DK Street Journal

Applied Digital, a Dallas developer that builds entire data-center campuses and leases them whole to a single tenant, told investors on 27 July that it now holds 1.4 gigawatts of contracted critical computing load. The leases carry about $36bn of contracted revenue, rising to roughly $86bn if every renewal option is exercised.

About 175 megawatts of that is energized — 100MW delivered at its Polaris Forge 1 campus in North Dakota in October 2025 and a further 75MW at the end of June. Roughly one-eighth of the contracted book is live and collecting rent. The rest is a construction program that must be financed before it earns anything.

The rent is contracted; the money is not

That is why the cost of capital, not the tenant list, is doing the work. Applied Digital funded a third building at Polaris Forge with a $300m senior secured bridge led by Goldman Sachs at the Secured Overnight Financing Rate (SOFR) plus 275 basis points, and a subsidiary raised $2.35bn in private notes. Campus equity comes largely from structurally senior preferred stock: the company lifted its Series G capacity to $2.0bn and drew a further $562.5m under a Macquarie facility of up to $5.0bn, which sits ahead of common holders in every campus.

That channel is repricing. Two of the last three commercial mortgage bond deals funding data centers had to widen pricing to attract buyers, and high-yield data-center spreads have widened since June as investors questioned whether AI revenue can carry long-dated lease obligations. On 18 August, the 30-year Treasury yield topped 5.3%, a 19-year high; the same session CoreWeave, Applied Digital's largest tenant, fell 12.1% on debt-financing concerns.

The counterparty itself has improved. CoreWeave holds all three Polaris Forge 1 leases — 400MW, about $11bn, roughly 15-year terms — and now backs them with springing guarantees and a $50m letter of credit after earning an A3 rating. Three new 15-year leases with a single high investment-grade hyperscaler have cut CoreWeave to about 31% of the contracted book.

The numbers behind the de-rating are mixed rather than bad. Fiscal 2026 revenue reached $611.3m, up 183.7%, but the fourth quarter's gross margin fell to 15.7% from 42.5% because $152.4m of the beat was one-time tenant fit-out fees. Consensus does not model a full-year profit before fiscal 2028. Trailing price-to-sales is 14.3x against 28.7x in early May, 10.0x forward; on enterprise value, the implied price per contracted megawatt has fallen to roughly $6.5m from about $9.2m while contracted capacity rose from about 1GW to 1.4GW. Cheaper against itself, not against anything absolute.

China is not the same trade

GDS Holdings, China's largest carrier-neutral colocation operator, raised its full-year sales target to 1GW on 13 August after a record 470MW of first-half bookings, with backlog at 757MW and capex guidance up to about $1.4bn. It funds projects onshore at 60% debt to 40% equity against stabilized yields of 10-11%, in a bank market management calls highly supportive. The catch is present-tense: second-quarter revenue grew 6.2% and gross margin fell to 19.5% from 23.8%, as legacy contracts reprice and mix shifts to cheaper inland markets. It trades at 14.0x trailing enterprise value to EBITDA with net leverage of 4.7x.

VNET Group, a Beijing wholesale and colocation operator, is the diverging name. Capacity in service crossed a gigawatt for the first time, up 49.4%, utilization reached 73.9% and adjusted EBITDA margin widened to 33.0% from 30.1%. But it missed on earnings and set a full-year revenue midpoint below consensus on 18 August and lost 16.9% in that session — its entire monthly decline. At 9.7x trailing EV/EBITDA it is the cheapest of the three.

What binds the Chinese pair is silicon, not credit. SMIC's most advanced stable node runs above 93% utilization with every certified domestic designer competing for wafers; GDS puts its 2026 workload mix at roughly half CPU, half GPU, with the GPU share rising in 2027 as domestic supply catches up. Their capacity is gated by accelerators arriving, not by lenders.

The share-price damage has landed hardest on the smallest builders — Keel Infrastructure, WhiteFiber and SharonAI each fell more than 14% in a month, far more than Applied Digital's 4.7%. Applied Digital itself broke trend during the period: its 50-day average sat above the 200-day on 1 July and crossed below it by 20 August, even as the contracted book grew.

The setup

Where it stands — Applied Digital's contracted book grew while its sales multiple halved; the financing cost of building it grew too. Would confirm — Energized capacity moving above 175MW toward the 1.4GW contracted, with recurring lease revenue replacing one-time fit-out fees. Would invalidate — New campus debt or preferred priced materially wider than SOFR plus 275 basis points, compressing stabilized lease yields. Watch next — Applied Digital's fiscal first-quarter results in October, and whether GDS's backlog exceeds 1GW by year-end. Valuation — APLD 14.3x trailing sales, 10.0x forward, against 28.7x in early May; GDS 14.0x EV/EBITDA, VNET 9.7x.

Every increase in the data rate running across a server rack tightens the tolerance on the clock driving it. That is why SiTime, a Santa Clara company that builds timing chips out of silicon micro-electro-mechanical systems (MEMS) rather than the quartz crystals the industry has used for decades, told investors this month that hyperscale customers now add several hundred dollars of synchronization content per data-center rack. Its communications, enterprise and data-center line produced $101.2m in the June quarter — 64% of company revenue, and up 181% from a year earlier. That is the ninth consecutive quarter of triple-digit growth in the segment, and the first time it has crossed a $100m quarterly run rate.

The content story is not design-win language. Book-to-bill, average selling prices and order sizes all rose; customers are placing orders 12 to 18 months ahead; inventory stood at $104m to serve the September quarter. Company revenue growth has accelerated four quarters running — 45%, 66%, 88%, then 127% to $157.4m. Gross margin went from 53.5% to 63.0% over the same stretch, above the 57-60% band the company used to earn.

Two sessions, one layer

None of that stopped the repricing. From the 17 August close through 19 August, SiTime fell 20.6%, MACOM 19.5% and Semtech 19.5%, with MaxLinear down 23.2%, Fabrinet 22.3% and Credo 18.4%. Nvidia fell 2.1% and Qualcomm 0.4% over the same two days. This was not a rotation out of AI hardware; it was a repricing of the layer that sells content into the rack. The trigger was Fabrinet's 17 August print — record revenue up 45% year over year, but a much smaller implied sequential step-up — which dragged the optical group with it, compounded by a scare over AI-lab revenue and a 30-year Treasury yield near a 19-year high. The revenue read itself was hardly weak: Anthropic told investors its annualized run rate reached $65bn in July.

Where the multiples sit

SiTime's earnings multiple is useless here — operating margin was 5.2% in the quarter, and trailing price-to-earnings sits above 1,000x — so gross profit is the honest lens across three companies earning 63%, 58% and 52% margins. On it, SiTime trades at 57.4x trailing and 30.7x forward, down from 82x as recently as 18 August. That is still the most expensive name in the interconnect complex: MACOM is at 30.9x, Fabrinet about 29x, Semtech 20.6x, Amphenol 17.3x. SiTime is leading the group on business momentum and on price simultaneously. It also carries the group's cleanest concentration risk — $22.8m of the quarter, about 14.5% of revenue, came from one large consumer customer. In July it closed the $1.5bn cash-plus-stock purchase of Renesas' timing division, removing one of the four suppliers that dominate a concentrated global timing market and tracking to roughly $340m annualized at about 70% gross margin.

MACOM, a Massachusetts maker of radio-frequency, microwave and photonic chips for telecom, defense radar and data centers, is the cleanest divergence. Data-center revenue grew 40% sequentially to $137.6m, book-to-bill hit a record 1.6:1 after 1.3x and 1.5x, and the company guided the September quarter to $415-425m with gross margin of 60-61%. Its top ten customers are now under 40% of revenue, down from a far narrower base two years ago. It fell with the group anyway, from 38x gross profit in May to 30.9x.

Semtech is the ambiguous one. Its data-center revenue was a record $71.6m, up 39%, but that is under a quarter of the $291m total; the rest grew about 10%, and corporate gross margin is stuck at 51.9% versus 63.3% in fiscal 2023. The AI mix has not yet lifted the P&L. What has changed is the balance sheet: net leverage cut to 1.5x from 7.2x, with annualized interest below $3m against $75m a year earlier. Its CopperEdge active copper chips, pitched on up to 90% less power than digital-signal-processor alternatives, are attacking a socket Credo holds with its HiWire cables and four hyperscaler customers above 10% of revenue each.

One control: MaxLinear raised full-year optical data-center guidance and is still down 31% over three months. Across the seven-name radio-frequency and wireless group, five of seven more than doubled over twelve months and the median gained 156% — the year was never one optics name's arithmetic. The three months since 20 May were a uniform de-rating, not a demand break.

The setup

Where it stands — Timing and short-reach content per rack is rising on shipped units, but only SiTime's P&L is majority-AI, and its multiple is the group's highest. Would confirm — SiTime's September quarter landing inside the $285-295m guide with gross margin near 68%. Would invalidate — Semtech's data-center line missing the guided 35% sequential step, or SiTime's segment growth falling below 100%. Watch next — MACOM's fiscal fourth quarter, guided to $415-425m, reports in November. Valuation — SiTime 57.4x trailing and 30.7x forward gross profit, down from 82x on 18 August, against MACOM 30.9x and Semtech 20.6x.

Vicor, a maker of brick-format DC-to-DC converters in Andover, Massachusetts, spent years watching a larger rival take the power socket inside Nvidia's flagship AI accelerator. In the June quarter it collected $30.4m in royalties from the industry that adopted the architecture it patented — about a fifth of core revenue, from customers who in many cases had to be compelled.

The compulsion is documented. In February 2025 the US International Trade Commission confirmed the validity and infringement of two Vicor patents and issued a limited exclusion order barring importation of unlicensed computing systems containing infringing power modules, with cease-and-desist orders naming Delta Electronics and Quanta entities. That order is why a hyperscaler and a set of original-equipment manufacturers now pay. A second enforcement case is expected to reach final determination in 2027.

The comp that hides the growth

Vicor's reported June quarter looks stagnant: revenue of $143.4m, up 1.6% year on year, with gross profit down 9.8%. The year-ago period contained a one-time $45m patent settlement. Adjusting for it, core product-plus-royalty revenue grew 49.3% and backlog rose 145% year on year to roughly $380m.

The forward book is the cleaner tell. One-year backlog climbed 26% sequentially to $379.7m, following a 70% sequential rise the quarter before, with orders coming from existing aerospace, defense, industrial and high-performance-computing customers rather than from licensing. Gross margin of 58.0% sits 730 basis points below last year but 280bp above the March quarter, and well clear of the 47.2% trough of early 2025. Management signed a new licence recognizing $15m in the quarter, structured as $60m over two years, raised its long-term target to $2.5bn of revenue at a 70% gross margin, and said a second chip fab would be needed to get there.

What makes the royalty line durable is arithmetic in the rack. Nvidia's rack generations step from 145kW to 330kW for Vera Rubin and on toward 570kW, with native 800-volt direct-current designs contemplated up to one megawatt. Currents that large make the old two-stage step-down physically wasteful and force vertical power delivery directly beneath the chip — Vicor's territory. Its second-generation parts reach 3 amps per square millimeter against roughly 1 for competing first-generation designs, though volume ramps are guided only to late 2027.

Same architecture, three different businesses

Monolithic Power, the fabless designer that took the Nvidia socket, is the one whose results already reflect all this. June-quarter revenue hit a record $980.6m, up 47.6% — the fourth straight quarter of acceleration — and operating margin widened to 31.0% from 24.8%. On its 30 July call it said enterprise-data revenue rose 45% sequentially with no concentrated customers, and lifted the growth floor for that business from 85% to 130%. Its shares fell 14.8% over three months regardless. At 35.8x trailing gross profit and 28.4x forward, it is cheaper than the 48.7x it commanded in early May but still the dearest name in analog.

Navitas is the case where the de-rating is earned. The Torrance, California designer of gallium-nitride and silicon-carbide power devices is one of about fifteen suppliers Nvidia has named for 800VDC, and it reported June revenue of $10.5m, down 27.3% year on year, with GAAP gross profit of minus $1.0m and a $27.2m operating loss. Trailing revenue is roughly $36.5m against a $3.3bn market value — 90.4x sales, with no earnings or gross-profit multiple available. Consensus does not see $200m of revenue before 2029. The shares sit 42% below the $21.89 at which the company raised $373m in the same quarter, and it is now fighting on two legal fronts: Wolfspeed sued in Delaware in July over five patents covering essentially the entire product line, and Navitas countersued Renesas over gallium-nitride patents in August.

The discount rate, not the demand

The selling was not gradual and was not about rack orders. Vicor dropped 19.2% in the 2 July session and 11.6% on 28 July; Navitas fell 21.0% the day after its results; Monolithic Power lost 10.4% on 5 June. Then, on 18 August, the 30-year Treasury yield topped 5.33%, a 19-year high, on inflation, a debt path approaching $40tn and questions about policy under new Federal Reserve chair Kevin Warsh — and the highest-multiple silicon de-rated together, with Microchip, Vicor, NXP, ON Semiconductor and Vertiv all rolling over from an uptrend within eight sessions. Eaton, the electrical-equipment maker selling into the same racks on a far lower multiple, rose 7.2% over the past month. Same content, different duration.

Vicor's trailing multiple on gross profit is now 34.7x, against 61.8x in mid-May, with the forward at 27.3x. Consensus asks for $603m of revenue this year and $939m next — numbers that require the backlog and the licence schedule to convert on time.

The setup

Where it stands — Vicor's order book and royalty stream are compounding while its multiple has roughly halved since May. Would confirm — One-year backlog holding above $380m in the September quarter with royalty revenue at or above $30m. Would invalidate — Gross margin slipping back toward the 55.2% of the March quarter, or licensees lapsing before the second trade case concludes. Watch next — Third-quarter results in October, and site selection for the second fab, guided within weeks of the July call. Valuation — 34.7x trailing gross profit and 27.3x forward, against 61.8x trailing in mid-May.

Nvidia has spent the past year pushing data-center builders toward racks that draw a megawatt and run on 800 volts of direct current, an architecture that promotes energy storage from a backup accessory in a room down the hall to a component of the rack's own power design. That is the single most-cited reason to own a battery company right now. It is worth checking who actually gets the order.

The answer, so far, is contract manufacturers. Flex, the Singapore-based electronics manufacturer, launched an 800 VDC Power Rack built with Nvidia for the Vera Rubin platform, offering battery backup units and capacitor backup units as options on it. Delta Electronics, the Taiwanese power-electronics maker, shows in-row 660 kW racks with an 80 kW battery backup unit embedded per shelf. The cells behind those shelves come from Asian suppliers. None of the three US-listed battery names most often attached to the theme are in that path.

The incumbent sells the room, not the rack

EnerSys makes stored-energy systems for industry: uninterruptible power supply strings and DC power for telecom and data centers, forklift batteries, and defense cells. Its business is genuinely inflecting. In the quarter ended 5 July, revenue rose 4.8% to $935.6m while gross profit climbed 23.7% and operating income 75.1%. Data-center orders were up 80% with 12-to-36-month visibility, and total book-to-bill ran 1.06x. Its answer to the lithium transition, the DataSafe Noir cabinet launched in June, claims more than twice the output power of published rival specifications — and generates no revenue until fiscal 2028. That is facility-level equipment competing with Vertiv and Eaton, not rack content.

Two caveats sit inside the earnings. A $31m tariff refund added $0.63 a share; strip it and the 510-basis-point gross-margin expansion becomes 180. And $1.25 of the $3.66 adjusted diluted EPS was a Section 45X manufacturing tax credit, with second-quarter guidance implying that share rises toward 38%. The credit phases down from 2030 and expires in 2033, and its domestic-content threshold tightens each year — which is what the roughly $500m Greenville, South Carolina lithium plant is for.

Pipeline is not an order book

Eos Energy builds zinc-based grid storage — cubes sited beside substations and solar farms, outside the data-center fence. Revenue rose 351% to $68.8m and backlog hit a record $807m, up about a quarter sequentially. Gross margin remains negative at -71%, though that is the fourth straight improvement from -203% a year ago. Data centers are 32% of a $24.6bn opportunity pipeline; qualification for on-site co-location is still under way and nothing is contracted. Meanwhile the diluted share count grew 43% year on year, and $364m of cash faces a quarterly adjusted-EBITDA loss near $71m. Management trimmed the top of 2026 guidance to $350m from $400m, blaming a production-line consolidation rather than demand; consensus sits at $309m.

Phones, not infrastructure

Enovix makes silicon-anode lithium cells, and its customers are smartphone, smart-eyewear and drone makers. Second-quarter revenue was $9.0m, up 20.8% — a sharp deceleration from 85% growth three quarters earlier. It shipped 2,100 eyewear cells against a 50,000-unit order, and the lead smartphone customer's final qualification is not due to finish until end-2026. Chief executive Raj Talluri then resigned on 13 August to run Kulicke & Soffa, with the finance chief taking over on an interim basis; the shares lost 30% in five sessions on volume ten times normal.

What the prices say

EnerSys has nearly doubled over twelve months, but the shares are down about 10% from their May high and have traded sideways for a month, with the 50-day average now flat against the 200-day. The move was a re-rating: 19.9x trailing earnings today against roughly 11x a year ago, versus 14.4x forward and a 10.3% trailing free-cash-flow yield. Eos has fallen 47% in three months and is priced at 4.56x trailing sales against 3.16x forward — a spread that assumes revenue triples again. Enovix, down 66% over a year, still carries 19.3x trailing sales and 18.0x forward, meaning the market prices essentially no revenue inflection at all. Long-duration, pre-profit equities have also been fighting a 10-year Treasury yield near 4.64% after a 20-month high.

The setup

Where it stands — Only EnerSys converts data-center demand into profit today, and it does so outside the rack, at facility level. Would confirm — A named 800 VDC rack design win, or DataSafe Noir revenue pulled into fiscal 2027. Would invalidate — EnerSys data-center orders decelerating below book-to-bill 1.0x, or Eos converting none of its $807m backlog. Watch next — EnerSys fiscal Q2 results in early November; Eos third-quarter shipments against $300–350m full-year guidance. Valuation — EnerSys 19.9x trailing and 14.4x forward earnings, against roughly 11x trailing a year ago.

Three chip suppliers sit between artificial-intelligence processors, moving data across a rack rather than doing arithmetic inside it. All three reported June quarters within a fortnight, and the results point in different directions. The share prices have not.

The one that is accelerating and de-rating

Coherent, a laser and photonics manufacturer that makes the indium-phosphide chips and optical transceivers inside data-center networks, posted record revenue of $2.05bn in the June quarter, up 33.7% year on year. Its Datacenter & Communications segment was 79% of that, and data-center revenue rose 66% year on year and 24% sequentially on 800-gigabit and 1.6-terabit transceiver demand — a third consecutive quarter of double-digit sequential growth. Gross margin has now expanded four quarters running, to 38.5%. The company produced roughly 80% more indium-phosphide lasers than a year earlier and expects to double internal capacity a quarter ahead of plan, moving to six-inch wafers that yield four times the output at half the cost.

The shares fell 24.8% over the same three months. Coherent trades at 20.8x trailing gross profit, down from about 30x in late May and below the 25x it carried in early May. Its trailing price-to-earnings ratio of 65.3 compares with 30.1x forward. In early May that trailing multiple was above 300 — the earnings arrived; the price did not follow.

The one priced for it

Astera Labs sells retimers, switches and modules that keep signals intact across the PCI Express and CXL fabrics linking accelerators to each other and to memory. Revenue reached $392.4m, up 104% year on year and 27% sequentially, at a 73.3% gross margin. Management guided the September quarter to $540m-$560m, roughly 40% sequential growth, with operating margin near 43%. Its Scorpio X switches entered volume production a quarter early and should be the largest product family by revenue this quarter, with more than ten customers engaged and content above $1,000 per accelerator.

That is the cleanest acceleration of the three, and it is fully in the price: 54.3x trailing gross profit, dearer than either peer and no cheaper than the 52.3x it carried in early May despite falling 21% from its August high.

The weak leg

Credo Technology makes active electrical cables and the signal-conditioning silicon inside them — copper links up to about seven metres that draw roughly half the power of optics, a market it holds some 80-88% of, ahead of Marvell and Broadcom. Fiscal 2026 revenue tripled to $1.335bn at a 68% gross margin, and operating margin reached 35.7%.

The headline growth rate is backward-looking. Sequential growth ran 51.9% in the January quarter, then 7.4% in the May quarter, and guidance implies about 7.6% again. Consensus has fiscal 2027 revenue at $2.46bn, which requires the remaining three quarters to average roughly 12-15% sequential growth — a re-acceleration the current guide does not contain. Concentration is extreme: the top three customers were 34%, 27% and 16% of the latest quarter. Credo's multiple, at 47.3x trailing gross profit, is essentially where it stood in early May, even as the growth path changed underneath it. Results land after the close on 1 September.

Why they fell together

Nothing in August came from these companies' order books. Early gains followed a report that the Federal Communications Commission (FCC) was drafting a ban on new Chinese optical transceiver imports. The reversal came on 18 August, when the 30-year Treasury yield hit a 19-year high of 5.33% and chip stocks sold off across the board. That transmission is newer than it looks: hyperscalers' incremental annual borrowing rose from 9% of capital spending in fiscal 2024 to 32% of the trailing twelve months by mid-2026, so long-bond yields now price their suppliers directly. A day earlier, contract optics assembler Fabrinet reported record revenue up 45% at a gross margin of 11.99% and fell 20%, dragging the complex with it.

Volatility here is not new. Since mid-June, Credo has moved 5% or more on 22 of 43 sessions, Coherent on 23, Astera Labs on 21 — against one such session for Nvidia. The unresolved question underneath is physics: copper links stop working past roughly two metres, and in March a group including Nvidia, Broadcom and Meta founded a consortium to define a common optical layer for rack-scale connections. Credo has bought its way toward optics through DustPhotonics; Astera Labs through aiXscale. Coherent already sells the lasers.

The setup

Where it stands — Coherent's business is accelerating into a falling multiple; Astera Labs' is accelerating into an unchanged one; Credo's is decelerating into an unchanged one. Would confirm — Credo guiding its October quarter below 12% sequential growth on 1 September. Would invalidate — Coherent's December-quarter data-center revenue growing under 10% sequentially, or gross margin slipping below 38.5%. Watch next — Credo's fiscal first-quarter results, after the close on 1 September, against guidance of $465m-$475m. Valuation — Coherent 20.8x trailing gross profit and 14.0x forward, versus 25x in early May; Credo 47.3x, Astera Labs 54.3x.

Duke Energy raised $1.75bn in mid-August by selling equity units — a hybrid instrument that pays holders a distribution now and converts into common stock later. The deal closed on 13 August at a 7.75% total annual distribution rate: 2.90% of contract adjustment payments on top of 4.85% interest on remarketable senior notes, with settlement due by 1 August 2029.

That number is worth holding against another one. Duke Energy Carolinas, the utility arm serving the Carolinas, settled its North Carolina rate case at a 9.8% authorized return on equity (ROE) on a 53% equity layer, having asked for 10.95%. The spread between what Duke pays for equity-like capital and what it is permitted to earn on the rate base that capital funds is now about 205 basis points, before tax and before execution risk. Duke serves roughly 8.2 million electric customers across six states and runs the largest regulated capital program in the industry, spending about $1bn a month.

The debt is not the problem

The cleanest public read on what this capital costs sits in a corner of the New York Stock Exchange most equity investors ignore: $25-par notes sold to retail buyers. CMS Energy, the Michigan parent of Consumers Energy with 1.9 million electric and 1.8 million gas customers, has three — 5.625% junior subordinated notes due 2078 and two 5.875% issues due 2078 and 2079. DTE Energy, which supplies about 2.3 million electric customers in southeastern Michigan, has three junior subordinated debentures maturing between 2077 and 2081. Duke has $500m of 5.625% debentures due 2078, deferrable for up to 40 quarters. Entergy Arkansas's 4.875% first mortgage bonds due 2066 are secured, and serve as the control.

At 19 August closes those eight yield between 6.15% and 6.94% on current-yield, with the Entergy secured bond cheapest to the issuer and CMS's 2079 notes dearest. Over twelve months their prices fell about 9.5% on average, deepest at DTE's 5.25% debentures, down 13.2%. Add back roughly six points of coupon and the holder lost about 3.5%.

But the decomposition matters more than the loss. Implied yields on the seven notes with a full year of history rose about 64 basis points. The 30-year Treasury rose roughly 46 basis points over the same window, touching 5.323% on 18 August, a 19-year high, against 4.89% at the end of July 2025. Some 70% of the repricing is therefore the curve. Only about 18 basis points is credit. Breckinridge Capital Advisors puts utility investment-grade spreads at 95 to 105 basis points against a five-year average nearer 105 to 110 — inside long-run norms, even as gross utility issuance runs toward a record, roughly $145bn this year.

The marginal cost of new debt confirms it. DTE issued $1bn of 6.200% junior subordinated debentures due 2058 in June. That is some 370 basis points below the 9.90% ROE the Michigan Public Service Commission granted Consumers Energy in March, and further below still after tax. CMS refiled on 2 June seeking $456m at a 10.25% ROE. Hybrid debt has not converged on authorized returns. Equity has narrowed the gap; debt has not closed it.

Who pays for the data centers

The second leg of the funding question is whether the utility or the customer carries the load-growth capital. Michigan has begun answering. The commission conditionally approved DTE's contracts to supply 1.4 GW to Oracle's Saline Township data center in December, attached protections for other customers, and ordered a large-load tariff filing within 90 days. DTE puts the Oracle contract at roughly $300m a year of benefit to existing customers and a pending 1.0 GW Google deal at about $1.7bn over its life, enough fixed-cost absorption to stay out of a rate case until 2028. CMS estimates its own large-load tariff delivers $7.50 a month of residential bill relief per gigawatt added.

That approval is not finished. Google's special contracts remain under review with a decision expected in September, and Earthjustice appealed the earlier approval to the Michigan Court of Appeals, filing its initial brief on 6 August.

The businesses agree; the shares do not

Duke reported second-quarter adjusted earnings per share of $1.43, up 14%, reaffirmed full-year guidance of $6.55 to $6.80, and carries 7.8 GW of signed data-center energy services agreements with a 15.4 GW late-stage pipeline it expects to convert by mid-2027. Its differentiator is unglamorous and hard to copy: 26 GE Vernova gas turbines on order, the first delivered to Person County in July, with gas supply secured into the early 2030s. It has also said it will not build large nuclear without additional protection against first-of-a-kind risk — a refusal to put construction risk on its own balance sheet.

DTE has 2.4 GW signed and says Oracle and Google alone underwrite "solidly 8%" earnings growth. CMS reaffirmed 2026 earnings per share of $3.83 to $3.90 despite a storm-driven margin hit and introduced 2027 guidance of $4.08 to $4.17, while exiting non-utility renewables and redirecting $1.7bn of that capital into the utility.

Over twelve months CMS shares are down 2.1%, DTE up 0.3% and Duke up 1.2%, against American Electric Power up 11.9% and NextEra up 13.0%. All three flipped decisively lower in the first half of August, their 50-day averages crossing below the 200-day; Duke's turn brackets the eleven sessions around its equity-unit pricing almost exactly.

The de-rating is real. Forward price-to-earnings now sits at 18.28x for CMS, 18.06x for DTE and 18.46x for Duke, against trailing multiples of 20.94x, 21.87x and 18.60x — and against roughly 20.2x to 20.9x trailing for all three in May. On the capital-structure-neutral measures that suit leveraged rate-base businesses, Duke is the cheapest at 11.56x trailing enterprise value to EBITDA and 1.76x book, and the only one of the three with positive trailing free cash flow, at a 1.6% yield; CMS runs at -8.6% and DTE at -6.7% as capex outruns operating cash. Credit, meanwhile, is not deteriorating: Duke targets funds from operations to debt of 14.5% this year rising to 15%, DTE around 15%, against BBB-tier downgrade thresholds nearer 12% to 13%.

The setup

Where it stands — The retail notes fell on Treasury duration, not credit, while the issuers' earnings accelerated and their shares stalled.

Would confirm — Utility investment-grade spreads holding at or inside 105 basis points as 2026 issuance completes.

Would invalidate — A new utility hybrid pricing above 8%, or funds from operations to debt guided below 13%.

Watch next — Michigan's decision on DTE's 1.0 GW Google special contracts, expected September 2026.

Valuation — Duke at 18.46x forward and 18.60x trailing earnings, 11.56x EV/EBITDA, versus roughly 20.3x trailing in May.

A review that ended quietly

On 9 February, Kyndryl's chief financial officer and general counsel left with immediate effect, the company delayed a quarterly filing, and it cut its free-cash-flow guidance for the year to $325-375m from $550m. The shares lost more than half their value in a session. The audit-committee review behind those departures, prompted by voluntary document requests from the Securities and Exchange Commission's enforcement division into cash-management practices and adjusted free-cash-flow disclosure, found no misstatements of previously issued financial statements and required no restatement. The amended annual report added disclosure, nothing more. The stock has traded between $10.62 and $14.69 in every month since.

Kyndryl is the 2021 spin-out of International Business Machines' managed-infrastructure arm: 72,000 people who operate other companies' mainframes, networks, cloud estates and digital workplaces for banks, telecom carriers, retailers and carmakers. It says it runs more than half the world's outsourced mainframes, staffed by 8,000 to 9,000 specialists — a skills pool that is not being replaced.

The business went the other way

Revenue in the year to March 2026 was $15.09bn, essentially flat. Gross profit rose 4.6% to $3.29bn, and gross margin reached 21.79%, up 92 basis points. That margin has climbed every year since the spin, from 11.3% in fiscal 2022, while revenue fell by $3.6bn — a five-year trade of volume for profitability, and it is still working: operating income rose 15.0% to $635m.

The forward book is the surprising part. Trailing twelve-month signings reached $14.2bn, now larger than annual revenue, with projected gross margin on that work of about 25% — above the margin of the business it replaces. Kyndryl Consult revenue grew 14% and its signings rose 50% in the June quarter; revenue tied to hyperscaler partnerships grew 48% over twelve months to $2bn. Forty deals above $50m were signed in the period, with new scope and new customers making up about 30% of large-deal value against 15% a year earlier.

The drag is disclosed and finite. Customer spending routed through IBM has halved from roughly $4bn annualized at separation to under $2bn, as clients buy IBM hardware and software directly and keep Kyndryl's services — a three-point revenue headwind that management says costs little in earnings. The June quarter also carried $152m of workforce-rebalancing charges, producing an adjusted pretax loss of $37m against $128m of income a year earlier. Roughly $200m of such charges are planned this year against about $200m of in-year savings and $400-500m annualized by fiscal 2028.

What it costs

Kyndryl trades at 6.7x forward earnings against 8.4x trailing, and 2.44x trailing EBITDA. Price to trailing gross profit is about 0.84x, down from roughly 2.3x a year ago — a 63% de-rating over a period in which gross-profit dollars grew. Management reaffirmed fiscal 2027 guidance of $600-700m adjusted pretax income and $400-500m of free cash flow, and kept fiscal 2028 targets above $1.2bn and $1bn respectively, against a market value of $2.75bn.

It is also acting like a buyer, not a target: a consortium of Apollo Global Management and Kyndryl has approached DXC Technology with an all-cash proposal reported at $22-25 a share, against a DXC price near $10.85. DXC, a $1.75bn outsourcer whose quarterly revenue fell 5.1%, is the declining asset here.

The channel split, and why

The idea that artificial-intelligence spending bypasses resellers and distributors is contradicted by the hardware data: ODM-direct's share of the server market fell from 64.1% to 50.2% in the first quarter of 2026 as enterprise and sovereign buyers chose established vendors, while server revenue grew 30.4%. What is splitting the channel is memory cost. Server DRAM contract prices are forecast to rise 13-18% in the third quarter, and cloud providers have locked long-term agreements — so the increase lands on enterprise buyers.

Ingram Micro, the global distributor that supplies resellers in more than 100 markets, is converting that: quarterly revenue rose 13.6% to $14.5bn and gross profit rose faster, up 14.2%, with GPU and AI infrastructure sales more than doubling. It trades at 8.0x forward earnings against 14.8x trailing. CDW, the largest US corporate and public-sector reseller, is on the other side: revenue rose 10.0% to a record, but gross profit grew 6.3% and gross margin fell 70 basis points to 20.1%, with management attributing personal-computer growth to higher selling prices rather than units. Year-to-date adjusted free cash flow was 42% of non-GAAP net income against an 80-90% target.

The setup

Where it stands — Kyndryl's margins, signings and consulting mix are improving while its equity trades below its annual gross profit.

Would confirm — Fiscal 2027 adjusted pretax income lands inside the reaffirmed $600-700m range with free cash flow of $400m or more.

Would invalidate — Signings fall back below revenue, or gross margin slips from 21.8%, as rebalancing charges outrun the promised savings.

Watch next — The September-quarter report, and whether the Apollo-Kyndryl approach to DXC becomes a firm offer.

Valuation — 6.7x forward earnings against 8.4x trailing; 0.84x trailing gross profit versus roughly 2.3x a year ago.

Cloudflare told investors on 6 August that more than half the traffic now crossing its network is generated by AI agents rather than people. The company — a cloud network that sits between the public internet and its customers' websites, selling denial-of-service protection, firewalls, content delivery and a serverless developer platform on one fabric — reported June-quarter revenue of $696.1m, up 35.9% year over year. That was the fourth consecutive quarter of accelerating growth.

Almost everything underneath it improved too. Dollar-based net retention reached 120%, up six points in a year. Customers spending more than $100,000 annually numbered 4,698, a 27% increase with a record 986 net additions. Non-GAAP gross margin rose 30 basis points sequentially to 73.1% — the first sequential gain in eight quarters, after two years of dilution from unpaid network traffic.

What Cloudflare is actually selling

The deals management named are not delivery bids. A five-year, $31.8m contract with a digital media company; an $11m European renewal that consolidated five legacy vendors; a $7.7m federal win. Cloudflare wins by replacing several point products at once, which is why it can grow while published content-delivery pricing runs between $0.002 and $0.085 a gigabyte depending on commitment and region.

What it does not sell, at least visibly, is AI inference revenue. Workers AI, AI Gateway, R2 storage and the pay-per-crawl toll on AI training bots — the products the valuation is credited to — carry no separate disclosed revenue line. What investors got instead was developer counts: 7.4 million on the Workers platform, about two million added in the quarter. In July the company switched pay-per-crawl to a pay-per-use model that pays publishers when content is used to answer a question. No figure attached.

Three sessions

Cloudflare's shares have held an uptrend since 5 May, its 50-day average above its 200-day for 71 straight sessions — the longest such run since 2025. Over those 71 sessions the stock rose 16.45%. Remove three of them and the other 68 compound to minus 12.18%. The three are 28 May, 1 June — when Nvidia's Computex keynote put agentic AI on an autumn timetable — and 7 July, its analyst day, when RBC framed more than $5bn of revenue before 2028.

Neither earnings print helped. The stock fell 24% on 8 May on a soft guide, a gross margin of 72.8% against a 75.1% estimate, and a plan to cut roughly a fifth of staff. It fell 3.7% on 6 August. Price-to-trailing-gross-profit has gone from 38.5x in February to 43.4x in May to 55.4x now, touching about 64x on 13 August.

Akamai earns its de-rating; Fastly does not

Akamai, the 1998-vintage delivery pioneer whose 4,400 points of presence sit inside carrier last-mile networks, now rents that footprint as GPU capacity. Revenue grew 5.4% to $1.10bn last quarter, but trailing gross profit went backwards by 0.5% and GAAP operating income fell 47% to $80.3m as capex hit 32% of revenue. Delivery shrank 6%; Cloud Infrastructure Services, at $99m, grew 39%. Against that sits $2.8bn of signed multi-year compute commitments, including a seven-year, $1.8bn contract with Anthropic — revenue that begins ramping in the fourth quarter. The shares are down 22.2% over three months and fetch 6.66x trailing gross profit, 16.7x forward earnings.

Fastly, the smaller edge vendor serving publishers, media and e-commerce, is the control that breaks the story. Its revenue grew 23.3% to $183.3m, the fastest in four years. Gross margin widened 1,193 basis points to 63.25%, so gross profit grew 51.9%. Net retention improved for a fifth quarter to 117%, and management described price erosion as mid-single-digit and driven by volume tiers, not competitive attack. Concentration is the flaw: the top ten customers are 37% of revenue and grew 48%, while everyone else grew 12%. The stock fell 12.7% on 19 August and 9.2% on 20 August with no company announcement to explain it, and at 8.48x gross profit remains dearer than in February.

The last week is about bonds

All three peaked within a session of one another, Cloudflare on 13 August, the other two on 14 August. On 18 August the 30-year Treasury yield topped 5.33%, a 19-year high. A higher discount rate falls hardest on companies whose earnings sit furthest out — which describes the most expensive of these three far better than the cheapest.

The setup

Where it stands — Cloudflare's fundamentals accelerated all year; its multiple expanded faster, and the expansion traces to three narrative sessions. Would confirm — A disclosed Workers AI, R2 or pay-per-use revenue line, or non-GAAP gross margin rising sequentially again in the third quarter. Would invalidate — Third-quarter revenue below the $736-737m guide, or net retention slipping back from 120%. Watch next — Akamai's fourth-quarter results, when the $2.8bn of compute backlog is scheduled to begin converting into revenue. Valuation — Cloudflare at 55.4x trailing and 48.5x forward gross profit, versus 38.5x in February and 6.66x for Akamai.