DK Street Journal

The companies that rent out armies of engineers, consultants and back-office staff to large corporations spent the past year being priced as the most obvious casualty of artificial intelligence: if software can write software and process invoices, the argument runs, a business that bills by the person has a problem. Over the twelve months to 7 August the five names in this group fell an average of about 20%.

Then, in the last full week of July, the trade briefly ran in reverse. Investors dumped semiconductor shares — the VanEck Semiconductor fund fell 5.1% in a single session and the Nasdaq lost 2.9% in the week to 17 July — and rotated into the laggards. India's technology index gained 16.7% in July while the Philadelphia semiconductor index fell 21%, the widest monthly gap since 1999. Jefferies upgraded Indian IT services from underweight to neutral on 27 July, calling it a tactical reversal after a 25% fall while still forecasting only low-to-mid single-digit revenue growth through fiscal 2028.

That is a flow story, not an earnings story. The five trading days from 22 to 29 July supplied nearly the whole month; the seven sessions since have gone nowhere, with Accenture up 1.5%, Tata Consultancy Services flat and Genpact down 6.6%.

Two of five are actually improving

Cognizant, the New Jersey-based outsourcer serving banks, insurers, drugmakers and media firms, is the clearest corroboration. It reported second-quarter revenue of $5.5bn and raised full-year adjusted earnings guidance to $5.70–$5.82, with financial-services revenue up 12% and $1.15bn of stock repurchased in the quarter. Operating margin expanded for a sixth consecutive quarter and trailing bookings reached $29bn, up 5%. It also announced a deepened Google Cloud partnership to deploy agent-style AI in client operations. Caveats: full-year revenue growth was revised to 4–5.5% on macro uncertainty, and management noted one in four large companies has paused AI deployments.

Genpact, the Bermuda-registered business-process manager running finance, claims and supply-chain operations for clients, grew revenue 7.1% to $1.343bn and raised full-year earnings growth guidance to at least 12%, with a thirteenth straight quarter of gross-margin expansion. The shares fell 5.2% the next session anyway.

Accenture, the Dublin-headquartered consulting and managed-services giant with 799,000 staff, diverges. Its May-quarter new bookings fell 2% in dollars and 3% in local currency, a book-to-bill near 1.03x, with full-year local-currency growth guided to 3–4%. Its bought-in growth continues: on 28 July it agreed to take a majority stake in UniCredit's technology venture from IBM, one of several deals in a $9bn acquisition push, while headcount grew just 1% against 6% dollar revenue growth — revenue per employee is rising.

Infosys cut fiscal-2027 revenue growth guidance to 1.5–3% on its 23 July call and conceded clients are demanding AI productivity pass-throughs at renewal and mid-contract. Tata Consultancy Services, India's largest, grew reported revenue 13.9% but only 0.4% sequentially in constant currency, with margin down to 24.0%.

Verdict on the business: SPLIT. Two confirm, three do not.

Where the discount is real

Valuation is the stronger leg. Accenture trades at 13.9x trailing and 12.7x forward earnings on an 11.7% free-cash-flow yield, against trailing multiples of 36.7x, 26.9x, 30.1x, 29.9x and 21.4x at its last five fiscal year-ends — its cheapest decade point, and analysts have already cut the mean price target 28% since May, to roughly $179. Cognizant is at 12.4x trailing and 10.0x forward against 17.1x and 18.2x at the last two year-ends. Genpact sits at 10.1x and 8.4x versus about 15x at both. Tata is the least dislocated at 17.8x trailing, 15.9x forward, 12.3x enterprise value to EBITDA. Verdict on valuation: CONFIRMS for the three Western names, inconclusive for Tata.

The structural risk is documented, not hypothetical: rival HCLTech has quantified 2–3% annual revenue compression from AI productivity handed back to clients, and warns of 3–5% next year. The offset is that headcount is falling faster: Tata ended March with 23,460 fewer employees than a year earlier while revenue grew.

The tape agrees only halfway. Infosys and Tata's 50-day averages crossed back above deep-downtrend territory into neutral on 21 July; Accenture and Cognizant only stabilised on 7 August, and no member of the group has reached an uptrend on any measure.

The setup

Where it stands — A five-session rotation out of chipmakers lifted the group; only Cognizant and Genpact raised guidance behind it. Would confirm — Accenture reporting fiscal fourth-quarter new bookings back in growth, and Cognizant holding full-year revenue growth at 4–5.5% or better. Would invalidate — Another guidance cut at Infosys, or Cognizant's margin-expansion streak breaking after six quarters. Watch next — Accenture's fiscal fourth-quarter results in late September; Infosys and Tata report December-quarter figures in January. Valuation — Accenture 13.9x trailing, 12.7x forward, against 21.4x–36.7x at its last five fiscal year-ends; Cognizant 12.4x and 10.0x.

A group that was never one business

The three companies grouped together as identity and access management — software that decides who, or what, may enter a corporate system — do not sell the same product, and over the past month the market has started pricing them separately. BlackBerry gave back nearly a fifth of its value; SailPoint added almost a quarter; Okta finished the month where it began. For a year those differences were buried inside a flattering average.

Okta: the profits confirm, the growth does not

Okta sells single sign-on, multi-factor authentication and lifecycle management to enterprises and governments, plus the Auth0 developer toolkit for logging customers and machines into applications. Workforce identity is about 59% of contract value, Auth0 about 41%.

The operating story is genuinely good. In the quarter ended 30 April, revenue was $765m, gross margin held at 77.8%, operating income of $56m grew 43.6% on 11.2% revenue growth, free cash flow was $271m at a 35% margin, and net revenue retention inflected up to 107%. Remaining performance obligations rose 16%. Management reported the beat on 28 May and the stock rose 21% in a session. Company disclosures also put its "Okta for AI Agents" suite at roughly 30% of new bookings in the January quarter, the only hard number anyone has attached to the argument that software agents multiply the credentials a company must issue.

The growth line contradicts it. Quarterly revenue growth has gone 12.7%, 11.6%, 11.6%, 11.2%. Annual growth has fallen from 42.9% in fiscal 2023 to 11.8% in fiscal 2026. And company guidance for the July quarter is $790–794m, about 9% growth, with committed backlog up about 11% — deceleration, not the re-acceleration the tape implies. Consensus models roughly 10% revenue growth through fiscal 2029.

Verdict on the business: mixed — CONFIRMS on margins, CONTRADICTS on growth. On valuation it is cleaner. Trailing price-to-sales went from 4.62x on 3 May to 8.22x today and market value from $13.48bn to $24.64bn — an 83% re-rating against 11% revenue growth. At 105.9x trailing and 38.6x forward earnings for roughly 10–12% forward earnings growth, and with the mean analyst target still near $120 against $148 spot even after target raises to $165 and $175 explicitly citing agentic-AI identity demand, the multiple CONTRADICTS the fundamentals.

SailPoint: the one genuine divergence

SailPoint governs which employees, contractors and machine accounts hold which permissions — the audit-and-revoke layer above the login. It reported annual recurring revenue of $1.163bn, up 26%, with cloud recurring revenue up 36% to $781m, and raised full-year guidance to $1.27bn revenue and about $200m of free cash flow. The shares are nonetheless down 2.8% over twelve months and the $10.58bn market value sits below the roughly $12.8bn at which the company relisted in February 2025.

The qualification is the profit and loss account: the GAAP operating margin worsened to −28.5%, gross margin fell to 64.5% (depressed by acquisition accounting), and the net loss widened 60% to $74.7m. At 9.44x trailing sales that is not obviously cheap. Management also said about 10% of customers have adopted its AI features and the agentic pipeline is doubling, while stressing it is not yet material. And the recent move has no earnings behind it: the eight-session, 22% run followed a Wells Fargo buy rating dated 27 July, with results due around 1 September. Business CONFIRMS; the July price move is INCONCLUSIVE.

BlackBerry: a strong quarter, and not an identity company

BlackBerry's June quarter split $72m of QNX embedded automotive software and $74m of secure government communications out of $153m total — about 95% of revenue, none of it identity governance. The quarter itself was strong: revenue up 25.6%, operating income up nearly seven-fold, guidance raised to $594–621m. But full-year 2026 revenue was $549m growing 2.7%, against 9.05x sales and 61x enterprise value to EBITDA, and management flagged that a large Canadian government deployment is unlikely to repeat. The stock fell about 32% in July, its worst month in over three years, on insider sales including 125,000 CEO shares. Business CONFIRMS; valuation CONTRADICTS.

What the tape adds

Okta's 50-day average crossed above its 200-day on 8 June, ten days after the earnings beat, and has stayed there — a trend dated to a result. BlackBerry's uptrend, intact since 8 May, was cut back on 6 August. SailPoint has held an uptrend for only six sessions and failed a similar attempt in June. Behind all three sits an unresolved argument that software multiples fell below the S&P 500's for the first time on fears AI agents break seat-based pricing — the same fear identity vendors claim to be immune from.

The setup

Where it stands — One year of shared gains has broken into three separate stories, only SailPoint showing improving fundamentals against a flat share price.

Would confirm — Okta's July-quarter committed backlog growth printing above 12%, versus the roughly 11% management guided.

Would invalidate — Okta revenue growth at or below 9% with net retention slipping back under 107%.

Watch next — Okta reports after the close on 26 August; SailPoint reports around 1 September.

Valuation — Okta 105.9x trailing and 38.6x forward earnings, 8.22x sales versus 4.62x in early May; SailPoint 9.44x sales.

Three of the four chipmakers grouped in this brief reported results in the ten days to 7 August, and all three set revenue records. Microchip Technology — which makes the small embedded-control chips, analog parts and memory devices that run cars, factory equipment and appliances — reported June-quarter net sales of $1.485bn, up 38% year over year, with adjusted earnings of $0.76 a share against a $0.68 consensus. Its shares rose 13.9% on 7 August on about four times a quiet day's volume.

Rambus, which sells the buffer and clock chips that sit between a server's processor and its memory modules and licenses a large memory-architecture patent portfolio, posted record revenue of $207.4m, up 20.4%, with product revenue of $99.2m up 22% — the shipping hardware growing faster than the licensing base. It ended the quarter with $824.9m in cash and began a roughly $100m accelerated share repurchase on 5 August. Everspin Technologies, an 85-person maker of magnetoresistive memory — chips that keep their contents when power is lost — reported record revenue of $18.7m, up 41.9%, helped by a $40m, two-and-a-half-year development subcontract with a US prime defence contractor. It still ran a $4.4m operating loss.

The label does not fit the businesses

The grouping's premise is Compute Express Link (CXL), a standard that lets servers pool memory chips and share them across processors, easing the shortage of memory bandwidth that starves artificial-intelligence accelerators. On the 27 July call, Rambus management said "CXL remains a fragmented ASIC market" and that the company is doing intellectual-property enablement rather than selling CXL products; its new DDR5 chipsets are a low-double-digit share of product revenue, with the higher-capacity MRDIMM module parts material only in 2027. Microchip's data-centre business — guided to roughly $1bn of calendar-2026 revenue, up about 69%, on 14 design wins for its PCIe Gen 6 switches — is interconnect and storage plumbing, not memory pooling. Everspin's CXL work is a September conference demonstration and a 4 August memorandum of understanding with MaxLinear to evaluate its chips in AI servers — a framework, not a supply contract. Allegro MicroSystems, which makes magnetic position and current sensors for carmakers and industrial customers, books under 5% of revenue from AI data centres.

Sector move, not segment move

The July decline was industry-wide. The Philadelphia semiconductor index fell as much as 6% on 28 July, with SK Hynix down 14.65% and more than $1 trillion of chip-sector value erased on doubts about circular AI financing. Memory principals fell harder than this group over the same 30 days — SanDisk down 29.8%, Western Digital 21.1%, Micron 7.5% — while Nvidia rose 9.7% and Broadcom 10.1%. Allegro's 15.0% decline is indistinguishable from NXP Semiconductors at 15.5%, STMicroelectronics at 18.2% and ON Semiconductor at 13.5%: it followed the automotive analog group, not anything to do with memory. Rambus's 50-day average crossed below its 200-day on 22 July and Allegro's uptrend weakened on 3 August, but every name has risen sharply since the 29 July low, and all four remain well below their 52-week highs — Everspin 62%, Rambus 41%, Allegro 37%, Microchip 18%.

What the numbers say

Business momentum: CONTRADICTS the decline at three names. Microchip's gross margin has improved five quarters running (53.6%, 55.9%, 59.6%, 61.0%, 63.2%) against an $18.6m loss a year earlier; consensus has fiscal-2027 revenue at $6.22bn, up 33.3%. Rambus guided September revenue to $210–216m and consensus carries $825.6m for 2026 and $987.9m for 2027. Allegro is the exception and INCONCLUSIVE: revenue up 27.5% to $259.2m, but a 9.8% operating margin, and price targets cut on 31 July — TD Cowen to $48 from $66, Mizuho to $58 from $67.

Valuation: a genuine de-rating at two, not at the third. Rambus trades at 45.7x trailing and 33.4x forward earnings, with price-to-sales down to 14.56x from 16.82x on 3 May and price-to-gross-profit 18.60x against 20.92x — multiples down 11–15% while revenue growth accelerated. Microchip's 117x trailing earnings is distorted by a trough year; the 26.5x forward multiple on $3.19 of expected fiscal-2027 earnings is the usable one, and its price-to-gross-profit fell to 14.91x from 20.14x in May as gross profit grew 62.8%. Everspin has no forward earnings multiple — consensus still shows a 2026 loss of $0.11 — and trades at 5.35x forward sales on 33% expected growth. Allegro, at 42.5x forward earnings, 79.7x trailing enterprise value to EBITDA and a 1.1% free-cash-flow yield, is the one name where the de-rating looks earned.

The setup

Where it stands — Three of four printed record quarters into a sector-wide July selloff; the group has snapped back but sits far below its highs. Would confirm — Rambus product revenue lands in its guided $110–116m range while royalties fall to $69–75m, as management said. Would invalidate — Microchip's gross margin fails to extend past 63.2% next quarter, breaking a five-quarter recovery. Watch next — Everspin's September memory-industry conference demonstration of its CXL proof-of-concept; Rambus reports September-quarter results in late October. Valuation — Rambus 45.7x trailing / 33.4x forward earnings versus 53.7x trailing on 3 May; Microchip 26.5x forward against 117x trailing.

Two of the biggest suppliers of technology services to the US government reported results in the first week of August that broke a year-long narrative of federal budget cuts, contract terminations and squeezed prices. CACI International, a Reston, Virginia firm of 25,000 people that builds signals-intelligence, cyber, electronic-warfare and secure enterprise computing systems for the Pentagon and the intelligence agencies, reported fiscal 2026 revenue of $9.57bn, up 10.9%, free cash flow up 68% per share to $735m, contract awards of $10.2bn and funded backlog — work with money actually appropriated against it — up 28.6% to $5.4bn. Guidance for fiscal 2027 of $10.65–10.85bn in revenue and $32.96–33.86 in adjusted earnings per share sits about 7% above the $31.13 analysts had modelled. The stock rose 22% in one session.

The day before, Leidos Holdings — a $17bn group of 47,000 employees running national-security systems, air-traffic-control modernisation for the Federal Aviation Administration and medical services for the Department of Veterans Affairs — posted adjusted earnings of $3.26 a share against $2.91 expected and raised its full-year revenue and profit guidance. It closed up 10%.

The move is led from the top, not the bottom

Over the 30 days to 7 August, CACI gained 30.5% and Leidos 28.4%; Science Applications International (SAIC), a $5.3bn engineering and IT-modernisation contractor also based in Reston, added 11.7%. BigBear.ai, a 579-person artificial-intelligence analytics company in Columbia, Maryland, fell 1.2% and did not participate at all. Weighting by size gives +25.6% against +17.4% equal-weighted — the two largest names led, which is the opposite of a low-quality, small-cap-driven bounce. The twelve-month picture is the mirror image: the group's -8.5% average year is almost entirely BigBear.ai, down 50.5%, against CACI up 28.9%.

Does the business explain it?

For CACI, CONFIRMS. Fourth-quarter revenue rose 17.6% with operating margin at 10.1% against 9.0% a year earlier, and roughly 83% of expected fiscal 2027 revenue comes from programmes already won, with 9% from recompetes.

For Leidos, CONFIRMS with a caveat. Revenue growth accelerated to 7.2% from 3.7% the prior quarter, the Defense segment booked 2.2x as much new work as it billed, and free cash flow set a record. But operating margin compressed to 11.1% from 13.4% and operating income fell 11.2%. Management flagged suspended incentive payments on veterans' disability examinations pending an administrative review, and the Defense Health Agency taking electronic-health-record integration work in-house. Top line accelerating while profit shrinks is not yet a clean recovery.

For SAIC, INCONCLUSIVE. Revenue rose 1.5% last quarter — the first increase after four declines — and net income jumped 69% on operating margin of 8.8% versus 6.4%. But bookings of $2.1bn produced a book-to-bill of just 1.1x, consensus still models revenue falling 1.2% this fiscal year, and the diluted share count has shrunk 13% in two years. This is margins and buybacks, not growth.

For BigBear.ai, CONTRADICTS in reverse: revenue grew 13% to $36.7m with gross margin up 7.8 points to 32.8% and backlog of $270m, and the shares kept falling — because operating margin is still -74% and the stock trades at 11.9x sales against Leidos at 0.98x and SAIC at 0.73x.

Where the multiples sit

Leidos is the only one whose valuation has not repaired: 12.7x trailing and 11.1x forward earnings, with a 12.5% free-cash-flow yield, against 15.6x at the end of 2024, 16.2x at the end of 2025 and 17.2x at its November peak. It remains about 31% below that high, after falling roughly 44% year-to-date into late June on soft results, a lost health-record integrator role and a Jefferies downgrade with a $140 target. CACI, at 26.5x trailing, has gone through the top of its own 19.2–23.9x range in three sessions; 20.7x forward is mid-range, so the case now rests on delivering the fiscal 2027 guide. SAIC at 14.2x trailing compares with 12.4x in late June and roughly 11.5x last November — on flat revenue.

The tape, and the overhang

The trend measures lag the news: Leidos's 50-day average has been below its 200-day since 6 March and still is, even though the shares are 33.6% off their 30 June low; CACI's crossed back only on 7 August; SAIC's has been positive since 29 May. The chart is confirming late, not early.

What none of the four has neutralised is procurement reform. The General Services Administration's OneGov programme directs agencies to buy as one customer with pre-negotiated pricing across some $490bn of annual spend, explicitly reducing bespoke contract opportunities, with about 1,600 contracts allowed to expire in fiscal 2025, while the Office of Management and Budget wants IT contractor pricing and utilisation data shared between agencies. And the money is not settled: the House Appropriations Committee approved a $1.07trn defence bill on 24 June, but no floor vote is scheduled and Senate appropriators are deadlocked ahead of the 30 September deadline, with experts warning the much-cited $1.5trn defence budget is unlikely to materialise.

The setup

Where it stands — Two beat-and-raise quarters have re-rated CACI and SAIC; Leidos has rallied 34% off its low but its multiple has not. Would confirm — Leidos holding full-year adjusted EPS at $12.20–12.50 with operating margin back above 12% next quarter. Would invalidate — CACI missing its $10.65–10.85bn fiscal 2027 revenue guide, or funded backlog reversing the 28.6% gain. Watch next — 30 September 2026 appropriations deadline; SAIC's next quarterly report in early September. Valuation — Leidos 12.7x trailing / 11.1x forward vs 16–17x in 2025; CACI 26.5x trailing vs a 19–24x two-year range.

Natural gas in the United States is cheap again, and for entirely physical reasons. Producers pumped an average 110.6 billion cubic feet a day in July, matching the record set last December and running 4.3% above year-ago levels, with working storage at 3,117 billion cubic feet, 6.7% above the five-year average. Demand did not keep up: flows to export terminals averaged 17.2 Bcf/d in July, down from 17.4 in June on scheduled maintenance at the Freeport plant in Texas, and cooler-than-normal August forecasts cut expected air-conditioning load. The front-month contract closed at $2.66.

The four producers held roughly flat through that. Their peers did much the same — CNX rose 5.9% over the month, EQT 1.0%, Antero Resources 0.1%, Gulfport slightly negative — so this is a sector-wide refusal to follow the commodity down, not something specific to these four names. The interesting question is whether the underlying businesses justify the composure.

One company is growing

Range Resources, an Appalachian driller with about 794,000 net acres in Pennsylvania and only 564 employees, is the outlier. Second-quarter revenue of $834m rose 19.1% year on year while operating income rose 73.9% — operating margin widened to 39.1% from 26.8%. It reported record production of 2.3 billion cubic feet equivalent per day, tracking to 2.5 by year-end and 2.6 by end-2027, with 2027 capital spending held flat at roughly $700m. It returned $489m in the first half through buybacks, dividends and debt repayment, about 5.5% of its market value, and has signed a 10-year, 75 million cubic feet a day supply deal for a Midwest power plant — contracted demand that does not wait on new pipeline capacity. Verdict A: CONTRADICTS. The business accelerated; the stock did not.

On valuation Range is the quality name rather than the cheap one: 9.3x forward against 10.5x trailing earnings, on consensus 2026 earnings of $4.11 a share versus $2.74 actual in 2025, with the cohort's highest trailing enterprise-value-to-EBITDA at 7.0x and a 13.1% free-cash-flow yield. INCONCLUSIVE — growth is priced, discount is not.

Three companies are shrinking

Expand Energy, the $21.5bn producer formed when Chesapeake Energy renamed itself in 2024 and now the largest US gas driller, saw revenue fall 19.7% to $2.96bn and operating income fall 47.9%. Its balance sheet did the opposite: total debt cut to $3.7bn, leverage near 0.5x, $850m of stock repurchased year to date and another $1bn authorised. It trades at 3.7x trailing EV/EBITDA, cheapest here, on an 11.8% free-cash-flow yield — but its forward price/earnings of 10.1x sits above its 7.9x trailing, because consensus expects earnings to fall. It has also had no permanent chief executive since the board replaced Domenic Dell'Osso with chairman Michael Wichterich in February.

Comstock Resources, a $3.9bn Haynesville shale pure-play in Louisiana and East Texas run by 252 people, is the broken one. Revenue fell 24.9% to $353m, operating income 74.8%, and diluted earnings were three cents a share. Realised gas of $2.93 only beat the $2.54 unhedged price because 63% of volume was hedged. Free cash flow is negative 18.7% of market value, leverage 3x, and forward earnings multiple 32.0x against 7.6x trailing — the cheap trailing figure is an artefact. Management said 2027 activity is undecided pending higher prices. The one bright spot is external: Sixth Street paid $600m for 27% of its Pinnacle midstream arm, implying a $2.2bn value. Verdict B: JUSTIFIED DE-RATING.

Tourmaline Oil, Canada's largest gas producer at C$23bn, grew revenue 25.5% but lost 55.8% of operating income as gross margin fell to 28.0% from 44.7%. It produced 594,000 barrels of oil equivalent a day against guidance of 595,000-605,000 — deliberately, injecting gas into storage to sell later — cut 2026 capital spending by C$350m and set 2027 at C$2.3bn. Realised gas was C$3.12; its hedge book at C$4.97 is carrying results. At 13.1x forward earnings and a 0.9% free-cash-flow yield it is the most expensive name here on forward numbers.

The catalyst is not in this quarter

The shortage argument is genuine: Plaquemines feedgas has passed 4.1 Bcf/d, and the combined ramp of Plaquemines, Golden Pass, Corpus Christi Stage 3 and Calcasieu Pass Phase 2 should add roughly 8 Bcf/d over 30 months. But the timetable is long. The Energy Information Administration expects Henry Hub near $3.70 in 2026, and the forward curve is flat at $3.41 for 2026 and $3.51 for 2027 — a level at which Tourmaline says cash flow only just covers maintenance, growth and the dividend. The deficit thesis these four names sit inside is dated to the second half of 2028.

The setup

Where it stands — Four gas producers held flat while the commodity fell 17%; only Range's second quarter improved. Would confirm — Storage surplus narrowing toward Range's projected 38 days of supply by year-end, from 40. Would invalidate — Dry gas production holding above 111 Bcf/d into winter while the 2027 strip stays below $3.60. Watch next — Third-quarter results in late October, and whether Expand names a permanent chief executive. Valuation — Expand 3.7x trailing EV/EBITDA but 10.1x forward earnings vs 7.9x trailing; Comstock 32.0x forward vs 7.6x trailing.

Extreme Networks closed its fiscal year on 30 June with revenue up 13% and adjusted earnings per share up 26% to $1.06, and beat the quarter on profit — $0.32 against $0.30 expected. Then it guided fiscal 2027 revenue to $1.38–1.40bn, or about 8–8.5% growth, blaming a transition in which older service contracts run off and are replaced by its newer Platform ONE subscription. The stock fell 12.5% before the open on 5 August and kept going, from $32.34 on 4 August to $23.97 on 7 August.

That matters beyond one $3.1bn company, because Extreme is grouped with the three vendors that build the switches carrying traffic inside artificial-intelligence data centres — and it has essentially no exposure to them. Its book is campus and edge gear for healthcare, education and retail. In an equal-weighted average of four names, a company one-seventy-sixth Arista's size drags the whole group negative while the other three sit within 2% of cycle highs.

The business is real; the price is the argument

Arista Networks, which designs high-performance Ethernet switches and the operating software that runs them for cloud giants and banks, posted its first $3bn quarter, revenue up 37.7% year on year and accelerating from 35.1% the quarter before. Multiyear purchase commitments tripled to $9.7bn from $3.6bn, deferred revenue reached $6.9bn, and its Etherlink AI fabric switches now count more than 100 customers versus four or five in 2024. Management raised full-year guidance to $12.6bn, $2.1bn above the target set at its own analyst day. Gross margin fell to 62.9% from 65.2% on cloud-customer mix; operating margin still rose to 45.4%.

The business CONFIRMS the move. The valuation does not settle it: 58.8x trailing and 46.9x forward earnings, 19.1x forward sales, falling to 37.3x on 2027 consensus earnings of $5.05. The stock popped 6.6% on the print and gave the whole thing back in two sessions, ending 1.0% below where it started.

Cisco Systems — enterprise switching, routing, security and, since the Splunk deal, observability software — is not decelerating. Revenue growth ran 7.5%, then 9.7%, then 12.0% across the last three quarters. Total product orders grew 35% with five new hyperscaler design wins, and AI-infrastructure orders from hyperscalers hit $1.9bn in the quarter and $5.3bn year-to-date, against a full-year target lifted to roughly $9bn from $5bn. Gross margin gave up 194 basis points to 63.6% on that mix; operating margin still expanded to 25.0%.

But Cisco's price-to-sales went from 5.99x in early May to 7.88x now — a 32% expansion against a 35% market-cap gain. Almost the entire advance is re-rating. At 40.2x trailing and 25.4x forward 2027 earnings on 9–11% consensus revenue growth, the business CONFIRMS and the multiple is doing the heavy lifting.

HPE: the cheapest, and the most re-rated

Hewlett Packard Enterprise sells servers, storage and — since the $14bn Juniper Networks purchase — a full networking stack. April-quarter revenue rose 40.7% to $10.7bn, gross margin improved 887 basis points to 36.5%, and operating margin more than doubled to 7.0%. Networking delivered $2.7bn, with data-centre networking up 233% and full-year segment growth guidance raised to 72–75%; cumulative AI-networking orders are guided to $1.7–1.9bn by year-end. At 15.6x forward earnings, 1.57x forward sales and an 8.0% trailing free-cash-flow yield it is the cheapest of the four — but net leverage sits at 2.6x from the term loans that funded Juniper, and price-to-sales has gone 1.05x to 1.81x in three months.

The threat under all of it

Demand is not the issue: the five largest cloud buyers have committed to $660–690bn of capital spending in 2026, roughly three-quarters of it AI infrastructure. Share is. NVIDIA's Spectrum-X Ethernet switch revenue grew 192.7% to $2.1bn and took the top revenue position at 21.5%, ahead of Arista's ~20.7%. Meanwhile hyperscalers keep running their own network software on contract-built "white box" switches, and Dell'Oro names contract manufacturers among the primary beneficiaries of a switch market that has doubled in three years. Arista's management calls white box "tactical." The revenue line says it is more than that.

The setup

Where it stands — Three of four sit near highs on genuine order growth; only Arista's advance came from earnings rather than an expanding multiple.

Would confirm — Cisco printing fourth-quarter revenue inside its $16.7–16.9bn guide with hyperscaler AI orders tracking to $9bn.

Would invalidate — Arista's third-quarter revenue landing below its ~$3.3bn guide, or gross margin under the 62–64% full-year range.

Watch next — Cisco reports fourth-quarter results after the close on Wednesday 12 August 2026.

Valuation — Arista 58.8x trailing / 46.9x forward; Cisco 40.2x / 25.4x; HPE 15.6x forward; Extreme 18.4x forward, down from 24.8x pre-print.

The worry is the funding, not the demand

The companies in this story are landlords to the artificial-intelligence boom rather than participants in it: they own or lease the halls, wire the power and rent capacity to the firms that buy the chips. Through late July they were repriced on a question that has nothing to do with tenant appetite. Alphabet's 22 July results raised 2026 capital-spending guidance to $195-205bn from $180-190bn and flagged a "significant" increase for 2027 — and the shares fell 7%, a reversal of the reflex that had rewarded every spending raise. Days later, reporting that Nvidia was in talks to guarantee roughly $250bn of financing for an OpenAI data-centre project helped wipe about $1.3 trillion off chip-sector value on concerns about circular funding and mounting debt.

The resulting split runs along balance sheets, not business models. Over the 30 days to 7 August, Microsoft rose 31.4%, established landlords Digital Realty and Equinix gained 9.9% and 2.6% — while junk-rated, equity- and lease-funded builders fell: TeraWulf 25.3%, Nebius 13.5%, Core Scientific 10.6%, IREN 8.5%. Demand itself is not in question: hyperscaler 2026 budgets run near $200bn at Amazon and $110-120bn at Microsoft, and colocation operators are signing 10-15 year build-to-suit leases on whole campuses.

The group average is two micro-caps

Equal-weighted, the six names fell 17.4% in 30 days. Weighted by market value they fell 9.2%. SharonAI (-34.6%) and WhiteFiber (-33.0%), together 7% of the group's ~$20.6bn of combined value, account for roughly 65% of the decline. Applied Digital fell 7.3% and GDS 2.6%.

The twelve-month picture is equally split: Keel Infrastructure is up 203.1% and Applied Digital 104.6%, while GDS is down 14.2% and VNET 10.1%. And the one-month frame flatters the tape — Applied Digital is down 29.4% over 90 days and 41.3% from its 52-week high; GDS is down 27.7% over 90 days, with its 50-day average beneath its 200-day every session since 18 June.

Applied Digital: the business and the tape point opposite ways

Applied Digital, a Dallas developer of purpose-built AI data centres that pivoted out of crypto-mining hosting, grew fiscal-2026 revenue (year to 31 May) 183.7% to $611.3m, with gross margin widening from 10.5% to 25.8%. It reported on 27 July a continuing-operations net loss of $249.2m, or $0.91 a share. Contracted revenue is the striking number: a 15-year, 300-megawatt take-or-pay lease with a U.S. investment-grade hyperscaler worth about $7.5bn lifted total contracted lease revenue to roughly $31bn across four campuses and 1,200 megawatts, from roughly $16bn in May.

On business momentum the tape CONTRADICTS the fundamentals. On valuation it is INCONCLUSIVE. Price-to-sales compressed from 28.7x in early May to 14.4x trailing and 10.1x forward — but price-to-gross-profit barely moved, 63.3x to 64.4x trailing, so investors are paying the same for each dollar of actual profit. Two caveats sit under the lease story: fourth-quarter gross margin fell to 15.7% from 42.5% the prior quarter, and diluted shares rose from 201.2m to 275.2m, alongside 2030 debt and a $5bn preferred-equity facility. Anchor tenant CoreWeave, on roughly 400 megawatts, saw its five-year credit-default swaps top about 855 basis points — implying roughly a coin-flip default probability over five years.

China's two hosts are growing into a falling tape

GDS Holdings, Shanghai-based and China's largest listed data-centre operator, grew first-quarter revenue 23.6% to RMB 3,367m with gross margin up to 33.6% from 23.7% and operating margin at 27.0%. It guided 2026 revenue to RMB 12.4-12.9bn and adjusted EBITDA to RMB 5.75-6.0bn, with RMB 30-50bn of capital spending over three years. It trades at 13.75x trailing enterprise value to EBITDA. VNET, the Beijing hosting and colocation group, grew revenue 19.8% to RMB 2,691m — a fifth straight quarter near 20% — but gross margin slipped to 22.9% from 25.2%, free cash flow yield is -45.5%, and it has guided to another 450-500 megawatts and RMB 10-12bn of 2026 capex. Both are diverging from their own results, and both carry U.S. delisting risk. Read them on EV/EBITDA, not the headline price-to-sales figures, which divide a dollar market value by yuan revenue.

The tail: one earner, one shrinker, one shell

Keel Infrastructure, the former Bitfarms rebuilt as an AI-power and data-centre developer, is the case where price CONFIRMS the business: first-quarter revenue fell 44.7% to $37.0m at a -71.1% gross margin, consensus sees fiscal-2026 revenue of $112.5m (-59%), and price-to-sales rises from 12.7x trailing to 22.6x forward because the denominator is shrinking; shares outstanding grew 45% in five quarters. SharonAI, a 25-person GPU-cloud startup, booked $294,014 of quarterly revenue against a $513m market value — 334x trailing sales — with the 3.4x forward figure resting on two analysts modelling a 98-fold revenue jump off a $1.25bn five-year contract with ESDS Software Solutions that this desk previously reported as disputed. WhiteFiber, spun out of Bit Digital last August, is the quiet operator: revenue up 30.7% to $21.9m at a 60.2% gross margin and the only positive free-cash-flow yield in the group, at 5.97%, on 19.2x trailing and 12.8x forward price-to-gross-profit.

The setup

Where it stands — A financing-driven de-rating is hitting all six, but only Keel's shrinking revenue justifies it; Applied Digital's contracted book doubled while its shares fell.

Would confirm — Applied Digital converting Polaris Forge leases into cash with gross margin recovering above the 25.8% full-year level next quarter.

Would invalidate — A CoreWeave payment deferral or renegotiation on its roughly 400 megawatts, or further equity issuance beyond the 36.8% share growth already booked.

Watch next — Applied Digital's fiscal first-quarter results in October; GDS and VNET second-quarter reports and any change to 2026 capex guidance.

Valuation — Applied Digital 14.4x trailing and 10.1x forward sales versus 28.7x in May; GDS 13.75x and VNET 9.58x trailing EV/EBITDA.

Between 22 July and 3 August, the companies that own the pipes carrying natural gas across North America reported second-quarter results, and almost all of them told shareholders the year was running ahead of plan. Williams Companies, which owns Transco — the largest US gas pipeline by volume — raised full-year adjusted EBITDA guidance to $8.3–8.5bn and lifted its 2025–30 annual growth target to 11%-plus from 10%-plus, having guided 9% in May. Kinder Morgan, which moves roughly 40% of all US gas across some 83,000 miles of pipeline, guided full-year EBITDA at least 5% above budget, about $430m of extra cash profit, and raised its dividend 2%. TC Energy, the Calgary group with 93,300 km of gas pipe and about 4,300 megawatts of power generation, reported comparable EBITDA up 12% to C$2.95bn and earnings of C$0.94 a share against C$0.83 expected. DT Midstream, a $13.4bn pure-play with 588 employees whose LEAP pipeline gathers gas in the Haynesville shale of Louisiana, reaffirmed 2026 EBITDA of $1.155–1.225bn.

The shares went the other way. Every one of the five grew revenue and operating income year over year — revenue up 10.2% at Williams, 10.8% at Kinder Morgan, 6.7% at TC Energy, 11.0% at DT Midstream, and 78.4% at Energy Transfer, the partnership whose marketing arm inflates the top line but whose operating income still rose 47%.

The contracts are mostly signed, with dates on them

The re-rating of this group over the past year rests on gas-fired power for data centres, so the quality of the contracts matters. Most are executed and capacity-specified. Williams closed a $5.34bn investment from Blackstone Credit & Insurance, with Apollo and KKR vehicles, for 49% of five named behind-the-meter power projects, keeping control; the first, Socrates, delivers 200MW to a Meta campus in Ohio, and the largest, Neo at 682MW, is due in the second half of 2028. Energy Transfer's Oasis pipeline committed up to 450,000 million British thermal units a day of firm gas — about 1.2 gigawatts of power — for at least ten years to a Texas data-centre campus. DT Midstream's LEAP Phase 5 adds 200 million cubic feet a day for 2028 under long-term contracts with two producers. Kinder Morgan's sanctioned backlog stands at $9.6bn, down only because $650m of projects entered service, with federal certificates expected on a $3.5bn Southeast expansion and a $1.7bn Mississippi line.

Two softer edges: Kinder Morgan's Project 219 South rests on a non-binding open season, and TC Energy's flagship Crossroads expansion has precedent agreements but no final investment decision until the fourth quarter — the least-executed marquee project in the group, run against the highest leverage target, 4.75 times EBITDA.

Cheaper than in May, on higher earnings

Against this desk's own recorded multiples from 3 May, all four corporates have de-rated while earning more. Williams' trailing price-to-earnings ratio went from 34.1x to 27.9x, Kinder Morgan's from 21.3x to 19.8x, Energy Transfer's from 16.0x to 12.5x, DT Midstream's from 32.5x to 28.6x. Energy Transfer is cheapest on every lens — 9.5x enterprise value to EBITDA, a 9.7% free-cash-flow yield — and the only one that rose over the month. DT Midstream is dearest at 15.2x and fell hardest, with net income up just 4.7% and margins slipping. TC Energy compresses from 26.5x trailing to 16.8x forward, the widest gap here. Fundamentals CONFIRM the year; valuation is INCONCLUSIVE on the month — nothing in the multiples explains why guidance raises met selling.

Rates, not gas

The tape points elsewhere. Most of Williams' and Kinder Morgan's monthly loss landed in a single session, 27 July, and long-dated Treasury yields set fresh 2026 highs that week after a hawkish Federal Reserve hold drew three dissents for a hike, the ten-year printing about 4.74%, its highest since January 2025, before easing to 4.65% on 7 August. Long-duration, yield-paying infrastructure trades against that line. All five remain in uptrends, with Energy Transfer and Kinder Morgan strongest and DT Midstream downgraded a notch on 30 July.

The feared contagion from upstream has not arrived. Henry Hub futures fell nearly 15% in July and Appalachian producers EQT, Antero, Range and Expand sit in deep downtrends — yet Kinder Morgan's gathering volumes rose 26%, with its Haynesville system up 54% and transport up 7%. Fee-based contracts do not care much about the commodity price.

The one number that cuts against the group: guided 2026 EBITDA growth of roughly 6–12% sits well below twelve-month share gains of 15–28%. Multiple expansion did about half the work.

The setup

Where it stands — Guidance rose across four of five names in late July; the shares fell on rates, not disclosure. Would confirm — Third-quarter volumes and reaffirmed or raised full-year EBITDA at Williams and Kinder Morgan. Would invalidate — A guidance cut, or gathering volumes turning negative year over year as Appalachian output falls. Watch next — TC Energy's Crossroads final investment decision, expected in the fourth quarter of 2026. Valuation — Williams 27.9x trailing and 29.1x forward earnings, against 34.1x trailing on 3 May.

The catalysts that did not arrive

Three things were supposed to mark the end of crypto's year-long drawdown this summer: a Federal Reserve turning toward rate cuts, US legislation setting clear rules for digital-asset markets, and a return of buying through the exchange-traded funds (ETFs) that hold bitcoin and ether directly. In the five weeks since, all three moved the wrong way.

The Fed's 30 July decision to hold rates was read as hawkish by bitcoin analysts, who then split on what follows. The US 10-year Treasury yield held near 4.68%, with markets pricing roughly 42% odds of a quarter-point rate rise in September as inflation runs above target for a fifth year. The Senate confirmed it would not vote on the Digital Asset Market Clarity Act — the bill dividing oversight of crypto between securities and commodities regulators — before its summer break, returning on 14 September with three weeks to find 60 votes; three Democratic senators formally opposed the merged draft in mid-July. On Coinbase's 30 July earnings call, chief executive Brian Armstrong had put the odds of a pre-recess vote at about 30%.

Fund flows were worse. US spot bitcoin ETFs drew about $205m in July, the smallest monthly intake since the products launched in January 2024, following $2.43bn of redemptions in May and $4.52bn in June. Ether funds broke an eight-week outflow streak in mid-July, then turned negative again at about -$70m in the week to 30 July even as the token rose. Bitcoin, near $64,300, sits roughly 49% below its October 2025 record while the S&P 500 prints all-time highs.

The month belongs to one buyback

BitMine Immersion Technologies, a seven-employee holding company that runs an ether treasury alongside legacy mining-equipment consulting, gained 26.6% in 30 days and supplied roughly 69% of the group's average advance. The reason is disclosed: BitMine held 5.8 million ether as of 2 August — about 4.8% of supply — worth $11.3bn with cash, against a $10.72bn market value, and has repurchased 16.1 million shares since 1 July under a $4bn authorisation. At 0.895x book value, that is stock retired below asset value. Against it: diluted shares still rose to 551.8m from 200.6m a year earlier, a 175% increase, and the operating business lost $11.9m on $46.5m of revenue last quarter. Verdict: INCONCLUSIVE — a supported advance in the narrow mechanical sense, on a business that barely exists.

Strategy's discount is earned

Strategy, the software company turned levered bitcoin holder, trades at 0.757x book — a 24% discount, the balance-sheet version of the sub-1.0 net-asset-value multiple crypto treasuries are judged on, which fell to roughly 0.72x in the June correction, matching the 2022 low. The deterioration behind it is real. Strategy reported 843,775 bitcoin, $17.06bn raised year-to-date through drip-feed share sales, and $218.4m of bitcoin sold under a new monetization program; it sold a further $105m in the week to early August to fund $52.4m of preferred dividends. Those dividends now annualise near $1.763bn, payable whatever bitcoin does. Diluted shares rose 15.2% year over year; the software arm grew 6.9% to $122.4m, immaterial against an $8.22bn quarterly net loss. Verdict: CONFIRMS — the business explains the de-rating.

Coinbase is the outlier

Coinbase, the largest US-listed crypto exchange, is the one name where tape and business diverge. Revenue is still shrinking, but the rate has more than halved: -54.6%, then -30.5%, then -18.5% year over year. Gross margin expanded to 84.4% from 71.5%, and gross profit fell only 3.8% on an 18.5% revenue drop. Second-quarter subscription and services revenue reached $555.1m — 48% of net revenue — including $292m from stablecoins, and the company took a record 10.3% share of global crypto trading volume, a third straight quarterly gain, on $207.8m of adjusted earnings before interest, tax, depreciation and amortisation. Management said bitcoin trading is now 12% of revenue, down from over 50% historically. Yet the stock fell 3.6% in the month, and at 7.32x trailing sales its forward multiple of 7.49x sits above trailing — consensus models 2026 revenue down 24.8% to $5.41bn before a 27.7% rebound in 2027. Verdict: CONTRADICTS on the business, INCONCLUSIVE on valuation until that V begins to print.

What the tape says

Bitcoin's 50-day average crossed back above the deeper downtrend threshold on 28 July, nine sessions ago — the fifth such attempt since January, and the previous four all reverted within weeks. BitMine's is two sessions old, and its April version lasted 42 days before failing. Coinbase has held the deepest downtrend band for 39 consecutive sessions and Strategy for 26; both trade below their 50-day averages, and all five members sit below their 200-day, from bitcoin at -11.0% to Strategy at -35.8%. A de-rated complex has stopped falling. That is not the same as a bottom.

The setup

Where it stands — Bitcoin near $64,300 and flat on the month; the group's gain is one company retiring its own shares below asset value.

Would confirm — Spot bitcoin ETF net inflows above roughly $1bn in August, reversing July's record-low $205m.

Would invalidate — Bitcoin losing its nine-session trend upgrade, as the prior four attempts since January did.

Watch next — The Senate returns 14 September with three weeks to vote on the market-structure bill.

Valuation — Coinbase 7.32x trailing sales, 7.49x forward; Strategy 0.757x book; BitMine 0.895x book.

Texas's governor on 3 August ordered state regulators and the grid operator to audit every data-center project waiting in the interconnection queue before any further approvals, after connection requests swelled to 474 gigawatts — roughly five times the state's record peak demand, with data centers about 90% of the total. The grid operator has suspended its scheduled large-load classification notices and halted the related transmission study until the review finishes, and will seek an exception at a 20 August regulatory meeting. BloombergNEF estimates the pause could delay 49.8 gigawatts of data-center load and cost projects up to $15bn.

That matters because the entire re-rating of Texas merchant power in 2025 rested on those queue numbers being real demand. The audit does not cancel anything; it postpones the moment anyone finds out.

The split runs on geography, not on the theme

This is not the artificial-intelligence power trade unwinding wholesale. Over the past 30 days Constellation Energy, the largest US nuclear fleet owner and a PJM-region operator, is up 10.4%, and Talen Energy, another PJM generator, is down only 5.5%. The losses are concentrated in the two Texas-weighted names: NRG Energy, a Houston company selling electricity to about 6 million retail customers under the Reliant, Direct Energy and Green Mountain brands while owning gas, coal, solar and nuclear generation, and Vistra, an Irving, Texas integrated retailer-generator with roughly 38,700 megawatts serving 4.3 million customers in 20 states. The equipment side has kept its premium entirely: GE Vernova, which builds the turbines both companies have booked, is up 53% over twelve months.

Vistra: the business contradicts the tape

Vistra reported quarterly revenue of $5.003bn, up 17.7%, with generation segment adjusted profit up 68% to $994m on hedging, higher PJM capacity revenue and the Lotus acquisition. Operating income was $1.525bn against $515m a year earlier. Management reaffirmed 2026 adjusted EBITDA guidance of $6.8–7.6bn and held the 2027 range of $7.4–7.8bn, while conceding it is biased toward the lower end absent a move in power curves. Neither the pending $4bn Cogentrix acquisition — 10 gas plants totalling 5,496 MW, closing mid-to-late 2026 — nor the Meta contract sits in that 2027 number; together management puts them at about $700m.

On contract quality, the bearish premise fails. Vistra's Meta agreements are signed 20-year power purchase agreements for 2,609 MW of nuclear capacity, with dated deliveries beginning late 2026. Business verdict: CONTRADICTS the move. Two caveats: per-share growth is heavily buyback-assisted — the diluted count fell from 482m in 2021 to 340m in 2025 — and reported gross profit was negative $943m last quarter, a hedge-accounting artefact that makes the 60.3x trailing price-to-earnings ratio meaningless.

Valuation verdict: POSSIBLE DISLOCATION. At 15.5x 2026 consensus earnings of $9.06, 13.0x 2027's $10.93 and 11.2x 2028's $12.69, with trailing enterprise-value-to-EBITDA at 13.7x and a 1.96% free-cash-flow yield depressed by growth spending, the shares carry a shrinking multiple against unchanged guidance.

NRG: the de-rating is doing work

NRG's adjusted EBITDA rose 34% to $1.2bn, but that came from the first full quarter of the LS Power assets ($370m of East-segment contribution), not the base business. Adjusted earnings per share fell to $1.49 from $1.73 against roughly $1.82 expected; the diluted share count rose from 196m to 210m, so this profit growth was bought rather than earned. Texas segment profit fell $131m as Houston power averaged $33 per megawatt-hour, well under the $52 planning assumption. Virginia's re-entry into a regional carbon program added a $70m cost drag. Full-year 2025 operating income had already fallen 23.7% to $1.85bn. Guidance of $7.90–$9.90 was reaffirmed but flagged as landing below the midpoint, and the shares hit a 52-week low. Its headline 1.2 GW data-center project is at "aligned principal commercial terms" with an unnamed counterparty — pre-land, pre-final-investment-decision — against Vistra's executed contract. Growth spending of $721m has pushed NRG's 3x net-leverage target from 2028 to 2029. Business verdict: INCONCLUSIVE, tilting negative. Valuation verdict: JUSTIFIED DE-RATING — 13.3x forward, 30.9x trailing, 11.4x EV/EBITDA, 1.40% free-cash-flow yield.

One structural point cuts against the bear case for both. PJM's capacity auction for 2028/29 cleared at the $325 per megawatt-day regulatory cap for a third straight time, versus $555 in PJM's own uncapped simulation, and procured 138,318 MW — 5.6% below its own reliability requirement. Scarcity is being suppressed administratively, not resolved.

On the tape: both names have been below trend since spring, but the 30-day damage is two sessions — 4 August, when NRG fell 15.5% and Vistra 8.2%, and Vistra's late-July print. That is event risk, not drift.

The setup

Where it stands — Texas's connection freeze hit the two Texas-weighted generators; Vistra's numbers improved through it, NRG's did not. Would confirm — NRG's full-year adjusted earnings per share landing in the lower half of $7.90–$9.90 with Texas realisations again below $52 per megawatt-hour. Would invalidate — Vistra guiding 2027 adjusted EBITDA above the $7.4–7.8bn range once Cogentrix and Meta are consolidated. Watch next — The Texas regulator's 20 August open meeting on the grid operator's good-cause exception request. Valuation — Vistra 15.5x 2026 and 13.0x 2027 earnings; NRG 13.3x forward against 30.9x trailing, down from 38.6x in May.