DK Street Journal

Arteris sells no chips. It licenses network-on-chip interconnect — the internal plumbing that moves data between processor cores, accelerators and memory controllers inside a system-on-chip — to automotive, consumer and data-center designers, and it collects money twice: a licence fee when a customer starts a design, then a royalty on every unit shipped, often three or four years later. In the June quarter the first half of that model worked unusually well. Revenue rose 46.2% to $24.1m, the fourth consecutive quarter of acceleration from 18.3% a year earlier. Annual contract value plus royalties reached a record $99.5m, up 44%. Remaining performance obligations — contracted work not yet recognized — hit a record $135m. Management raised full-year revenue guidance to $95-98m and reiterated a target of non-GAAP operating profit by the fourth quarter.

The shares have gone the other way with unusual violence. Arteris closed at an all-time high of $48.59 on 30 June and at $24.83 on 19 August — down 48.9%, the steepest fall in a group that includes Astera Labs, Arm, SanDisk and Cadence. The first leg was company-specific: a 21% single session in early July on aggressive insider selling, a disclosed CFO departure and valuation after a rally of more than 350%. The second leg was not. On 18 August a closely watched semiconductor gauge fell 5% on fears that AI infrastructure spending is peaking, the same day the 30-year Treasury yield touched 5.323%, a 19-year high.

Why duration is the whole argument

A higher long rate hurts distant cash flows most, and Arteris's cash flows are about as distant as semiconductors get. Trailing royalties were $8.6m — growing 65%, but still under 9% of the bookings base. Consensus does not show positive earnings per share until 2027. Data-center and AI infrastructure reached roughly 29% of bookings in the first half, on the back of a hyperscale cloud company standardizing on its infrastructure silicon, but those sockets pay per-unit royalties years from now.

The bear case is not demand. It is dilution and cost. The GAAP operating loss widened to $13.9m from $8.2m, gross margin slipped to 85.1% from 89.4%, and the diluted share count rose 13% to 47.3m after a $72m at-the-market equity raise struck above $35 — roughly 30% above today's price. Arteris holds $123m of cash and no debt, so runway is not the question; the question is what the equity was worth when it was sold. At 15.6x trailing gross profit the stock is well below May's roughly 23.5x, but still about 55% above the roughly 10.1x it carried in February. The de-rating has retraced two-thirds of the spring re-rating, not all of it.

What Arteris does own is position. It is the second-largest chip-infrastructure IP vendor behind Arm, with an estimated 31.3% share of interconnect IP, competing as a single-component specialist against Synopsys and Cadence, whose full design flows come bundled. No customer exceeded 10% of licence revenue in the first half — the first such period in over six months.

The same layer, with silicon attached

Rambus shows what happens when royalties are already cash. It licenses memory and interconnect IP but also sells DDR5 register clock drivers and companion chips that ship one per server memory module, and it holds roughly 40% of that market against Renesas and Montage Technology. June-quarter revenue was a record $207.4m, up 20.4%, at a 79.8% gross margin, with record product revenue of $99.2m. It guided the September quarter to $210-216m and disclosed a next-generation high-bandwidth-memory design win at a Tier 1 US hyperscaler.

It still fell 32.0% over three months, to 16.6x trailing gross profit from roughly 26.1x on 20 May — while that gross profit grew 28.3%. Over the same three months Micron rose 28.0% and SanDisk 12.7%. The companies whose modules carry Rambus's chips re-rated; the licensing layer that supplies them did not. One caveat sits under the product story: registered-DIMM bit supply is forecast to grow 15-20%, a ceiling on units no design win removes.

The setup

Where it stands — Both chip-IP licensors are compounding faster than before while their multiples compress, with the August leg driven by rates and AI-capex fear. Would confirm — Arteris exiting 2026 with bookings of $102-106m and a positive non-GAAP operating result in the fourth quarter. Would invalidate — Arteris royalties stalling near $8.6m or fresh equity issued below $25 a share. Watch next — Sarab Sinha becomes Arteris chief financial officer on 8 September; Rambus reports the September quarter in late October. Valuation — Arteris: 15.6x trailing, 14.0x forward gross profit, versus 23.5x in May and 10.1x in February; Rambus 16.6x trailing.

Bill Holdings runs the plumbing small American companies use to pay their suppliers — accounts-payable and receivable automation, sold largely through accounting firms — and it reported the June quarter on Wednesday evening. Payment volume reached $98bn, up 14% year over year across 37 million transactions. The more interesting figure is the sliver it keeps. Transaction fees grew 17%, faster than volume, so the implied take rate widened to roughly 33.1 basis points from about 32.3bp a year earlier.

Take-rate compression is the standard bear case on payment networks. This was the reverse. Underneath, subscription fees were $76.2m, transaction fees $324.3m and float revenue — interest on customer money in transit — $35.7m, which makes float about 8% of the line and transactions, not seats, the engine. Core revenue rose 16% to $400.5m and non-GAAP operating income rose 80% to $101.6m, with adjusted earnings of $0.84 a share against a $0.71 consensus. The shares fell 3.0% — the only one of thirteen back-office and workflow software names down that session. The stated reason was the outlook: fiscal 2027 core revenue of $1.669-1.719bn, or 11-14% growth, a step down from the 16% just delivered.

The month was one session

That matters because of what the rest of the group did on no numbers at all. Averaged across thirteen names — Workday, Paycom, Bill, ADP, Paychex, Paylocity, DocuSign, Monday.com, Atlassian, Asana, AppFolio, Workiva and Freshworks — the month to 19 August produced a 23.5% gain against 3.6% for the S&P 500 tracker. Nearly all of it arrived twice. On 28 July every one of the ten members with price data rose between 8.7% and 20.6% while the index was flat, after Nvidia's Jensen Huang told investors that AI agents increase demand for software tools rather than replace them. Remove that day and the group's month is 10.7%. Remove Atlassian's 7 August earnings session too and it is 4.9%, with five of the thirteen negative.

Workday, which sells human-capital, financial and spend management software to large enterprises, gained 34.8% in that month having disclosed nothing since 21 May; it does not report until after the close on 27 August. The last numbers it published were good ones — 12-month subscription backlog up 15.5% to $8.806bn, ahead of subscription revenue growth of 14.3%, with total backlog up 10.9% to $27.294bn — and operating margin has widened from 1.7% to 13.3% across four quarters. Backlog growing faster than revenue is the single cleanest rebuttal to the claim that AI agents are eating enterprise seats. It is also three months old.

Paycom, which sells the same category of software to mid-sized American employers at a price per employee per month, is the group's clearest case of profit without volume. Revenue grew 9.8% to $531.2m last quarter; operating income rose 50%, lifting margin from 23.2% to 31.7%, on more than $100m of annual research savings from its own automation and a $30m cut in third-party AI token fees. It bought back a fifth of the company for $1.4bn this year, taking diluted shares from 56.3m to 45.9m. Management said client employment was stable with no acceleration — and US nonfarm payrolls actually fell by 23,000 in July, which is the denominator for every per-employee biller here. Consensus has Paycom earnings growing 30.8% this year on revenue growth of 7.7%.

One premise worth retiring: interest on client payroll balances is not a headwind. Paycom's guide includes about $105m of interest on funds held for clients, on average daily balances of roughly $2.9bn, up 9%, and the Federal Reserve held at 3.50-3.75% on 29 July with three dissenters preferring a hike.

What the prices now assume

Because Bill runs a GAAP loss, earnings multiples are useless for comparison; gross margins across the three run 76-83%, so price against trailing gross profit is the usable common measure. Workday has gone from 4.61x in mid-May to 6.76x, against 9.17x a year ago — half the de-rating recovered before the confirming print. Paycom has gone from 4.23x to 5.95x against 7.94x, at 18.4x forward earnings and a 7.5% free-cash-flow yield, on 7-8% guided revenue growth. Bill sits at 3.49x, versus 3.04x in May and 3.58x a year ago: no dearer than before the sector's trough, despite gross profit compounding 12.4%, at 14.2x forward earnings and an 8.9% free-cash-flow yield.

The trend has, in fact, turned. Seven of the thirteen — including Workday and Paycom — ended 19 August with their 50-day average above the 200-day, after bearish readings in May. The twelve-month picture is still ugly: the group is down 11.5% on average, Monday.com by 48.3%. What separates Bill is that its moat is not a seat. Payments cannot be routed around money-transmitter licensing, know-your-customer and anti-money-laundering compliance, or bank partnerships, and rivals are arriving by function rather than scale — Ramp and Brex pushing from corporate cards into payables, Tipalti in global mass payments, Melio in small-business bill pay. Automation there adds throughput rather than removing billable units.

The setup

Where it stands — Bill delivered the group's only new numbers and was the only member sold on them; Workday and Paycom rallied on a rotation and a buyback. Would confirm — Take rate holding above 33bp with payment volume growth at or above 14% in the September quarter. Would invalidate — Core revenue growth falling toward the low end of the 11-14% guide, or net customer adds below the ~1,800 just posted. Watch next — Workday reports fiscal second-quarter results after the close on 27 August; 12-month backlog growth above 15% is the test. Valuation — Bill at 3.49x trailing gross profit and 14.2x forward earnings, against 3.04x in May and 3.58x a year ago.

Fabrinet, a Thailand-based contract manufacturer that assembles optical transceivers, tunable lasers and switching modules to other companies' designs, closed its fiscal year with revenue of $4.64bn. That was growth of 35.7%. Its gross margin for the year was 11.99%, against 12.09% the year before — scale, in this business, buys volume but not price.

That single line explains more about the past week in optical networking than any demand datapoint. Fabrinet's fiscal fourth quarter, reported on 17 August, set a revenue record at $1.32bn, up 45% and above its own guidance, with non-GAAP earnings of $4.10 a share against $2.65 a year earlier. Management then guided the current quarter to $1.375-1.425bn, above the $1.321bn consensus. The shares fell 20.25% the next day, the largest single-session decline across a 110-name list of AI hardware suppliers, and pulled the rest of the optics complex down with them.

Build-to-print has no operating leverage

Fabrinet does not own the designs it builds. It qualifies processes, then manufactures at the margin its customers permit, and its fiscal 2026 annual report discloses four customers at 10% or more of revenue — Cisco at 20%, NVIDIA 16%, Nokia 11% and Amazon 11%, or 57.4% between them. Revenue growth accelerated through the year, from 21.6% to 44.6%, while gross profit stayed pinned near a twelfth of it. Chief executive Seamus Grady told the call he sees "no end in sight" to data-center demand. Consensus nonetheless models fiscal 2027 revenue growth of 23.6%, a marked step down.

The systems and the glass tell a different story

Ciena, which sells coherent optical transport platforms and network software to carriers and, increasingly, to cloud operators, is the counterweight. April-quarter revenue rose 39.5% to $1.57bn, operating margin reached 15.1% against 3.1% a year earlier, and backlog climbed more than $600m sequentially to $7.7bn as the company raised its full-year outlook to about $6.3bn. Roughly 46% of revenue now comes from cloud and hyperscale customers.

Corning is not an optics pure-play but a materials group spanning display glass, ceramics, solar and labware. Its Optical Communications sales rose 32% to $2.07bn in the June quarter, with segment net income up 77% to $438m and enterprise optical — the data-center line — up 65% to $1.27bn. It has signed long-term agreements with Meta, worth up to $6bn, plus NVIDIA and Amazon, and will expand US optical connectivity capacity tenfold under the NVIDIA deal. Management says the margin gains come from connectivity design, not fiber pricing, and that demand still exceeds supply. Its weak spots are elsewhere: solar lost $7m on a plant shutdown, and glass for handsets grew 1%.

What actually moved

On 4 August, reports that the Federal Communications Commission was drafting a ban on imports of new Chinese optical transceiver models added 8-16% to these names in a session; Counterpoint Research put the loser at China's Innolight, near 27% share. Between 17 and 19 August that rally reversed entirely — Fabrinet down 22.3%, Corning 12.0%, Ciena 10.3% — helped by Anthropic's reported $65bn revenue run rate against an $80bn expectation and a 30-year Treasury yield at 5.323%, a 19-year high.

On price to trailing gross profit — the honest lens for a thin-margin assembler and a mixed conglomerate alike — Fabrinet has fallen from about 48x in mid-May to 29.3x, Ciena from 35.5x to 23.6x, Corning from 25.6x to 21.3x. In all three cases trailing gross profit was still growing. The caveat is the anchor: over twelve months Ciena remains up 350%, Corning 136% and Fabrinet 59%. This is a de-rating from an extreme. Co-packaged optics, the structural bear case, is not the cause — LightCounting and Cignal AI both expect no material displacement of pluggable modules for three years, and Corning frames inside-the-box photonics as a $10bn platform where it currently has no content at all.

Fabrinet's downtrend predates the print: its 50-day average crossed below its 200-day in early July. Corning and Ciena only turned in mid-August, after the businesses reported.

The setup

Where it stands — Three interconnect suppliers with accelerating revenue have de-rated by a third or more since May on multiples, not results. Would confirm — Fabrinet's October quarter landing inside its $1.375-1.425bn guide with gross margin at or above 12%. Would invalidate — Ciena reporting backlog below $7.7bn, or Corning trimming its Optical Communications growth outlook. Watch next — Ciena's fiscal third quarter, before the US open on 3 September 2026. Valuation — Fabrinet at 29.3x trailing gross profit and 23.9x forward, against roughly 48x in mid-May.

EnerSys, which builds the reserve-power systems that keep data halls, telecom sites and warehouse forklifts running when the grid or the shift changes, reported its fiscal first quarter on 13 August. Revenue reached $935.6m, up 4.8%. Operating income rose 75% to $151.4m. Orders from data-center customers were up 80% from a year earlier with 12 to 36 months of visibility, and total book-to-bill came in at 1.06x — more ordered than shipped. The shares are down 18% over three months.

That looks like a divergence. Read one line further down and it is mostly an accounting question.

The credit inside the margin

EnerSys reported gross margin of 33.5%, up 510 basis points, and 28.5% excluding the Internal Revenue Code Section 45X advanced-manufacturing credit. The credit is therefore worth roughly five points of gross margin. On the bottom line it is starker: adjusted earnings of $3.66 a share fall to $2.41 without it — about a third of the total. Strip out a $30.9m tariff refund as well and the underlying figure is $1.78, still up 42%.

The dependence is growing, not fading. Second-quarter guidance of $3.15-3.25 embeds $42-47m of 45X benefit; ex-credit, the guide is $1.95-2.05. And the credit's terms have changed. Under last year's tax law, a primary component integrated into a secondary one at the same plant now requires that at least 65% of direct material costs come from US-sourced primary components, for tax years beginning after 31 December 2026 — which reaches EnerSys's fiscal 2028.

So does the price make sense? A year ago the stock closed at $99.13 against fiscal 2025 earnings of $8.99 a share, near 11x. Today it is close to $194 on trailing earnings of about $9.34 — 20.3x. The multiple almost doubled in twelve months; the three-month fall gives back a slice of that. Forward, 14.7x on consensus of $13.18 for fiscal 2027 is the counterweight, but that consensus contains the credit.

The competitive position is narrower than the order growth suggests. EnerSys sells batteries into data-center power systems dominated by Schneider Electric, Vertiv and Eaton at the integrated-platform level. Its incumbency is in lead-acid, which still holds about 62% of the data-center battery installed base while lithium grows fastest — so the company must win a lithium socket it is only now sampling. Its answer, a high-density lithium cabinet claiming 2.5 times a rival's volumetric density, has 100 units on order and revenue expected in fiscal 2028.

What actually fell

The five smaller names classified alongside EnerSys did not sell off together. Over the past month Enovix fell 29.5% and T1 Energy 25.4%; the other four averaged -2.2%, and EnerSys rose slightly.

Enovix, a Fremont maker of silicon-anode cells for phones, smart eyewear and drones, dropped 18% in one session on 17 August after chief executive Raj Talluri resigned to run Kulicke & Soffa, followed by a William Blair downgrade. Remove those two sessions and its month is roughly flat. Its business went the other way: revenue of $9.0m, up 21%, a fifth straight quarter of growth, positive gross margin at 14.4%, and a defense and drone pipeline up 41% sequentially to $183m. It holds $552m of cash against a $31.4m quarterly outflow — about four years — meaning cash is three-quarters of the market value and the operating business is priced near $190m.

T1 Energy is not a storage company at all. It makes solar cells and modules domestically, and it fell after disclosing on 28 July that capital costs at its Austin cell fab rose about 20% to $510m with first production slipping to the first quarter of 2027, then priced $120m of 4.75% convertible notes.

Eos Energy, which sells zinc-halide batteries for long-duration grid storage, is the one genuine storage pure-play, and it is diverging in both directions at once: revenue of $68.8m, up 351%, record backlog of $807m, against gross margin of -71% and a diluted share count up 43% year on year to 339.8m. Its sales multiple has collapsed to 5.05x trailing from 18.85x in early May. Amprius, making silicon-nanowire cells for drones and aerospace, is the quiet performer — revenue of $34.0m at 27.3% gross margin, near-breakeven earnings before interest, taxes, depreciation and amortization, and 13.8x sales against Enovix's 20.7x. Solid Power, a pre-commercial solid-electrolyte developer for electric vehicles, booked negative revenue in the quarter on an accounting reversal.

Demand is not the problem

US operators added a record 15 GW of utility-scale storage in 2025 and plan roughly 24 GW in 2026. What changed is cost and policy: the Section 301 duty on Chinese non-EV lithium cells rose from 7.5% to 25% on 1 January, in a market where China supplies over 80% of US storage cells, and battery-grade lithium carbonate has roughly tripled from its 2025 trough even as pack prices set fresh lows. Over three months every name here fell, averaging -31%, alongside Fluence at -45% and Vertiv at -31% while the S&P 500 rose. That is the whole power complex losing multiple, not storage losing customers.

The setup

Where it stands — EnerSys's order book is growing while its multiple deflates, with a third of profit resting on a federal credit. Would confirm — Fiscal second-quarter adjusted earnings ex-45X landing at or above the guided $1.95-2.05. Would invalidate — Data-center orders falling below book-to-bill of 1.0x, or the lithium cabinet slipping past fiscal 2028. Watch next — EnerSys fiscal second-quarter results, due in November; Eos's second-year Department of Energy loan tranche. Valuation — 20.3x trailing and 14.7x forward, against roughly 11x trailing a year ago.

Analog Devices completed a $1.5bn all-cash purchase of Empower Semiconductor on 7 July. Empower builds integrated voltage regulators and silicon capacitors — parts that sit on or beside the processor package and perform the final voltage conversion before current reaches the chip. ADI, a Wilmington, Massachusetts maker of data converters, amplifiers and power-management parts for industrial, automotive and communications customers, said when it announced the deal in May that Empower's regulators cut system power consumption by roughly a fifth. On its 19 August call management put the prize in dollars: about 100 gigawatts of data-center construction between 2026 and 2031, worth $1-1.5bn of analog content per gigawatt.

That is the same few square inches where Monolithic Power Systems, a fabless designer of DC-to-DC converters and vertical power modules, earns its valuation. And it is contested. Nvidia's named partner ecosystem for next-generation high-voltage direct-current racks lists ten suppliers — ADI and Monolithic Power alongside Infineon, Renesas, ON Semiconductor, Texas Instruments, STMicroelectronics, ROHM, Navitas and Innoscience.

The concentration fear its own numbers refute

The discount attached to Monolithic Power traces to a specific episode: in November 2024 a broker note argued its allocation in Nvidia's Blackwell line was at risk, with Renesas and Infineon absorbing share, and the shares fell about 25% in a session.

The June quarter says something else. Revenue reached a record $981m, up 47.6%, the fourth straight quarter of acceleration from 18.9% a year ago. Operating income grew 84%, lifting operating margin to 31.0% from 26.5%. On the 30 July call management said enterprise-data revenue rose 45% sequentially "with no concentrated customers," raised the floor of its full-year growth guidance for that business from 85% to 130%, and put its share of central-processor power above 30%. It is sampling 800-volt AC-DC parts built on its own silicon carbide, and doubled its buyback authorization to $1bn.

Where the premium sits now

Gross margins differ by twelve points across these three, so price against trailing gross profit is the comparison that holds. Monolithic Power stands at 35.4x, down from 48.7x three months ago even as trailing gross profit grew 17.3%. Its forward earnings multiple of 47.5x has slipped below the 50-60x range that had it labeled stretched in May. It remains the most expensive of the group by a wide margin — ADI is at 19.9x, Microchip 13.6x, and NXP Semiconductors, the Dutch automotive chipmaker, 7.7x.

ADI's own quarter was its first above $4bn: revenue of $4.02bn, up 39.6%, with operating margin at 40.1% against 30.7% three quarters earlier. Communications grew 84%, and data center is now 80% of that segment. Industrial, still 49% of revenue, grew 53% — the AI build is additive, not the whole story. Free cash flow ran at $4.9bn over twelve months, 36% of revenue. The shares fell on both the day before and the day of the print.

Microchip Technology, the Chandler, Arizona microcontroller house, is the recovery case rather than the AI case. Gross margin has expanded five quarters running, 53.6% to 63.2%, and operating margin swung from 3.0% to 22.7%. Bookings were the strongest in about four years and distributor inventory sits at 25 days, the low end of its range — a channel drained, with restocking still ahead. Its own fabs are far from full, with $450m of equipment idle. September revenue is guided up about 40% year on year.

One mechanism sits under all three margin lines: analog suppliers are recapturing two years of absorbed input costs through outright price increases, ADI averaging 15% across its line from 1 February. Price falls almost entirely into gross profit.

What actually happened to the shares

The de-rating is a three-month event, not a slow bleed: ADI is down 12.5% since mid-May, Monolithic Power 19.4%, Microchip 20.6%. Two sessions did most of it. On 28 and 29 July chip stocks shed more than $1trn on evidence of Chinese progress in advanced chipmaking and doubts about AI capital spending — SK Hynix fell 14.65% in a day — taking the three down 8.6%, 10.7% and 16.1% in forty-eight hours. Both earnings gaps inside the window ran the other way: Microchip rose 13.9% on 7 August, Monolithic Power 8.4% on 31 July. All three now trade with their 50-day averages below where they sat in June.

The setup

Where it stands — Monolithic Power's premium has compressed by a quarter while ADI bought its way into the same power socket. Would confirm — Monolithic Power holding enterprise-data growth above its raised 130% full-year floor at the next print. Would invalidate — ADI or Renesas naming a design win on a next-generation rack that displaces Monolithic Power's vertical modules. Watch next — Monolithic Power's September-quarter report, late October, and Microchip's guided $1.6bn September revenue. Valuation — Monolithic Power 35.4x trailing gross profit and 47.5x forward earnings, versus ADI 19.9x and 30.1x, Microchip 13.6x and 21.1x.

Three cooling makers, three unrelated quarters

Within a fortnight this summer, the three largest US-listed makers of air conditioning equipment reported results that had almost nothing in common. AAON, which engineers semi-custom rooftop units in Tulsa and, through its BasX division, builds air handlers and liquid-cooling systems for data halls, said revenue had doubled. Lennox International, which sells furnaces and heat pumps to American homeowners through independent dealers, cut its profit outlook and pushed the expected residential recovery into 2027. Trane Technologies, the $100bn maker of applied chillers and Thermo King transport refrigeration, raised guidance on record orders.

The shares behaved as though all three had reported the same bad quarter. AAON is down 37.8% since 19 May and Lennox 21.4% over the past month. Trane fell 2.7%. Because Trane is roughly 83% of the trio's combined value, the group has in fact barely moved; the damage is concentrated in two names, for two entirely different reasons.

The business is accelerating; the multiple is not

AAON's second-quarter revenue was $627m, up 101% from a year earlier — the fourth consecutive quarter of accelerating growth, from 17% to 42% to 54% to triple digits. Operating income rose 192%, and operating margin expanded to 11.0% from 7.6%. Backlog reached about $2.0bn, up 98% year on year, with the BasX book up 185%. The legacy rooftop business, sold into a soft commercial market, still grew 40%.

One line went the wrong way. Gross margin fell 223 basis points to 24.3%, and management cut full-year gross margin guidance to 25–26% from 27–28% while raising the sales guide to 55–60% growth. The cause is mechanical and disclosed: overhead at the new Memphis plant ran $18.1m in the quarter against $3m a year earlier. Strip that allocation out and the Oklahoma core segment's margin rose 60 basis points to 31.2%, with the BasX segment at 30.0%. Management expects a modest sequential lift in the third quarter and a meaningful one in the fourth, as late-2025 price increases now embedded in backlog flow through production.

Still, AAON's 24.3% gross margin is now the lowest of the three, against 34.9% at Lennox and 35.6% at Trane — an awkward fact for a company long described as earning structurally better margins than mass-market rivals.

Five sessions, two of them somebody else's

The decline was not a grind. Five days — 1 and 2 July, 28 and 29 July, and 10 August — compound to -38.7%, more than the entire move since 30 June, meaning every other session was net positive. Two of those five were peer earnings days. Carrier reported on 28 July and fell 8.9% on margin compression, even as its Americas residential orders rose about 70%; Lennox reported the next morning. AAON's own results did not arrive until 10 August.

Lennox's fall, by contrast, is one event. The company cut full-year adjusted earnings guidance to $23.00–$24.00 and moved the demand recovery to 2027. Home Comfort Solutions revenue fell 7% with unit volumes off 12%. There is a real mechanism behind it: the 30-year Treasury yield reached about 5.33% on 18 August, its highest since 2007, and financed replacement demand follows housing turnover. Less noticed, Lennox's commercial arm grew 24% and its growth guidance was raised to roughly 20%.

Trane is diverging from both. Organic bookings rose 37% to a record $12.1bn backlog, up 70% with Americas commercial HVAC orders at an all-time high, and it raised full-year guidance and its dividend. It is also arming directly at AAON's socket: its Stellar modular chiller platform is tracking $500m of revenue, coolant-distribution and LiquidStack acquisitions are ahead of plan, and capacity is being quadrupled over three years under long-term agreements with hyperscalers.

Where the prices sit

AAON trades at 13.4x trailing gross profit, against roughly 25x in May and 18.4x in early May, while trailing gross profit itself grew from $386m to $494m. Forward earnings put it at 35.8x, or 23.4x on 2027 consensus. The offsets are genuine: free cash flow yield is negative at -1.8% on $103m of year-to-date capital spending, BasX bookings were sequentially lighter than four prior quarters that ran near three times book-to-bill, and no customer above the 10% revenue threshold is named in the latest 10-Q, leaving concentration in a hyperscaler-weighted backlog unquantified.

Lennox at 8.3x trailing gross profit, down from 11.1x in February on flat gross profit, and a 5.1% free cash flow yield, has largely absorbed its own bad news. Trane's 12.8x has hardly moved in six months, and its 30.1x forward multiple sits just below the 32x that has been the standing caution on the name. AAON's trend turned down only in August; the business, so far, has not.

The setup

Where it stands — AAON's orders and operating leverage are improving while its multiple has compressed by nearly half since May. Would confirm — Third-quarter gross margin above 24.3% sequentially, with total backlog holding at or above $2.0bn. Would invalidate — BasX backlog falling sequentially, or full-year gross margin guidance cut below 25%. Watch next — AAON's third-quarter results, due early November, the first test of the promised margin recovery. Valuation — 13.4x trailing gross profit versus about 25x in May; 35.8x forward earnings, 23.4x on 2027 consensus.

Three quarters, three different businesses

In the first week of August the three largest listed quantum-computing companies reported June quarters that had almost nothing in common. IonQ, a trapped-ion computer maker in College Park, Maryland that sells machine time through Amazon's Braket and Microsoft's Azure Quantum, booked revenue of $80.05m — nearly four times the year-earlier figure — and raised full-year guidance to $280m-$290m. Rigetti Computing, a 163-person builder that fabricates its own superconducting processors, grew revenue 183% to $5.14m. D-Wave Quantum, which sells quantum annealers, a narrower class of machine built for optimization problems rather than general computation, reported $3.08m, marginally below the prior year.

The divergence runs deeper than the growth rates. IonQ's expansion came from shipping boxes: gross margin fell to 24.9% from 56.5% in the fourth quarter of 2025. Management said 60% of the quarter's revenue was non-government and half was international, and reported remaining performance obligations of $485m against $122m a year earlier. Rigetti was the only one of the group whose margin improved, to 42.6% from 31.4%, on sales of its Novera on-premises processors — but its 2025 revenue of $7.09m was 34% below 2024, so the base is not yet compounding.

D-Wave's number is the one that needs explaining. Its recurring cloud-subscription line grew 50% to $1.9m, but first-half revenue fell 67% because the prior-year half contained $13.7m from the company's first outright system sale. Bookings of $35.5m for the half look spectacular until you remove the $20m agreement signed in January with Florida Atlantic University for a single Advantage2 machine. Second-quarter bookings alone were $2.1m. Adjusted losses before interest, taxes, depreciation and amortization widened 85%, to -$37.1m.

What the forward multiples assume

All three are loss-making, so earnings multiples are meaningless; sales and gross profit are the honest lenses. D-Wave trades at 571 times trailing sales and 165 times forward sales. That compression is not a business improving — it is the estimate assuming two unshipped annealers land by December at $20m-$40m apiece. Rigetti's 423 times trailing falls to 239 times forward only if it books $14.1m in the second half, roughly 60% of which is a single $8.4m order for a 108-qubit system from India's Centre for Development of Advanced Computing, recognized on installation and acceptance.

IonQ is the cheapest of the three on sales, at 65.7 times trailing and 55.9 times forward. On gross profit it is 181 times trailing and 154 times forward. The gap between those two ratios is the arithmetic price of a 25% margin.

The cost of the cash

None of these companies funds itself from operations, and the share counts show it. IonQ's diluted count rose 46.5% year on year to 367.7m, before roughly 24m more shares went out to close its $1.8bn purchase of SkyWater Technology, the only US quantum foundry running at 200mm scale. Its $3.0bn of cash and investments falls to about $2.0bn after that deal — near $5.10 a share, or 12% of the equity value.

D-Wave's diluted count rose 25.9%; it holds $546.2m, about $1.47 a share, after a $400m at-the-market offering last June that followed a $150m one in January 2025. Rigetti holds $541.3m with no debt and burned about $28m in the quarter — a long runway, but $1.62 of cash per share against a $17.00 price. Quantum Computing Inc., which draws 70%-80% of revenue from government subcontracting, is the odd one out: roughly 64% of its $2.02bn market value is cash, and it trades at 1.27 times book, yet its June-quarter gross margin was negative and its share count rose 58.9%.

Washington became the marginal funder

In May the Commerce Department announced $2.013bn of letters of intent with nine quantum companies — IBM, GlobalFoundries, D-Wave, Rigetti, Infleqtion, PsiQuantum, Quantinuum, Atom Computing and Diraq — taking a minority equity stake in each as a condition of the money. Rigetti's slice runs to $100m over three years, and the terms contemplate issuing the government new shares priced at the lowest close on three specified dates, discounted a further 15%. IonQ and Quantum Computing Inc. are not on the list. Separately, the Defense Advanced Research Projects Agency advanced 11 firms to Stage B of its Quantum Benchmarking Initiative, IonQ among them; Rigetti says its qualification is still in progress.

The shares, briefly

From 20 July to 19 August, Infleqtion — which listed in February through a merger with Churchill Capital X and took in more than $550m — rose 38.1%, IonQ 26.6%, Rigetti 19.3%, D-Wave 15.6% and Quantum Computing Inc. 6.3%. The gains were not a grind: two sessions around IonQ's 5 August print delivered roughly 19 of its 26.6 points. All four of the older names remain below their 31 December closes, D-Wave by 26.1%. And the advance happened while the 30-year Treasury yield topped 5.33%, a 19-year high — the longest-duration equities in technology rising straight through a discount-rate shock.

The setup

Where it stands — A partial repair of a year-long drawdown, with three businesses whose revenue engines have almost nothing in common. Would confirm — D-Wave shipping both annealing systems by December and Rigetti's $8.4m C-DAC installation being accepted in the fourth quarter. Would invalidate — A third straight D-Wave quarter near $3.1m, or a fresh at-the-market equity program at any of the three. Watch next — IonQ's investor day at the New York Stock Exchange on 8 September, when combined IonQ-SkyWater guidance is expected. Valuation — D-Wave 571x trailing sales, 165x forward; Rigetti 423x and 239x; IonQ 65.7x and 55.9x, or 154x forward gross profit.

The largest single line item at the companies that assemble and brand enterprise servers and storage arrays is no longer the processor. It is memory. And on 28 May, NetApp — which sells all-flash storage arrays running its ONTAP data-management software, plus first-party storage services inside Azure, Amazon Web Services and Google Cloud — told investors that the cost of that memory would take roughly two points off its gross margin this fiscal year.

That is a guidance cut, delivered in the same breath as record results. NetApp's revenue rose 12.5% to $1.95bn in the January-to-April quarter, a fourth consecutive acceleration from just 1.2% three quarters earlier. All-flash array revenue hit a record $1.2bn, up 18%. Yet reported gross margin has now slipped two quarters running, from 71.96% to 70.58% to 70.07%, and the fiscal 2027 guide of 68.5–69.5% sits below all of them. Management says pricing actions are being taken to protect profitability. The guide says they are not yet enough.

Why the bill exploded

Memory makers have been shifting constrained wafer capacity toward high-bandwidth memory for artificial-intelligence accelerators, starving the conventional server DRAM that ordinary enterprise machines need. Contract prices for that conventional DRAM rose roughly 93–98% quarter on quarter in the first quarter of 2026, according to TrendForce, with NAND flash following. The cost of a memory-heavy 512-gigabyte, dual-socket server build has roughly doubled since January 2025. The rate is now decelerating — TrendForce expects 13–18% for server DRAM in the third quarter — but the level is set.

Dell absorbed it; opex saved the quarter

Dell, the largest of the three and the volume leader in AI servers with roughly a fifth of the market, is the clearest case. Revenue in the quarter to 1 May rose 87.5% to $43.8bn. Gross profit rose only 57.6%. Gross margin fell 337 basis points to 17.75%, meaning the incremental margin on $20.5bn of brand-new revenue was about 13.9 cents on the dollar. Operating margin still expanded, to 8.34% from 5.27% — but only because operating expense grew 11.4% against revenue growth of 87.5%. That is expense leverage, not pass-through.

Dell is trying. It raised PowerEdge server list prices in December and lifted commercial PC and monitor prices 10–30%. Chief operating officer Jeff Clarke told investors the company is "repricing, it feels like, every day", with quotes valid for the shortest period ever. Demand is not the issue: Dell entered the quarter with a $51.3bn AI-server backlog, booked $24.4bn of new orders and lifted full-year revenue guidance to $165–169bn.

HPE is a different story wearing the same clothes

Hewlett Packard Enterprise sells ProLiant servers and, since the Juniper deal, a large switching and routing business. Its gross margin expanded 894 basis points to 36.52% — the opposite of Dell. But that is acquisition mix, not pricing power: Networking revenue rose 148.2% to $2.7bn while the segment's own operating margin fell to 21.6% from 25.0%. Server revenue grew 32.7% to $5.5bn. The genuine progress is financial: net debt to adjusted EBITDA improved to 2.6x from a pro-forma 3.1x, with free-cash-flow guidance raised to at least $2.0bn, per its 10-Q. Barron's midyear panelists judged Dell's hyperscaler servers superior to HPE's on total cost of ownership — HPE is following, not leading, on hardware.

The multiple did the work

Because the three earn 18%, 37% and 70% gross margins, price against gross profit is the only lens that compares them. Dell has gone from 3.69x trailing gross profit in mid-February to 6.88x in May to 11.37x now; its trailing gross-profit dollars grew 19.1% over that span. HPE went 2.87x to 5.50x on 29.7% more gross profit. NetApp went 4.28x to 7.79x on 5.1%. The advance is overwhelmingly re-rating.

Nor was it a quiet grind. Dell has printed 27 sessions of 5% or more since 31 March, including a 32.8% gain the day after results and an 11.3% drop on 28 July. HPE jumped 24.0% on 2 June, NetApp 22.4% on 29 May. All three fell together on 19 August inside a broad selloff in AI hardware and memory names, and all three remain in uptrends, 50-day averages above 200-day.

The setup

Where it stands — Revenue and backlog are validated at all three; the memory cost is landing on gross margin and being offset below it. Would confirm — Dell reporting flat or higher gross margin on sequentially higher AI-server revenue. Would invalidate — NetApp narrowing its fiscal 2027 gross-margin guide back toward 71%, or Dell's incremental gross margin rising above 20%. Watch next — Dell reports 1 September; HPE and NetApp both report after the close on 2 September. Valuation — Dell 11.37x trailing gross profit against 8.81x forward, versus 3.69x in February; HPE 5.50x/4.75x; NetApp 7.79x/7.17x.