DK Street Journal

Sterling Infrastructure, the Woodlands, Texas contractor that grades the ground and pours the foundations under new data centers, told investors on 4 August that signed backlog had reached $4.3bn, more than double a year earlier. It raised full-year guidance to revenue of $4.00-4.15bn and adjusted earnings of $19.70-20.30 a share. The shares fell 9.4% that day and roughly 14% intraday, investors focusing on the margin implied by the added revenue.

By then most of the damage was already done. Sterling fell 25% over four sessions into 28 July, before it reported anything. That week belonged to the customers, not the builders: Alphabet's 23 July capital-spending raise sent its own shares down 7% and set up Amazon, Meta and Microsoft to meet skeptical investors, with Meta dropping nearly 10% after guiding 2026 capital spending to $125-145bn on costlier memory, chips and data-center components. On 29 July, chip stocks shed more than $1 trillion of market value. The contractors were downstream of that.

Proximity to the pad, not distance from it

Which names fell inverts the intuitive reading. Over the past month Sterling is down 21.1%, MasTec 19.7% and Argan — the gas-fired power plant builder — 17.3%. Dycom, the West Palm Beach specialty contractor that places and splices fiber for carriers and cable operators, is down 3.2%. Quanta Services is up 3.2%. Widen the frame and the pattern holds: EMCOR gained 5.3% over the same month and IES Holdings 7.1%. This was not a construction selloff. It was a selloff in the companies whose order books read most directly as a bet on new campuses being started.

Only one of the falls has a disclosed cause inside the business, and it is not a data-center one. MasTec, the Coral Gables infrastructure builder, lost 18.9% in a single session on 31 July after lowering its communications outlook on a wireless slowdown and wireline projects deferred into 2027, with a raised earnings midpoint of $9.30 that still sat below where analysts had been. Its 18-month backlog hit a record $21.4bn, clean energy up 58%.

Dycom was supposed to be the control at the wrong end of this trade — a telecom builder with no compute load to serve. It is instead the most resilient member and the cheapest, at 9.6x trailing and 7.9x forward gross profit, 13.4x trailing enterprise value to EBITDA. Its last report, on 27 May, showed revenue up 56.1%, backlog of $11.9bn and raised fiscal 2027 guidance of $7.38-7.65bn; the honest caveat is that only 24.7% of that growth was organic and gross margin compressed about a point, to 14.0%. Long-haul fiber between campuses is compute infrastructure too.

What Sterling actually conceded

Sterling's second quarter was not ambiguous. Revenue rose 90.1% to $1.168bn, the fourth consecutive quarter of acceleration. Gross margin widened to 24.18% from 23.29%, and operating income grew 113.8% — faster than revenue, which is what operating leverage looks like. More than 92% of E-Infrastructure backlog is mission-critical work: data centers, semiconductor plants and manufacturing.

The concession is mix. Sterling's acquired electrical business is growing 140% at roughly 12% operating margins, against site development in the high-20s, pulling segment guidance toward the mid-20s. Management also flagged softer third-quarter award timing and warned backlog could fall sequentially. That is a genuine reason to pay less per dollar of revenue — but the market has repriced the whole franchise, not the increment. Price per dollar of trailing gross profit has gone from 43.2x in mid-May to 19.6x, with 16.6x forward; trailing earnings are 36.9x against 78x earlier in the year. Consensus models $4.06bn of 2026 revenue and $19.78 of earnings — estimates that rose with the results.

Sterling's durable advantage is incumbency: once it wins the first phase on a campus it negotiates the next ones directly rather than re-bidding, and it is now selling electrical work alongside site work on three or four projects, up from none a year ago. The binding constraint is electricians, not orders.

Quanta is the different animal here. The Houston contractor reported record backlog of $53.4bn, up 50%, with remaining performance obligations of $33.6bn, revenue up 41.1% and gross margin at 16.17% against 13.21%. But technology and large-load work is 15-20% of revenue — four-fifths of that backlog is transmission, renewables and pipelines. It self-performs 80-85% of its work and spends $250m a year training craft against a four-year journeyman pipeline. It also carries the group's steepest absolute price: 72.4x trailing earnings, 33.7x trailing EV/EBITDA.

The channel that matters next

The transmission from AI sentiment to contractor order books runs through owners' financing costs, and it is tightening. Morgan Stanley notes investment-grade data-center secured bonds have widened 40 basis points since mid-June and high-yield equivalents 120 basis points, against $360bn of global fixed-income supply this year. That is a cost of capital for the people who sign the pads, and it shows up in awards long before it shows up in revenue.

A second leg down hit every name in the week to 21 August — Sterling off 14.6%, Argan 14.5%, MasTec 11.9%, Quanta 8.8% — and Sterling's 50-day average crossed below its 200-day on 20 August, ending an uptrend that had run since spring. Dycom's crossed on 19 August despite the shallowest decline of the five. Twelve-month gains are dented, not erased: Argan is still up 136.1%, Sterling 86.4%, Quanta 71.5%.

The setup

Where it stands — Sterling's backlog and guidance rose through a selloff that halved what buyers pay per dollar of its gross profit. Would confirm — Third-quarter E-Infrastructure margin holding at or above 24% with combined backlog flat or higher. Would invalidate — A sequential backlog decline that management attributes to cancelled or deferred campuses rather than award timing. Watch next — Dycom's fiscal second-quarter report in late August, its first disclosure since 27 May. Valuation — Sterling at 19.6x trailing and 16.6x forward gross profit, against 43.2x in mid-May; Quanta 20.4x, Dycom 9.6x.

Every large listed vendor of the plumbing that carries a business phone call, a text message or a one-time passcode has been repriced upward since February. Only one of them has a meter that visibly sped up. That gap between what the shares did and what the businesses did is now wide enough to price.

Four companies, one rerating

Twilio sells application programming interfaces (APIs) that let software developers embed voice, messaging, email and identity verification into their own products, and bills by the message and the minute. Bandwidth does nominally the same job but owns its carrier network and sells enterprise voice and SIP trunking, mostly in the United States; it is about 22 times smaller. Five9 sells cloud contact-center software to companies running customer-service teams, and RingCentral sells cloud phone systems and meetings to businesses replacing on-premise handsets.

Between 20 February and 20 August, Five9 rose 67% and RingCentral 66%. Their businesses did not move that way: Five9 grew revenue 10.3% in its June quarter with gross profit up 7.3% and gross margin slipping to 53.4%, and RingCentral grew 5.9% with gross profit up 7.0%. Whatever lifted this group, growth was not it.

Only Twilio's meter moved

Twilio's reported revenue growth accelerated for four consecutive quarters — 14.7%, 14.3%, 20.0% and 22.0% — to $1.5bn in the June quarter. The harder test is gross profit, which accelerated in parallel to 20.4%, trailing revenue by under two points. Dollar-based net expansion, the spend of existing customers a year on, reached 116%.

That matters because a large slice of the industry's reported growth is not growth at all. US carriers raised their application-to-person messaging surcharges again for 2026, with T-Mobile's revised schedule effective 19 January adding an inbound-message fee and lifting outbound rates. Those fees pass straight through: they add to revenue and an equal amount to cost of revenue. Twilio absorbed $71m of them in the quarter and guides to roughly $250m for the year. Strip them out and messaging grew about 18% rather than 28% — but gross profit still compounded at 20%, because the faster-growing pieces are software, not raw texts. Voice grew above 20%, with identity verification past 30%.

Bandwidth is where the same headline produced the opposite result. Revenue rose 22.2% to $219.9m, gross profit only 9.6%, and GAAP gross margin fell 410 basis points to 35.7%. Its cloud-communications line — the part that is not pass-through — grew 12%, and its owned-network enterprise voice grew 9%, less than half Twilio's rate on the identical product. The shares fell 29.3% on 29 July despite a revenue beat and raised guidance.

Bandwidth is not broken. Its gross-profit growth has accelerated four quarters running off a low base, adjusted EBITDA margin hit a record 18.3%, net leverage is steady at 1.6 times, and it now carries production traffic for Salesforce's Agentforce agent platform — revenue excluded from guidance. What it lacks is a cushion. It issued $316m of zero-coupon convertible notes due 2032 on 18 June at a conversion price near $72.64, about 37% above today's price, and bought back stock this year at an average $57.06. Both were struck against a share price that no longer exists; the stock peaked at $78.44 on 9 July and has fallen 41% since.

What the price now assumes

Twilio's trailing price/earnings ratio of 29 times is meaningless — a one-off $1.07bn net income sits in the quarter, which is why the forward multiple, 37 times, is higher than the trailing one. On price to trailing gross profit, the one lens comparable across four different margin structures, Twilio is at 12.4 times, against 4.9 for Bandwidth, 3.8 for Five9 and 3.0 for RingCentral. Six months ago Twilio itself was near 7 times, and in late May near 11. Free cash flow yield is 3.3%.

The counter-anchor is that on enterprise value to next year's revenue, roughly 4.7 times, Twilio still sits below its own ten-year median near 7.2 times. Nearly all of the advance arrived in three sessions — 1 May, 1 June and 7 August, the last following an 11% earnings beat — and outside those gaps Twilio lagged both Five9 and RingCentral over three months. Management raised full-year guidance twice, then guided September-quarter organic growth to 11-12% from 17% and warned that 5% beats should not be the norm. Consensus has revenue decelerating to 11.7% next year. The business earned its rerating; the multiple has moved further than the guidance.

The setup

Where it stands — Twilio's growth and gross profit both accelerated four quarters running, and it trades at three to four times peer multiples on gross profit.

Would confirm — September-quarter organic revenue growth printing at or above the 11-12% guide with gross profit growth again within two points of revenue.

Would invalidate — Gross-profit growth falling materially behind reported revenue growth, signalling the acceleration was carrier pass-through after all.

Watch next — Twilio's third-quarter results in early November; Bandwidth's, with the Salesforce Agentforce ramp still outside guidance.

Valuation — Twilio at 12.4 times trailing and 11.5 times forward gross profit, versus roughly 7 times in February and 4.9 for Bandwidth.

When Amkor Technology told investors on 27 July that its biggest end market would shrink in the third quarter, it also told them its data-center business would grow about 30% sequentially off a record quarter. The market took the first half of that sentence. The shares have since fallen to $50.09, some 42% below their 30 June peak of $86.23, including an 18.3% slide in the four sessions to 21 August. ASE Technology, which does more of the same work than anyone in the world, has barely moved.

Both are outsourced assembly and test houses — the rung of the supply chain where a finished silicon die is bumped, joined to its memory, encased and tested. Amkor, based in Tempe, Arizona, is the largest US-listed one, with about 30,800 employees and a $12.4bn market value. ASE, of Kaohsiung, is the global leader at $79.8bn, and also runs a large contract-electronics arm that dilutes its packaging economics. ChipMOS Technologies, a $1.9bn Hsinchu company, is the useful control: it tests and assembles memory and display-driver chips and almost nothing else.

The business did not deteriorate

Amkor's second quarter was the best in its recent history. Revenue reached $1.9bn, gross profit rose 75% to $318.6m, and operating income more than doubled. Factory utilization went from the low 50s to the high 70s in percentage terms, and management guided third-quarter gross margin to 18.5–19.5%. Its computing segment set a record, up 20% sequentially, on data-center demand and a new high-density fan-out processor ramp.

The damage came from communications — chips for phones. Amkor guided that book down by a high single-digit percentage against normal seasonality, on the migration of system-in-package work from Korea to Vietnam, changed customer build patterns and a shortage of memory components, and said the drag runs into the first half of 2027. Third-quarter revenue was guided to $1.95–2.05bn against consensus near $2.12bn. The stock fell roughly a quarter that day.

The memory shortage doing the damage is the same force reflating the rest of the industry. Standard DRAM contract prices rose 90–95% quarter on quarter in the first quarter of 2026 as Samsung, SK Hynix and Micron converted lines to high-bandwidth memory, where revenue per wafer is several times that of conventional DDR5. ChipMOS is what that looks like on the income statement: gross margin has risen every quarter for a year, from 6.6% to 18.0%, and operating margin from 0.4% to 12.8%.

Two prices for the same capacity

Amkor's trailing gross profit is 23% higher than six months ago, at $1.16bn. Its price per dollar of that profit has gone the other way — 12.7 times in February, 16.1 times in May, 10.8 times now. It trades at 22.4 times trailing and 19.8 times forward earnings, 9.7 times EV/EBITDA, and yields 4.1% on trailing free cash flow. The one genuine forward problem is Arizona: management expects depreciation and underutilization there to dilute margins, and consensus has it, with earnings per share of $2.52 this year but only $2.75 next.

ASE has not been repriced at all. Its multiple of trailing gross profit was 15.5 times in February, 19.1 times in May and 18.6 times today — roughly 1.7 times Amkor's — on 42 times trailing earnings and 6.6 times book. The business supports a lot of that. Its assembly, test and materials segment produced 66% of revenue but 94% of operating profit last quarter at a 27.3% gross margin, up 5.4 points, while the contract-electronics arm earned a 2.4% operating margin. Leading-edge packaging revenue is tracking above its $3.5bn guide and is targeted to double next year, and ASE raised advanced-packaging quotes by more than 20% in July. What it does not have is cash generation: capex was raised three times this year to a record $10.5bn, and free cash flow is guided negative into 2027.

The volume is being outsourced

The fear that foundries would keep advanced packaging for themselves is running the wrong way. Nvidia has reserved roughly 800,000 to 850,000 wafers of TSMC's chip-on-wafer-on-substrate (CoWoS) capacity for 2026, more than half the total, so overflow has to go outside. TSMC signed a ten-year agreement in June to buy packaging and test from Amkor's Peoria, Arizona campus; Nvidia widened its own Amkor deal in July with a prepayment to fund US capacity; and TrendForce reported on 5 August that TSMC is expanding outsourcing of the chip-on-wafer front-end step to outside assemblers.

The selling is also not an industry verdict. Amkor's 50-day average slipped below its 200-day on 10 August, the only one of the group to turn down; ASE and ChipMOS merely cooled, and Taiwan-listed Powertech fell 1.3% in a month. What sets Amkor apart is a phone book that shrinks while it borrows — $1.15bn of zero-coupon convertibles — to build American capacity that earns nothing until 2027 or 2028.

The setup

Where it stands — Amkor's AI and automotive packaging set records while its phone business shrank, and the shares carry the phone outcome. Would confirm — Third-quarter gross margin printing inside the guided 18.5–19.5% range with computing revenue up about 30% sequentially. Would invalidate — Computing growth stalling, or fourth-quarter guidance extending the communications decline beyond the flagged first half of 2027. Watch next — Amkor's third-quarter report in late October; ASE's fourth-quarter guide on whether packaging gross margin clears 30%. Valuation — Amkor at 22.4 times trailing and 19.8 times forward earnings, 10.8 times trailing gross profit against 16.1 times in May; ASE 18.6 times.

The week the cost of capital picked the winner

Investors spent last week re-sorting cloud companies by a question that had not mattered much before: who has to buy the hardware. On 18 August the 30-year US Treasury yield topped 5.33%, its highest since June 2007, driven by sticky inflation, deficit worry and a surge of corporate bond issuance from technology firms funding data centers. Operators that put graphics processing units (GPUs) on their own balance sheets were sold together — CoreWeave and Nebius among them. Nebius then fell again on a $4.5bn convertible note offering, new paper punished rather than applauded.

The three small companies in the rung of cloud below Amazon, Microsoft and Google split cleanly along that line, and not in the direction their results would suggest.

Nutanix: the multiple moved, the business did not

Nutanix sells the software layer of a private data center — virtualization, storage and networking bundled as the Acropolis stack, sold by subscription and run on servers the customer owns. Roughly 30,000 enterprises use it, and almost every new one arrives from VMware, whose pricing was overhauled after Broadcom bought it. Chief executive Rajiv Ramaswami says that migration wave "has years to run", with 500 to 1,000 new customers landing each quarter.

Because Nutanix owns no chips, its economics are the mirror image of a GPU landlord's: gross margin of 86.9%, and a trailing free-cash-flow yield of 4.25% against DigitalOcean's 0.11%. That is why it was flat in the week everything levered broke.

The difficulty is what has happened to the price in the absence of news. Revenue growth has decelerated four quarters running — 19.2%, then 13.5%, then 10.4%, then 10.0% on $703.1m in the February-to-April quarter. On 4 August the company told the Securities and Exchange Commission it would cut about 5% of its workforce, some 390 jobs, at a pre-tax cost of $33m to $43m. The genuine offset is operating leverage: operating income rose 51% year over year, lifting operating margin to 10.4% from 7.6%.

Against that, the shares have re-rated hard. The price paid per dollar of trailing gross profit was 4.68x on 3 May and is 7.57x today — a 62% expansion with no financial disclosure since the third-quarter report on 27 May, which showed annual recurring revenue of $2.43bn, up 15%. Forward earnings are capitalized at 30.5x while consensus has revenue growing 12.9% next fiscal year. Nutanix reports fourth-quarter and full-year results after the close on 26 August.

DigitalOcean: the numbers accelerated and the shares fell anyway

DigitalOcean rents computing by the hour to individual developers, start-ups and small businesses, on flat published prices, with 1,462 employees. Its June-quarter revenue reached $281.2m, a fourth consecutive acceleration to 28.6% growth. Annual recurring revenue from artificial-intelligence customers rose 212% to $234m, with inference — running trained models, not building them — now more than 70% of it. Remaining performance obligations, the contracted revenue not yet delivered, went to $894m from $71m a year earlier after the company signed its first nine-figure annual commitment, stretching average contract life past three years from 1.6.

The debit is on the cost line, and it is the whole argument. Gross margin has fallen every quarter for five, to 55.0% from 61.4%, because data-center expansion lifts depreciation and colocation costs months before those facilities earn anything. So gross profit grew 18.1% while revenue grew 28.6%, and reported operating income fell 17.5%. Management also cut its adjusted free-cash-flow margin guide to 11-13%. With roughly 155 megawatts of committed capacity against CoreWeave's more than a gigawatt, DigitalOcean cannot compete for training work and is betting on inference price-performance instead.

The shares ran to $135.40 on 17 August, nine trading days after the beat, then fell four straight sessions on rising volume. That has taken the price per dollar of trailing gross profit from 26.0x to about 22.7x. It still costs three times Nutanix on that measure while its gross profit compounds less than twice as fast.

Rackspace: the control, and not an AI story

Rackspace manages public and private cloud for enterprises from San Antonio. Second-quarter revenue was $670m, up 1%, with a $68m net loss, $2.79bn of debt and $202m of liquidity against an $836m equity value. Gross margin improved to 19.4% from 13.7%, and the loss widened anyway. On 9 July the company cut full-year revenue guidance by $150m and announced a $250m equity offering; the stock fell 33.6% that day. Its 179% twelve-month gain is a bounce off $1.17, and it has round-tripped. Consensus has revenue down 6.4% this year — the only one of the three pointing lower.

What would make Nutanix's price right

There is a real mechanism underneath the repatriation case. An enterprise survey commissioned by Cloudian found the share of companies using public cloud as their primary home for production inference fell to 41% from 56% in a year, with 56% now running or planning private-cloud inference on cost, latency and data-sovereignty grounds. Nutanix has shipped bare-metal Kubernetes aimed at GPU density to catch exactly that work. None of it is yet visible in the disclosed growth rate, which remains a VMware replacement story running at 10%.

The setup

Where it stands — Nutanix has re-rated 62% since early May on no new results; DigitalOcean has de-rated while every operating metric improved except margin.

Would confirm — Nutanix fiscal-fourth-quarter revenue growth above 12% with fiscal-2027 guidance ahead of the $3.20bn consensus.

Would invalidate — Growth printing near 10% again on 26 August, or DigitalOcean's gross margin falling below 55% next quarter.

Watch next — Nutanix reports after the US close on Wednesday 26 August; DigitalOcean guided the September quarter to $304-307m.

Valuation — Nutanix: 7.57x trailing and 6.50x forward gross profit, against 4.68x on 3 May; DigitalOcean 23.1x trailing, 19.9x forward.

Rambus, the San Jose designer of the small companion chips that sit between a server processor and its memory modules, crossed a line last quarter that it has been approaching for a decade: it earned more from selling silicon than from licensing patents. Product revenue reached a record $99.2m in the June quarter, up 22% from a year earlier, overtaking the royalty line that has defined the company since the 1990s.

That is a better business than the one the market is currently paying for — and a more cyclical one.

The annuity stopped growing

Total June-quarter revenue was a record $207.4m, up 20.4% year over year and an acceleration from 8.1% in the March quarter. Gross margin widened to 79.8% from 74.9%. But the royalty stream underneath has been flat for four quarters: $69.6m in March, an $84.2m spike in June on contract timing, and September guided back to $69-75m. All of the growth is now chips.

The chip growth has a clean mechanism. Rambus is paid per memory module, and the number of memory channels per server socket is rising — AMD at twelve, Intel moving from eight to twelve, the roadmap pointing to sixteen — with each channel adding a register clock driver and power-management part. Management's own long-term product gross-margin model is 60-65%, and it is running inside that; the headline 79.8% reflects the zero-cost royalty line, not the silicon.

The bear case attacks the denominator, not demand. Because Rambus earns a fee per module rather than a slice of the DRAM price, it needs modules to be built — and TrendForce expects registered-DIMM bit supply to grow only 15-20% as memory makers divert wafers to high-bandwidth memory for AI accelerators. Baird downgraded Rambus to Neutral on exactly that logic while keeping a $120 target. The market has priced the same idea across the tier: over 90 days Micron rose 26% and Marvell 24%, while Rambus fell 36.5%. Scarcity accrues to whoever owns the bits.

Microchip is not a classification error

Microchip Technology, a maker of microcontrollers and analog parts for cars and factories that also owns a PCI Express switch franchise from its Microsemi acquisition, has come out of a brutal inventory correction. June-quarter revenue was $1.485bn, up 38%. Gross margin hit 63.2% against 53.6% a year earlier, a fifth straight quarter of expansion, and operating margin recovered to 22.7% from a 3.0% trough. Inventory fell to 175 days, distributor stock sits at 25 days, and June was the strongest booking quarter in about four years.

Its data-center exposure is real: sales there were about $591m in calendar 2025, roughly 14% of revenue, and the company has guided its Data Center Solutions unit alone toward $500m in 2026 within a total data-center number approaching $1bn. Fourteen PCIe Gen6 design wins are booked; none ships in volume yet. Set against that: underutilization charges still run about $30.5m a quarter, and net debt of $5.2bn absorbs all free cash beyond the dividend.

The selloff belongs to analog, not memory

Over the past month Texas Instruments fell 10.4%, NXP 19.6%, ON Semiconductor 19.5% and STMicroelectronics 23.7%. Rambus, down 13.3%, and Microchip, down 10.7%, sit inside that band. On 18 August analog names sold off together with no company-specific catalyst.

The controls confirm it. Allegro MicroSystems, which sells magnetic position sensors to carmakers and has no memory product at all, grew revenue 27.5% to $259.2m with margin up — and fell hardest of the four, down 25.5% in a month and 47% below its 30 June high. Everspin Technologies, which makes magnetoresistive memory, is the one name up over 30 days; its 90-day collapse is the unwind of a spring spike on a $40m defense subcontract, after Kerrisdale Capital disclosed a short position arguing its largest end market is casino gaming.

What the price now assumes

Rambus trades at 16.7 times trailing gross profit, against 20.9 times in February and 28.0 times in May — while trailing gross profit rose from $537.6m to $592.0m over that stretch. It is cheaper on that measure than it was six months and a 28% twelve-month gain ago. Microchip's trailing price-to-earnings of 104x is trough-earnings noise; on gross profit it has gone from 17.5 times in February to about 13.6 times, with the profit base up 27% to $3.08bn.

Rambus peaked at $170.66 on 3 June and has lost 47% since, its 50-day average below its 200-day since late July; Microchip's averages crossed the same way on 11 August. Allegro alone still holds an uptrend. In each case the shares turned well after the numbers improved.

The setup

Where it stands — Two improving semiconductor businesses have de-rated alongside analog peers with no memory exposure, on multiples below February's. Would confirm — Rambus September product revenue lands in the guided $110-116m range, up about 20% year over year. Would invalidate — Rambus royalties print below $69m, or Microchip's December bookings fall back under 1.0 book-to-bill. Watch next — Rambus reports the September quarter in late October; Microchip's Hailo acquisition closes in September. Valuation — Rambus at 41.0x trailing and 30.0x forward earnings, 16.7x trailing gross profit versus 20.9x in February.

When GitLab widened access to its Duo Agent Platform in the quarter ended April 30, it did not attach a price to it. The agents went to users on the company's free tier, alongside automated security remediation and pipeline configuration, according to the company. GitLab sells a single application for planning, building, securing and releasing software, and it charges by the developer seat. So an agent that writes code is not a new thing to bill for. It is a reason for the seat to survive.

JFrog made the opposite choice, because its product forced it. JFrog runs Artifactory, the universal registry where enterprises store, version and distribute the packages, containers and binaries their software actually ships as. Volume through the registry is part of what customers pay for.

The meter is registering

JFrog's June-quarter revenue was $163.8m, up 28.7% — the fourth consecutive quarter of acceleration, from 25.5% a year ago. Cloud revenue rose 53% to $87.5m and now makes up 53% of sales, against 45% a year earlier. Management's explanation is unusually specific: customers are consuming above their contractual minimums because AI coding tools and new AI-native package types, including model-context-protocol servers, push more binaries through the registry. The company excludes that overage from guidance.

The corroborating detail sits below the revenue line. Trailing four-quarter net dollar retention reached 121%, three points higher than a year ago. Customers spending more than $1m a year went to 97 from 61, the company reported. Remaining performance obligations grew 38%, ahead of revenue. And gross margin widened to 77.9% from 76.3% — the hosting-cost squeeze that was supposed to arrive with a rising cloud mix has not. JFrog also added a fourth foundational artificial-intelligence lab as a customer, displacing an incumbent that could not handle the binary volumes.

The seat is eroding

GitLab's disclosed trajectory runs the other way. Revenue growth has decelerated four quarters running — 29.2%, 24.6%, 23.2%, 23.1% — and gross margin compressed 2.5 points to 85.8%, so gross profit grew 19.6% against revenue growth of 23.1%. Dollar-based net retention was 117% for the April quarter, down from 118% for the prior fiscal year and 123% the year before that, per the company's results. Remaining performance obligations grew 18%, slower than revenue — the mirror image of JFrog.

The structural problem is that source code is portable in a way a build pipeline is not. GitLab is the third platform in its own market, behind Microsoft's GitHub and Atlassian's Bitbucket, and the agentic workflow is consolidating around the work-tracking layer rather than the repository. Atlassian's June-quarter shareholder letter lists GitLab as one of several repositories feeding "code context" into its Teamwork Graph, while Atlassian ships a Jira coding agent that returns ready-to-review pull requests.

What the shares did

GitLab rose about 32% over the past month without disclosing anything. The moves were analyst target raises — Bank of America to $45 from $38 while keeping a neutral rating, UBS to $40 and Truist to $40 — and a 50-day moving average crossing above the 200-day on August 11. JFrog rose 10% and its trend was downgraded on August 13, a week after its best quarter; it remains 11% below its July 6 high. Over twelve months the ranking inverts: JFrog is up 87%, GitLab down 7%.

Both are now expensive against their own recent history. JFrog costs 23.4x trailing gross profit, up from 11.0x in February, and 92x forward earnings. GitLab, which is still loss-making on a reported basis, costs 8.00x trailing gross profit against 5.09x in May — a 57% expansion in three months — on 50.8x forward earnings, while consensus models revenue growth falling further, to 17.8% this fiscal year.

Atlassian, the group's largest name, supplied roughly a fifth of the segment's monthly advance in a single session after reporting revenue of $1.766bn, up 27.6%, and its first properly profitable quarter. The same report guided the coming year's revenue growth to about 13%, half of last year's, with data-center revenue declining.

GitLab's July-quarter results land after the close on September 1, the first test of whether free agents defend a paid seat.

The setup

Where it stands — JFrog's metered registry is accelerating on AI-generated volume; GitLab's per-seat platform is decelerating and gave its agents away free. Would confirm — GitLab's September 1 net retention falling below 117%, or JFrog holding cloud growth above 45% next quarter. Would invalidate — GitLab reporting reaccelerating revenue growth above 23.1% with separately priced Duo revenue disclosed. Watch next — GitLab Q2 fiscal 2027 results, September 1; JFrog's swampUP conference in September with Anthropic and Cursor keynoting. Valuation — JFrog 23.4x trailing gross profit and 21.6x forward, against 11.0x in February; GitLab 8.00x trailing, 7.21x forward, from 5.09x in May.

Fabrinet builds high-speed optical transceivers under contract, mostly in Thailand, for the networking vendors that sell them to hyperscale data centers. On 17 August it reported record quarterly revenue of $1.32bn, up 45% year on year, with data-center work now more than half the total — and then guided the December quarter to about 6% sequential growth. Its shares fell more than 20% the next session, and the semiconductor companies whose silicon goes inside those modules fell with it.

What did not move is the more useful fact. Over the five sessions to 21 August, Nvidia lost 3.6%, Broadcom 7.4% and Qualcomm 1.7%. Over the same stretch Fabrinet fell 26.3%, Coherent 17.4%, Credo 18.2% and Applied Optoelectronics 19.2%. The drawdown stopped at the interconnect layer. It coincided with the 30-year Treasury yield topping 5.31%, a 19-year high, a duration shock that lands hardest on the longest-dated growth stories — which is exactly what these are.

Three businesses, all accelerating

SiTime replaces quartz crystals with silicon oscillators and clock chips built on micro-electromechanical systems, and its content per AI server dwarfs its content per phone. Revenue grew 126.5% in the June quarter to $157.4m, the fourth consecutive quarter of acceleration from 44.8% a year earlier. Gross margin reached 63.0%, up roughly 950 basis points over that span. Its communications, enterprise and data-center line hit $101.2m, or 64% of revenue, up 181%. On 1 July it closed the $1.5bn purchase of Renesas's Timing Product Division, removing its most direct clock-chip rival and adding a business it says runs at 70% gross margin. The risks are its own: one consumer customer supplied $22.8m of revenue last quarter, and the carve-out still depends on Renesas for manufacturing and test.

MACOM sells analog chips across radio frequency, microwave and lightwave into three end markets, and that mix is the point. Data center reached $137.6m in the July quarter — about 40% of revenue — alongside a record $133.4m of industrial and defense and $71.3m of telecom. Total revenue grew 35.8% to $342.2m, gross margin expanded for a fourth straight quarter to 58.3%, and operating income more than doubled on that 36% revenue growth. Book-to-bill printed a record 1.6-to-1, with 800-gigabit and 1.6-terabit platforms dominating orders. Top-ten customers are now under 40% of revenue.

MaxLinear sells communications chips for broadband modems, Wi-Fi and optical transceivers. Its infrastructure line grew about 145% to roughly $85m on Keystone, a digital signal processor for 100-gigabit-per-lane optical links now in volume production at US and Asian hyperscalers; only three vendors ship such parts at volume. Guidance for the September quarter of $210-220m compared with roughly $174m of consensus.

What the price now pays for

Earnings multiples are useless here — SiTime's trailing price/earnings ratio is above 1,000x on near-zero GAAP profit and MaxLinear's is negative — so the honest comparison is price per dollar of trailing gross profit, and all three earn between 58% and 63%.

MACOM trades at 30.8x, against roughly 49x three months ago and 34x six months ago, while its trailing gross profit has grown 18% since February. That is the one place where price has moved decisively against the business. SiTime, at 58.6x trailing and 31.3x forward, is a third cheaper than in May but back to where it stood in February, and remains the most expensive name in the group; its forward figure also embeds acquired revenue. MaxLinear is the cheapest at 18.0x trailing and 13.8x forward, but it stood near 6x in February before a 324% twelve-month run — against that anchor, August unwinds a re-rating rather than creating a gap.

The path matters too. Of MaxLinear's 26% fall over the past month, 24 points came on 18 and 19 August, and the shares had risen 25% in the week before, from $69.16 on 11 August to $86.46 on 17 August, with no news of their own; the quarter was reported on 23 July. SiTime is still up 3.1% over thirty days despite surrendering its entire post-earnings gain. This was two sessions, not a grind.

The setup

Where it stands — Three accelerating businesses were repriced together on a contract manufacturer's guidance and a bond-yield shock, not on their own results. Would confirm — MACOM's October quarter landing inside its $415-425m guide with data-center revenue up another 35% sequentially. Would invalidate — Book-to-bill falling below 1.0 at MACOM, or SiTime's data-center line failing to grow sequentially in Q3. Watch next — MACOM's fiscal fourth-quarter report in early November; SiTime guided Q3 to $285-295m at about 68% gross margin. Valuation — MACOM 30.8x trailing gross profit and 27.0x forward, against about 49x in May; SiTime 58.6x; MaxLinear 18.0x.

The machines that push natural gas out of the rock and into a pipeline are mostly rented, not owned. Three companies control most of that rented horsepower in the United States, and in the first week of August they reported quarters that shared almost no characteristics — one shrinking, one setting records, one digesting an acquisition. The market has since treated them as a single trade, and a badly damaged one.

Three quarters, three businesses

Archrock, the largest US contract compressor, rents compression packages to gas gatherers and processors on multi-year fee contracts and runs an aftermarket parts arm alongside. Its second-quarter revenue was $371.2m, down 3.1% year on year — the first decline in two years — with operating income down 53.7%. Operating horsepower fell to 4.5m from 4.7m, reflecting the sale of roughly 165,000 non-strategic units, and period-end utilization slipped to 94.4%. Management tightened full-year adjusted EBITDA guidance to $865m–$885m from $865m–$915m, blaming lubricating-oil inflation and the cost of readying idle units for redeployment rather than demand. The fee-based core held: contract operations revenue rose 3% at a 71% margin, the company signed an eight-year contract covering about 665,000 horsepower with a midstream customer, leverage fell to 2.6x, and the dividend rose 10% at 3.1x coverage.

Kodiak Gas Services, the Permian-weighted specialist in large-horsepower units, did the opposite. Revenue rose 21.1% to $391.1m, adjusted EBITDA hit a record $217m, fleet utilization ran at 98.2% and pricing rose 4.5% to $23.80 per horsepower per month. It raised full-year guidance, and leverage of 3.1x is the lowest in its history after an $836m equity raise in May. Its newer power arm — 405 megawatts of distributed gas generation — earned a 64.5% gross margin on $33m of quarterly revenue, and Kodiak issued a limited notice to proceed on a sub-100 megawatt behind-the-meter project in West Texas for a hyperscaler, its first contracted data-center power deal. The pipeline of such projects doubled in a month.

USA Compression, an Energy Transfer-sponsored master limited partnership with the largest fleet by horsepower, grew revenue 36.8% to $342.1m on its J-W acquisition. Distributable cash flow reached $125.3m against $89.9m a year earlier, lifting distribution coverage to 1.65x from 1.40x — and it still deferred any distribution increase to fund horsepower growth.

The week they stopped being oilfield stocks

Between 22 and 28 July, Kodiak fell 15.7%, Natural Gas Services 9.6% and Archrock 8.0%. Schlumberger rose 4.8% and Baker Hughes 3.3% in the same days. What fell alongside the compressors was the power complex: Vistra down 10.9%, Vertiv 10.5%, Quanta Services 8.5%. These stocks were being sold as derivatives of the artificial-intelligence power build, not as oilfield services.

The difference is what came next. By 21 August, Quanta, Eaton and Constellation Energy had recovered to roughly their pre-selloff levels. Archrock closed at $31.79 against $38.36 on 22 July; Kodiak at $59.47 against $67.00. Three things happened in between: Texas Governor Greg Abbott ordered a freeze on new data-center approvals on 3 August pending an audit of their energy and water use; Archrock's guidance cut knocked 7.4% off the shares in three sessions; and the 30-year Treasury yield topped 5.33% on 18 August, a 19-year high.

The rate story is only half right. USA Compression is the most levered name at 3.72x and yields about 8% — precisely what a long-bond repricing should punish — yet it fell 1.1% over the past month while the growth names fell hardest. This is a growth-multiple unwind wearing a credit story's clothes.

What the volumes say

Demand shows no compression-specific crack. Caterpillar engine lead times sit near 195 weeks, so Archrock is ordering for 2029 delivery; Kodiak is already half-contracted on its 2027 units and USA Compression has secured about half of its own. Engine scarcity, not customer appetite, is the binding constraint, which is why incumbents with vendor relationships are the only firms able to add capacity. Haynesville output rose 7% in the first half, and more than 5.25 Bcf/d of new Permian egress arrives by late 2026 — pipe that requires compression to fill. The soft spots are real but modest: Permian gas growth is decelerating to roughly 4% a year, and July LNG feedgas averaged 17.4 Bcf/d, below the 18–19 Bcf/d seen earlier in the year.

What the price now assumes

Archrock trades at 17.1x trailing earnings and 18.5x forward — the forward multiple sits above trailing because consensus expects 2026 earnings per share of $1.72 against $1.84 delivered in 2025. Against 21.8x in mid-May, that is a de-rating the numbers largely justify. Kodiak's 9.83x trailing enterprise value to EBITDA is indistinguishable from Archrock's 9.84x and below USA Compression's 10.07x, despite the 24-point gap in revenue growth between the two. The offsetting caveat is Kodiak's 28.1x forward earnings, which requires 2026 EPS to nearly double to $2.11.

Archrock's 50-day average crossed below its 200-day on 19 August, and USA Compression has been in a clear downtrend since mid-July; Kodiak alone still holds an uptrend. The shares have kept falling since the wider power trade stabilized — which is the fact worth resolving.

The setup

Where it stands — Archrock's contraction is priced; Kodiak's record quarter carries the same EBITDA multiple as its shrinking rival.

Would confirm — Kodiak signing the firm long-term West Texas behind-the-meter power contract before year-end, as management guided.

Would invalidate — Kodiak utilization falling below 96% or pricing per horsepower turning negative sequentially in Q3.

Watch next — Third-quarter results in early November, and any lifting of the Texas data-center approval freeze.

Valuation — Kodiak: 9.83x trailing EV/EBITDA, 28.1x forward earnings; Archrock 9.84x and 18.5x, versus 21.8x trailing in May.

At its May-quarter results, Accenture stopped telling investors how much generative-AI work it was signing. The Dublin-headquartered consultancy — strategy, systems integration, cybersecurity and outsourced operations for large enterprises, 799,000 employees — had turned its "advanced AI" bookings line into the single number the market used to size enterprise AI demand. Management withdrew it on the grounds that AI is now pervasive across everything the firm sells.

What stayed on the page argued the other way. Total new bookings fell 2% to $19.3bn, a book-to-bill of roughly 1.03. The company then narrowed full-year revenue guidance to 3.0-4.0% in local currency, citing weak discretionary information-technology spending. Reported growth decelerated from 8.3% in the February quarter to 5.6% in the May quarter, and gross margin slipped to 32.77% from 32.87%. There is a genuine large-deal story underneath — 104 client bookings above $100m year to date, up 13%, and a seven-year, roughly €200m NATO secure-cloud contract won in July — but fewer, bigger programs are not the same as more revenue.

Since 22 July the shares are up 34%.

The two that raised numbers

Cognizant, the Teaneck, New Jersey outsourcer that delivers most of its work from India, is the only one of the three to lift its earnings outlook: full-year adjusted earnings per share guidance went to $5.70-$5.82 from a range implying 7-9% growth. Adjusted operating margin rose about 40 basis points to 16%, a sixth consecutive quarter of expansion, and trailing bookings grew 5% to $29bn with seven contracts above $100m signed in the quarter. The qualifications matter: revenue growth slowed to 4.5%, gross margin fell 30 basis points, net income declined 1.4% to $636m, and $84m of the quarter's costs went to a restructuring program that is part-financing the margin gain through severance.

Genpact is the surprise. Spun out of General Electric, it runs accounts payable, invoice-to-cash and financial planning desks for banks and insurers — per-seat process work, the first thing an AI agent is supposed to delete. Instead its revenue has accelerated for four straight quarters, to 7.1% in the June quarter, with its Advanced Technology Solutions line up 24.1% to $363m while the legacy back office grew 1.9%. Gross margin widened 62 basis points to 36.5%, the highest of the three, and the company raised both its AI-segment growth target, to at least 25%, and its earnings growth guidance. Genpact generates roughly $37,200 of revenue per employee against Accenture's $91,500 — the lowest-priced labor in the group, and the best numbers.

The offshore control

If AI work were simply accruing to listed services firms, Infosys and Tata Consultancy Services would be moving too. They are not. Over the same 30 days Infosys rose 6.6% and TCS 4.2%, and both sit below their 29 July closes while the American names kept climbing. Infosys in July cut its guidance for the coming year to 1.5-3.0% and named the reason: pricing improvement below plan. Its AI-first services reached 8.2% of revenue, about $417m in the quarter, while AI simultaneously raised client productivity expectations and compressed the traditional portfolio. Both halves are true at once.

So the buyer is not paying for AI revenue. The buyer is leaving AI infrastructure. Chip stocks shed more than $1 trillion around 29 July, and on 19 August the Nasdaq Composite fell 1.0% on another chip selloff and firm long yields — the session Accenture rose 3.8%, Cognizant 5.5% and Genpact 5.1%. In the four sessions to 21 August the three added 9.6%, 10.1% and 12.0%.

What the price now assumes

Accenture trades at 14.8x trailing and 13.5x forward earnings. Holding today's consensus constant, that forward multiple was 10.1x on 22 July, 12.8x in May and 15.5x in February: the estimates came down and the multiple went up. Evercore ISI's initiation makes the case explicitly on reversion to a historical average, not on upgrades. Cognizant at 10.9x forward and Genpact at 9.0x are both below where they traded in February on today's earnings, and both raised numbers — a different transaction.

One complication runs through all of it. A $100,000 fee on new H-1B petitions filed from abroad hits these firms hardest, and Forrester expects the effect to be higher onshore prices and more offshore delivery. Rising realized pricing next year may say nothing about AI at all. Cognizant, meanwhile, reports that customers are asking for "AI-infused rate cards" that embed model costs — the billable hour being replaced by outcome pricing rather than simply deflating — while one in four large enterprises has paused AI deployments outright.

The setup

Where it stands — All three US names have re-rated hard off June lows; only Cognizant and Genpact raised guidance behind it. Would confirm — Accenture's next quarter showing new bookings back above $20bn and full-year growth guidance held at 4%. Would invalidate — Genpact's Advanced Technology Solutions growth falling below 20%, or Cognizant's adjusted operating margin ending its expansion streak. Watch next — Accenture reports fiscal fourth-quarter results in late September, its first without a separate AI bookings line. Valuation — Accenture 14.8x trailing, 13.5x forward, against 15.5x forward in February; Cognizant 10.9x forward, Genpact 9.0x.

A billion-dollar quarter nobody paid for

Bloom Energy, which builds solid-oxide fuel cells that convert natural gas into electricity without combustion and sells them to data-center operators as on-site prime power, booked its first billion-dollar quarter in the three months to June. Revenue was $1.065bn, up 166% from a year earlier — faster than the 130% growth it posted in March. Gross margin widened to 33.4% from 30.0% the quarter before. Operating margin reached 17.1%, against 1.5% a year ago, and net income has now been positive for three consecutive quarters after a full-year 2025 loss of $88.4m.

Management raised full-year revenue guidance to $3.9–4.2bn and lifted the operating income target to $800–900m from $425–450m. It also disclosed that Brookfield's project-financing framework had been expanded to as much as $25bn from $5bn. Free cash flow was $175m in the quarter, so the Fremont and Newark factory expansions are being funded out of the business rather than by selling stock.

Why the fuel cell wins the socket

The demand mechanism is a queue. Roughly 2,600 GW of generation now waits for US grid interconnection, with median time to commercial operation approaching five years and nearly 80% of projects eventually withdrawing. The obvious alternative is no faster: heavy-duty H-class gas turbines run 54 to 60 months from slot reservation to delivery, with slots booked through 2029. Bloom quotes 90 days. That gap, not the chemistry, is the pricing power. On its July call the company said every major US hyperscaler and more than a dozen neocloud and colocation operators have approved its systems, that 80% of last year's bookings were repeat orders, and that its share of data-center fuel cells is in the "very high 90s."

What actually broke

Two things, neither of them orders. On 8 July the short-seller Hunterbrook Media alleged Bloom depends on Chinese scandium contrary to management's statements; Bloom rejected the claims and said its supply visibility covers 25 GW a year. Then on 24 July the stock fell 14.9% in a session on permitting delays at Oracle and American Electric Power projects, including a second regulatory rejection of a pipeline serving Oracle's Project Jupiter in New Mexico.

That is the durable question. A gigawatt of Bloom's latest servers consumes roughly 150 million cubic feet of gas a day. Close to 20 GW of distributed gas and fuel-cell capacity has been proposed across the US, which would be 4-5 billion cubic feet a day of new demand, drawn from the same production base that must also feed liquefied natural gas exports scaling from about 15 toward 35 billion cubic feet a day. The megawatts are contracted. The molecule, and the pipe carrying it, are not.

The price

Trailing gross profit compounded from $625m through December to $973m through June. Over the same stretch the price per dollar of that gross profit fell to 60.3x from about 119x in late May, and sits below the 69x of six months ago. Price-to-sales is 18.85x, against 31.81x in early May. This is a de-rating, not a downgrade — but 46x forward gross profit, 75x forward earnings and a 1.07% free cash flow yield leave no room for the gas problem to become chronic. Consensus still models revenue of $6.71bn in 2027.

The other two are not in this business

Plug Power sells hydrogen fuel cells for forklifts and electrolyzers for industrial projects, with no AI load to serve. June-quarter revenue was $178.3m, up 2.5%, at a gross margin of minus 0.9% and a net loss of $188.2m; diluted shares have gone from 858m two years ago to 1.39bn. It fell 1.3% over the past month — it has not tracked Bloom at all.

FuelCell Energy builds molten-carbonate plants and does have a data-center foot in the door: a June agreement with Fit Energy USA for up to 380 MW, of which only an initial 30 MW carries a deposit and a delivery date. Its April-quarter revenue fell 4.9% to $35.6m at a minus 36.3% gross margin, and backlog slipped to $1.14bn from $1.26bn. Its forward price-to-sales of 10.0x exceeds its trailing 9.2x, because analysts expect this year's revenue to come in below the last twelve months'.

Bloom closed at $200.03, some 42% below its 22 June peak, and its 50-day average crossed beneath its 200-day on 20 August — the same session a broad fuel-cell selloff hit all three names with no company news behind any of them, on Treasury yields near 4.71%. Over twelve months the shares are still up more than fourfold.

The setup

Where it stands — Bloom's revenue and margins accelerated through June while its multiple halved; the selling is about fuel permitting, not orders.

Would confirm — Third-quarter revenue lands inside the raised $3.9–4.2bn full-year path with gross margin at or above 33%.

Would invalidate — Further pipeline or permit rejections at Oracle or AEP sites, or guidance cut on delivery slippage.

Watch next — Bloom's third-quarter results, due late October, and any update on Project Jupiter's gas supply.

Valuation — 60.3x trailing gross profit and 46.2x forward, against about 119x three months ago and 69x six months ago.