DK Street Journal

Centrus Energy exists to enrich uranium. In the three months to June it did less of that than a year earlier — separative work unit (SWU) volumes fell 23% — and still reported revenue growth of 14%, to $176.1m, because it resold more natural uranium to utilities at slimmer margins. Enrichment pricing rose 3% over the year. Cost of sales in the low-enriched uranium (LEU) segment rose 36%. Gross margin fell to 28.3% from 34.9%, gross profit fell 7.4% to $49.9m, and operating income fell 69% to $10.4m.

That matters because the entire re-rating of this corner of the market rests on a physical bottleneck. Enrichment sells for roughly $160 per SWU today against about $40 before 2022, a 167% rise driven by a US ban on Russian enriched uranium that becomes a full prohibition in 2028 — against a Western capacity base that cannot meaningfully expand before then. Centrus, headquartered in Bethesda, Maryland, is the only US-owned commercial enricher and the only listed pure play on that squeeze. Its backlog now stands at $4.5bn stretching to 2040, of which $3.7bn is enrichment and uranium sales and $2.4bn sits under definitive agreements rather than contingent terms. A $900m Department of Energy task order is excluded from that figure.

The backlog is real. The earnings are going the other way. Consensus has Centrus earning $2.50 a share this year against $3.90 delivered in 2025, and the capital to build the Piketon, Ohio, plant is coming partly from shareholders: diluted shares have gone from 16.4m to 21.9m in eight quarters, roughly 33% dilution, with $53.9m raised through the at-the-market program last quarter alone. Capex guidance of $350-500m for 2026 brackets the revenue guidance of $450-500m. Free cash flow yield is -6.3%.

The month belonged to the pounds, not the reactors

From 22 July to 21 August the gains ran through the miners. Uranium Energy Corp, a US developer that booked no revenue at all last quarter against a $40.8m operating loss, rose 22.5%. Cameco, the world's largest listed uranium producer, rose 13.4%. Centrus rose 6.9%. The two pre-revenue reactor developers did not participate: NuScale Power, whose small modular design is the only one with US regulatory certification, gained 7.3% on quarterly revenue that fell to $75,000 from $8.05m; Oklo, developer of the Aurora powerhouse, fell 4.7% despite holding $3bn of cash.

This was not a discount-rate rally either. The 30-year Treasury yield topped 5.33% on 18 August, a 19-year high. Three days later, at the Global 2026 conference, Urenco USA said it would expand American enrichment capacity by nearly 50% with a new plant, and Ur-Energy shipped its first uranium from Shirley Basin. Cameco closed up 7.24% that session, Centrus up 5.75% — Centrus rising, notably, on news that its scarcity is being addressed by a competitor. Spot uranium was $88.29 a pound on 20 August, little changed since February, while the long-term contract price reached $90 in the first quarter, its highest since 2008.

Cameco got more expensive as it earned less

Cameco's second quarter went backwards. Revenue fell 7.2% to $814.1m, gross margin fell to 21.1% from 29.3%, and gross profit fell a third. The swing factor was Westinghouse Electric, 49%-owned by Cameco, which reported a $10m net loss on Cameco's share against $126m of earnings a year earlier; Westinghouse has confidentially filed for a US listing. Production guidance of 19.5-21.5 million pounds was left intact.

The price paid for that stream has widened. Cameco costs 64.2x trailing gross profit today against 60.3x in May, having fallen from 72.0x in February — the multiple re-expanded over three months while the profit underneath it shrank. Forward earnings multiple: 66.4x.

BWX Technologies is the reverse. It builds the US Navy's nuclear reactors and fuel on a sole-source basis and makes commercial reactor vessels and TRISO fuel. Revenue grew 18% to $901.6m, backlog reached $8.4bn, up 40%, and on 3 August it raised every 2026 guidance line, including adjusted earnings to $4.70-4.80 a share. The shares fell 10.5% over the month to a 52-week low. Its price per dollar of trailing gross profit has compressed from 25.9x in February to 18.5x — a 28% de-rating while that gross profit grew 6%.

Cameco and BWXT have both traded with their 50-day average below their 200-day since early July. Centrus emerged from that condition on 12 August. On the numbers, the ranking is inverted: the cheapest business is the one growing fastest.

The setup

Where it stands — Enrichment scarcity is being priced into Centrus and Cameco while the only member with rising backlog and raised guidance, BWXT, sits at a 52-week low. Would confirm — Centrus third-quarter SWU volumes returning to growth with enrichment pricing up more than 3% year over year. Would invalidate — Centrus 2026 revenue landing below the $450m guidance floor, or Cameco cutting its 19.5-21.5 million pound production range. Watch next — Centrus and Cameco third-quarter results, early November; the Westinghouse listing terms whenever the S-1 goes public. Valuation — Centrus at 32.0x trailing and 32.5x forward gross profit, versus 32.9x in February; Cameco 64.2x, BWXT 18.5x.

A doubling with no news behind it

Palo Alto Networks, which sells firewalls and a wide stack of security subscriptions to large enterprises, service providers and governments, has not published a financial result since late April. Its fiscal fourth quarter and full-year figures are scheduled for 1 September. Everything the market has done to the shares since — a rise of about 141% from 20 February to 21 August — rests on that April print and on what investors expect to follow it.

Measure the price against what the company keeps on each dollar of sales. Trailing four-quarter gross profit rose from $7.27bn to $7.63bn over those six months, a gain of 4.9%. The price paid for each dollar of that gross profit went from roughly 16.7x in February to 28.3x in May and 38.2x now. The business grew by a twentieth; the price attached to it grew by more than half again.

The April quarter itself is harder to read than the headline suggests. Revenue of $3.00bn was up 31.1%, and next-generation security annual recurring revenue (ARR) grew 60% to $8.1bn — but roughly $388m of revenue and $1.63bn of that ARR arrived with the CyberArk and Chronosphere acquisitions, leaving organic growth near 14% and organic ARR near 28%. CyberArk, an identity-security specialist bought for $21.1bn in February with 112m newly issued shares, also pushed purchase accounting through cost of revenue: gross margin fell to 67.6% from 74.2% two quarters earlier, and the company posted an operating loss of $183m. Diluted shares rose 4.9% in a single quarter. Management trimmed fiscal 2026 adjusted earnings guidance to $3.65–3.70 a share from $3.80–3.90 to absorb the deal.

The platform story underneath is real but young. Palo Alto added 110 net new platformized customers in the quarter, reaching about 2,280, and guided the June quarter to next-generation security ARR of $8.90–8.95bn. Prisma AIRS, its dedicated product for securing AI systems, ended the quarter with more than 300 customers, up from 100, with management targeting $100m of ARR within a couple of quarters — about 1.2% of the current base. The AI-security line is a rounding error on today's revenue and the entire justification for tomorrow's multiple.

Cloudflare's numbers agree; its disclosure does not

Cloudflare runs a global network that sits between the public internet and its customers' sites, selling denial-of-service protection, bot mitigation, content delivery and a serverless developer platform on one fabric. Its June quarter is the cleaner result of the two. Revenue of $696.1m grew 35.9%, the fourth consecutive acceleration. Dollar-based net retention reached 120%, up six points on the year. Customers spending more than $100,000 annually rose 27% to 4,698. Non-GAAP gross margin ticked up to 73.1%, the first sequential gain in eight quarters, after a long slide management blames on free-traffic mix. Free cash flow of $56.4m was up 69%, and diluted shares grew just 2.0%.

What is missing is the money. Cloudflare added roughly 2 million developers to its Workers platform in the quarter and says more than half the traffic crossing its network now comes from AI agents rather than people. It discloses no revenue figure for Workers AI, its AI gateway, or pay-per-crawl, the mechanism that makes Cloudflare the merchant of record when an AI crawler pays for a page. It has told AI companies to separate search crawlers from training crawlers by 15 September or be blocked by default on ad-bearing pages — a strong negotiating position with no invoice attached to it yet. Meanwhile network capex, 7% of revenue in the quarter, is guided to 14–15% for the full year as graphics-processor capacity lands at the edge. At 57.0x trailing gross profit against 38.5x in February, the shares already carry the outcome.

The leaders became the laggards

Cloudflare has traded above both its 50-day and 200-day averages, with the shorter above the longer, on every session since 5 May; Palo Alto since 26 May. Those are the longest such runs among the large security and enterprise-software names. Yet over the 30 sessions to 21 August, when that group gained roughly 14%, the two were at the bottom: Cloudflare up 7.7% and Palo Alto up 4.6%, against Zscaler's 21.1%. Both peaked on 13 August and have since given back 11.4% and 9.6%.

The apparent laggards deserve a correction. Zscaler, which sells cloud-delivered secure access, is only cheap-looking because it fell 31.5% in one session on 27 May — the largest single-day move in the group — on a leadership change and soft guidance; its recovery is what powered the last month. CrowdStrike, the endpoint-security vendor, looks broken only because of a four-for-one stock split effective 2 July; adjusted for it, the shares are up roughly 96% over six months. It reports on 26 August.

Nor is this a rate trade. On 18 August, when the 30-year Treasury yield hit a 19-year high, both names fell. They rose on 21 August when Jerome Powell opened the door to a September cut, but no faster than the Nasdaq. For Cloudflare, five sessions supplied more than three times its entire net gain for the month. That is a market repricing an idea, session by session, well ahead of the disclosure that would settle it.

The setup

Where it stands — Both shares carry AI-era multiples; only Cloudflare's reported growth has accelerated to match, and neither discloses AI revenue in dollars. Would confirm — Palo Alto's initial fiscal 2027 next-generation security ARR guidance above $11bn, implying organic growth holding near 30%. Would invalidate — Cloudflare's non-GAAP gross margin falling back below 72% as second-half capex lands, or net retention slipping from 120%. Watch next — Palo Alto reports fiscal fourth quarter on 1 September; CrowdStrike reports 26 August. Valuation — Palo Alto at 38.2x trailing gross profit and 29.3x forward, against 16.7x in February; Cloudflare at 57.0x and 50.0x, against 38.5x.

Bitcoin rose 22% in the week to 21 August after the US Treasury doubled its long-duration bond buybacks to at least $4bn per operation and President Trump pushed Congress on crypto legislation, forcing roughly $2.5bn of short liquidations in a day. The three companies most often described as the leveraged way to own that move did not participate. Across the sharpest two sessions, when bitcoin gained 14.7%, Hut 8 fell 1.5%, Cipher Mining rose 3.1% and TeraWulf 6.6%.

That is because the underlying business has changed and the label has not. All three began as bitcoin miners; all three now sign long-dated leases handing energized, grid-connected megawatts to artificial-intelligence tenants. The proper comparison is no longer hashprice. It is the cost of thirty-year money.

What each one actually earns

TeraWulf, which operates the Lake Mariner campus in western New York, is the only one where the conversion shows up as money. High-performance computing leasing brought in $31.9m in the second quarter, up 52% sequentially and now 71% of revenue. Gross margin reached 72.3%, from 53.6% a year earlier. After quarter-end it signed a 401 MW, 20-year lease with Anthropic in Hawesville, Kentucky, worth roughly $19bn, with first capacity due in the second half of 2027. Against that, it had 102 MW of operating capacity entering the third quarter — about a tenth of what it has contracted — and diluted shares up 25.5% year over year, to 485.7m.

Hut 8, cast by most coverage as the mining holdout, holds the largest contract book of the three: about 949 MW across its River Bend and Beacon Point campuses and roughly $26.6bn of base-term value. Its second Beacon Point lease, signed 20 July, runs 15 years over 352 MW with a 3% annual escalator. Yet 97% of second-quarter revenue — $72.5m of $74.9m — still came from mining. It funded the buildout with $7.5bn of non-recourse, fully amortizing project notes, and its diluted share count was flat.

Cipher Mining is the one where the decline has support in the numbers. Revenue fell 43% to $24.8m, at a gross margin of minus 209%, because the Black Pearl mining fleet was switched off before rent began. Its tenants are strong — Amazon Web Services holds a 15-year, 300 MW lease at Barber Lake — and three executed leases should yield about $793m of average annual net operating income from October. But consensus expects 2026 revenue to fall 14.1% before jumping 273% in 2027.

Why the shares fell anyway

The damage was done in earnings week, 3 to 10 August: Cipher lost 32%, Hut 8 22.6%, TeraWulf 13.6%. A second leg arrived on 18 August, when the 30-year Treasury yield touched 5.323%, its highest since 2007. Cipher fell 12.1% that session, TeraWulf 11%, Hut 8 7.5% — alongside CoreWeave's 12% drop. Nearly all the value in these companies sits in cash flows starting in 2027 and running to 2040, financed by debt; data-center paper with less-established tenants prices at 275 to 325 basis points over investment-grade benchmarks. Cipher's own project notes span 6.000% to 7.125% across roughly $4.54bn of non-recourse borrowings.

This was not a repricing of data-center property generally. Over the same 30 days Digital Realty rose 8.5% and Equinix 5.1%. It was a repricing of contracted-but-unbuilt megawatts carrying leverage.

Where the price sits

All three lose money, so earnings multiples say nothing. On price to gross profit, Hut 8 trades at 124 times trailing — and 124 times forward, no relief at all, with enterprise value at 209 times EBITDA, after a 231% twelve-month gain against a bitcoin price that fell 18.5%. TeraWulf compresses from 67.7 times to 42.3 times forward, the only one where growth does real work. Cipher's trailing gross profit is negative; its price-to-sales barely moves, from 33.7 to 30.1 times.

One credit sits behind a striking share of it: Anthropic is the named counterparty at both TeraWulf's Kentucky campus and Hut 8's River Bend. Both stocks had been in a strong uptrend in May; by July that had faded to something milder, and it has not recovered.

The setup

Where it stands — Three converted miners fell through August while bitcoin rallied, repriced as leveraged landlords rather than crypto proxies. Would confirm — TeraWulf's leasing revenue exceeding $40m in the third quarter with gross margin holding above 70%. Would invalidate — Shares tracking bitcoin roughly one-for-one over the next month, restoring the crypto-beta reading. Watch next — Cipher's Barber Lake Phase 1 rent, about 168 MW, is due to begin in October 2026. Valuation — Hut 8 at 124x gross profit trailing and forward; TeraWulf 67.7x falling to 42.3x; Cipher's is negative.

The chips that step grid power down to the sub-1-volt, thousand-amp rails an AI accelerator actually draws are having the best year in the history of the category. Monolithic Power Systems, which sells DC-to-DC conversion chips into servers, cars and industrial equipment, grew revenue 47.6% in the June quarter to a record $980.6m. Analog Devices, the broad-line analog chipmaker, reported its first $4bn quarter on 19 August. Both stocks then fell.

The fall is four days old. At the 17 August close all three of the companies below were still trading above where they stood a month earlier. Then, between 18 and 21 August, the six largest names in the group dropped an equal-weighted 9.9% — more than 100% of the decline recorded over the full 30 days. The trigger sits in the bond market, not the order book: on 18 August the 30-year Treasury yield topped 5.33%, a fresh 19-year high, on a widening federal deficit and sticky inflation, and the move was global — Japan's 10-year yield reached a 30-year high, Germany's 30-year bund its highest since 2011. A higher long rate mechanically shrinks the present value of earnings that arrive years out, which is precisely what these multiples are built on. Nvidia rose 1.4% over the same month. Nothing in the customer's shares says accelerator demand was repriced.

The socket incumbent

Monolithic Power is the name whose entire AI case rests on content per accelerator rising faster than hyperscalers can second-source it — a risk that cost it share in 2024. Its Q2 disclosure answers that directly. Enterprise data revenue was $380.6m, up 164.3% year over year and 38.8% of the company. Management raised its full-year enterprise-data growth forecast to 130% from 85%, said there is no single-customer concentration in that line, and put CPU share above 30%. Gross margin was 55.18%, inside a 26-basis-point band across five quarters. Operating income grew 84.4% on 47.6% revenue, widening operating margin by 620 basis points to 31%. Inventory was described as very low.

What it does not own is the next architecture uncontested. Nvidia's 2026 partner list for 800-volt direct-current power conversion runs to at least fifteen silicon suppliers, Analog Devices and Texas Instruments among them.

Analog Devices is a rival, not a bystander

It is tempting to file Analog Devices as an industrial-and-automotive cyclical. Its own quarter refuses. Communications revenue grew 84%, with data center now 80% of that segment and data-center power alone more than doubling. Management called intermediate-to-core conversion for 6,000-amp, sub-1-volt processors one of the fastest and largest growing analog opportunities. On 7 July it closed a $1.5bn all-cash purchase of Empower Semiconductor, which makes voltage regulators that sit under the accelerator and feed current vertically. Gross margin reached 67.3% and operating margin 40.1%, up 1,170 basis points. The shares went from $393.33 on 17 August to $370.24 on 20 August — down into the print and through it.

The selling was upside down

If this were AI power content being marked down, exposure would predict damage. It predicted the opposite. Over 30 days onsemi — silicon-carbide and power chips for electric vehicles and factories, with no meaningful data-hall business — fell 19.5%, and 32.2% over three months. Microchip Technology, a microcontroller maker still working through an automotive and industrial inventory correction, fell 10.7%. Texas Instruments, the largest analog supplier and the most catalog-driven, fell 10.4%. Monolithic Power fell 5.8% and Analog Devices 3.3%.

Vicor diverges

Vicor, the Andover, Massachusetts maker of modular power converters that pioneered 48-volt-to-point-of-load conversion and lost the Nvidia H100 socket to Monolithic Power, is the one name whose trend genuinely broke; its 50-day average crossed below its 200-day on 12 August. Headline revenue of $143.4m grew 1.6%, and gross margin fell 734 basis points to 58.0% — against a prior-year quarter containing a $45m patent settlement. Sequentially the picture inverts: revenue up 26.9%, margin up 280 basis points, advanced products up 45% to $94.2m, and one-year backlog up 26% to $379.7m. Cash rose $49.4m to $453.6m, so the cash-burn worry is misplaced; $15m of the quarter, though, was licensing income under a $60m agreement.

What the price now assumes

Monolithic Power trades at 48.0x forward earnings against 80.3x trailing, whose base is distorted by a one-off 2024 tax item; 35.8x trailing gross profit, down from 38.5x on 18 August. Analog Devices is cheapest on every lens — 29.1x forward, 19.9x trailing gross profit against 23.1x a week ago, a de-rating produced by the new quarter entering the base rather than by the price. Vicor's 62.9x trailing versus 58.4x forward is a spread of barely four points, against Monolithic Power's thirty-two: the market is not paying for this year's growth there but for 2027, when consensus has revenue rising 55.6%.

The setup

Where it stands — Two of the three have accelerating revenue, expanding margins and falling multiples at once; the drawdown began with a rate shock. Would confirm — Monolithic Power's enterprise data revenue growing sequentially again in the September quarter with gross margin held near 55%. Would invalidate — A disclosed second source at a major accelerator customer, or Monolithic Power inventory days rising as gross margin slips. Watch next — Monolithic Power's third-quarter report in late October; Vicor's second-fab site selection, promised within weeks of 21 July. Valuation — Monolithic Power 48.0x forward versus 80.3x trailing; Analog Devices 29.1x forward; Vicor 58.4x forward against 62.9x trailing.

The most consequential disclosure in enterprise software this month was a contract renewal. On its 6 August call, Datadog — which sells cloud monitoring software that tracks whether a customer's applications, servers and logs are behaving, and bills by volume consumed — told investors that its largest artificial-intelligence customer, a nine-figure account running 17 of its products, had signed again at reduced usage.

Nothing was lost. The customer stayed, on the same platform, using the same tools. It simply decided to send fewer bytes. Under consumption pricing that is indistinguishable from a cancellation for the quarter in question: Datadog cut third-quarter revenue guidance to $1.135–1.145bn, implying 28–29% growth against the 35.6% it had just delivered. The shares fell about 19% that day despite a revenue beat and a raised full-year outlook.

The quarter underneath

Strip that account out and Datadog's operating record is the strongest it has been in two years. Revenue of $1.121bn grew 35.6%, a fourth consecutive acceleration from 28.4% three quarters earlier. Non-artificial-intelligence revenue — the ordinary business of watching ordinary software — reached high-20s growth, up from 18% a year ago. Customers spending more than $100,000 a year rose to 4,720 from 3,850 and now supply 91% of annual recurring revenue. Net revenue retention held in the low 120s.

Profitability turned as well. GAAP operating margin reached +0.49% against −4.13% a year earlier, free cash flow was $279m, and the diluted share count rose just 1.9% — unusually little dilution for a company of this compensation profile. The one genuine deterioration is gross margin, down 180 basis points to 78.6% from 80.4% two quarters ago, as Datadog's own AI features push inference costs into cost of revenue.

The moat argument is unchanged: when Datadog is unavailable, engineering teams lose visibility into every production system at once, and more than a thousand integrations across twenty-odd products make it painful to unpick. Gartner's most recent observability ranking places it among eight leaders alongside Dynatrace, Splunk and Elastic, and the category is consolidating — Palo Alto Networks bought Chronosphere for $3.35bn in January. But a moat governs whether a customer leaves, not how much it spends. Management said outright that priorities have shifted from validating AI to optimizing its cost.

What the price already assumed

Datadog costs 27.3 times trailing gross profit today, against 27.1 times in May and 15.4 times in February. The re-rating happened before this quarter, not because of it; the past three months added nothing to the multiple even as gross profit grew 33%. That figure still sits far above Dynatrace at roughly 8.4 times and Elastic at 6.8 times. Trailing free cash flow yield is 1.38%, so cash generation is not anchoring anything. Datadog's trend has been intact since 21 May — its 50-day average above its 200-day on every session since, including through the 20% fall.

Snowflake is the opposite arrangement. Its cloud data warehouse, where customers pay for compute consumed rather than seats occupied, delivered product revenue of $1.33bn, up 34%, with remaining performance obligations of $9.21bn, up 38%, and net revenue retention of 126% in its April quarter, and lifted full-year product revenue guidance to $5.84bn. That was 27 May. Since then it has published nothing, and price per dollar of trailing gross profit has gone from 18.0 times to 32.9 times — an 83% expansion against trailing gross profit growth of 7.4%. Its AI evidence remains account counts, not dollars: Cortex across 9,100-plus accounts, no quantified revenue line. It is still GAAP loss-making, at −$295.6m last quarter, with gross margin flat at 66.6%. Databricks, its closest private rival, reached roughly $6.9bn annualized revenue by mid-2026. Snowflake reports its July quarter on 2 September, and will face the same optimization question Datadog just answered badly.

The company keeping this group afloat isn't an AI story

Elastic, which sells the search and logging stack behind Elasticsearch, rose 46.6% over thirty days with no earnings report at all — six straight advancing sessions in mid-August on a security-platform upgrade and price-target raises. That is roughly half the group's entire monthly gain, from its cheapest and slowest-growing member: revenue grew 16.0% last quarter, decelerating.

Strategy, the bitcoin treasury that still carries an analytics label, contributed the rest — and only in the final three sessions, rising 12.7%, 7.8% and 7.0% from 19 to 21 August as bitcoin cleared $68,000 on crypto-policy optimism. Its software arm booked $122.4m of revenue last quarter, about 0.3% of market value, while the GAAP loss was $8.22bn on bitcoin marks. It trades at 0.90 times book, with third-party trackers putting its premium to bitcoin holdings at 0.68–0.71 times in early August against roughly 3.4 times in November 2024, and an average coin cost near $75,419 versus spot around $73,755. A sixth member, Clearwater Analytics, stopped trading entirely: Permira and Warburg Pincus took it private in June at $24.55 a share.

One further point about the rally's cause: this month's software strength is being attributed to second-quarter earnings beats rather than to the interest-rate path, with Federal Reserve officials floating increases rather than cuts. The bid is for results. Datadog produced them and was sold anyway, on a single customer's spreadsheet.

The setup

Where it stands — Datadog's growth accelerated to a two-year high while one AI customer's usage cut pulled next quarter's guide seven points lower.

Would confirm — Third-quarter revenue landing above the $1.145bn top of guidance with net revenue retention still above 120%.

Would invalidate — A second large AI account disclosed as optimizing, or gross margin falling below 78% as inference costs build.

Watch next — Snowflake reports its July quarter after the close on 2 September, with product revenue guided to $1.415–1.42bn.

Valuation — Datadog at 26.6x trailing and 23.6x forward gross profit, against Dynatrace near 8.4x and Elastic at 6.8x.

EMCOR Group, the electrical and mechanical contractor that installs power distribution, chilled-water piping, clean-room ventilation and low-voltage cabling inside a data hall, told investors on 30 July that its order book had reached a record $17.14bn — up from $11.91bn a year earlier, with the largest growth in network and communications. It also told them something less comfortable. Only about three-quarters of that book will convert to revenue within twelve months, against 85% historically, because the projects have gotten bigger.

Quanta Services, which builds and maintains high-voltage transmission lines and substations for utilities and does much of it on live wire, said the same thing in a different form. Its record backlog reached $53.4bn, up 50%, but management noted that the big 765- and 500-kilovolt corridors are still in engineering, with field execution starting in the second half of 2027, and that roughly 95% of its generation work is not in backlog at all and runs past 2030.

That is a description of a very long-dated asset. On 17 and 18 August, the 30-year Treasury yield topped 5.31% and then 5.323%, its highest since 2007, on inflation, fiscal supply and uncertainty about policy under the new Federal Reserve chair. In the four sessions from that close, Quanta fell 8.9% and EMCOR 8.7%. Neither had spoken to the market since 30 July.

The businesses went the other way

Quanta's revenue growth accelerated for four straight quarters — 15.6%, then 19.6%, 26.3% and 41.1% — reaching $9.56bn in the June quarter. Gross margin widened 296 basis points year over year to 16.2%, and gross profit grew 72.8% on 41.1% more revenue, which is what operating leverage looks like when it is real. Two honest deductions: four acquisitions closed in the quarter will contribute $1.2–1.4bn of this year's revenue, so part of the growth is bought; and Quanta does not disclose data-center work as a dollar line anywhere in its statements. The "15–20% of revenue" figure for technology and load-center work is spoken commentary on a call.

EMCOR does disclose it. Electrical construction revenue of $1.66bn rose 24% at a 13.9% operating margin, up 210 basis points, on network-and-communications work up 45%. Mechanical construction revenue of $2.3bn rose 31% with data-center revenue more than doubling — but margin there fell 110 basis points to 12.5% as the mix shifted toward guaranteed-maximum-price and construction-manager contracts, which cap the upside. Company-wide operating margin still set a record at 10.6%.

Against that, both multiples compressed. Quanta's price per dollar of trailing gross profit is 20.33x, against roughly 26.8x on 21 May and 22.7x in February, even as trailing gross profit grew to $4.73bn. EMCOR's is 9.37x, against about 11.0x at both earlier dates. Gross profit is the usable lens here because Quanta's 72x trailing earnings are distorted by acquisition amortization; on enterprise value to EBITDA it is 33.7x against EMCOR's 14.9x, so the two are not priced as the same business.

What the group number hides

The five specialist builders most often grouped together fell about 15.5% on average over thirty days, but that average is three unrelated events. Sterling Infrastructure broke 31% in six sessions in late July on its own print. MasTec fell 18.5% on 31 July despite record backlog, after cutting its communications outlook on wireless slowdowns and wireline work deferred into 2027. Quanta, over the identical thirty days, was up 1.5%.

Dycom Industries is the control that decides which story this is. It builds fiber for carriers and cable operators rather than content inside data halls, and its 50-day average crossed below its 200-day on 19 August — five sessions before it reports, so the move cannot be its own news. Its fiscal first quarter grew 56.1%, but gross margin slipped to 14.0% from 15.0%, and MasTec's cut is a live read-across into Dycom's 26 August call.

One constraint sits under all of it, and it is not demand. The binding limit is craft labor — electricians, pipe fitters, supervisors. EMCOR spent roughly $750m in the quarter buying five union electrical contractors with $625m of revenue between them. It was buying crews.

The setup

Where it stands — Both protagonists raised guidance in July and de-rated in August on no company news, alongside a 19-year high in long yields.

Would confirm — Quanta and EMCOR report September-quarter backlog and remaining performance obligations still growing double digits year over year.

Would invalidate — EMCOR's twelve-month burn rate falls below 75%, or either company's book-to-bill drops under 1.

Watch next — Dycom reports fiscal second-quarter results on 26 August, guided to $1.94–2.01bn of contract revenue.

Valuation — Quanta at 20.33x trailing gross profit and 16.92x forward, versus about 26.8x in May; EMCOR 9.37x and 8.57x, versus 11.0x.

Vacancy in the largest North American data-center markets ended last year at 1.4%, a record low, and 0.3% in Northern Virginia. Record space under construction is 88% pre-leased before it opens. Asking prices for 3-to-10-megawatt requirements rose 12.5% in a year. That is the whole mechanism: leases signed at the pricing of five years ago are expiring into a market with almost nothing available, so rent resets sharply upward without a single new customer being won.

The question is not whether the two landlords that dominate the listed market are capturing it. It is what they are paying to build the next tranche.

Two ways to buy a megawatt

Digital Realty is the wholesale developer — 309 facilities leased in large blocks to cloud and enterprise tenants, with about 3.0 gigawatts in service. Its June-quarter revenue reached $1.92bn, up 28.9% year on year, an acceleration from 16.2% in the prior quarter. Operating margin widened to 25.9% from 14.2% a year earlier. Backlog of signed-but-not-commenced leases hit a record $1.9bn at full share, roughly 30% of in-place data-center revenue. Renewals above one megawatt repriced 66.7% higher in cash terms.

The construction pipeline is $20bn, double the first half of the year, covering 1.4 gigawatts that is 63% pre-leased at an expected 11.5% stabilized yield. Paying for it has meant shares. Digital Realty bought Blackstone out of three fully leased Northern Virginia sites for $3.5bn, of which $2.3bn was stock, and is buying Columbia Capital and more of South Africa's Teraco with equity too. Diluted shares rose from 345.7m to 361.5m. Yet core funds from operations per share still grew 14%, and leverage fell to 4.7x with a 2.9% average coupon and a maturity wall that peaks in 2029, not next year.

Equinix sells the opposite product: cabinets, power density and cross-connects inside neutral campuses where networks meet. Revenue grew 16.4%, the fourth straight quarterly acceleration. Interconnection revenue reached $453m, up 11.3%, on a record 9,700 net new interconnections. Monthly recurring revenue per cabinet reached $2,538, up 6%, and churn fell to 1.8%. Adjusted funds from operations per share rose 18% alongside the largest guidance raise in the company's history.

Equinix is doubling capital spending to $5-6bn this year and $5-7bn annually through 2029. It will fund that with a 50% payout ratio and debt — barely 1.1% dilution — and has told investors the consequence: leverage up about one turn, and blended cost of capital up roughly 150 basis points. That is the line the bond market now presses on.

The duration test, and what it showed

In the week the 30-year Treasury yield touched 5.32%, its highest since 2007, both landlords fell — Digital Realty 4.5%, Equinix 3.0%. But the actual bond proxies did not: American Tower rose 1.2%, Realty Income 0.3%, Prologis was flat, and the S&P 500 fell 1.5%. Over twelve months the three listed tower REITs fell between 15% and 26% while these two compounded. Whatever is moving them, it is not the yield curve alone.

What has moved is the price of the revenue. Both stocks are higher over a year — Digital Realty 16.8%, Equinix 38.5% — and both are flat-to-slightly-down over three months while sales compounded. Equinix now trades at 20.7 times trailing gross profit and 19.8 times forward, against 21.6 times in mid-May. Digital Realty is at 10.3 times sales against 11.1 times then. Growth is outrunning the multiple, not following it. Consensus still models 16.9% revenue growth at Digital Realty this year, below the 28.9% it just reported.

Digital Realty's shares only regained a firm uptrend in mid-August; they spent late November through January in a sustained decline. The year was not a straight line.

The competitive floor differs too. Digital Realty bids for hyperscale megawatts against CyrusOne, owned by KKR and Global Infrastructure Partners, and QTS, owned by Blackstone — permanently capitalized private balance sheets that never answer to a quarterly share price. Equinix's defense is the ecosystem itself: eight of the ten largest AI model providers run workloads on its platform, and a cross-connect is worth what the other tenants in the building are worth.

The setup

Where it stands — Both landlords are accelerating into a 1.4% vacancy market, and both are cheaper per dollar of revenue than three months ago. Would confirm — Digital Realty's backlog rising above $1.9bn with cash renewal spreads holding in the raised 9-11% range. Would invalidate — Equinix's blended cost of capital rising more than the guided 150 basis points, or AFFO per share growth falling below 9%. Watch next — Third-quarter results in late October, where Equinix guided softly despite raising the full year. Valuation — Equinix at 20.7x trailing gross profit and 19.8x forward, versus 21.6x in mid-May; Digital Realty near 23.4x 2026 core FFO guidance.

The geothermal company that made money last quarter did not make most of it from geothermal.

Ormat Technologies, a Reno-based operator of geothermal and recovered-energy power plants across the United States, Kenya, Indonesia and Chile, reported second-quarter revenue of $258.8m, up 10.6%. Gross profit rose 20.8% to $68.7m and consolidated gross margin widened 220 basis points to 26.5%. The engine was Energy Storage, where revenue jumped 195% to $42.8m at a 56.2% segment gross margin, earned on merchant prices in PJM Interconnection, the mid-Atlantic grid. Management expects that margin to settle at 40–50% for the year.

The Electricity segment — the contracted geothermal baseload at the center of the data-center power argument — grew 5.8% to $169.3m, helped by the acquired Blue Mountain plant, higher Puna rates and less curtailment. Its full-year guide was cut by roughly $5m to $710–725m on one-to-two-month commissioning delays at two Caribbean projects. Below the gross-profit line, capital intensity bites: operating income fell 3.2% and net income 3.4%, on $3.4bn of debt at a 3.9% average rate, or 4.3x adjusted earnings before interest, taxes, depreciation and amortization (EBITDA).

What actually moved Fervo

Fervo Energy, a Houston developer with 199 employees that borrows horizontal drilling and fiber-optic sensing from shale to mine heat from hot rock, is the opposite balance sheet. It has 658 megawatts under binding power purchase agreements (PPAs) representing a $7.2bn revenue backlog — including two 15-year contracts totaling 320 MW with Southern California Edison — and raised its 2030 capacity target to 1.1 gigawatts. Its widely cited 3-gigawatt framework with Google is described in Fervo's own registration statement as non-binding, with no obligation on Google to buy power.

Quarterly revenue was $113,000. On 12 August Fervo reported a $55.9m net loss, or -$0.38 a share against estimates nearer -$0.09, and guided 2027 revenue to $60–80m; the shares fell 16.6% that session. The guide is not an execution failure — it reflects curtailment on third-party transmission lines leaving Cape Station after the line's operator added another asset. GeoBlocks 1 and 2 are mechanically complete, with first power targeted for the fourth quarter, and the drilling curve is improving: a record 19,500-foot well drilled in 21 days.

Over the 30 days to 21 August, Fervo fell 39% while Ormat rose 4.6%. In the four sessions around Fervo's print, Ormat moved -0.4%. The long bond explains only the tail: the 30-year Treasury yield reached 5.311% on 17 August, a 19-year high, and from 14 to 21 August Ormat lost 5.5% and Fervo 13.9%, against 1.2% to 3.1% declines at regulated utilities. Nor is tax policy the culprit: under the 2025 federal budget law, geothermal keeps the full credit if construction starts by 2033, while wind and solar face 2026 deadlines.

The divergence that matters

Ormat's de-rating happened in June and July, not August. Shares fell 22.6% from 16 June, the day Bernstein initiated coverage at Underperform with a $115 target, doubting it could compete with Fervo in enhanced geothermal. Ormat's answer is procurement, not laboratory work: a collaboration with oilfield-services firm SLB whose Desert Peak pilot drills in the fourth quarter, a $25m stake in Sage Geosystems, and a new 100-megawatt modular turbine unit for large-scale enhanced systems. Its data-center exposure is already signed — roughly 13 MW to Switch on a 20-year term and up to 150 MW through NV Energy to serve Google's Nevada operations.

The price paid for that has fallen while the profit grew. Ormat costs 20.1x trailing gross profit, against about 25.9x in May and 26.5x in February, even as trailing gross profit rose 21.7% to $331.9m. Enterprise value to trailing EBITDA is 22.0x — the fairer lens for a leveraged generator than 52.1x trailing earnings. The catch sits in consensus: analysts model FY2026 revenue of $1.177bn, then $1.176bn in 2027. Flat.

Fervo cannot be valued on earnings. It trades at 1.78x book, and $2.1bn of cash is 43% of its market value, leaving enterprise value near $4.5m per contracted megawatt. Second-half capex of $850–900m against that cash implies roughly fourteen months of runway before project debt or equity is needed. The May initial public offering raised $2.04bn at $27.00 a share, putting a standard lockup expiry in mid-November — inside the window.

The setup

Where it stands — One profitable incumbent re-rated cheaper on growing profit; one developer marked down on its own loss and a transmission constraint. Would confirm — Fervo delivers first power from GeoBlock 1 by 31 December and announces a behind-the-meter data-center contract. Would invalidate — Ormat's Electricity segment misses the reduced $710–725m guide, or storage margin falls below the 40% floor. Watch next — Fervo's IPO lockup expires mid-November 2026; Ormat's third-quarter report follows in early November. Valuation — Ormat: 20.1x trailing gross profit versus 25.9x in May, 42.9x forward earnings; Fervo: 1.78x book, no earnings.