a licensing business that runs near 97%. Annual gross margin has already slipped to 92.5% from 94.9%, operating margin to 18.5% from 20.6%, and GAAP operating income fell 8.4% year on year even as revenue grew 22%. Arm will break the silicon business out as a segment only once it exceeds 10% of revenue, expected in fiscal 2028. Until then the dilution is visible in the margin line and nowhere else.

The de-rating arrived in four sessions

Arm closed at $244.21 on Friday, down about 14% in a month, and essentially all of that loss landed between 17 and 21 August. On 18 August alone Arm fell 9.3%, the day the 30-year Treasury yield topped 5.33%, a 19-year high and the Philadelphia Semiconductor index dropped 5.4%. Long-dated royalty streams are the most duration-sensitive assets in semiconductors, and they were marked accordingly. CEVA fell 10.8% that session and is down about 31% over the month, thirteen days after its raised guidance.

What the selloff did not do is take the prices back to where they started the year. Arm trades at 53.1x trailing gross profit — the fitter lens, since almost all its revenue is gross profit and its 251x trailing earnings multiple is distorted by tax and investment items — against 68.4x in May but 30.1x in February. Forward earnings are 109x. CEVA is at 7.6x trailing gross profit against 10.9x in May, the cheapest licensor in the group, on 51x forward earnings. Astera Labs, which sells the retimers, cable modules and fabric switches that move data between accelerators inside an AI rack, is the control: revenue grew 104.5% to $392.4m at a 73.3% gross margin, September revenue is guided to $540-560m, and management put content for its Scorpio X switch alone above $1,000 per accelerator. Its multiple fell from 67x trailing gross profit to 54x in the same four sessions. Nothing in its business changed.

The setup

Where it stands — Arm's royalty rate per chip is rising and its unit base is not; the de-rating tracked long yields, not the quarter. Would confirm — September-quarter royalty revenue growing in the mid-teens or better with data-center royalties again up more than 50%. Would invalidate — Royalty growth printing below 10%, or company-wide gross margin falling under 90% as silicon shipments scale. Watch next — Arm's fiscal second-quarter results, guided to $1.38bn of revenue; CEVA's third quarter, guided to $30.5-34.5m. Valuation — 53.1x trailing gross profit, 45x forward, versus 68.4x in May and 30.1x in February; 109x forward earnings.

Amplitude, which sells software that tells app and website owners which features people actually use, told investors on 5 August that June-quarter revenue grew 21.2% to $100.9m. That was an acceleration from 16.9% in the March quarter, and the first genuine reacceleration in more than a year at a company that had spent two years decelerating. The shares, which had fallen 21.4% in a single session after the March report, have since gone from $6.52 in late May to $12.74.

The reacceleration was, in large part, purchased. Annual recurring revenue reached $410m, up 22%, but only $19m of the $36m added in the quarter was organic — the other $17m came with Statsig, the experimentation and feature-flagging startup Amplitude acquired. Of the 824 customers now paying more than $100,000 a year, up 30%, forty arrived in the same transaction. Net dollar retention of 105% is quoted on a pro forma basis that includes Statsig's book.

What the acquisition costs to run

The bill shows up one line down. Gross margin fell to 68.5% from 73.0% in March and 72.6% a year earlier, so gross profit grew 14.3% while revenue grew 21.2%. Management named three causes: Statsig runs on Google Cloud at margins currently in the low 50s pending optimization, AI tooling is generating elevated inference costs, and customer data ingestion is at an all-time high. Margins are guided to stay in the low 70s near term.

That last item is the mechanism the whole usage-priced software story rests on — more data ingested means more billed — arriving as a cost before it arrives as revenue. The rest of Amplitude's quarter is stronger than the headline suggests: free cash flow was a record $23.7m, the non-GAAP operating loss narrowed to $1.5m, remaining performance obligations rose 35% to $483m, and the company bought back $69m of stock, shrinking the diluted share count. On a GAAP basis the operating loss widened to $35.2m.

Amplitude is the number-two product-analytics vendor by deployments, at 28.1% of tracked installations against PostHog's 45.6% and Mixpanel's 10.5%. Management's own stated near-term risk is that customers do not know what it now sells — many were unaware of the Statsig deal or the wider suite.

The price ran further than the business

Amplitude now costs 6.51x its trailing gross profit and 5.94x the forward figure, against 3.31x in May and 3.73x in February. That is a 97% multiple expansion in three months against trailing gross profit that grew 6.8%. As recently as late July the shares fetched 4.42x. The post-earnings continuation came not from new disclosure but from a run of analyst price-target increases — Piper Sandler and UBS to $13, Morgan Stanley to $16 and Bank of America to $14.

Dynatrace, which monitors cloud applications for large enterprises and diagnoses what breaks, is the other side of the comparison. Its ARR reached $2.14bn, up 17% in constant currency, with net new ARR of $85m; trailing-twelve-month organic net-new-ARR growth hit 17% against 12% a quarter earlier, a fourth straight quarter of acceleration. Adjusted free cash flow runs at 28% of revenue, the diluted share count is down 3.5% year over year after $275m of buybacks in one quarter, and full-year revenue and margin guidance were both raised. The shares jumped nearly 11% premarket on the print. At 8.43x trailing gross profit it remains the cheapest large observability name — Datadog trades at 26.3x — and at $48.88 it is exactly flat against $48.81 a year ago.

Not a telemetry re-rating

The tempting reading is that software priced by data volume re-rated while software priced by seats did not. It does not survive the dates. Dynatrace's entire 90-day advance sits in three sessions — 5 August's results, 26 June and 1 July. Strip those and the ninety days are worth 0.4%. Two of the three are shared: Progress Software rose 11.2% and 16.6% on those same dates, Sprinklr 5.1% on the first.

Progress, which sells developer and infrastructure software including the OpenEdge platform and MOVEit file transfer, had its own catalyst — June-quarter earnings of $1.62 a share against a $1.49 consensus, with full-year cash-flow guidance raised. It grew revenue 6.8% and still trades at 2.24x gross profit, a round trip to where it stood in February. Sprinklr, which sells customer-experience management software to large brands, is the weakest business of the four: revenue up 6.8%, gross margin down to 65.2% from 69.5%, gross profit essentially flat, and a chief executive who calls this a transition year spent paying down technical debt. Its multiple rose anyway, to 3.07x from 2.31x.

Nor is this falling rates. The 10-year Treasury yield sat at 4.682% in mid-August with investors shifting from expecting cuts to weighing increases. What moved was capital: the Philadelphia semiconductor index peaked above 14,600 in June and fell toward 11,200 in July, and money rotated into software, the side of the AI trade that had been priced as the loser. That bid is indiscriminate — Monday.com rose 7%, HubSpot 6% and Intuit 5% in a single session on 18 August, none of them priced by data ingested.

The setup

Where it stands — Amplitude's growth reaccelerated on an acquisition while its multiple nearly doubled; Dynatrace's accelerated organically and its multiple rose a quarter as much.

Would confirm — Amplitude organic net new ARR above $19m in the September quarter with gross margin recovering into the low 70s.

Would invalidate — Gross margin holding at or below 68.5% while net dollar retention slips back under 100% once Statsig laps.

Watch next — Amplitude's September-quarter report in November, its first full quarter with Statsig; Dynatrace's fiscal second half, when 70% of subscription renewals reset.

Valuation — Amplitude at 6.51x trailing and 5.94x forward gross profit, against 3.31x in May; Dynatrace 8.43x and 7.61x, against Datadog's 26.3x.

Asana has told investors nothing about its business since late May, and will say nothing more until 3 September. Its shares have risen nearly 30% in a month anyway. Over the same stretch monday.com, which did report, published record enterprise growth and a beat on both revenue and earnings — and fell 9.4% in the three sessions afterwards.

Both sell cloud software for coordinating work: Asana a project-and-goals platform used by marketing, operations and engineering teams, monday.com a configurable "Work OS" from which customers assemble their own project, sales and support applications. Both are priced per seat, which is why both spent the past year being sold as the software most exposed to AI agents doing the work a seat used to do. Asana is down 30% over twelve months; monday.com is 58% below its 52-week high.

What actually moved them

Roughly two-thirds of each stock's monthly gain landed in the last week of July, when Asana jumped from $6.88 to $8.19 and monday.com from $73.91 to $87.09. Neither company published anything. What changed was the argument: Jensen Huang told investors that companies "will rent agents that use those tools, which means the software industry in the future will be much larger than the software industry of today" — a direct inversion of the seat-destruction thesis. The second burst came on 13 August, when Silver Lake was reported in talks to acquire Workday for about $51bn, putting a private-equity floor under de-rated software multiples.

This was not a bond-market bid. The 10-year Treasury yield sat near 4.7% on 20 August, within a few basis points of a 20-month high. Rates rose while the shares did. And the rally was broad: thirteen comparable seat-priced software names ran from 8% at BILL to 93% at Atlassian, leaving Asana's 29.5% and monday.com's 25.4% squarely mid-pack.

Only one of the two has the numbers

Asana's April-quarter revenue was $205.1m, up 9.5%, the fourth straight quarter in the nine-percent range against 19.2% two fiscal years ago. Gross profit grew just 6.9%, because gross margin fell 211 basis points to 87.6% as inference costs landed in cost of revenue. Dollar-based net retention was 96% — and also 96% among customers spending $100,000 or more, of which there were 817, up 12%. Management guided fiscal 2027 to $855.5m–$863.5m, growth of 8.2% to 9.2%. Its AI Studio product is disclosed as 17% of net new annual recurring revenue, a bookings-mix share rather than a reported dollar line. Adjusted free cash flow was $84.5m last fiscal year against stock compensation guided to the low 20s as a percentage of revenue — roughly twice the cash generated — and the diluted share count still grew 1.4%.

monday.com's June quarter grew 21.9% to $364.6m, with a 17% non-GAAP operating margin and annual recurring revenue past $1.5bn. Customers above $100,000 of ARR reached 2,019, up 37%, and those above $500,000 reached 114, up 68%. AI revenue doubled sequentially to 17% of net new ARR on a seat-plus-credit model launched in May. The blemishes are real: net retention of 109% was the lowest the company has ever reported, guided toward 108%, third-quarter growth was set at 16–17%, and 20% of staff were cut on 22 July. But retention among customers above $100,000 was 115%, the diluted share count shrank 16.6% year on year, and free cash flow yield is 7.8% against Asana's 4.9%. The $870m buyback authorisation is now fully spent, removing a bid that supported the shares through the fall.

The price of the same dollar

Because Asana loses money, earnings multiples are useless; gross profit is the honest denominator, and it tells a clean story. Asana's price per dollar of trailing gross profit was 2.14x on 3 May and 2.72x on 29 July. It is 3.18x today — a 49% re-rating in fifteen weeks with no new financial disclosure in between. monday.com's went 3.15x, 3.18x, 3.16x: its share price rose about as fast as its gross profit grew and its share count shrank. On forward earnings the gap is starker still — Asana at 25.7x against monday.com at 16.6x. Atlassian, the profitable comparator on the same rung, grew 27.6% last quarter and commands 8.07x trailing gross profit, more than double either.

The technical picture confirms the asymmetry rather than the fundamentals. Asana's 50-day average crossed above its 200-day on 7 August and its uptrend strengthened on 14 August; monday.com only turned upward on 12 August and has stayed there more tentatively. The likeliest mechanism is mechanical: short interest in Asana ran near 38.8% of shares in mid-August, and the move has been described as a short-covering rebound. Analysts model Asana's revenue at $929m in fiscal 2028, up 7.9% — the Street is not underwriting a re-acceleration.

The setup

Where it stands — Asana's multiple has re-rated 49% since early May on no disclosure; monday.com's is unchanged while its gross profit grew 20%. Would confirm — Asana's 3 September quarter showing net retention at or above 100% and revenue growth above 10%. Would invalidate — Retention holding at 96% with fiscal 2027 guidance unchanged at 8–9% growth. Watch next — Asana reports second-quarter fiscal 2027 results on Thursday 3 September; monday.com's third quarter guides to 16–17% growth. Valuation — Asana 3.18x trailing and 2.99x forward gross profit versus 2.14x in May; monday.com 3.16x and 2.94x, flat since May.

A month of gains with nothing to report

Asana, which sells cloud software for coordinating projects, campaigns and company goals, has not communicated with the market since 28 May. Its next results land on 3 September. In between, the shares rose 37% from 22 July to 21 August. Nothing the business did explains it.

What happened instead was a rescue of the entire de-rated software complex. Nvidia's Jensen Huang told investors that agentic artificial intelligence expands rather than destroys software demand — "those agents are going to use more tools than ever" — after which ServiceNow rose 10% and Atlassian and HubSpot 6% each. Three Federal Reserve cuts through end-2025 did the rest, pulling money into long-duration, low-profitability equities of exactly this type. Asana's largest single advance came in the sessions to 28 July, when it gained 19% alongside Atlassian. Its second came on 7 August, Atlassian's earnings day, when that stock gapped 35.8% on 20.1m shares.

The pair did not lead the rebound; they trailed it. RingCentral rose 77% over the same window, Wix 68%, Appian 65% and Paycom 64%.

The number the whole case rests on

The argument for Asana is that AI Studio, sold by consumption rather than by user, breaks the link between revenue and headcount. It is a real disclosed line, not a slide: the company exited fiscal 2026 with over $6m of AI annual recurring revenue, growing more than 50% a quarter, with eight customers spending above $100,000 a year on it alone. Against $791m of total revenue, that is 0.8%.

The seat base it is meant to offset is still contracting. Dollar-based net retention was 96% in the April quarter, and 96% again among customers spending $100,000 or more — the largest accounts are shrinking as fast as the average one. The count of those accounts reached 817, up 12%, so what growth exists comes from new logos crossing the line, not expansion behind it. Revenue grew 9.5%, and has sat within half a point of that mark for four consecutive quarters. Guidance for the year is 7.5-8.5%. Gross margin fell 212 basis points to 87.6% as inference costs settled into cost of revenue.

The cost line is where Asana has genuinely changed. Non-GAAP operating margin hit a record 11.5%, up 720 basis points, and adjusted free cash flow was $34.4m against $9.9m a year earlier. Diluted shares grew 1.4%, a fraction of prior years. This is a business being run for cash, not a business whose demand has turned.

That matters for what the price now assumes. Asana trades at 3.18x trailing gross profit, against 2.72x in late July and roughly 2.1x three months ago. Forward gross profit takes it only to 2.99x — a 6% spread, meaning almost no gross-profit growth is expected. Atlassian, competing for the same enterprise work-coordination budget with Cloud revenue up 31% and remaining performance obligations up 44%, commands 7.1x. The market pays more than twice as much per dollar of gross profit for the name whose seats are growing.

The one that fell on its own news

Monday.com, an Israeli company selling a configurable "Work OS" from which customers assemble their own apps, is the inverse trade. It reported on 10 August and the shares fell 5.6%. Revenue of $364.6m grew 21.9%, down from 26.2% three quarters earlier, and third-quarter guidance implies 16-17%. Net dollar retention slipped to 109%, its lowest ever, with management calling 108% the floor. In July it cut about 630 jobs, a fifth of staff, for roughly $100m of annual savings.

The other half of the print was strong. It crossed $1.5bn in annual recurring revenue, with customers above $100,000 up 37% and above $500,000 up 68%. AI revenue doubled in a quarter to 17% of net new bookings — twenty times Asana's share. Operating income turned positive at $19.9m. Diluted shares fell 16.6% on an $870m buyback that is now fully spent.

And its multiple never moved: 3.16x trailing gross profit, unchanged since late July, for a company growing twice as fast as Asana at the same price per dollar of gross profit. Forward earnings are 16.6x against 38.0x trailing, trailing free-cash-flow yield is 7.8%, and the shares sit 59% below their high.

The setup

Where it stands — A sector-wide rebound lifted both names; only Asana's multiple re-rated, and only monday.com's numbers were tested by an earnings print.

Would confirm — Asana's 3 September report showing net retention back above 100% and AI revenue past 2% of total.

Would invalidate — Asana guiding fiscal 2027 revenue growth below 7.5% with net retention still at 96%.

Watch next — Asana reports second-quarter results after the close on Thursday 3 September.

Valuation — Asana 3.18x trailing gross profit, 2.99x forward, versus 2.72x in late July and 7.1x at Atlassian.

On 18 August the 30-year Treasury yield topped 5.33%, its highest in 19 years, on a global bond selloff and worry about a US debt load nearing $40 trillion. Long-duration assets sold off. But the companies that own the buildings artificial-intelligence computing runs inside did not sell off together, and the gap between them was wide enough to rule out a common cause.

Equinix, which rents rack space, power and network cross-connects in 273-plus data centers across 36 countries to more than 10,000 customers, closed down 1.1% that session. Digital Realty, the wholesale landlord that leases whole halls to hyperscale tenants from a fleet of 309 facilities holding roughly 3.0 gigawatts of live capacity, fell 1.3%. Applied Digital, a former crypto miner that now builds single-tenant AI campuses in North Dakota, dropped 8.6%. Core Scientific, which converted mining sites to compute hosting after emerging from Chapter 11, fell 6.7%.

Same tenants, different balance sheets

The two rungs sell adjacent products to overlapping customers. What differs is who carries the financing risk. Digital Realty funds at investment-grade spreads, sits at 4.7x debt-to-EBITDA against a 5.5x internal ceiling, and has about $6bn of liquidity. Applied Digital's equity is a levered residual on project debt raised against contracts with tenants that are themselves borrowers. That distinction became visible in the credit market this spring: CoreWeave's $3.1bn delayed-draw facility closed in May was collateralized by two non-investment-grade customer contracts and priced at a higher rate than the prior facility backed by Meta-anchored paper. CoreWeave, Applied Digital's anchor tenant, disclosed $35.6bn of total debt in its second quarter.

When the long end moves, the landlord's spread narrows. The developer's financing question reopens.

The landlords' numbers are accelerating

Equinix's revenue growth ran 5.2% in the third quarter of 2025 and reached 16.4% in the second quarter of 2026, on revenue of $2.625bn. Operating margin widened to 25.3% from 21.9%. Annualized gross bookings were $424m, up 23%, and churn improved to 1.8%. Management raised 2026 guidance by the largest increment in company history, to 11-12% revenue growth and 10-12% growth in adjusted funds from operations (AFFO) per share. Diluted shares rose 1.1%.

Digital Realty is the sharper case. Cash rents on renewals rose a record 25.4%, and on leases above one megawatt they rose 66.7% — the pricing power sits in the hyperscale sockets, not the retail ones. Backlog reached a record $1.9bn of annualized rent at full share, up 75% since January and equal to roughly 30% of in-place data-center revenue. It has $20bn under construction, of which 1.4 gigawatts is 63% pre-leased at an 11.5% average stabilized yield — six points of spread over the 5.32% long bond. That cushion matters because construction costs keep climbing: the average global build ran $10.7m per megawatt in 2025 against $7.7m in 2020, with a further 6% rise forecast for 2026.

What the developers are earning

Applied Digital's fiscal fourth-quarter revenue rose 580.7% to $258.7m, and it holds about 1,410 megawatts of contracted load worth roughly $36bn over 15-year take-or-pay terms. Only 175 megawatts is live — 12%. Gross margin fell to 15.7% from 42.5% the prior quarter, the full-year operating loss was -$236m, and trailing free cash flow runs at -35.4% of market value. Diluted shares went from 201.2m to 275.2m in one fiscal year, alongside $450m of convertible notes. Core Scientific grew revenue 108.8% to $164.2m and lost $1.155bn in the quarter. Neither has an earnings multiple to anchor; Applied Digital trades at 60.7x trailing gross profit.

The rate explanation fails a simpler test. American Tower and Crown Castle, which own cell towers and no AI load at all, fell 15.3% and 26.2% over twelve months. Equinix rose 38.5% and Digital Realty 16.8%. A shared duration factor cannot open a 50-point gap.

Where the price sits

Equinix's guided 2026 AFFO is $42.69-$43.29 per share; at $1,069.50 that is about 24.9x the midpoint, at or just below the bottom of the 25-30x forward multiple the stock has historically commanded as the sector's most expensive name. The trailing price/earnings ratio of 68.3x is a depreciation artifact. Digital Realty's guided 2026 core funds from operations is $8.15-$8.20, putting it near 23.4x against a historic 22-25x — but its share count rose 4.6% year on year on the Blackstone and Columbia Capital deals, so per-share growth is being bought.

Over 90 days Applied Digital fell 40.0% and Core Scientific 28.1%, while Equinix and Digital Realty were within half a point of unchanged. The AI leasing story did not weaken. The cost of financing it without an investment-grade rating did.

The setup

Where it stands — The investment-grade data-center landlords held through the bond selloff; the project-debt developers absorbed six to eight times the hit.

Would confirm — Digital Realty converts backlog on schedule with stabilized development yields staying above 11% into 2027.

Would invalidate — Applied Digital energizes a large share of its 1,410 contracted megawatts and gross margin recovers toward 40%.

Watch next — Equinix's third-quarter report in late October, against 11-12% guided 2026 revenue growth.

Valuation — Equinix 24.9x guided 2026 AFFO versus a 25-30x history; Digital Realty 23.4x core FFO versus 22-25x.

In the three months to June, AAON's BasX division — which builds purpose-built cooling units and pre-packaged outdoor mechanical rooms for data centers — sold more than three times what it did a year earlier. The rest of the company, which makes rooftop air-conditioning units for ordinary commercial buildings, grew 40%. It did so into a US market where non-residential construction spending fell 7% in nominal dollars over the first five months of 2026; the American Institute of Architects' forecast panel now expects a 0.3% decline for the full year, and about a 1% fall excluding data centers. Whatever is driving this corner of the heating and cooling business, it is not the building cycle.

Four sessions did all the damage

Eight listed companies carry the thermal load of the AI build — from cold plates inside the rack to the pipefitters who weld the central plant. As a group they are down about 6.8% over 30 days. Almost all of it arrived between 17 and 21 August, when the same eight fell 7.7% on an equal-weighted basis. Before that week nVent was up 9.7% on the month and EMCOR up 12.9%.

No member reported results in that window. What did happen is that the 30-year Treasury yield topped 5.33% on 18 August, a 19-year high, in a global bond selloff. The one name that rose was Ecolab, the water-treatment and cleaning-chemicals group, up 2.5% — and its cooling exposure is a $4.75bn purchase of CoolIT Systems that closed on 2 July, not a multi-year order book. Long-dated backlogs got marked down; a just-bought earnings stream did not. That is what a repricing of duration looks like, not a repricing of demand.

nVent built its growth rather than buying it

nVent Electric makes the enclosures, busbars and liquid-cooling hardware that sit inside and beside the rack. June-quarter revenue rose 52.8% to $1.47bn, of which 47 points were organic — the EPG acquisition added roughly seven. Its infrastructure vertical, overwhelmingly data centers, grew more than 100% organically and now accounts for close to 60% of first-half sales, against 12% when nVent was spun out. Operating margin widened 417 basis points to 20.4%.

Management guides to more than $2bn of data-center revenue this year on a $2.5bn backlog, and is building a third liquid-cooling plant in three years, 160,000 square feet, for early 2027. It is a challenger, not the incumbent: Vertiv holds over 11% of the liquid-cooling market and the top five vendors only about 35% between them. Ecolab's CoolIT now competes for the same socket. nVent costs 14.0x trailing gross profit, down from 16.6x in May while that gross profit grew 21.7% since February. The forward price/earnings ratio of 29.9x against 41.2x trailing, though, already assumes a 53% earnings jump — and consensus has growth halving in 2027.

EMCOR sells hours, and its order book is full

EMCOR Group is a labor business: 44,000 electricians and pipefitters installing central plant, high-purity piping and power distribution. Its remaining performance obligations hit a record $17.14bn at 30 June, up 43.9%, and 95% of that is organic. Bookings of roughly $6.7bn against $5.16bn of revenue imply a book-to-bill near 1.3. Network and communications led the increase — mechanical revenue in that category more than doubled — though institutional, healthcare and water work contributed materially too.

The mechanism is rack density. Management describes campus projects growing from 20 megawatts in 2019 to 100-200 megawatts now, with mechanical revenue per employee running up to twice the normal rate on AI jobs. Operating margin reached 10.6%, above the high-single-digit range EMCOR has historically earned — 6.96% in 2023, 9.84% last year — even as the mechanical segment gave back 110 basis points to 12.5% accepting guaranteed-maximum-price contracts. EMCOR costs 9.66x trailing gross profit against roughly 11.0x in both February and May, and against 20.2x for Comfort Systems, the closest listed comparable, which grew 50.3% on a richer 25.9% gross margin.

AAON's problem is real; the de-rating is twice its size

AAON's June revenue doubled to $627m and operating income nearly tripled. But gross margin fell 230 basis points to 24.3% and full-year gross-margin guidance was cut 200 basis points to 25-26%, sending the shares down 5.9% on the print. The cause is capacity: Memphis overhead ran $18.1m in the quarter against $3m a year earlier, while BasX's own segment gross margin improved to 30.0%. Free cash flow is negative on $102.6m of year-to-date capital spending. Total backlog is near $2.0bn but slipped sequentially as large engineered orders lumped. The market was entitled to price that. It priced considerably more: 25.8x trailing gross profit in May, 13.1x now — below February's 21.7x — with the last 9.3% arriving in a week that carried no AAON news at all.

The setup

Where it stands — Three accelerating businesses re-rated downward in one bond-market week, with only AAON carrying a disclosed margin deterioration.

Would confirm — EMCOR's September-quarter remaining performance obligations hold above $17bn with network and communications again the largest increment.

Would invalidate — nVent's 2026 data-center revenue guidance falls below $2bn, or AAON's BasX backlog declines year on year.

Watch next — Third-quarter results: EMCOR and nVent in late October, AAON in early November.

Valuation — EMCOR at 9.66x trailing gross profit versus 11.0x in May and 20.2x for Comfort Systems; nVent 14.0x, AAON 13.1x.

Schrödinger sells the physics-based simulation software that drug chemists use to decide which molecules are worth synthesizing, and in the June quarter that business got smaller. Software revenue fell 10%, to $32.5m. Total revenue still rose 7.5% to $58.9m, because drug-discovery revenue jumped to $23.0m from $13.9m on a single $10m collaboration milestone triggered by Eli Lilly's acquisition of Ajax Therapeutics, a company Schrödinger helped found. Management raised full-year drug-discovery guidance on that one payment.

Since 22 July the shares are up 34%. Certara, which sells biosimulation software and modelling services used in regulatory filings, is up 25%. Veeva Systems, the cloud software vendor built exclusively for life-sciences commercial and clinical operations, and the only one of the three with no discovery-compute workload, is up 34% — the leader.

What actually improved

The genuinely encouraging number at Schrödinger is annual contract value, which grew 27% year on year to $29.6m as biotech funding thawed; management points to an initial public offering pace running at twice last year's. But the full-year contract-value guide of $218-228m implies 10-15% growth, embedding a sharp second-half slowdown. The reported software decline is partly mechanical: hosted licensing reached 47% of software revenue from 31% a year earlier, and hosted contracts are recognized ratably rather than upfront, costing roughly $2-3m of revenue for each point of mix shift. Software gross margin still slipped to 71% from 76%.

Below that, the operating loss was $41.5m. The $6.0m of net income came from below the operating line, flattered by a gain tied to the Ajax transaction. Cash and marketable securities stood at $418.8m against a trailing burn near $32m — years of runway — but diluted shares grew to 75.8m from 73.4m, with no repurchase.

Certara's picture inverts. Revenue fell 10.8% to $93.3m and missed consensus of $98.6m, sending the stock down about 16% before the open; most of the decline is the May carve-out of its regulatory and medical-writing unit, which had carried $19.2m of quarterly revenue. Operating margin swung to -0.2% from +9.1%. Yet software bookings rose 9% to $50.7m, and trailing-twelve-month software bookings grew 7% excluding the Chemaxon acquisition against 0.8% at the end of 2025. Software is now 53% of revenue, up from 40% two years ago. Services bookings fell 6%. Certara also shrank its share count 4%, to 154.4m.

Its moat is regulatory, not computational: every one of the 13 novel therapies the Food and Drug Administration approved in the quarter came from a Certara client. Its artificial-intelligence attachment — an NVIDIA collaboration integrating the BioNeMo agent toolkit — carries no disclosed revenue, as Schrödinger's Bunsen agentic co-scientist, now deployed at Bristol Myers Squibb, does not either.

The control name is winning

Veeva grew revenue 16.3% to $882.9m last quarter at a 30.9% operating margin, with operating income up 16.8%. It is the only one of the three whose forward multiples sit below trailing: 26.8x forward earnings against 42.1x trailing, 14.4x forward gross profit against 15.9x. Consensus expects Veeva's gross profit to grow and both protagonists' to shrink — Schrödinger's price-to-gross-profit is 10.0x trailing and 10.7x forward; Certara's 5.5x and 5.9x, with consensus revenue down 9% this year.

The timing points away from a discovery-compute story. Two sessions carry two-thirds of Schrödinger's month: 6 August, the day after earnings, and 19 August. On that second day the S&P 500 fell 0.87% and the Nasdaq 1% as yields rose, while enterprise software names rallied on a rotation out of capital-hungry AI hardware. The other force is customer budgets: the S&P Biotech ETF has risen more than 50% since bottoming on 11 May.

What has changed most is price. Schrödinger costs about 10.2x trailing gross profit versus roughly 6.7x at the May close, and Certara 5.6x versus 3.4x — expansion of 50% and 66% in three months against businesses whose recurring lines are, respectively, shrinking and only just reaccelerating off a very low base.

The setup

Where it stands — Both discovery-software names have re-rated hard on bookings and one-off milestones while reported recurring revenue fell at each. Would confirm — Schrödinger software revenue returning to year-on-year growth, and Certara total bookings growing faster than 1%. Would invalidate — Schrödinger's full-year contract value landing inside the 10-15% guide, or Certara services bookings falling again. Watch next — Veeva reports 26 August; both smaller names report third quarters in early November. Valuation — Schrödinger 10.0x trailing gross profit and 10.7x forward, against 6.7x in May; Certara 5.5x and 5.9x, against 3.4x.

Cummins told investors on its August call that its power-generation growth target of 15% to 25% is limited by how many engines it can build, not by how many customers want them. The company is adding 20 gigawatts of incremental global capacity phased across 2027 to 2030, and is currently taking orders for its largest 95-litre generator sets through the second half of 2028. That is a supplier rationing output — an unusual position for a diesel-engine maker whose share price is normally set by North American trucking.

The engine maker's mix is changing faster than its cycle

Cummins builds diesel and natural-gas engines, powertrains and components for trucks, construction and marine customers, and through its Power Systems segment the standby and prime-power generators that sit outside data-center campuses. Power Systems is now roughly a quarter of group revenue and it grew fastest: revenue of $2.3bn in the June quarter, up 19%, at a 24.5% EBITDA margin, 170 basis points better than a year earlier. Engine grew 6%, Components 7%, Distribution 9%. In China, power generation revenue nearly doubled on data-center construction. The truck business, meanwhile, is merely steady — the company's North American heavy-duty forecast is 240,000 to 250,000 units in a market it expects down 4%.

Group revenue was a record $9.46bn and full-year guidance was raised for the second time this year, to 10–13% growth. The stock fell 2% that day. Cummins is a clear second to Caterpillar in data-center generators, ahead of Rolls-Royce mtu and Mitsubishi, and is expanding its Fridley, Minnesota plant to shorten 18-month lead times. At 19.8x forward earnings against 29.9x trailing, the market is discounting roughly 30% earnings growth already; price to trailing gross profit is 9.2x, against 10.3x three months ago. The bear case is not demand but the supply response — David Giroux of T. Rowe Price argued at Barron's midyear roundtable that the capacity Caterpillar and Cummins are adding will erode the economics of 2030 data-center power.

Fluence has the orders and not the factory

Fluence integrates grid-scale battery storage systems, a business spun out of a Siemens and AES joint venture, and it is the one genuine break here. June-quarter revenue rose 7.9% to $649.8m, missing consensus by roughly 29%, while gross margin collapsed to 5.1% from 14.8% and gross profit fell 62.7%. The cause is manufacturing: a 15 GWh automated facility in Houston slipped a quarter on construction and equipment problems, and an international line needed rework. Full-year revenue guidance went to about $3.0bn and EBITDA to a loss.

The order book says the demand is real. Intake tripled to $1.44bn, including a first $850m of data-center awards, and backlog reached a record $6.4bn. But Fluence is a $2.1bn company that may need $300m to $500m of extra working capital next year against $863m of liquidity, and it now ranks fourth among global storage integrators behind Sungrow, Tesla and CATL. Its worst session was its own results day.

Cummins' losses arrived on somebody else's news

Four sessions account for more than Cummins' entire monthly decline; across the other seventeen the stock rose. On 29 July, Baird downgraded Caterpillar on local opposition to data-center construction — 75 projects worth about $130bn delayed or blocked in one quarter — and Cummins fell 5.5% alongside it. On 18 August the 30-year Treasury yield hit 5.337%, its highest since 2007, and everything that finances a data center repriced.

That is not a generic equipment de-rating. Generac, the standby-generator maker, booked more than $100m of data-center revenue in the quarter and guides to nearly $450m for the year, and it fell less than 4%; Plug Power, which has no data-center revenue, was flat. The names with the most AI content fell hardest.

Bloom Energy, which sells solid-oxide fuel cells that energize a data hall without waiting for a utility interconnection, sits in the middle. Revenue grew 166% to a first $1.065bn quarter and gross margin widened to 33.4%. Its price to trailing gross profit was about 129x in May and is 61x now — a full round trip to where it stood six months ago, achieved partly by a 23% rise in the diluted share count over two quarters.

The setup

Where it stands — Cummins' generator business is capacity-constrained into 2028 while its shares de-rated on rates and a rival's downgrade. Would confirm — Power Systems revenue growth holding at or above 15% in the September quarter with EBITDA margin above 24%. Would invalidate — A cut to the 15–25% power-generation growth guide, or hyperscaler order deferrals disclosed on the next call. Watch next — Cummins' third-quarter results in early November; Fluence's fiscal fourth quarter and Houston start-up in late November. Valuation — Cummins at 19.8x forward and 29.9x trailing earnings; 9.2x trailing gross profit versus 10.3x three months ago.